Datasets:

nickypro

/

minipile-split

like 0

--- abstract: | Due to their large dynamical mass-to-light ratios, dwarf spheroidal galaxies (dSphs) are promising targets for the indirect detection of dark matter (DM) in $\gamma$-rays. We examine their detectability by present and future $\gamma$-ray observatories. The key innovative features of our analysis are: (i) We take into account the [*angular size*]{} of the dSphs; while nearby objects have higher $\gamma$ ray flux, their larger angular extent can make them less attractive targets for background-dominated instruments. (ii) We derive DM profiles and the astrophysical $J$-factor (which parameterises the expected $\gamma$-ray flux, independently of the choice of DM particle model) for the classical dSphs [*directly*]{} from photometric and kinematic data. We assume very little about the DM profile, modelling this as a smooth split-power law distribution, with and without sub-clumps. (iii) We use a Markov Chain Monte Carlo (MCMC) technique to marginalise over unknown parameters and determine the sensitivity of our derived $J$-factors to both model and measurement uncertainties. (iv) We use simulated DM profiles to demonstrate that our $J$-factor determinations recover the correct solution within our quoted uncertainties. Our key findings are: (i) Sub-clumps in the dSphs do [*not*]{} usefully boost the signal; (ii) The sensitivity of atmospheric Cherenkov telescopes to dSphs within $\sim\!20$kpc with cored halos can be up to $\sim\!50$ times worse than when estimated assuming them to be point-like. Even for the satellite-borne Fermi-LAT the sensitivity is significantly degraded on the relevant angular scales for long exposures, hence it is vital to consider the angular extent of the dSphs when selecting targets; (iii) [*No*]{} DM profile has been ruled out by current data, but using a prior on the inner dark matter cusp slope $0\leq\gamma_{\rm prior}\leq1$ provides $J$-factor estimates accurate to a factor of a few if an appropriate angular scale is chosen; (iv) The $J$-factor is best constrained at a critical integration angle $\alpha_c=2r_\mathrm{half}/d$ (where $r_\mathrm{half}$ is the half light radius and $d$ is the distance to the dwarf) and we estimate the corresponding sensitivity of $\gamma$-ray observatories; (v) The ‘classical’ dSphs can be grouped into three categories: well-constrained and promising (Ursa Minor, Sculptor, and Draco), well-constrained but less promising (Carina, Fornax, and Leo I), and poorly constrained (Sextans and Leo II); (vi) Observations of classical dSphs with Fermi-LAT integrated over the mission lifetime are more promising than observations with the planned Cherenkov Telescope Array for DM particle mass $\lesssim\! 700$ GeV. However, even Fermi-LAT will [*not*]{} have sufficient integrated signal from the classical dwarfs to detect DM in the ‘vanilla’ Minimal Supersymmetric Standard Model. Both the Galactic centre and the ‘ultra-faint’ dwarfs are likely to be better targets and will be considered in future work. author: - | A. Charbonnier$^{1}$, C. Combet$^{2}$, M. Daniel$^{4}$, S. Funk$^{5}$, J.A. Hinton$^{2}$[^1], D. Maurin$^{6,1,2,7}$, C. Power$^{2,3}$, J. I. Read$^{2,11}$, S. Sarkar$^{8}$, M. G. Walker$^{9,10}$, M. I. Wilkinson$^{2}$\ $^1$Laboratoire de Physique Nucléaire et Hautes Energies, CNRS-IN2P3/Universités Paris VI et Paris VII, 4 place Jussieu, Tour 33, 75252 Paris Cedex 05, France\ $^2$Dept. of Physics and Astronomy, University of Leicester, Leicester, LE1 7RH, UK\ $^3$International Centre for Radio Astronomy Research, University of Western Australia, 35 Stirling Highway, Crawley, Western Australia 6009, Australia\ $^4$Dept. of Physics, Durham University, South Road, Durham, DH1 3LE, UK\ $^5$W. W. Hansen Experimental Physics Laboratory, Kavli Institute for Particle Astrophysics and Cosmology, Department of Physics and SLAC National Accelerator Laboratory, Stanford University, Stanford, CA 94305, USA\ $^6$Laboratoire de Physique Subatomique et de Cosmologie, CNRS/IN2P3/INPG/Université Joseph Fourier Grenoble 1,53 avenue des Martyrs, 38026 Grenoble, France\ $^7$Institut d’Astrophysique de Paris, UMR7095 CNRS, Université Pierre et Marie Curie, 98 bis bd Arago, 75014 Paris, France\ $^8$Rudolf Peierls Centre for Theoretical Physics, University of Oxford, 1 Keble Road, Oxford, OX1 3NP, UK\ $^9$Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge, CB3 0HA, UK\ $^{10}$Harvard-Smithsonian Center for Astrophysics, 60 Garden St., Cambridge, MA 02138, USA\ $^{11}$Institute for Astronomy, Department of Physics, ETH Zürich, Wolfgang-Pauli-Strasse 16, CH-8093 Zürich, Switzerland bibliography: - 'jbar\_project.bib' date: 'Accepted Xxxx. Received Xxxx; in original form Xxxx' title: | Dark matter profiles and annihilation in dwarf spheroidal galaxies: prospectives for present and future $\gamma$-ray observatories\ I. The classical dSphs --- \[firstpage\] astroparticle physics — (cosmology:) dark matter — Galaxy: kinematics and dynamics — $\gamma$-rays: general — methods: miscellaneous Introduction\[sec:intro\] ========================= The detection of $\gamma$-rays from dark matter (DM) annihilation is one of the most promising channels for indirect detection [@1978ApJ...223.1015G; @1978ApJ...223.1032S]. Since the signal goes as the DM density squared, the Galactic centre seems to be the obvious location to search for such a signal [@1987ApJ...313L..47S]. However, it is plagued by a confusing background of astrophysical sources . For this reason, the dwarf spheroidal galaxies (dSphs) orbiting the Milky Way have been flagged as favoured targets given their potentially high DM densities and small astrophysical backgrounds [@1990Natur.346...39L; @Evans:2003sc]. Despite the growing amount of kinematic data from the classical dSphs, the inner parts of their DM profiles remain poorly constrained and can generally accommodate both cored or cuspy solutions [e.g. @koch07b; @2007ApJ...669..676S; @2009ApJ...704.1274W]. There are two dSphs|Fornax and Ursa Minor|that show indirect hints of a cored distribution [@2003ApJ...588L..21K; @2006MNRAS.368.1073G]; however, in both cases the presence of a core is inferred based on a timing argument that assumes we are not catching the dSph at a special moment. Theoretical expectations remain similarly uncertain. Cusps are favoured by cosmological models that model the DM alone, assuming it is cold and collisionless [e.g. @navarro96]. However, the complex dynamical interplay between stars, gas and DM during galaxy formation could erase such cusps leading to cored distributions [e.g. @1996MNRAS.283L..72N; @2005MNRAS.356..107R; @2008Sci...319..174M; @2010ApJ...725.1707G; @2010Natur.463..203G; @2011arXiv1105.4050C]. Cores could also be an indication of other possibilities such as self-interacting dark matter [e.g. @2000PhRvD..62f3511H; @Moore:2000fp]. Knowledge of the inner slope of the DM profile is of critical importance as most of the annihilation flux comes from that region. Lacking this information, several studies have focused on the detectability of these dSphs by current $\gamma$-ray observatories such as the satellite-borne Fermi-LAT and atmospheric Cherenkov telescopes (ACTs) such as H.E.S.S., MAGIC and VERITAS, using a small sample of cusped and cored profiles (generally one of each). Most studies rely on standard core and cusp profiles fitted to the kinematic data of the dSph of interest . Other authors use a ‘cosmological prior’ from large scale cosmological simulations [e.g. @2010AdAst2010E..45K]. Both approaches may be combined, such as in @2007PhRvD..75h3526S and @2009JCAP...06..014M who rely partially on the results of structure formation simulations to constrain the inner slope and then perform a fit to the data to derive the other parameters. However such cosmological priors remain sufficiently uncertain that their use is inappropriate for guiding observational strategies. There have been only a few studies [e.g., @2009PhRvD..80b3506E] which have *not* assumed strong priors for the DM profiles. In this work, we revisit the question of the detectability of dark matter annihilation in the classical Milky Way dSphs, motivated by ambitious plans for next-generation ACTs such as the Cherenkov Telescope Array (CTA). We rely [*solely*]{} on published kinematic data to derive the properties of the dSphs, making minimal assumptions about the underlying DM distribution. Most importantly, we do not restrict our survey of DM profiles to those suggested by cosmological simulations. We also consider the effect of the spatial extent of the dSphs, which becomes important for nearby systems observed by background-limited instruments such as ACTs. This paper extends the earlier study of @letter_dsph which showed that there is a critical integration angle (twice the half-light radius divided by the dSph distance) where we can obtain a robust estimate of the $J$-factor (that parameterises the expected $\gamma$-ray flux from a dSph independently of the choice of dark matter particle model; see Section \[sec:method\]), regardless of the value of the central DM cusp slope $\gamma$. Here, we focus on the full radial dependence of the $J$-factor. We consider the effect of DM sub-lumps within the dSphs, discuss which dSphs are the best candidates for an observing programme, and examine the competitiveness of next-generation ACTs as dark matter probes. This paper is organised as follows. In Section \[sec:method\], we present a study of the annihilation $\gamma$-ray flux, focusing on which parameters critically affect the expected signal. In Section \[sec:detectability\], we discuss the sensitivity of present/future $\gamma$-ray observatories. In Section \[sec:mcmc\], we present our method for the dynamical modelling of the observed kinematics of stars in dSphs. In Section \[sec:results\], we derive DM density profiles for the classical dSphs using an MCMC analysis, from which the detection potential of future $\gamma$-ray observatories can be assessed. We present our conclusions in Section \[sec:conclusions\].[^2] This paper includes detailed analyses from both high-energy astrophysics and stellar dynamical modelling. To assist readers from these different fields in navigating the key sections, we suggest that those who are primarily interested in the high-energy calculations may wish to focus their attention on Sections \[sec:method\], \[sec:detectability\] and \[sec:results\] before moving to the conclusions. Readers from the dynamics community may instead prefer to read Sections \[sec:method\], \[sec:mcmc\] and \[sec:results\]. Finally, those who are willing to trust the underlying modelling should proceed to Section \[sec:results\] where our main results regarding the detectability of dSphs are presented in Figs. \[fig:JdSph\_Jdm\_bkgd\],  \[fig:Jd\_gamma\_fixed\],  \[fig:min\_detect\_sigmav\] and \[fig:mssm\]. The dark matter annihilation signal: key parameters {#sec:method} =================================================== The $\gamma$-ray flux --------------------- The $\gamma$-ray flux $\Phi_{\gamma}$ (photons cm$^{-2}$ s$^{-1}$ GeV$^{-1})$ from DM annihilations in a dSph, as seen within a solid angle $\Delta\Omega$, is given by (see Appendix \[app:defs\] for definitions and conventions used in the literature): $$\frac{\mathrm{d}\Phi_{\gamma}}{\mathrm{d}E_{\gamma}}(E_{\gamma},\Delta\Omega) = \Phi^{\rm pp}(E_{\gamma}) \times J(\Delta\Omega)\,, \label{eqn:dmanihil}$$ The first factor encodes the (unknown) particle physics of DM annihilations which we wish to measure. The second factor encodes the astrophysics [*viz.*]{} the l.o.s. integral of the DM density-squared over solid angle $\Delta\Omega$ in the dSph — this is called the ‘$J$-factor’. We now discuss each factor in turn. ### The particle physics factor {#sec:dNdE_xsec} The particle physics factor ($\Phi^{\rm pp}$) is given by: $$\Phi^{\rm pp}(E_{\gamma})\equiv \frac{\mathrm{d}\Phi_{\gamma}}{\mathrm{d}E_{\gamma}} = \frac{1}{4\pi}\frac{\langle\sigma_{\rm ann}v\rangle}{2m_{\chi}^{2}} \times \frac{dN_{\gamma}}{dE_{\gamma}}\;,$$ where $m_{\chi}$ is the mass of the DM particle, $\sigma_{\rm ann}$ is its self-annihilation cross-section and $\langle \sigma_{\rm ann} v\rangle$ the average over its velocity distribution, and $\mathrm{d}N_{\gamma}/\mathrm{d} E_{\gamma}$ is the differential photon yield per annihilation. A benchmark value is $\langle \sigma_\mathrm{ann} v\rangle\sim 3\times 10^{-26}\, {\rm cm}^3\, {\rm s}^{-1}$ [@1996PhR...267..195J], which would result in a present-day DM abundance satisfying cosmological constraints. Unlike the annihilation cross section and particle mass, the differential annihilation spectrum ($\mathrm{d}N_{\gamma}/\mathrm{d} E_{\gamma}(E_\gamma)$) requires us to adopt a specific DM particle model. We focus on a well-motivated class of models that are within reach of up-coming direct and indirect experiments: the Minimal Supersymmetric Standard Model (MSSM). In this framework, the neutralino is typically the lightest stable particle and therefore one of the most favoured DM candidates (see e.g. @2005PhR...405..279B). A $\gamma$-ray continuum is produced from the decay of hadrons (e.g. $\pi^0 \rightarrow \gamma\gamma$) resulting from the DM annihilation. Neutralino annihilations can also directly produce mono-energetic $\gamma$-ray lines through loop processes, with the formation of either a pair of $\gamma$-rays ($\chi\chi\rightarrow \gamma\gamma$; [@1997NuPhB.504...27B]), or a $Z^0$ boson and a $\gamma$-ray ($\chi\chi\rightarrow \gamma Z^0$; [@1998PhRvD..57.1962U]). We do not take into account such line production processes since they are usually sub-dominant and very model dependent [@2008JHEP...01..049B]. The differential photon spectrum we use is restricted to the continuum contribution and is written as: $$\frac{\mathrm{d} N_{\gamma}}{\mathrm{d} E_{\gamma}} (E_\gamma) = \sum_i b_i \, \frac{\mathrm{d} N_\gamma^i}{\mathrm{d} E_\gamma} (E_\gamma, m_\chi) \; , \label{eq:3contribs}$$ where the different annihilation final states $i$ are characterised by a branching ratio $b_i$. Using the parameters in @2004PhRvD..70j3529F, we plot the continuum spectra calculated for a 1 TeV mass neutralino in Fig.  \[fig:phi-susy\]. ![Differential spectra (multiplied by $x^2$) of $\gamma$-rays from the fragmentation of neutrino annihilation products (here for a DM particle mass of $m_\chi = 1$ TeV). Several different channels are shown, taken from @2004PhRvD..70j3529F and an average parametrisation @1998APh.....9..137B is marked by the black solid line; this is what we adopt throughout this paper. The black dashed line is the benchmark model BM4 [@2008JHEP...01..049B] which includes internal bremsstrahlung and serves to illustrate that very different spectra are possible. However, the example shown here is dominated by line emission and therefore highly model dependent; for this reason, we do not consider such effects in this paper.[]{data-label="fig:phi-susy"}](fig1.eps){width="\linewidth"} Apart from the $\tau^+\tau^-$ channel (dash-dotted line), all the annihilation channels in the continuum result in very similar spectra of $\gamma$-rays (dashed lines). For charged annihilation products, internal bremsstrahlung (IB) has recently been investigated and found to enhance the spectrum close to the kinematic cut-off (e.g., [@2008JHEP...01..049B]). As an illustration, the long-dashed line in Fig. \[fig:phi-susy\] corresponds to the benchmark configuration for a wino-like neutralino taken from @2008JHEP...01..049B. However, the shape and amplitude of this spectrum are strongly model dependent [@2009JCAP...01..016B] and, as argued in @2010PhRvD..81j7303C, this contribution is relevant only for models (and at energies) where the line contribution is dominant over the secondary photons. We wish to be as model-independent as possible, and so do not consider internal bremsstrahlung. In the remainder of this paper, all our results will be based on an [ *average*]{} spectrum taken from the parametrisation ([@1998APh.....9..137B], solid line in Fig. \[fig:phi-susy\]): $$\frac{\mathrm{d} N_{\gamma}}{\mathrm{d} E_{\gamma}} = \frac{1}{m_\chi}\frac{\mathrm{d} N_{\gamma}}{\mathrm{d} x} = \frac{1}{m_\chi} \frac{0.73 \, e^{-7.8x}}{x^{1.5}}\, , \label{eq:param-berg}$$ with $x \equiv E_{\gamma}/m_{\chi}$. Finally, in order to be conservative in deriving detection limits, we also do not consider the possible ‘Sommerfeld enhancement’ of the DM annihilation cross-section [@2004PhRvL..92c1303H; @2005PhRvD..71f3528H].[^3] This depends inversely on the DM particle velocity, and thus requires precise modelling of the velocity distribution of the DM within the dSph; we will investigate this in a separate study. ### The J-factor The second term in Eq. (\[eqn:dmanihil\]) is the astrophysical [*J-factor*]{} which depends on the spatial distribution of DM as well as on the beam size. It corresponds to the l.o.s. integration of the DM density squared over solid angle $\Delta\Omega$ in the dSph: $$J = \int_{\Delta\Omega}\int \rho_{\rm DM}^2 (l,\Omega) \,dld\Omega. \label{eq:J}$$ The solid angle is simply related to the integration angle $\alpha_{\rm int}$ by $$\Delta\Omega = 2\pi\cdot(1-\cos(\alpha_{\rm int})) \,.$$ The $J$-factor is useful because it allows us to rank the dSphs by their expected $\gamma$-ray flux, independently of any assumed DM particle physics model. Moreover, the knowledge of the relative $J$-factors would also help us to evaluate the validity of any potential detection of a given dSph, because for a given particle physics model we could then scale the signal to what we should expect to see in the other dSphs. All calculations of $J$ presented in this paper were performed using the publicly available [CLUMPY]{} package (Charbonnier, Combet, Maurin, in preparation) which includes models for a smooth DM density profile for the dSph, clumpy dark matter sub-structures inside the dSph, and a smooth and clumpy Galactic DM distribution. [^4] ### DM profiles For the DM halo we use a generalised ($\alpha,\beta,\gamma$) Hernquist profile given by [@hernquist90; @1993MNRAS.265..250D; @zhao96]: $$\rho(r)=\rho_\mathrm{s}\biggl (\frac{r}{r_\mathrm{s}}\biggr )^{-\gamma}\biggl [1+\biggl (\frac{r}{r_\mathrm{s}}\biggr )^{\alpha}\biggr ]^{\frac{\gamma-\beta}{\alpha}}, \label{eq:hernquist1}$$ where the parameter $\alpha$ controls the sharpness of the transition from inner slope, $\lim_{r\rightarrow 0}\mathrm{d}\ln(\rho)/\mathrm{d}\ln(r)= -\gamma$, to outer slope $\lim_{r\rightarrow \infty}\mathrm{d}\ln(\rho)/\mathrm{d}\ln(r)= -\beta$, and $r_\mathrm{s}$ is a characteristic scale. In principle we could add an additional parameter in order to introduce an exponential cut-off in the profile of Eq (\[eq:hernquist1\]) to mimic the effects of tidal truncation, as proposed in e.g., the Aquarius [@2008MNRAS.391.1685S] or Via Lactea II [@2008Natur.454..735D] simulations. However, the freedom to vary parameters $r_s$, $\alpha$ and $\beta$ in Eq (\[eq:hernquist1\]) already allows for density profiles that fall arbitrarily steeply at large radius. Moreover, given that our MCMC analysis later shows that the outer slope $\beta$ is unconstrained by the available data and that the J-factor does not correlate with $\beta$, we choose not to add further shape parameters. For profiles such as $\gamma\geq1.5$, the quantity $J$ from the inner regions diverges. This can be avoided by introducing a saturation scale $r_{\rm sat}$, that corresponds physically to the typical scale where the annihilation rate $[\langle \sigma v\rangle \rho(r_{\rm sat})/m_\chi]^{-1}$ balances the gravitational infall rate of DM particles $(G\bar{\rho})^{-1/2}$ [@1992PhLB..294..221B]. Taking $\bar{\rho}$ to be about 200 times the critical density gives $$\rho_{\rm sat}\approx 3 \times 10^{18} \left(\frac{m_\chi}{100~\rm GeV}\right) \times \left( \frac{10^{-26} {\rm cm}^3~{\rm s}^{-1}} {\langle \sigma v\rangle}\right) M_{\odot}~{\rm kpc}^{-3}. \label{eq:rho_sat}$$ The associated saturation radius is given by $$r_{\rm sat} = r_\mathrm{s} \left(\frac{\rho_\mathrm{s}}{\rho_{\rm sat}}\right)^{1/\gamma}\ll r_\mathrm{s}\;. \label{eq:r_sat}$$ This limit is used for all of our calculations. Motivation for a generic approach and reference models ------------------------------------------------------ In many studies, the $\gamma$-ray flux (from DM annihilations) is calculated using the point-source approximation [e.g., @2006PhRvD..73f3510B; @2010AdAst2010E..45K]. This is valid so long as the inner profile is steep, in which case the total luminosity of the dSph is dominated by a very small central region. However, if the profile is shallow and/or the dSph is nearby, the effective size of the dSph on the sky is larger than the point spread function (PSF) of the detector, and the point-source approximation breaks down. For upcoming instruments and particularly shallow DM profiles, the effective size of the dSph may even be comparable to the field of view of the instrument. This difference in the radial extent of the signal does matter in terms of detection (see Section \[sec:detectability\]). Hence we do not assume that the dSph is a point-source but rather derive sky-maps for the expected $\gamma$-ray flux. ### Illustration: a cored vs cusped profile Fig. \[fig:cumul\_generic\_norm\] shows $J$ as a function of the integration angle $\alpha_{\rm int}$ for a dSph at 20kpc (looking towards its centre). The black solid line is for a cored profile ($\gamma=0)$ and the green dashed line is for a cuspy profile ($\gamma=1.5$); both are normalised to unity at $\alpha_{\rm int}=5^{\circ}$. For the cuspy profile, $\sim\!100\%$ of the signal is in the first bin while for the cored profile, $J$ builds up slowly with $\alpha_{\rm int}$, and $80\%$ of the signal (w.r.t. the value for $\alpha_{\rm int}=5^{\circ}$) is obtained for $\alpha^{80\%}\approx 3^\circ$. ![Finite size effects: $J$ as a function of the integration angle $\alpha_{\rm int}$ for a dSph at 20kpc (pointing towards the centre of the dSph). The black solid line is for a cored profile ($\gamma=0)$ and the green dashed line is for a cuspy profile ($\gamma=1.5$); both are normalised to unity at $\alpha_{\rm int}=5^{\circ}$.[]{data-label="fig:cumul_generic_norm"}](fig2.eps){width="\linewidth"} This is also indicated by the symbols which show the contribution of DM [*shells*]{} in two angular bins — whereas the (green) hollow squares have a spiky distribution in the first bin ($\gamma=1.5$), the (black) filled circles ($\gamma=0)$ show a very broad distribution for $J$. The integration angle required to have a sizeable fraction of the signal depends on several parameters: the distance $d$ of the dSph, the inner profile slope $\gamma$, and the scale radius $r_\mathrm{s}$. Small integration angles are desirable since this minimises contaminating background $\gamma$-ray photons and maximises the signal to noise. Thus the true detectability of a dSph will depend on its spatial extent on the sky, and thus also on $d$, $\gamma$ and $r_\mathrm{s}$. ### Generic dSph profiles {#sec:generic} As will be seen in Section \[sec:results\], the errors on the density profiles of the Milky Way dSphs are large, making it difficult to disentangle the interplay between the key parameters for detectability. Hence we select some ‘generic profiles’ to illustrate the key dependencies. The most constrained quantity is the mass within the half-light radius $r_\mathrm{\rm half}$ (typically a few tenths of a kpc), as this is where most of the kinematic data come from [e.g., @2009ApJ...704.1274W; @2010MNRAS.406.1220W]. For the classical Milky Way dSphs, the typical mass within $r_\mathrm{\rm half} \sim 300$pc is found to be $M_{300} \sim 10^7 M_\odot$ [@2008Natur.454.1096S — see also the bottom panel of Fig. \[fig:cor\_par5\]]. If the DM scale radius is significantly larger than this ($r_\mathrm{s} \gg r_\mathrm{\rm half}$) and the inner slope $\gamma\gtrsim 0.5$, we can approximate the enclosed mass by: $$M_\mathrm{300} \simeq \frac{4 \pi \rho_\mathrm{s} r_\mathrm{s}^3}{3-\gamma} \left(\frac{300\,\mathrm{pc}}{r_\mathrm{s}}\right)^{3-\gamma} \approx 10^7 M_\odot \,. \label{eq:M300}$$ The parameter $\rho_\mathrm{s}$ is thus determined completely by the above condition, if we choose the scale radius $r_\mathrm{s}$ and cusp slope $\gamma$. Table \[tab:generic\_model\_definition\] shows, for several values of $r_\mathrm{s}$ and $\gamma$, the value required for $\rho_\mathrm{s}$ to obtain the assumed $M_{300}$ mass. We fix $\alpha=1, \beta=3$ but our results are not sensitive to these choices. [^5] The values of $r_\mathrm{s}$ are chosen to encompass the range of $r_\mathrm{s}$ found in the MCMC analysis (see Section \[sec:results\]). To further convince ourselves that the generic profiles we present here are a possible description of real dSphs, we checked (not shown) using typical stellar profiles and properties of these objects (i.e., halflight radius of a few 100 pc), that a flat $\sim 10$ km s$^{-1}$ velocity dispersion profile within the error bars is recovered. We also study below the effect of moving these dSphs from a distance of 10kpc to 300kpc, corresponding to the typical range covered by these objects. -------------------------------------------- ------ ------ ------- -- $\gamma~~\backslash~~r_\mathrm{s}$ \[kpc\] 0.10 0.50 1.0 0.00 224 25.8 16.02 0.25 196 18.6 10.22 0.50 170 13.4 6.47 0.75 146 9.5 4.06 1.00 125 6.7 2.52 1.25 106 4.7 1.54 1.50 88 3.2 0.92 -------------------------------------------- ------ ------ ------- -- : The required normalisation $\rho_\mathrm{s}$ to have $M_{300}=10^7 M_{\odot}$ for a sample of $(1,3,\gamma)$ profiles with varying scale radius $r_\mathrm{s}$.[]{data-label="tab:generic_model_definition"} ### Sub-structures within the dSph {#sec:sub_reference} Structure formation simulations in the currently favoured $\Lambda$CDM (cold DM plus a cosmological constant) cosmology find that DM halos are self-similar, containing a wealth of smaller ‘sub-structure’ halos down to Earth-mass halos [e.g. @2005Natur.433..389D]. However, as emphasised in the introduction, such simulations typically neglect the influence of the baryonic matter during galaxy formation. It is not clear what effect these have on the DM sub-structure distribution. For this reason, we adopt a more generic approach. We assess the importance of clumps using the following recipe:[^6] 1. we take a fraction $f=20\%$ of DM mass in the form of clumps; 2. the spatial distribution of clumps follows the smooth one; 3. the clump profiles are calculated [*à la*]{} @2001MNRAS.321..559B (hereafter B01), i.e. an ‘NFW’ profile [@navarro96] with concentration related to the mass of the clumps. 4. the clump mass distribution is $\propto M^{-a}$ ($a=-1.9$), within a mass range $M_{\rm min}-M_{\rm max} = [10^{-6}-10^{6}]~M_\odot$. Although these parameters are very uncertain, they allow us to investigate the impact of substructures on the J-factor. They are varied within reasonable bounds in Section \[sec:boost\] (and Appendix \[app:boost\]) to determine whether the sub-clump contribution can boost the signal. Note that a 20% clump mass fraction is about twice as large as the fraction obtained from numerical simulations (see, e.g., @2008MNRAS.391.1685S). This generous fraction does not affect our conclusions, as discussed below. $J_{\rm sm}$ and $J_{\rm subcl}$ for the generic models {#sec:gen_dep} ------------------------------------------------------- As an illustration, we show in Fig. \[fig:generic\_2D\_Jsm\_and\_subcl\] one realisation of the 2D distribution of $J$ from a generic core profile ($\gamma=0$) with $r_{s}=1$ kpc (sub-clump parameters are as described in Section \[sec:sub\_reference\]). The dSph is at $d=100$ kpc. We note that our consideration of a $\gamma=0$ smooth component with NFW sub-clumps is plausible if, e.g., baryon-dynamical processes erase cusps in the smooth halo but cannot do so in the sub-subhalos. The total $J$ is the sum of the smooth and sub-clump distributions. The centre is dominated by the smooth component, whereas some graininess appears in the outskirts of the dSph. ![2D view ($x$ and $y$ axis are in degrees) of $J$ for the generic dSph with $\gamma=0$ and $r_\mathrm{s}=1$ kpc at $d=100$ kpc ($M_{300} = 10^7 M_\odot$). The sub-clumps are drawn from the reference model described in Section \[sec:sub\_reference\], i.e. f=20%, sub-clump distribution follows smooth, and sub-clump inner profiles have NFW with B01 concentration. From top to bottom panel: $\alpha_{\rm int}=0.1^\circ$, $0.05^\circ$, and $0.01^\circ$. For the sake of comparison, the same colour scale is taken for the three integration angles ($J$ is in units of $M_\odot^2$ kpc$^{-5}$).[]{data-label="fig:generic_2D_Jsm_and_subcl"}](fig3a.eps "fig:"){width="\linewidth"} ![2D view ($x$ and $y$ axis are in degrees) of $J$ for the generic dSph with $\gamma=0$ and $r_\mathrm{s}=1$ kpc at $d=100$ kpc ($M_{300} = 10^7 M_\odot$). The sub-clumps are drawn from the reference model described in Section \[sec:sub\_reference\], i.e. f=20%, sub-clump distribution follows smooth, and sub-clump inner profiles have NFW with B01 concentration. From top to bottom panel: $\alpha_{\rm int}=0.1^\circ$, $0.05^\circ$, and $0.01^\circ$. For the sake of comparison, the same colour scale is taken for the three integration angles ($J$ is in units of $M_\odot^2$ kpc$^{-5}$).[]{data-label="fig:generic_2D_Jsm_and_subcl"}](fig3b.eps "fig:"){width="\linewidth"} ![2D view ($x$ and $y$ axis are in degrees) of $J$ for the generic dSph with $\gamma=0$ and $r_\mathrm{s}=1$ kpc at $d=100$ kpc ($M_{300} = 10^7 M_\odot$). The sub-clumps are drawn from the reference model described in Section \[sec:sub\_reference\], i.e. f=20%, sub-clump distribution follows smooth, and sub-clump inner profiles have NFW with B01 concentration. From top to bottom panel: $\alpha_{\rm int}=0.1^\circ$, $0.05^\circ$, and $0.01^\circ$. For the sake of comparison, the same colour scale is taken for the three integration angles ($J$ is in units of $M_\odot^2$ kpc$^{-5}$).[]{data-label="fig:generic_2D_Jsm_and_subcl"}](fig3c.eps "fig:"){width="\linewidth"} In this particular configuration, the ‘extended’ signal from the core profile, when integrated over a very small solid angle, could be sub-dominant compared with the signal of NFW sub-clumps that it hosts. The discussion of cross-constraints between detectability of sub-halos of the Galaxy vs. sub-clumps in the dSph is left for a future study. In the remainder of the paper, we will replace for simplicity the calculation of $J_{\rm subcl}(\alpha_{\rm int})$ by its mean value, as we are primarily interested in ‘unresolved’ observations. Hence clumps are not drawn from their distribution function, but rather $\langle J_{\rm subcl}\rangle$ is calculated from the integration of the spatial and luminosity (as a function of the mass) distributions (see Appendix \[sec:sub-clumps\]). ### Radial dependence $J(\theta)$ {#sec:radial_dep} The radial dependence of $J$ is shown in Fig. \[fig:generic\_Jtheta\_sm\_and\_subcl\] for four values of $\gamma$ (for an integration angle $\alpha_{\rm int}=0.01^\circ$). The dashed lines show the result for the smooth distribution; the dotted lines show the sub-clump contribution; and the solid lines are the sum of the two. The peak of the signal is towards the dSph centre. As long as the distribution of clumps is assumed to follow the smooth one, regardless of the value of $\gamma$, the quantity $(1-f)^{2}J_{\rm sm}(0)$ always dominates (at least by a factor of a few) over $\langle J_{\rm subcl}(0)\rangle$. (Recall that in our generic models, all dSphs have the same $M_{300}$.) The scatter in $J_{\rm tot}(0)$ is about 4 orders of magnitude for $\gamma\in[0.0-1.5]$, but only a factor of 20 for $\gamma\in[0.0-1.0]$. Beyond a few tenths of degrees, $\langle J_{\rm subcl}\rangle$ dominates. The crossing point depends on a combination of the clump mass fraction $f$, $\gamma$, $r_\mathrm{s}$, $d$, $\alpha_{\rm int}$. The dependence of $J$ on the two latter parameters are discussed in Appendix \[app:dep\_generic\]. The radial dependence is as expected: the smooth contribution decreases faster than that of the sub-clump one, because the signal is proportional to the squared spatial distribution in the first case, but directly proportional to the spatial distribution in the second case. Halving $f$ to match the fraction from N-body simulations would have a  25% effect on $(1-f)^2J_{\rm sm}$, but decrease $J_{\rm subcl}$ by a factor $4$, so that the cross-over between the two components would occur at a larger angle in Fig. \[fig:generic\_Jtheta\_sm\_and\_subcl\]. ![$J$ as a function of the angle $\theta$ away from the dSph centre for a dSph at $100$ kpc with $r_\mathrm{s}=0.5$ kpc ($\rho_\mathrm{s}$ is given in Table \[tab:generic\_model\_definition\]). The integration angle is $\alpha_{\rm int}=0.01^\circ$. For the four inner slope values $\gamma$, the various contributions to $J$ are shown as solid (total), dashed (smooth), and dotted lines (sub-clumps).[]{data-label="fig:generic_Jtheta_sm_and_subcl"}](fig4.eps){width="\linewidth"} ### Boost factor {#sec:boost} Whether or not the signal is boosted by the sub-clump population is still debated in the literature . As underlined in the previous sections, the sub-clump contribution towards the dSph centre never dominates over the smooth one if the spatial profile of the sub-clumps follows that of the smooth distribution, and if the integration angle remains below some critical angle discussed below. Let us first define properly the parameters with respect to which this boost is calculated, as there is sometimes some confusion about this. Here, we define it with respect to the integration angle $\alpha_{\rm int}$ (the pointing direction is still towards the dSph centre): $$B(\alpha_{\rm int}) \equiv \frac{ (1-f)^2 J_{\rm sm}(\alpha_{\rm int}) + J_{\rm subcl}(\alpha_{\rm int})}{J_{\rm sm}(\alpha_{\rm int})}\,. \label{eq:fig:generic_boost}$$ In most studies, the boost has been calculated by integrating out to the clump boundary (i.e., $\alpha_{\rm int}^{\rm all} = R_{\rm vir}/d$). But the boost depends crucially on $\alpha_{\rm int}$ (the radial dependence of the smooth and sub-clump contributions differ, see Section \[sec:radial\_dep\]). ![Boost factor as a function of $\alpha_{\rm int} \times (d/100$ kpc) for profiles sub-clumps [*follow*]{} smooth (see Section \[sec:sub\_reference\]): the dSph is at $d=100$ kpc (lines) or $d=10$ kpc (symbols).[]{data-label="fig:generic_boost_default"}](fig5.eps){width="\linewidth"} We plot in Fig. \[fig:generic\_boost\_default\] the boost for different inner slopes $\gamma$, where a direct consequence of Eq. (\[eq:rescale\_Jalpha\]) is the $\alpha_{\rm int}\times d$ rescaling. For $r_\mathrm{s}\lesssim 0.1$ kpc (regardless of $\gamma$), or for $\gamma\gtrsim1.5$ (regardless of $r_{\rm s}$), the signal is never boosted. [^7] For small enough $\alpha_{\rm int}$, $B$ is smaller than unity, and if $\gamma$ is steep enough, $B\approx(1-f)^2$. For large values, a plateau is reached as soon as $\alpha_{\rm int}d\gtrsim R_{\rm vir}$ (taken to be 3 kpc here). In between, the value of the boost depends on $r_\mathrm{s}$ and $\gamma$ of the smooth component. Going beyond this qualitative description is difficult, as the toy model formulae of Appendix \[sec:sub-clumps\] gives results correct to only a factor of $\sim\!2$ (which is inadequate to evaluate the boost properly). To conclude, the maximum value for sub-clump [*follows*]{} smooth is $\lesssim 2$, and this value is reached only when integrating the signal out to $R_{\rm vir}/d$. The boost could still be increased by varying the sub-clump properties (e.g., taking a higher concentration). Conversely, if dynamical friction has caused the sub-clump population to become much more centrally concentrated than the smooth component, then the boost is decreased. This is detailed in Appendix \[app:boost\]. For the most realistic configurations, there is *no* significant boost when a clump mass fraction f = 20% is used. Naturally this result is even more true for the smaller $f$ found in N-body simulations so we disregard the boost for the rest of this paper and consider only the smooth contribution. Sensitivity of present/future $\gamma$-ray observatories {#sec:detectability} ======================================================== Major new ground-based $\gamma$-ray observatories are in the planning stage, with CTA [@2010arXiv1008.3703C] and AGIS [@agis:website] as the main concepts. As the designs of these instruments are still evolving, we adopt here generic performance curves (described below), close to the stated goals of these projects. For the Large Area Telescope (LAT) of the Fermi $\gamma$-ray satellite, the performance for 1 year observations of point-like, high Galactic latitude sources is known [@fermi:website], but no information is yet available for longer exposures or for extended objects. We therefore adopt a toy likelihood-based model for the Fermi sensitivity, tuned to reproduce the 1 year point-source curves. We note that whilst this approach results in approximate performance curves for both the ground- and space-based instruments, it captures the key differences (in particular the differences in collection area and angular resolution) and illustrates the advantages and limitations of the two instrument types, as well as the prospects for the discovery of DM annihilation in dSphs within the next decade. Detector models --------------- The sensitivity of a major future $\gamma$-ray observatory based on an array of Cherenkov Telescopes (FCA in the following, for ‘Future Cherenkov Array’) is approximated based on the point-source differential sensitivity curve (for a $5\sigma$ detection in 50 hours of observations) presented by @2008ICRC....3.1469B. Under the assumption that the angular resolution of such a detector is a factor 2 better than HESS [@2008ApJ...679.1299F] and has the same energy-dependence, and that the effective collection area for $\gamma$-rays grows from $10^{4}$m$^2$ at 30 GeV to 1 km$^2$ at 1 TeV, the implied cosmic-ray (hadron and electron) background rate per square degree can be inferred and the sensitivity thus adapted to different observation times, spectral shapes and source extensions. Given that the design of instruments such as CTA are not yet fixed, we consider that such a simplified response, characterised by the following functions is a useful tool to explore the capabilities of a generic next-generation instrument: $${\rm LS} = -13.1-0.33\,X+0.72\,X^{2},$$ $${\rm LA} = 6 + 0.46\,X - 0.56\,X^{2},$$ $$\psi_{68} = 0.038+\exp{-(X+2.9)/0.61},$$ where $$X = \log_{10} {\rm (Photon Energy / TeV)},$$ LS = log$_{10}$(Differential Sensitivity/ergcm$^{-2}$s$^{-1}$), LA = log$_{10}(\mathrm{Effective Area}/ {\rm m}^{2})$, and $\psi_{68}$ is the 68% containment radius of the point-spread-function (PSF) in degrees. For the Fermi detector a similar simplified approach is taken, the numbers used below being those provided by @fermi:website. The effective area changes as a function of energy and incident angle to the detector, reaching a maximum of $\approx 8000\,\mathrm{cm}^2$. The effective time-averaged area is then $\epsilon A\Omega/4\pi$ and the data-taking efficiency $\epsilon \approx 0.8$ (due to instrument dead-time and passages through the South Atlantic Anomaly). The point spread function again varies as a function of energy (with a much smaller dependence as a function of incidence angle) varying from 10 degrees to a few tenths of a degree over the LAT energy range. A rate of $1.5 \times 10^{-5}$ cm$^{-2}$ s$^{-1}$ sr$^{-1}$ ($>100$ MeV) and a photon index of 2.1 are assumed for the background. The sensitivity is then estimated using a simplified likelihood method which provides results within 20% of the sensitivity for a one year observation of a point-like source given by @fermi:website. Whilst both detector responses are approximate, the comparison is still useful. Our work incorporates several key aspects not considered in earlier studies, including the strong energy dependence of the angular resolution of both ground and space based instruments in the relevant energy range of 1 GeV to 1 TeV and hence the energy-dependent impact of the angular size of the target region. Relative performance for generic halos {#sec:RelativePerformance} -------------------------------------- ![The cone angle encompassing 80% of the annihilation flux at as function of the inner slope $\gamma$. Several different values of $r_\mathrm{s}$ and distance $d$ are shown for each $\gamma$, all scaled by (1 kpc/$r_\mathrm{s}$ and 100 kpc/$d$). The best-fit curve is also shown, corresponding to Eq. (\[eq:angsiz\]). []{data-label="fig:alpha_80"}](fig6.eps){width="\linewidth"} Using the results from Section \[sec:generic\] and the detector performance models defined above we can begin to investigate the sensitivity of future ACT arrays and the Fermi-LAT detector (over long observation times) to DM annihilation in dSphs. The detectability of a source depends primarily on its flux, but also on its angular extent. The impact of source extension on detectability is dealt with approximately (in each energy bin independently) by assuming that the opening angle of a cone which incorporates 80% of the signal is given by $$\theta_{80} = \sqrt{\psi_{80}^{2}+\alpha_{80}^{2}}\,, \label{eq:PSF}$$ where $\psi_{80}=1.25\,\psi_{68}$ is assumed for the FCA and interpolated from values given for 68% and 95% containment for the LAT @fermi:website; here $\alpha_{80}$ is the 80% containment angle of the halo emission. The validity of this approximation (at the level of a few percent) has been tested (see Appendix \[app:PSF\]) by convolving realistic halo profiles with a double Gaussian PSF as found for HESS [@hesspsf]. An 80% integration circle is close to optimum for a Gaussian source on a flat background (in the background limited regime). Fig. \[fig:alpha\_80\] shows the 80% containment radius of the annihilation flux of generic halos as a function of the inner slope $\gamma$. This result can be parametrised as: $$\alpha_{80} = 0.8^{\circ}\,(1-0.48\gamma-0.137\gamma^2) \left(\!\frac{r_\mathrm{s}}{\rm 1\,kpc}\!\right)\left(\!\frac{d}{\rm 100\,kpc}\!\right)^{-1}. \label{eq:angsiz}$$ It is clear that for a broad range of $d$, $\gamma$ and $r_\mathrm{s}$ the characteristic angular size of the emission region is [*larger*]{} than the angular resolution of the instruments under consideration. It is therefore critical to assess the performance as a function of the angular size of the dSph as well as the mass of the annihilating particle. ![ Approximate sensitivities of Fermi-LAT (blue lines), HESS (black lines) and the FCA described above (red lines) to a generic halo with $J=10^{12}~M_{\odot}^2$ kpc$^{-5}$, as a function of the mass of the annihilating particle and for the annihilation spectrum of Eq. (\[eq:param-berg\]). [**Top**]{}: The impact of observation time is illustrated: dashed lines give the 1 year and 20 hour sensitivities for Fermi and FCA/HESS respectively while the solid lines refer to 10 year (200 hour) observations. [ **Middle**]{}: the impact of analysis methods is considered for 5 year (100 hour) observations using Fermi (FCA). Solid lines show likelihood analyses in which the mass and spectrum of the annihilating particle are known in advance, while dashed and dotted lines show simple integral flux measurements above fixed thresholds of 1 GeV (dashed) and 100 GeV (dotted). Note that the 1 GeV cut implies accepting all events for the FCA (where the trigger threshold is $\approx$20 GeV). [**Bottom**]{}: the impact of the angular extension of target sources, as given by the halo profile in Fig. \[fig:alpha\_80\] is illustrated. The solid lines reproduce the likelihood case from the middle panel for a point-like source, and with values of $\alpha_{80}$ of 0.1$^{\circ}$ (dashed) and 1$^{\circ}$ also shown.[]{data-label="fig:sensitivity"}](fig7a.eps "fig:"){width="0.97\linewidth"} ![ Approximate sensitivities of Fermi-LAT (blue lines), HESS (black lines) and the FCA described above (red lines) to a generic halo with $J=10^{12}~M_{\odot}^2$ kpc$^{-5}$, as a function of the mass of the annihilating particle and for the annihilation spectrum of Eq. (\[eq:param-berg\]). [**Top**]{}: The impact of observation time is illustrated: dashed lines give the 1 year and 20 hour sensitivities for Fermi and FCA/HESS respectively while the solid lines refer to 10 year (200 hour) observations. [ **Middle**]{}: the impact of analysis methods is considered for 5 year (100 hour) observations using Fermi (FCA). Solid lines show likelihood analyses in which the mass and spectrum of the annihilating particle are known in advance, while dashed and dotted lines show simple integral flux measurements above fixed thresholds of 1 GeV (dashed) and 100 GeV (dotted). Note that the 1 GeV cut implies accepting all events for the FCA (where the trigger threshold is $\approx$20 GeV). [**Bottom**]{}: the impact of the angular extension of target sources, as given by the halo profile in Fig. \[fig:alpha\_80\] is illustrated. The solid lines reproduce the likelihood case from the middle panel for a point-like source, and with values of $\alpha_{80}$ of 0.1$^{\circ}$ (dashed) and 1$^{\circ}$ also shown.[]{data-label="fig:sensitivity"}](fig7b.eps "fig:"){width="0.97\linewidth"} ![ Approximate sensitivities of Fermi-LAT (blue lines), HESS (black lines) and the FCA described above (red lines) to a generic halo with $J=10^{12}~M_{\odot}^2$ kpc$^{-5}$, as a function of the mass of the annihilating particle and for the annihilation spectrum of Eq. (\[eq:param-berg\]). [**Top**]{}: The impact of observation time is illustrated: dashed lines give the 1 year and 20 hour sensitivities for Fermi and FCA/HESS respectively while the solid lines refer to 10 year (200 hour) observations. [ **Middle**]{}: the impact of analysis methods is considered for 5 year (100 hour) observations using Fermi (FCA). Solid lines show likelihood analyses in which the mass and spectrum of the annihilating particle are known in advance, while dashed and dotted lines show simple integral flux measurements above fixed thresholds of 1 GeV (dashed) and 100 GeV (dotted). Note that the 1 GeV cut implies accepting all events for the FCA (where the trigger threshold is $\approx$20 GeV). [**Bottom**]{}: the impact of the angular extension of target sources, as given by the halo profile in Fig. \[fig:alpha\_80\] is illustrated. The solid lines reproduce the likelihood case from the middle panel for a point-like source, and with values of $\alpha_{80}$ of 0.1$^{\circ}$ (dashed) and 1$^{\circ}$ also shown.[]{data-label="fig:sensitivity"}](fig7c.eps "fig:"){width="0.97\linewidth"} Fig. \[fig:sensitivity\] shows the relative sensitivity of Fermi and an FCA within our framework as a function of the mass of the annihilating particle, adopting the annihilation spectrum given in Eq. (\[eq:param-berg\]), with the several panels illustrating different points. From Fig. \[fig:sensitivity\] top (the case of a point-like signal for different observation times) it is clear that Fermi-LAT has a considerable advantage for lower mass DM particles ($m_\chi\,\ll$ 1 TeV) on the timescale for construction of an FCA (i.e. over a 5-10 year mission lifetime) in comparison to a deep ACT observation of 200 hours. Furthermore, Fermi-LAT is less adversely affected by the angular extent of the target regions (see Fig. \[fig:sensitivity\] bottom), due to its modest angular resolution in the energy range where it is limited by background, meaning that the source extension is well matched to the PSF of the instrument. The middle panel of this figure illustrates the impact of different approaches to the analysis. In the case that there is a DM candidate inferred from the discovery of supersymmetry at the LHC (quite possible on the relevant timescale) a search optimised on an assumed mass and spectral shape can be made (solid curves). However, all instruments are less sensitive when a generic search is undertaken. Simple analyses using all the photon flux above a fixed energy threshold (arbitrarily set to reduce background) are effective only in a relatively narrow range of particle mass. For example keeping only $>$100 GeV photons works well for ACTs for 0.3-3 TeV particles; whereas keeping all photons $>$1 GeV works moderately well in the 0.1-0.2 TeV range, but is much less sensitive than the higher threshold cut over the rest of the candidate dark matter particle mass range. The features of these curves are dictated by the expected shape of the annihilation spectrum. From Eq. (\[eq:param-berg\]) the peak photon output (adopting the average spectrum for DM annihilation) occurs at an energy which is an order of magnitude below the particle mass – effective detection requires that this peak occurs within (or close to) the energy range of the instrument concerned. ![Relative DM annihilation detection sensitivity for a 100 hour FCA observation, as a function of dSph distance for different inner slopes $\gamma$ and with $r_{s}$ fixed to 1 kpc. The sensitivity for a realistic approach using $\theta_{80}$ is given relative to the sensitivity to a point-like source with the same flux. Larger values correspond to poorer performance (larger values of the minimum detectable flux). The assumed spectral shape is again as given by Eq. (\[eq:param-berg\]) with $m_{\rm \chi}=300$ GeV. This sensitivity ratio depends on the strategy used to estimate the background level at the dSph position. The dashed lines show the impact of using an annulus between 3.5$^{\circ}$ and 4.0$^{\circ}$ of the dSph centre as a background control region. The solid line assumes that the background control region lies completely outside the region of emission from the dSph. []{data-label="fig:rel_sensitivity"}](fig8.eps){width="\linewidth"} The total annihilation flux from a dSph increases at smaller distances as $1/d^{2}$ for fixed halo mass, making nearby dSphs attractive for DM detection. However, as Fig. \[fig:sensitivity\] shows, the increased angular size of such nearby sources raises the required detection flux. Fig. \[fig:rel\_sensitivity\] illustrates the reduction in sensitivity for an FCA with respect to a point-like source for generic dSph halos as a function of distance, for inner slopes, $\gamma$, of zero and one and with $r_\mathrm{s}$ fixed to 1kpc, relative to the assumption of the full annihilation signal and a point-like source. Even for $\gamma=1$, the point-like approximation leads to an order of magnitude overestimate of the detection sensitivity for nearby ($\sim\!20$ kpc) dSphs. A further complication is how to establish the level of background emission arising from the residual non-$\gamma$-ray background. A common method in ground-based $\gamma$-ray astronomy is to estimate this background from an annulus around the target source [see, e.g., @berge_bg]. The dashed lines in Fig. \[fig:rel\_sensitivity\] show the impact of estimating the background using an annulus between 3.5$^{\circ}$ and 4.0$^{\circ}$ from the target. This approach has a modest impact on sensitivity and is ignored in the following discussions as it reduces the detectable flux but also $\theta_{80}$ and leads to a small improvement in some cases. Jeans/MCMC analysis of dSph kinematics {#sec:mcmc} ====================================== dSph kinematics with the spherical Jeans equation {#sec:jeans} ------------------------------------------------- Extensive kinematic surveys of the stellar components of dSphs have shown that these systems have negligible rotational support [with the possible exception of the Sculptor dSph, see @2008ApJ...681L..13B]. If we assume that the dSphs are in virial equilibrium, then their internal gravitational potentials balance the random motions of their stars. In order to estimate dSph masses, we consider here the behaviour of dSph stellar velocity dispersion as a function of distance from the dSph centre (analogous to rotation curves of spiral galaxies). Specifically, we use the stellar kinematic data of @2009AJ....137.3100W for the Carina, Fornax, Sculptor and Sextans dSphs, the data of @mateo08 for the Leo I dSph, and data from Mateo et al. (in preparation) for the Draco, Leo II and Ursa Minor dSphs. @2009ApJ...704.1274W [W09 hereafter] have calculated velocity dispersion profiles from these same data under the assumption that l.o.s. velocity distributions are Gaussian. Here we re-calculate these profiles without adopting any particular form for the velocity distributions. Specifically, for a given dSph we divide the velocity sample into circular bins containing approximately equal numbers of member stars,[^8] and within each bin we estimate the second velocity moment (squared velocity dispersion) as: $$\langle \hat{V}^2\rangle = \frac{1}{N-1}\displaystyle\sum_{i=1}^N[(V_i-\langle \hat{V}\rangle)^2 - \sigma_i^2], \label{eq:vdisp}$$ where $N$ is the number of member stars in the bin. We hold $\langle V\rangle$ fixed for all bins at the median velocity over the entire sample. For each bin we use a standard bootstrap re-sampling to estimate the associated error distribution for $\langle \hat{V}^2\rangle$, which is approximately Gaussian. Fig. \[fig:momentprofiles\] displays the resulting velocity dispersion profiles, $\langle \hat{V}^2\rangle^{1/2}(R)$, which are similar to previously published profiles. {width="0.66\linewidth"} In order to relate these velocity dispersion profiles to dSph masses, we follow W09 in assuming that the data sample in each dSph a single, pressure-supported stellar population that is in dynamical equilibrium and traces an underlying gravitational potential dominated by dark matter. Implicit is the assumption that the orbital motions of stellar binary systems contribute negligibly to the measured velocity dispersions.[^9] Furthermore, assuming spherical symmetry, the mass profile, $M(r)$, of the DM halo relates to (moments of) the stellar distribution function via the Jeans equation: $$\frac{1}{\nu}\frac{d}{dr}(\nu \bar{v_r^2})+2\frac{\beta(r)\bar{v_r^2}}{r}=-\frac{GM(r)}{r^2}, \label{eq:jeans}$$ where $\nu(r)$, $\bar{v_r^2}(r)$, and $\beta_r\equiv\beta(r)\equiv 1-\bar{v_{\theta}^2}/\bar{v_r^2}$ describe the 3-dimensional density, radial velocity dispersion, and orbital anisotropy, respectively, of the stellar component. Projecting along the l.o.s., the mass profile relates to observable profiles, the projected stellar density $I(R)$ and velocity dispersion $\sigma_p(R)$, according to [@bt08 BT08 hereafter] $$\sigma_p^2(R)=\frac{2}{I(R)}\displaystyle \int_{R}^{\infty}\biggl (1-\beta_r\frac{R^2}{r^2}\biggr ) \frac{\nu \bar{v_r^2}r}{\sqrt{r^2-R^2}}\mathrm{d}r. \label{eq:jeansproject}$$ Notice that while we observe the projected velocity dispersion and stellar density profiles directly, the l.o.s. velocity dispersion profiles provide [*no*]{} information about the anisotropy, $\beta(r)$. Therefore we require an assumption about $\beta(r)$; here we assume $\beta = $ constant, allowing for nonzero anisotropy in the simplest way. For constant anisotropy, the Jeans equation has the solution (e.g., @mamon05): $$\nu\bar{v^2_r}=Gr^{-2\beta_r}\displaystyle\int_r^{\infty}s^{2\beta_r-2}\nu(s)M(s)\mathrm{d}s. \label{eq:jeanssolution}$$ We shall adopt parametric models for $I(R)$ and $M(r)$ and then find values of the parameters of $M(r)$ that, via Eqs. (\[eq:jeansproject\]) and (\[eq:jeanssolution\]), best reproduce the observed velocity dispersion profiles. ### Stellar Density {#subsec:stellar} Stellar surface densities of dSphs are typically fit by @plummer11, @king62 and/or @sersic68, profiles (e.g., @ih95). For simplicity, we adopt here the Plummer profile: $$I(R)=\frac{L}{\pi r_{\rm half}^2}\frac{1}{[1+R^2/r_{\rm half}^2]^2}, \label{eq:plummer}$$ which has just two free parameters: the total luminosity $L$ and the projected[^10] half-light radius $r_{\rm half}$. Given spherical symmetry, the Plummer profile implies a 3-dimensional stellar density (BT08) of: $$\nu(r)=-\frac{1}{\pi}\int_r^{\infty}\frac{\mathrm{d}I}{\mathrm{d}R}\frac{\mathrm{d}R}{\sqrt{R^2-r^2}}=\frac{3L}{4\pi r_{\rm half}^3}\frac{1}{[1+r^2/r_{\rm half}^2]^{5/2}}. \label{eq:abel}$$ Since we assume that DM dominates the gravitational potential at all radii (all measured dSphs have central mass-to-light ratios $\ga 10$, e.g., @mateo98), the value of $L$ has no bearing on our analysis. We adopt values of $r_{\rm half}$ (and associated errors) from Table 1 in the published erratum to W09; these data originally come from the star count study of @ih95. We have checked that a steeper outer slope or a steeper inner slope for the light profile leaves unchanged the conclusions (see Appendix \[app:light\_profile\]). ### Dark matter halo {#subsec:halomodel} For the DM halo we follow W09 in using a generalised Hernquist profile, as given by Eq. (\[eq:hernquist1\]). In terms of these parameters, i.e, the density $\rho_\mathrm{s}$ at scale radius $r_\mathrm{s}$, plus the (outer,transition,inner) slopes $(\alpha,\beta,\gamma)$, the mass profile is: $$\begin{aligned} M(r)=4\pi\displaystyle\int_{0}^{r}s^2\rho(s)\mathrm{d}s=\frac{4\pi \rho_\mathrm{s}r_\mathrm{s}^3}{3-\gamma}\biggl (\frac{r}{r_\mathrm{s}}\biggr )^{3-\gamma}\hspace{0.6in}\\ _2F_1\biggl [\frac{3-\gamma}{\alpha},\frac{\beta-\gamma}{\alpha};\frac{3-\gamma+\alpha}{\alpha};-\biggl (\frac{r}{r_\mathrm{s}}\biggr )^\alpha\biggr ],\nonumber \label{eq:hernquist2}\end{aligned}$$ where $_2F_1(a,b;c;z)$ is Gauss’ hypergeometric function. Eq. (\[eq:hernquist1\]) includes plausible halo shapes ranging from the constant-density ‘cores’ ($\gamma=0$) that seem to describe rotation curves of spiral and low-surface-brightness galaxies (e.g., @deblok10 and references therein) to the centrally divergent ‘cusps’ ($\gamma>0$) motivated by cosmological N-body simulations that model only the DM component. For $(\alpha,\beta,\gamma) = (1,3,1)$ Eq. (\[eq:hernquist1\]) is just the cuspy NFW [@navarro96; @navarro97] profile. Markov-Chain Monte Carlo Method {#subsec:mcmc} ------------------------------- For a given halo model we compare the projected (squared) velocity dispersion profile $\sigma^2_p(R)$ (obtained from Eq. \[eq:jeansproject\]) to the empirical profile $\langle \hat{V}^2\rangle(R)$ (displayed in Fig. \[fig:momentprofiles\]) using the likelihood function $$\zeta=\displaystyle \prod_{i=1}^N \frac{1}{\sqrt{2\pi\mathrm{Var}[\langle \hat{V}^2\rangle (R_i)]}}\exp\biggl [-\frac{1}{2}\frac{(\langle \hat{V}^2\rangle (R_i)-\sigma_p^2(R_i))^2}{\mathrm{Var}[\langle \hat{V}^2\rangle (R_i)]}\biggr ], \label{eq:likelihood}$$ where $\mathrm{Var}[\langle \hat{V}^2\rangle (R_i)]$ is the variance associated with the empirical mean square velocity, as estimated from our bootstrap re-sampling. In order to explore the large parameter space efficiently, we employ Markov-chain Monte Carlo (MCMC) techniques. That is, we use the standard Metropolis-Hastings algorithm [@metropolis53; @hastings70] to generate posterior distributions according to the following prescription: 1) from the current location in parameter space, $S_n$, draw a prospective new location, $S'$, from a Gaussian probability density centred on $S_n$; 2) evaluate the ratio of likelihoods at $S_n$ and $S'$; and 3) if $\zeta(S')/\zeta(S_n)\ge 1$, accept such that $S_{n+1}=S'$, else accept with probability $\zeta(S')/\zeta(S_n)$ and $S_{n+1}=S_n$ with probability $1-\zeta(S')/\zeta(S_n)$. In order to account for the observational uncertainty associated with the half-light radius adopted from @ih95, for each new point we scatter the adopted value of $r_{\rm half}$ by a random deviate drawn from a Gaussian distribution with standard deviation equal to the published error. This method effectively propagates the observational uncertainty associated with the half-light radius to the posterior distributions for our model parameters. Solutions of the Jeans equations are not guaranteed to correspond to physical models, as the associated phase-space distribution functions may not be everywhere positive. [@2006ApJ...642..752A] have derived a necessary relation between the asymptotic values of the logarithmic slope of the gravitational potential, the tracer density distribution and the velocity anisotropy at small radii. Models which do not satisfy this relation will not give rise to physical distribution functions. In terms of our parametrisation, this relation becomes $$\gamma_{\rm tracer} \gtrsim 2\beta_{\rm aniso}. \label{eq:anevans}$$ We therefore exclude from the Markov Chain those models which do not satisfy this condition. Because the Plummer profiles we use to describe dSph surface brightness profiles have $\gamma_{\rm tracer}=0$, this restriction implies $\beta_{\rm aniso}\la 0$. Given our assumption of constant velocity anisotropy, this disqualifies all radially anisotropic models. Relaxing this condition affects the results on the $J$-factors, but the difference is contained within their CLs (see Appendix \[app:light\_profile\]). For this procedure we use the adaptive MCMC engine CosmoMC [@lewis02]. [^11] Although it was developed specifically for analysis of cosmic microwave background data, CosmoMC provides a generic sampler that continually updates the probability density according to the parameter covariances in order to optimise the acceptance rate. For each galaxy and parametrisation we run four chains simultaneously, allowing each to proceed until the variances of parameter values across the four chains become less than 1% of the mean of the variances. Satisfying this convergence criterion typically requires $\sim 10^4$ steps for our chains. We then estimate the posterior distribution in parameter space using the last half of all accepted points (we discard the first half of points, which we conservatively assume corresponds to the ‘burn-in’ period). Detectability of Milky Way dSphs {#sec:results} ================================ This section provides our key results. For the benefit of readers who start reading here, we summarise our findings so far. In Section \[sec:method\], we focused on generic $(1,3,\gamma)$ profiles, to show that, most of the time, the sub-structure contribution is negligible, and to check that the only relevant dSph halo parameters are the density normalisation $\rho_\mathrm{s}$, the scale radius $r_\mathrm{s}$, and the inner slope $\gamma$ (because $J_{\rm dSph}\propto r_\mathrm{s}^{2\gamma}\times (\alpha_{\rm int}d)^{3-2\gamma}$, see also Appendix \[app:toyJ\]). In Section \[sec:detectability\], we provided the sensitivity of present and future $\gamma$-ray observatories, showing how it is degraded when considering ‘extended’ sources (e.g. a flat profile for close dSphs), and an instrument response that varies with energy. In Section \[sec:mcmc\], we presented our method to perform a Markov-Chain Monte Carlo analysis of the observed stellar kinematics in the 8 classical Milky Way dSphs under the assumptions of virial equilibrium, spherical symmetry, constant velocity anisotropy, and a Plummer light distribution. The analysis uses the observed velocity dispersion profiles of the dSphs to constrain their underlying dark matter halo potentials, parametrised using the five parameter models of Eq. (\[eq:hernquist1\]). {width="0.8\linewidth"} 6-parameter MCMC analysis|varying $\gamma$ {#subsec:gamma-free} ------------------------------------------ Our kinematic models have six free parameters, for which we adopt uniform priors over the following ranges: $$\begin{aligned} -\log_{10}[1-\beta_{\rm ani}]: [-1,+1];\\ \log_{10}[\rho_\mathrm{s}/(M_{\odot}\mathrm{pc}^{-3})]: [-10,+4];\\ \log_{10}[r_\mathrm{s}/\mathrm{pc}]: [0, 4];\\ \alpha: [0.5, 3];\\ \beta: [3, 7].\\ \gamma: [0, 2]~~{\rm or}~~[0, 1];\\ \label{eq:priors}\end{aligned}$$ The anisotropy parameter $\beta_{\rm ani}$ does not enter directly the profile/mass/J calculation, although it is of fundamental importance for the fit as it can correlate with the DM profile structure parameters (so with the mass and the $J$-factor). We have not checked explicitly the details of these correlations, but we have checked that restricting the range of possible $\beta_{\rm ani}$ does not significantly impact on the results for the $J$ calculation. Hence, we do not discuss this parameter further below. ### Parameter correlations Fig. \[fig:cor\_par6\] shows the marginalised probability density functions (PDFs) of the profile parameters and the joint distributions of pairs of parameters. The features of these plots are driven by the fact that most of the stellar kinematic data lie at radii of up to few hundred parsecs (see Fig. \[fig:momentprofiles\]). For instance, the outer slope $\beta$ is not at all constrained (i.e. the fit is insensitive to the value of $\beta$), because only tracers beyond a radius of $r\gtrsim 1$ kpc are sensitive to this parameter and these radii are sparsely sampled by the observations. The transition slope $\alpha$ and then the inner slope $\gamma$ are the two other least constrained parameters. In terms of best-fit models, as seen in Fig. \[fig:momentprofiles\], the match to kinematics data is equally good for varying $\gamma$ (black) models and models in which we fix the value to $\gamma=0$ (blue), or $\gamma=1$ (red). In the following, we will not discuss further the best-fit values. The more meaningful quantity, in the context of an MCMC analysis providing PDFs, is the [*median*]{} of the distribution. Several groups have shown recently that in a Jeans analysis, the observed flatness of dSph velocity dispersion profiles [@walker07b] leads to a constraint on $M(r_{\rm half})$—the mass enclosed within a sphere of radius $r_{\rm half}$—that is insensitive to assumptions about either anisotropy or the structural parameters of the DM halo [@2009ApJ...704.1274W; @2010MNRAS.406.1220W]. Using for the appropriate radius the mass estimate Eq. (\[eq:M300\]) and the above constraint leads to a relation between the profile parameters $$\log(\rho_\mathrm{s}) + \gamma \log(r_\mathrm{s}) \approx \mathrm{constant}.$$ This relation explains the approximately linear correlations between these parameters seen, for instance, in the bottom left panel of Fig. \[fig:cor\_par6\]. ### From $\rho(r)$ to $J(\alpha_{\rm int})$: uncertainty and impact of $\gamma_{\rm prior}$ Fig. \[fig:best\_CL\_profiles\] shows the density profile for Draco as recovered by our MCMC analysis. It is noticeable that the confidence limits are narrower for radial scales of a few hundreds pc|this is a common feature of the density profile confidence limits for all the dSphs we have considered. As discussed above, this is partly due to the fact that these are the radii at which the majority of the kinematic data lie. The least constrained $\rho(r)$ (less pronounced narrowing of the confidence limits) is that of Sextans, for which the range where useful data can be found is clearly the smallest compared to other dSphs (see Fig. \[fig:momentprofiles\]). The variation of the constraints on $\rho(r)$ as a function of radius impacts directly on the behaviour of $J$. Complications arise because it is the profile squared that is now integrated along a l.o.s. (given the integration angle $\alpha_{\rm int}$, see Eq. \[eq:J\]). The median value and 95% CL on $J$ as a function of the integration angle $\alpha_{\rm int}$ is plotted in Fig. \[fig:cumul\_draco\] (top), for two different priors on $\gamma_{\rm prior}$. The bottom panel gives the corresponding PDF for two integration angles. The prior has a strong impact on the result: the median (thick solid curves and large symbols – top panel) is changed by $\sim 50\%$ for $\alpha_{\rm int}\gtrsim 0.1^\circ$, but by a factor of ten for $\alpha_{\rm int}\sim 0.01^\circ$. However, the most striking feature is the difference between the CLs: for the prior $0\leq\gamma_{\rm prior}\leq2$, the typical uncertainty is 3 to 4 orders of magnitude (red dotted curves), whereas it is only $\lesssim$ than one order of magnitude for the prior $0\leq\gamma_{\rm prior}\leq1$ (blue dotted curves).[^12] The bottom panel of Fig. \[fig:cumul\_draco\] shows that $\log_{10}J$ has a long and flat tail (associated to large $\gamma$ values). This tail is responsible for the large upper limit of the $J$-factor CLs for $0\leq\gamma_{\rm prior}\leq2$. ![[**Top**]{}: $J(\alpha_{\rm int})$ for Draco as a function of the integration angle. Solid lines correspond to the median model, dotted lines to the 95% lower and upper CL. The to sets of curves correspond to two different $\gamma_{\rm prior}$ for the MCMC analysis on the same data. [**Bottom:**]{} PDF of the J-factor for $\alpha_{\rm int}=0.01^\circ$ (grey) and $\alpha_{\rm int}=0.1^\circ$ (black), when the range of the inner slope prior is $[0-1]$ (solid lines) or $[0-2]$ (dashed lines).[]{data-label="fig:cumul_draco"}](fig11a.eps "fig:"){width="\linewidth"} ![[**Top**]{}: $J(\alpha_{\rm int})$ for Draco as a function of the integration angle. Solid lines correspond to the median model, dotted lines to the 95% lower and upper CL. The to sets of curves correspond to two different $\gamma_{\rm prior}$ for the MCMC analysis on the same data. [**Bottom:**]{} PDF of the J-factor for $\alpha_{\rm int}=0.01^\circ$ (grey) and $\alpha_{\rm int}=0.1^\circ$ (black), when the range of the inner slope prior is $[0-1]$ (solid lines) or $[0-2]$ (dashed lines).[]{data-label="fig:cumul_draco"}](fig11b.eps "fig:"){width="\linewidth"} In Appendix \[app:biases2\], a detailed analysis of the impact of these two priors is carried on artificial data (for which the true profile is known). We find that the prior $0\leq \gamma_{\rm prior} \leq 2$ satisfactorily reconstructs $\rho(r)$ and $J(\alpha_{\rm int})$, i.e. the MCMC CLs bracket the true value. This is also the case when using the prior $0\leq \gamma_{\rm prior} \leq 1$. However, two important points are noteworthy: - this prior obviously performs better for $0\leq\gamma_{\rm true}\leq1$ profiles where it gives much tighter constraints on $J$; - for cuspier profiles (e.g., $\gamma_{\rm true}=1.5$), this prior succeeds slightly less (than the prior $0\leq\gamma_{\rm prior}\leq2$) in reconstructing $\rho(r)$, but it does surprisingly better on $J$ in terms of providing a value closer to the true one (see details and explanations in Appendix \[app:biases2\]). DM simulations and observations do not favour $\gamma>1$, although steeper profiles can still fit the kinematic data in a Jeans analysis. Indeed, the Aquarius simulations indicate values of $\gamma$ slightly smaller than 1, and although some recent simulations [@2010ApJ...723L.195I] have argued for cuspy profiles, this happens for micro-haloes only. Given that the $J$-factor for the cuspier profiles are only marginally more (or even less) reliable when using the prior $0\leq\gamma_{\rm prior}\leq2$, we restrict ourselves to the $0\leq \gamma_{\rm prior} \leq 1$ prior below. Note that other sources of bias exist. First, the reconstruction of $\rho(r)$ or $J(\alpha_{\rm int})$ is affected by the choice of binning used in the estimation of the empirical velocity dispersion profiles. Appendix \[app:binning\] shows that we obtain slightly different results when we apply our method to empirical velocity dispersion profiles calculated from the same raw kinematic data, but using different numbers of bins. We find that the effects of binning add an extra factor of a few uncertainty on $J$ for the least well measured (in terms of radial coverage) dSphs, for which more measurements are desirable. (On the other hand, Fornax and Sculptor are found to provide robust results against different binnings.) Second, we note that the analysis presented here uses a fixed profile for the light distribution which, when combined with our assumption of constant velocity anisotropy, restricts the possible halo profiles we can recover. Our constraints on $\rho(r)$ and $J(\alpha_{\rm int})$ are therefore sensitive to these assumptions [see, e.g., @2010MNRAS.408.2364S for an example of fitting the dSph kinematic data with cusped profiles when the light profile is also allowed to be cusped], although this does not change our conclusions (see Appendix \[app:light\_profile\] where different light profiles are used). This situation is set to change over the coming years as new distribution function-based models will permit constraints to be placed on the slope of the DM density profiles (Wilkinson et al. 2011, in prep.). ### Best constraints on $J$: median value and CLs ---------------------- --------- --------- --------- --------- --------- --------------------- -------------------------------------- ---------------------------------- ---------------------------------- ---------------------------------- -- dSph long. lat. d $2r_h$ $\phi$ $\alpha_\mathrm{c}$ $M_{300}$ $\log_{10}[J(0.01^\circ)]$ $\log_{10}[J(0.1^\circ)]$ $\log_{10}[J^\star(\alpha_c)]$ \[deg\] \[deg\] \[kpc\] \[kpc\] \[deg\] \[deg\] $[10^7 M_{\odot}]$ Ursa Minor$\!\!\!\!$ 105.0 +44.8 66 0.56 100.6 0.49 $1.54_{-0.21(-0.42)}^{+0.18(+0.33)}$ $10.5_{-0.6(-1.2)}^{+0.8(+1.5)}$ $11.7_{-0.3(-0.6)}^{+0.5(+0.8)}$ $12.0_{-0.1(-0.2)}^{+0.3(+0.5)}$ Sculptor 287.5 -83.2 79 0.52 88.0 0.38 $1.34_{-0.13(-0.23)}^{+0.12(+0.23)}$ $10.0_{-0.5(-0.8)}^{+0.5(+0.9)}$ $11.3_{-0.2(-0.3)}^{+0.2(+0.4)}$ $11.7_{-0.1(-0.1)}^{+0.1(+0.2)}$ Draco 86.4 +34.7 82 0.40 87.0 0.28 $1.22_{-0.14(-0.28)}^{+0.15(+0.28)}$ $9.8_{-0.5(-0.8)}^{+0.5(+0.9)}$ $11.2_{-0.2(-0.3)}^{+0.2(+0.4)}$ $11.6_{-0.1(-0.2)}^{+0.1(+0.2)}$ Sextans 243.5 +42.3 86 1.36 109.3 0.91 $0.61_{-0.31(-0.43)}^{+0.38(+0.96)}$ $9.4_{-1.2(-1.8)}^{+1.7(+2.9)}$ $10.7_{-0.8(-1.1)}^{+1.1(+1.9)}$ $11.1_{-0.4(-0.6)}^{+0.7(+1.5)}$ Carina 260.1 -22.2 101 0.48 99.2 0.27 $0.59_{-0.07(-0.14)}^{+0.10(+0.60)}$ $9.3_{-0.4(-0.8)}^{+0.3(+0.8)}$ $10.5_{-0.1(-0.2)}^{+0.2(+0.4)}$ $10.9_{-0.1(-0.1)}^{+0.1(+0.1)}$ Fornax 237.1 -65.7 138 1.34 102.9 0.56 $1.01_{-0.17(-0.28)}^{+0.30(+0.60)}$ $9.5_{-0.5(-0.8)}^{+0.5(+1.1)}$ $10.8_{-0.2(-0.3)}^{+0.2(+0.5)}$ $10.5_{-0.2(-0.4)}^{+0.3(+0.7)}$ LeoII 220.2 +67.2 205 0.30 107.2 0.08 $0.94_{-0.18(-0.29)}^{+0.26(+0.50)}$ $11.6_{-0.8(-1.5)}^{+0.8(+1.7)}$ $11.7_{-0.6(-0.9)}^{+0.7(+1.6)}$ $11.7_{-0.6(-0.9)}^{+0.7(+1.6)}$ LeoI 226.0 +49.1 250 0.50 117.1 0.11 $1.22_{-0.21(-0.36)}^{+0.24(+2.52)}$ $9.7_{-0.2(-0.5)}^{+0.3(+1.0)}$ $10.7_{-0.1(-0.2)}^{+0.1(+0.3)}$ $10.7_{-0.1(-0.2)}^{+0.1(+0.3)}$ ---------------------- --------- --------- --------- --------- --------- --------------------- -------------------------------------- ---------------------------------- ---------------------------------- ---------------------------------- -- \ $^\star$ Note that the values for $\log_{10}[J(\alpha_c)]$ differ from those quoted in @letter_dsph as the MCMC analysis is slightly different here. As validated by the simulated data, we are now able to provide robust (although possibly not the best achievable with current data) and model-independent constraints on $J(\alpha_{\rm int})$ for the 8 classical dSphs. The results are summarised in Table \[tab:res\_par6\] in terms of the median, and 68% and 95% CLs. The $J$-factor is calculated for $\alpha_{\rm int}=0.01^{\circ}$ (an angle slightly better than what can be achieved with FCA), $\alpha_{\rm int}=0.1^{\circ}$ (typical of the angular resolution of existing GeV and TeV $\gamma$-ray instruments), and for $\alpha_{c}=2r_\mathrm{half}/d$ (as proposed in @letter_dsph). We do not report the values of $\rho_\mathrm{s}$ and $r_\mathrm{s}$ as these vary across a large range|and therefore do not give additional useful information|nor the value of $\gamma$ as it is forced in the range $0\leq\gamma_{\rm prior}\leq1$ to give the least biased $J$ value. There is no simple way to provide unambiguously the best target, as their relative merit depends non trivially on their distance, their mass and the integration angle selected. As proposed in @letter_dsph, since the most robust constraint on $J$ is obtained for $\alpha_{\rm int}=\alpha_c$, having different integration angles for each dSph can be a good starting point to establish a relative ranking. The situation is complicated further for background-limited instruments such as CTA, as some loss of sensitivity can occur (see, e.g. Figure 4 of @letter_dsph). This is discussed, taking into account the full detail of the instruments, in Section \[sec:5.3\]. However, in this respect, the best target for future instrument may eventually become Leo II, which despite a quite large uncertainty outshines all other dSphs at $\alpha_{\rm int}=0.01^\circ$ (see also Fig. \[fig:JdSph\_Jdm\_bkgd\]). We note however that it is the dsph with the smallest amount of kinematic data at present (so it has the most uncertain $J$-factor). ### dSphs in the diffuse galactic DM signal: contrast The uncertainties in $J$ are illustrated from a different viewpoint in Fig.\[fig:JdSph\_Jdm\_bkgd\]. It shows, in addition to the mean, 68% and 98% CLs on the $J$s, the latitudinal dependence of the Galactic DM background (smooth and galactic clump contribution) for the same integration angle.[^13] For a typical present-day instrument resolution (integration angle $\alpha_{\rm int}\sim0.1^\circ$), we recover the standard result that the Galactic Centre outshines all dSphs. ![Galactic contributions to $J$ for the smooth (blue-dashed line), mean clump (red-dotted line) and sum (black-solid line) vs the angle from the Galactic centre. The symbols show $J$ for the dSphs, assuming a prior of $0 \leq \gamma_{\rm prior} \leq 1$ on the central DM slope. The central point corresponds to the median values, the solid bars to the 68% CLs, and the dotted bars to the 95% CLs. The integration angle is, from top to bottom $0.01^\circ$, $0.1^\circ$, and $1^\circ$. The Galactic contributions $J_{\rm sm}$ and $\langle J_{\rm subcl}\rangle$ scale as $\alpha_{\rm int}^2$, but $J_{\rm dSph}$ does not, changing the contrast of the dSphs w.r.t. to the DM Galactic background (see text for details).[]{data-label="fig:JdSph_Jdm_bkgd"}](fig12a.eps "fig:"){width="\linewidth"} ![Galactic contributions to $J$ for the smooth (blue-dashed line), mean clump (red-dotted line) and sum (black-solid line) vs the angle from the Galactic centre. The symbols show $J$ for the dSphs, assuming a prior of $0 \leq \gamma_{\rm prior} \leq 1$ on the central DM slope. The central point corresponds to the median values, the solid bars to the 68% CLs, and the dotted bars to the 95% CLs. The integration angle is, from top to bottom $0.01^\circ$, $0.1^\circ$, and $1^\circ$. The Galactic contributions $J_{\rm sm}$ and $\langle J_{\rm subcl}\rangle$ scale as $\alpha_{\rm int}^2$, but $J_{\rm dSph}$ does not, changing the contrast of the dSphs w.r.t. to the DM Galactic background (see text for details).[]{data-label="fig:JdSph_Jdm_bkgd"}](fig12b.eps "fig:"){width="\linewidth"} ![Galactic contributions to $J$ for the smooth (blue-dashed line), mean clump (red-dotted line) and sum (black-solid line) vs the angle from the Galactic centre. The symbols show $J$ for the dSphs, assuming a prior of $0 \leq \gamma_{\rm prior} \leq 1$ on the central DM slope. The central point corresponds to the median values, the solid bars to the 68% CLs, and the dotted bars to the 95% CLs. The integration angle is, from top to bottom $0.01^\circ$, $0.1^\circ$, and $1^\circ$. The Galactic contributions $J_{\rm sm}$ and $\langle J_{\rm subcl}\rangle$ scale as $\alpha_{\rm int}^2$, but $J_{\rm dSph}$ does not, changing the contrast of the dSphs w.r.t. to the DM Galactic background (see text for details).[]{data-label="fig:JdSph_Jdm_bkgd"}](fig12c.eps "fig:"){width="\linewidth"} The three panels illustrate the loss of contrast (signal from the dSph w.r.t. to the diffuse Galactic DM signal) as the integration angle is increased. This is understood as follows: the integrand appearing in Eqs. (\[eq:J2\]) and (\[eq:J3\]) is mostly insensitive to the l.o.s. direction a few tens of degree away from the Galactic centre, so that Eq. (\[eq:alphint2\_dep\]) holds, giving an $\alpha_{\rm int}^2$ dependence. For detectability (see also Sec \[sec:detectability\]), the naïve approach of maximising the integration angle (to maximise $J_{\rm dSph}$) must be weighed against the fact that an increased integration angle means more astrophysical $\gamma$-ray and cosmic-ray background. For large integration angles, dSphs also have poor contrast against the diffuse Galactic DM annihilation signal, indicating that the Galactic halo is a better target for any search on angular scales $\gtrsim$1 (see e.g. @2011PhRvL.106p1301A for such a search with H.E.S.S.). ### Comparison to other works Comparison between different works can be difficult as every author uses different definition, notations and units for the astrophysical factor. To ease the comparison, we provide in Appendix \[app:defs\] conversion factors between standard units (we also point out issues to be aware of when performing such comparisons). Below is a comparison to just a few of the works published on the subject, and only for the objects that these studies and the present one have in common: - The @Evans:2003sc values of $J/\Delta\Omega$ for Draco (with $\Delta\Omega=10^{-5}$ i.e. $\alpha_{\rm int}=0.1^\circ$) for all the profiles they explored (cored, $\gamma=0.5$, $\gamma=1$, $\gamma=1.5$) are larger (after correction by $\Delta\Omega$, given their definition of the astrophysical factor) than our 95% CL upper limit for this object shown in Table \[tab:res\_par6\]. The difference is probably related to our data set which is about twice as large as that used by @Evans:2003sc. - @2007PhRvD..75h3526S provide directly the $\gamma$-ray flux (i.e. including the particle physics term), so that we can only compare our respective rankings. These agree in general but for Sculptor we find a larger flux than Draco, conversely to these authors. - focused on Sextans, Carina, Draco and Ursa Minor. They found the latter to have the largest $J$ ($\Phi_{\rm cosmo}$ in their notation) of these 4 objects, followed by Draco, Carina and Sextans. But for the last two, this ranking is similar to ours. However, while their values of $J$ fall within our 68% (UMi, Sextans) or 95% (Carina) CL, their value for Draco is above our 95% CL upper limit. - @2009PhRvD..80b3506E also performed a statistical study on Draco and Ursa Minor, to determine their profiles from kinematic data and to derive the confidence levels on the J-factor. Given that their integration is performed on a slightly larger opening angle ($0.14^\circ$), our results appear to be in agreement. Their 90% CL limits are 2-3 times larger than the 95% CL limits given in Table \[tab:res\_par6\] but this may be due to the larger range they adopt for the prior on the inner slope (see App. \[app:biases1\]). - @2010AdAst2010E..45K gives the astrophysical factors of all the dSphs using a point-like approximation and a NFW DM profile, and integrated with a $\alpha_{\rm int}=0.15^\circ$ angular resolution. These can be compared to the median and confidence levels we derived in Table \[tab:res\_par6\] for $\alpha_{\rm int}=0.1^\circ$. The values of @2010AdAst2010E..45K (multiplied by $4\pi$ to match our definition of $J$) generally fall inside our 68% CL intervals, but for Leo II his value is just within our 95% CI while Draco and Carina cannot be accommodated at all. For these two objects, the values of @2010AdAst2010E..45K are much larger than the ones we find, and this is unlikely to be explained by the $0.05^\circ$ difference in integration angles. A simple explanation is that @2010AdAst2010E..45K does not use stellar kinematic data directly in his analysis, but stacks suitable Via Lactea halos ($M_{300} \approx 10^7 M_\odot$ and appropriate distances) and uses those averages to estimate $J$. Focusing on the ranking (without worrying about contrast to the background and the other instrumental constraints), both we and @2010AdAst2010E..45K agree that among the classical dSph UMi is a most promising target. However, while we find Sculptor and Draco to be the next most favorable targets, @2010AdAst2010E..45K names Draco and Carina from his ’simulation-based’ approach. For completness, we also compare our median values with the $J$ values used by different experimental groups: - The MAGIC collaboration published point source limits for Draco [@Albert2008] adopting the scheme of @2007PhRvD..76l3509S of a power law density profile, with an exponential cut-off. They examine two scenarios, a cored and a cusped model, but find no discernable difference when calculating $J$ for integration angles $<0.4^\circ$, i.e. larger than the MAGIC PSF. The value of $J$ they calculate for Draco is a higher than ours (after appropriate scaling of the integration region and unit conversion) by about a factor of 2. - The VERITAS collaboration also published limits on Draco and Ursa Minor [@2010ApJ...720.1174A]. They assume a NFW profile, take the density profiles from @2007PhRvD..75h3526S and follow @1998APh.....9..137B for the calculation of J. Whilst the range of density values in @2007PhRvD..75h3526S have a physical motivation the values used in @2010ApJ...720.1174A are rather arbitrarily chosen to be the midpoint of that range, which leads to consistently higher $J$-values than ours (by a factor of 3 for Draco and a factor of 1.2 for UMi). - The H.E.S.S. collaboration [@2011APh....34..608H] published limits on the southern sources Sculptor and Carina using NFW and isothermal profiles with a number of varying assumptions. This leads to a range of calculated $J$-values (rather than a single solution) that are consistent with our median value and estimated uncertainties. - The Fermi collaboration [@2010ApJ...712..147A] has published limits for a number of the sources studied here. They adopted a NFW profile within the tidal radius and following @2009JCAP...06..014M they calculated the $J$-value (using an MCMC approach on the observed stellar velocities) for a $1^\circ$ integration angle which is compatible with their high energy PSF. From this they find Draco to have a larger $J$ compared to the other dwarfs (a factor of $\sim 2$ higher than the next dwarf which is Ursa Minor), contrary to what we find in this study. 5-parameter MCMC analysis: $\gamma_{\rm prior}$ fixed {#sec:gamma_fixed_analysis} ----------------------------------------------------- Higher resolution numerical simulations following both DM and gas, additional kinematic data and new modelling techniques may help constraining the value of $\gamma$ in the near future. With the knowledge of $\gamma$, we should better constrain the radial-dependence of $J$, which is crucial to disentangle, e.g. dark matter annihilation from DM decay [@boyarsky06; @2010JCAP...07..023P]. The topic of decaying dark matter goes beyond the scope of this paper, and it will be discussed elsewhere. Below, we merely inspect the gain obtained on the $J$ prediction when having a strong prior on $\gamma$, and briefly comment on the possibility to disentangle $\gamma=0$ profiles from $\gamma=1.0$ profile in the case of annihilation (if this cannot be achieved, hopes for disentangling decay from annihilation would be quite low on a single object). ### Parameter correlations We repeat the MCMC analysis for fixed value of the inner slope $\gamma_{\rm prior}=0.$, 0.5, 1., and 1.5. The priors for the five other parameters are as given in Section \[subsec:gamma-free\]. Using Eq. (\[eq:M300\]) for the mass having a robust estimate of $M(r_\mathrm{half})$ [@2009ApJ...704.1274W; @2010MNRAS.406.1220W; @amorisco10] gives $\log(\rho_\mathrm{s}) + \gamma \log(r_\mathrm{s}) \approx \mathrm{constant}$ which reduces to $\log(\rho_\mathrm{s})\approx \mathrm{constant}$ for $\gamma=0$. As a result, we expect a strong correlation between $\rho_\mathrm{s}$ and $r_\mathrm{s}$ when $\gamma_{\rm prior}=1$ and none when $\gamma_{\rm prior}=0$. This is confirmed by the result of our MCMC analysis shown in Fig. \[fig:cor\_par5\] (here, for the Draco case). The half-light radius $r_\mathrm{half}$ for Draco is $\sim 200$ pc, but we choose to show the PDF for $M_{300}$ in the bottom panel of Fig. \[fig:cor\_par5\] as we wish to compare the mass of the dSphs among themselves (see Table \[tab:res\_par6\]). It confirms that the mass within an appropriate radius can be reliably constrained by the data regardless of the value of $\gamma$. ![[**Top:**]{} correlation and PDF of the profile parameters $\rho_\mathrm{s}$ and $r_\mathrm{s}$ from the 5-parameter MCMC analysis $\gamma_{\rm prior}=0.0$. [**Middle:**]{} same, but for $\gamma_{\rm prior}=1.0$. [**Bottom:**]{} PDF of $M_{300}$, the mass at 300 pc.[]{data-label="fig:cor_par5"}](fig13a.eps "fig:"){width="0.81\linewidth"} ![[**Top:**]{} correlation and PDF of the profile parameters $\rho_\mathrm{s}$ and $r_\mathrm{s}$ from the 5-parameter MCMC analysis $\gamma_{\rm prior}=0.0$. [**Middle:**]{} same, but for $\gamma_{\rm prior}=1.0$. [**Bottom:**]{} PDF of $M_{300}$, the mass at 300 pc.[]{data-label="fig:cor_par5"}](fig13b.eps "fig:"){width="0.81\linewidth"} ![[**Top:**]{} correlation and PDF of the profile parameters $\rho_\mathrm{s}$ and $r_\mathrm{s}$ from the 5-parameter MCMC analysis $\gamma_{\rm prior}=0.0$. [**Middle:**]{} same, but for $\gamma_{\rm prior}=1.0$. [**Bottom:**]{} PDF of $M_{300}$, the mass at 300 pc.[]{data-label="fig:cor_par5"}](fig13c.eps "fig:"){width="0.81\linewidth"} ### Uncertainties on the profile and on $J$ For any given $\gamma$, the uncertainty on $\rho(r)$ at small radii is related to the range of $r_\mathrm{s}$ values at which the asymptotic slope is reached (for each profile accepted by the MCMC analysis). For $\gamma_{\rm prior}=0$, the maximum uncertainty on $\rho(r)$ is directly related to the maximum uncertainty on $\rho_\mathrm{s}$ (since for $r\ll r_\mathrm{s}$, $\rho(r)$ is constant) which can be read off the PDF (top-left panel of Fig. \[fig:cor\_par5\]). This leads to an order of magnitude uncertainty on $\rho(r)$ for small $r$, which is consistent with the 95% CL shown in top panel of Fig. \[fig:rho\_gamma\_fixed\]. ![Median values (solid lines, filled symbols) and 95% CLs (dashed lines, empty symbols) from the fixed $\gamma_{\rm prior}$ MCMC analysis on Draco. [**Top:**]{} density profiles (the gray arrow indicates the value of $r_\mathrm{half}$). [**Bottom:**]{} $J$-factor (the gray arrow indicates $\alpha_c\approx 2 r_\mathrm{half}/d$).[]{data-label="fig:rho_gamma_fixed"}](fig14a.eps "fig:"){width="\linewidth"} ![Median values (solid lines, filled symbols) and 95% CLs (dashed lines, empty symbols) from the fixed $\gamma_{\rm prior}$ MCMC analysis on Draco. [**Top:**]{} density profiles (the gray arrow indicates the value of $r_\mathrm{half}$). [**Bottom:**]{} $J$-factor (the gray arrow indicates $\alpha_c\approx 2 r_\mathrm{half}/d$).[]{data-label="fig:rho_gamma_fixed"}](fig14b.eps "fig:"){width="\linewidth"} For $\gamma>0$, the uncertainty has to be read from the dispersion in the values of $\rho_\mathrm{s} r_\mathrm{s}^{\gamma}$, or equivalently, the mass $M_{300}$. The bottom panel of Fig. \[fig:cor\_par5\] shows that this mass is well-constrained, independently of $\gamma$ for the case of Draco (see however in Table \[tab:res\_par6\] for a larger spread for some dSphs), resulting in a smaller uncertainty for $\gamma_{\rm prior}=1.5$ than for $\gamma_{\rm prior}=0$ (top panel of Fig. \[fig:cor\_par5\]). We checked that the CLs obtained in Fig. \[fig:rho\_gamma\_fixed\] (in Appendix \[app:biases2\]) for the artificial data enclose correctly the range of reconstructed values: they are consistent with a larger reconstruction bias for $\gamma_{\rm prior}=0$ than for $\gamma_{\rm prior}=1.5$ at small radii. For the uncertainty on $J$, we can obtain a crude estimate by relying on the approximate formulae given in Appendix \[app:toyJ\]. For $\gamma>0$, $J\propto \rho_\mathrm{s}^2 r_\mathrm{s}^3$, and substituting the constant $M_{300}$ relationship leads to $J\propto r_\mathrm{s}^{3-2\gamma}$. The value of $r_\mathrm{s}$, as seen in its PDF in the top and middle panels of Fig. \[fig:cor\_par5\], varies by roughly a factor of 10. Because of the weighting power $3-2\gamma$, the uncertainty on $J$ is expected to be the smallest for $\gamma=1.5$, which is in agreement with the curves in Fig. \[fig:rho\_gamma\_fixed\] (bottom panel). However, the analysis of the artificial data in Appendix \[app:biases2\] shows that the typical CL on $J$ obtained in the bottom panel of Fig. \[fig:rho\_gamma\_fixed\] is likely to be underestimated for $\gamma_{\rm prior}=1.5$ (up to factor ${\cal O}(2)$, see Fig. \[fig:fake\_J\_par5\]). [^14] This happens for any integration angle. For this reason, we cannot rely of the $J$ value for $\gamma_{\rm prior}=1.5$ and focus only on the three cases $\gamma_{\rm prior}=0$, $\gamma_{\rm prior}=0.5$, and $\gamma_{\rm prior}=1.0$ below. ### $J(d)$ and departure from the $1/d^2$ scaling Fig. \[fig:Jd\_gamma\_fixed\] shows the $J$ median values, 65% and 95% CIs as symbols, dashed and solid error bars respectively, for an integration angle of $0.01^\circ$ (top), $0.1^\circ$ (middle), and $\alpha_\mathrm{c} \approx 2 r_\mathrm{half}/d$ [@letter_dsph] The $x$-axis is the distance to the dSph (in kpc). {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} For point-like sources, the $J$-factor of a single dSph scales as $1/d^2$, as illustrated by the blue-dashed line. Departure from this scaling is interpreted as a combination of a mass effect and/or a profile effect. For instance, Sextans and Carina are dSphs with smaller $M_{300}$ with respect to the other ones (see Tab. \[tab:res\_par6\]); consequently they are located below the dashed blue line in the top panel of Fig. \[fig:Jd\_gamma\_fixed\]. The exception is Leo II, which has a ‘small’ mass but is nevertheless above the dashed line. Although this analysis cannot constrain $\gamma$, we are tempted to interpret this oddity in terms of a ‘cuspier’ profile (w.r.t. those for other dSphs), which would be consistent with the fact that its $J$ remains similar in moving from $\alpha_{\rm int}=0.1^\circ$ (middle panel) to $0.01^\circ$ (top panel). However, an alternative explanation (which would be more consistent with the results obtained in this paper) could be the fact that Leo II has the smallest amount of kinematic data at present, and that its $J$ is overestimated (see Appendix \[app:binning\] to support this line of argument). We repeat that the relative brightness of the dSphs is further affected for background-dominated instruments (as described in Sec. \[sec:detectability\]), so that the ranking has to be based on Fig \[fig:min\_detect\_sigmav\] discussed in the next section. The bottom panel of Fig. \[fig:Jd\_gamma\_fixed\] shows the $J$ value for an ‘optimal’ integration angle $\alpha_\mathrm{c}$ that is twice the half-light radius divided by the dSph distance[^15] (this corresponds to the integration angle that minimises the CLs on $J$; see @letter_dsph). The yellow broken solid lines show the expected signal from the diffuse Galactic DM annihilation background, including a contribution from clumpy sub-structures (the extragalactic background, which also scales as $\alpha_{\rm int}^2$, has not been included). The total background may be uncertain by a factor of a few (depending on the exact Galactic (smooth) profile and local DM density). Its exact level|which depends on the chosen integration angle|determines the condition for the loss of contrast of the dSph signal, i.e. the condition for which looking at the DM halo (rather than at dSphs) becomes a better strategy. ### Conclusion for the fixed $\gamma_{\rm prior}$ analysis The analysis of simulated data shows that the analysis for $\gamma_{\rm prior}=1.5$ is biased by a factor of ${\cal O}(10)$ and that the CLs obtained on the real data are likely to be severely under-estimated in that case. But such steeply cusped profiles are neither supported by observations nor motivated by current cosmological simulations. For values of $\gamma_{\rm prior} \leq 1$, this bias is a factor of a few only, so that it shows that the results from a fixed $\gamma_{\rm prior}$ analysis of the 8 classical dSphs are robusts. However, this analysis shows that unless very small integration angles $\alpha_{\rm int}\lesssim 0.01^\circ$ are chosen (or if $\gamma_{\rm true}\gtrsim 1$), knowing the exact value of $\gamma$ does not help in improving the determination of $J$. Indeed, even using Draco, the stellar population of which is one of the most studied, the CLs of the three reconstructed fluxes ($\gamma_{\rm prior}=0$ in black full circles, $\gamma_{\rm prior}=0.5$ in red triangles, and $\gamma_{\rm prior}=1.0$ in blue stars) in Fig. \[fig:rho\_gamma\_fixed\] (bottom), overlap. Reversing the argument, if we do not know the inner slope, and if a $\gamma$-ray signal is detected from just one dSph in future, there will be little hope of recovering the slope of the DM halo from that measurement only. This means that the best way to improve the prediction of the $J$-factor in the future relies on obtaining more [*data*]{} and a more refined MCMC analysis; an improved prior on the DM distribution makes little difference. Sensitivity of $\gamma$-ray observatories to DM annihilation in the dSphs {#sec:5.3} ------------------------------------------------------------------------- {width="\linewidth"} {width="\linewidth"} {width="\linewidth"} The potential for using the classical dSph to place constraints on the DM annihilation cross-section, given the uncertainties in the astrophysical $J$-factor, can be seen in Fig. \[fig:min\_detect\_sigmav\]. Previous analyses have adopted the solid angle for calculation of the $J$-factor to be the angular resolution of the telescope for a point-like source, typically assuming a NFW-like profile [@2010ApJ...720.1174A; @2010ApJ...712..147A; @2011APh....34..608H]. By contrast our sensitivity plots take into account finite size effects: i) the $J$ values are based on the MCMC analysis with the prior $0\leq\gamma_{\rm prior}\leq1$, where the corresponding $J$ are shown in Fig. \[fig:JdSph\_Jdm\_bkgd\]; ii) the energy dependent angular resolution has also been taken into account assuming a standard $\gamma$-ray annihilation spectrum (see Section \[sec:dNdE\_xsec\]). Moreover for Fermi-LAT the background level assumed has been increased (resulting in a 25% worsening of the sensitivity above 100 MeV) to reflect the average situation in the directions of the classical dSph (the variation between the individual dSph is only 7% rms). A likelihood based analysis is used for both FCA and Fermi and a nominal observation zenith angle of 20$^{\circ}$ assumed for the FCA[^16] (see Section \[sec:RelativePerformance\]). The panels from top to bottom correspond to increasing DM (neutralino) masses. At low values, Fermi has a better sensitivity than FCA; at a mass of about 1 TeV the two are comparable, and for higher masses the FCA becomes the more sensitive instrument due to the vastly greater effective area at the photon energies at which the annihilation spectrum is expected to peak. Note that the precise value of $\langle\sigma v\rangle$ where the relative sensitivities of the two instruments cross depends on the form of the DM annihilation spectrum. Since we are examining the uncertainties in the astrophysical $J$-factor to the detectability of dSphs, we have used a conservative spectrum averaged over a number of possible annihilation channels (see Fig. \[fig:phi-susy\]) which results in the majority of produced $\gamma$-ray photons having energies $\simeq$10% of the DM particle mass. If we were to move from a relatively soft spectrum, such as $\mathrm{b\overline{b}}$ to a harder one, such as $\tau^{+}\tau^{-}$, this would benefit both instruments in different ways. For Fermi-LAT a harder spectrum makes the signal easier to distinguish above the diffuse $\gamma$-ray background; indeed the [@2010ApJ...712..147A] found that the detectable flux limit from a potential source could vary by a factor of 2–20 (with lower particle masses benefiting the most) between these different annihilation spectra. For the FCA, which has a very large effective area to photons $\geq100$GeV, the benefits of having more high energy photons is very apparent when it comes to flux sensitivity. For both observatories, an increased number of high energy photons needs to be balanced with the correspondingly better angular resolution, particularly if (e.g. for Fermi-LAT) a point-like source becomes spatially resolved. Our analysis places Ursa Minor as the best candidate for the northern sky (marginally better than Draco, which has long been a favourite target of northern hemisphere observatories) and Sculptor for the southern sky, when it comes to a favourable median and low uncertainty in the $J$-factor. It should be noted, however, that although the closest objects seem to be favoured, Leo II has the potential to yield a stronger signal, however more kinematic data are needed in order to constrain better its J-factor. In addition, it should be noted that the uneven sensitivity of the Fermi-LAT across the sky, caused in particular by the proximity of bright sources[^17] as well as the galactic diffuse background can change what is considered the favorite candidate. We emphasise that in our analysis the inner slope $\gamma$ has not been constrained, but that a better independent determination of $\gamma$ in future will not help providing a better determination of $J$ (see Fig. \[fig:Jd\_gamma\_fixed\]); this is discussed further in the Appendices. Carina, Fornax and Leo I are the targets least favoured. When compared to existing limits from Fermi-LAT [@2010ApJ...712..147A] or the current generation of ACTs [@2010ApJ...720.1174A; @2011APh....34..608H] it can be seen that our limits are not dissimilar from those that have already been published. For Fermi this is not surprising, since the source is unresolved and any difference should relate only to the assumed increase in exposure from 1 to 5 years, resulting in a factor of a few at best. The similarity in sensitivity between current and future ACTs is perhaps more surprising, but this as stated earlier relates to the naïve assumptions made on the form for the $J$-factor and the solid angle integrated over; in order to reach the currently claimed limits requires a deep exposure with an instrument as sensitive as CTA. {width="0.7\linewidth"} One last thing to note is that a common way to synthesise a deeper exposure is to stack observations of different sources together to provide an effective long exposure of a generic source. For a common universal halo profile this may be fine, however any analysis will have to take into account the different integration angles for each individual source correctly. If all dSphs do not share a common halo profile and hence have different $\gamma$ values, we have to rely on the varying-$\gamma$ analysis presented in the previous section and the relative ranking of potential targets would then be different. Discussion and conclusions {#sec:conclusions} ========================== We have revisited the expected DM annihilation signal from dSph galaxies for current (Fermi-LAT) and future (e.g. CTA) $\gamma$-ray observatories. The main innovative features of our analysis are that: (i) We have considered the effect of the [*angular size*]{} of the dSphs for the first time. This is important since, while nearby dSphs have higher $\gamma$ ray flux, their larger angular extent can make them sub-prime targets if the sensitivity is limited by cosmic ray and $\gamma$-ray backgrounds. (ii) We determined the astrophysical $J$-factor for the classical dSphs directly from photometric and kinematic data. We assumed very little about their underlying DM distribution, modelling the dSph DM profile as a smooth split-power law, both with and without DM sub-clumps. (iii) We used a MCMC technique to marginalise over unknown parameters and determine the sensitivity of our derived $J$-factors to both model and measurement uncertainties. (iv) We used simulated DM profiles to demonstrate that our $J$-factor determinations recover the correct solution within our quoted uncertainties. Our key findings are as follows: 1. Sub-clumps in the dSphs do [*not*]{} usefully boost the signal. For all configurations where the sub-clump distribution follows the underlying smooth DM halo, the boost factor is at most $\sim\!2-3$. Moreover, to obtain even this mild boost, one has to integrate the signal over the whole angular extent of the dSph. This is unlikely to be an effective strategy as the diffuse Galactic DM signal will dominate for integration angles $\alpha_{\rm int}\gtrsim 1^\circ$. 2. Point-like emission from a dSph is a very poor approximation for high angular resolution instruments, such as the next-generation CTA. For a nearby dSph, using the point-like approximation can lead to an order of magnitude overestimate of the detection sensitivity. In the case of a nearby cored profile consisting of very high mass DM particles, a point source approximation can be unsatisfactory even for the modest angular resolution of Fermi-LAT. 3. With the Jeans’ analysis, no DM profile can be ruled out by current data. The use of the MCMC technique on artificial data also shows that such an analysis is unable to provide reliable values for $J$ if the profiles are cuspy ($\gamma=1.5$). However, using a prior on the inner DM cusp slope $0\leq\gamma_{\rm prior}\leq1$ provides $J$-factor estimates accurate to a factor of a few. 4. The best dSph targets are not simply those closest to us, as might naïvely be expected. A good candidate has to combine high mass, close proximity, small angular size ($\lesssim 1^\circ$; i.e. not too close); and a well-constrained DM profile. With these criteria in mind, we find three categories: well-constrained and promising (Ursa Minor, Sculptor and Draco), well-constrained but less promising (Carina, Fornax and Leo I), and poorly constrained (Sextans and Leo II). Leo II may yet prove to be a viable target as it has a larger median $J$-factor than UMi, however more data are required to confirm its status. 5. A search based on a known DM candidate (from, e.g., forthcoming discoveries at the LHC) will do much to optimise the search strategy and, ultimately, the detection sensitivity for all $\gamma$-ray observatories. This is because the shape of the annihilation spectrum is a strong driver of the photon energy range that can provide the best information on the candidate DM particle mass. Fermi-LAT has great potential to probe down to the expected annihilation cross-section for particles of mass $\ll 700$ GeV, whereas a ground based instrument is more suited for probing particle masses above a few hundred GeV with a sufficiently deep exposure. However, even for 5 yr of observation with Fermi-LAT or 100 hrs with FCA, the sensitivity reach (Fig. \[fig:mssm\]) remains anywhere between 4 to 10 orders of magnitude above the expected annihilation cross-section for a cosmological relic (depending on the mass of the DM particle candidate). Improving these limits will require a harder annihilation spectrum than the conservative average we have adopted in this study, or a significant boost (e.g. from the Sommerfeld enhancement) to the $\gamma$-ray production. Finally, the ultra-faint dSphs have received a lot of interest in the community lately, as they could be the most-DM dominated systems in the Galaxy. We emphasise that the MCMC analysis we have performed for the classical dSphs cannot be applied ‘as is’ for these objects. First, the sample of stars observed is smaller. Second, the velocity dispersion is smaller and suffers from larger uncertainties than those for the classical dSphs. The robustness and systematic biases of the MCMC analysis will be discussed elsewhere (Walker et al., 2011, in preparation). Results concerning $J$ for the ultra-faint dSphs will be presented in a companion paper. Acknowledgments {#acknowledgments .unnumbered} =============== We thank the anonymous referee for their careful reading of the manuscript and useful comments. We thank Walter Dehnen for providing his code for use in generating artificial dSph data sets. MGW is supported by NASA through Hubble Fellowship grant HST-HF-51283, awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS 5-26555. CC acknowledges support from an STFC rolling grant at the University of Leicester. JAH acknowledges the support of an STFC Advanced Fellowship. MIW acknowledges the Royal Society for support through a University Research Fellowship. JIR acknowledges support from SNF grant PP00P2\_128540/1. SS acknowledges support by the EU Research & Training Network ‘Unification in the LHC era’ (PITN-GA-2009-237920). Part of this work used the ALICE High Performance Computing Facility at the University of Leicester. Some resources on ALICE form part of the DiRAC Facility jointly funded by STFC and the Large Facilities Capital Fund of BIS. Definitions, notation, conversion factors {#app:defs} ========================================= Studies of DM annihilations in the context of dSphs involves both particle physics and astrophysics. The obvious difference of scales between the two fields and habits among the two communities have given rise to a plethora of notations and unit choices throughout the literature. In this Appendix, we provide some explanatory elements and conversion factors to ease comparison between the different works published on the subject. As mentioned in §\[sec:method\], we define the differential $\gamma-$ray flux as integrated over the solid angle $\Delta\Omega$ as $$\frac{\mathrm{d}\Phi_{\gamma}}{\mathrm{d}E_{\gamma}}(E_{\gamma},\Delta\Omega) = \Phi^{\rm pp}(E_{\gamma}) \times J(\Delta\Omega)\,, \nonumber$$ where $$\Phi^{\rm pp}(E_{\gamma})\equiv \frac{\mathrm{d}\Phi_{\gamma}}{\mathrm{d}E_{\gamma}} = \frac{1}{4\pi}\frac{\langle\sigma_{\rm ann}v\rangle}{2m_{\chi}^{2}} \cdot \frac{dN_{\gamma}}{dE_{\gamma}}\nonumber\;,$$ and $$J(\Delta\Omega) = \int_{\Delta\Omega}\int \rho_{\rm DM}^2 (l,\Omega) \,dld\Omega. \nonumber$$ The solid angle is simply related to the integration angle $\alpha_{\rm int}$ by $$\Delta\Omega = 2\pi\cdot(1-\cos(\alpha_{\rm int})) \,.$$ In our work, the units of these quantities are as follows: - $\left[\mathrm{d}\Phi_{\gamma}/\mathrm{d}E_{\gamma}\right] = {\rm cm}^{-2}{\rm ~s}^{-1} {\rm ~GeV}^{-1}$; - $\left[ \Phi^{\rm pp}(E_{\gamma})\right] = {\rm cm}^{3}{\rm ~s}^{-1} {\rm ~GeV}^{-3} ({\rm ~sr}^{-1})$; - $[J] = M_\odot^2{\rm ~kpc}^{-5} ({\rm ~sr})$. First of all, note that the location of the $1/4\pi$ factor appearing in $\Phi^{\rm pp}$ is arbitrary. We followed and included it in the particle physics factor. In other works, it can appear in the astrophysical factor $J$ (e.g., @2009JCAP...01..016B). Therefore, to compare the astrophysical factors between several studies, one must first ensure to correct the value of $J$ by $4\pi$ if needed. In the text, we did not explicitly stated the solid angle dependence in the units of $J$ as it is dimensionless quantity. [^18] The conversion factor (once the $4\pi$ issue is resolved) from our $J$ units to that traditionally found in the literature are: - $1\; M_\odot^2{\rm ~kpc}^{-5}=10^{-15}~M_\odot^2{\rm ~pc}^{-5}$ - $1\; M_\odot^2{\rm ~kpc}^{-5}=4.45\times10^{6}{\rm ~GeV}^2{\rm ~cm}^{-5}$ - $1\; M_\odot^2{\rm ~kpc}^{-5} ~(\rm sr) = 1.44\times10^{-15}{\rm ~GeV}^2{\rm ~cm}^{-6} {\rm ~kpc} \;(\rm sr)$ Before comparing any number, one must also ensure that the solid angle $\Delta\Omega$ over which the integration is performed is the same. In most works, a $\alpha_{\rm int}=0.1^\circ$ angular resolution is chosen, corresponding to $\Delta\Omega=10^{-5}$ sr. However this is not always the case, as in the present study where we explore several angular resolutions. Note that the quantity $\bar J\equiv J/\Delta\Omega$ (in ${\rm GeV}^2{\rm ~cm}^{-5} {\rm ~sr^{-1}}$ for example) is also in use and the astrophysical factor is can be found under this form in some articles (e.g., [@Evans:2003sc]). Toy model for J (in dSphs) {#app:toyJ} ========================== The volume of the dSph is not always fully encompassed in the integration solid angle, as sketched in Fig. \[fig:sketch\_dsph\] (vertical hatched region) so that a numerical integration is required in general. ![Sketch of the integration regions contributing to the J factor: shown are the full integration region (vertical hatched) or a sub-region (cross-hatched) used for the toy calculations. The letter O shows the observer position, $\alpha_{\rm int}$ is the integration angle, $d$ is the distance of the dSph and $R_{\rm vir}$ its virial radius.[]{data-label="fig:sketch_dsph"}](figB1.eps){width="\linewidth"} However, a reasonable approximation for estimating the dependence of $J$ on the parameters of the problem, i.e. the distance to the dSph $d$, the integration angle $\alpha_{\rm int}$, and the profile parameters $\rho_\mathrm{s}$, $r_\mathrm{s}$ and $\gamma$), is to consider only the volume within the radius $$r_{\alpha_{\rm int}} = d\times \sin(\alpha_{\rm int}) \approx d\times \alpha_{\rm int} , \label{eq:Ralpha_alpha}$$ where the approximation is valid for typical integration angles $\alpha_{\rm int} \lesssim 0.1^{\circ}$. This volume corresponds to the spherical cross-hatched region in Fig. \[fig:sketch\_dsph\]. The toy model proposed below to calculate $J$ allows us to cross-check the results of the numerical integration for both the smooth and sub-clump contribution. We find that the model is accurate enough up to a factor of $2$ for $\gamma=0$ and $\gamma>0.5$, so can be used for gross estimates of any signal from a DM clump. For the smooth distribution --------------------------- About 90% of the clump luminosity is usually contained in a few $r_\mathrm{s}$, whatever the profile. The consequences are twofold. First, as can be read off Table \[tab:res\_par6\], $r_\mathrm{s}/d\ll 1$, so that the J factor amounts to a point like contribution $$J_{\rm point-like} = \frac{4\pi}{d^2}\int_{0}^{\min(r_{\alpha_{\rm int}},r_\mathrm{s})} r^2 \rho^2(r) dr. \label{eq:Japprox}$$ Secondly, it means that Eq. (\[eq:hernquist1\]) for the profile can be simplified into the approximate expression $$\rho_{\rm approx}(r) = \begin{cases} \rho_{\rm sat} &\text{if $r \le r_{\rm sat}$;}\\ \rho_\mathrm{s} \times \left(\frac{r}{r_\mathrm{s}}\right)^{-\gamma} &\text{if $ r_{\rm sat}<r \le r_\mathrm{s}$;}\\ 0 & \text{otherwise.} \end{cases} \label{eq:rhoapprox}$$ However, for all applications of our toy model, we will keep $\gamma< 3/2$, so that the saturation density above is never reached in the dSphs considered below. #### Various regimes {#various-regimes .unnumbered} The approximate formulae for $J$ is obtained by combining Eqs. (\[eq:Japprox\]) and (\[eq:rhoapprox\]): $$J_{\rm approx} =\frac{4\pi}{d^2}\int_{0}^{\min(r_{\alpha_{\rm int}},r_\mathrm{s})} r^2 \rho_{\rm approx}^2(r) dr. \label{eq:Japprox_final}$$ Using Eq. (\[eq:Ralpha\_alpha\]), this leads to $$J_{\rm approx} = \frac{4\pi}{d^2} \cdot \frac{\rho_\mathrm{s}^2 \,r_\mathrm{s}^{2\gamma}}{3-2\gamma}\cdot [\min(r_{\alpha_{\rm int}},r_\mathrm{s})]^{3-2\gamma}\,. \label{eq:Japprox2}$$ This formula gives satisfactory results for cuspy profiles (see below), but has to be modified in the following cases: - If $r_{\alpha_{\rm int}}\gtrsim r_\mathrm{s}$, the integration region encompasses $r_\mathrm{s}$. The $(1,3,\gamma)$ profiles decrease faster than $r^{-\gamma}$ for $r\sim r_\mathrm{s}$ hence integrating the toy model up to $r_\mathrm{s}$ is bound to overshoot the true result. We thus stop the integration at the radius $r_x$ such that $\rho_{\rm true}(r_x) = \rho_{\rm approx}(r_x)/x$, i.e. $$r_x = r_\mathrm{s} \cdot [x^{1/(3-\gamma)}-1] \,.$$ Taking $x=2$ gives a satisfactory fit to the full numerical calculation (see below). - If $r_{\alpha_{\rm int}}\gtrsim r_\mathrm{s}$ and $\gamma=0$, the integration can be performed analytically up to $R_{\rm vir}$ and is used instead. - If $r_{\alpha_{\rm int}}\lesssim r_\mathrm{s}$ and $\gamma=0$, the profile is constant, and integrating on the cross-hatched region (instead of the vertical hatched one, see Fig. \[fig:sketch\_dsph\]) undershoots the true result. A better approximation is to integrate on a conic section. For the same reason as given for the first item, we replace $r_\mathrm{s}$ by $r_x$ (with $x=2$) in the calculation of the cone volume. #### Resulting formula {#resulting-formula .unnumbered} To summarise, the final toy-model formula proposed for the smooth contribution of the dSph is: $$J_{\rm toy} = \frac{4\pi\rho_\mathrm{s}^2 }{d^2} \times \begin{cases} \displaystyle r_\mathrm{s}^{2\gamma}\cdot \frac{\min(r_x,r_{\alpha_{\rm int}})^{3-2\gamma}}{3-2\gamma} &\text{if $\gamma>0$;}\\ \displaystyle [I(r_{\alpha_{\rm int}})-I(0)] &\text{if $\gamma=$0, $r_{\alpha_{\rm int}}\!>\!r_x$;}\\ \displaystyle \frac{r_{\alpha_{\rm int}}^2 \cdot r_\mathrm{s}}{2} &\text{if $\gamma=$0, $r_{\alpha_{\rm int}}\!<\!r_x$;} \end{cases} \label{eq:J_final}$$ where $$\begin{aligned} r_{\alpha_{\rm int}} &=& \alpha_{\rm int} \cdot d, \nonumber\\ r_x &=& r_\mathrm{s} \cdot [x^{1/(3-\gamma)}-1], \nonumber\\ I(x)&=& -r_\mathrm{s}^6 (r_\mathrm{s}^2+5r_\mathrm{s}x+10x^2)/(30(r_\mathrm{s}+x)^5). \label{eq:J_final_def}\end{aligned}$$ #### Toy model [*vs*]{} numerical integration {#toy-model-vs-numerical-integration .unnumbered} Finally, we check the validity of this toy model by confronting it with the full numerical integration. Various inner slope $\gamma$ of the profile are considered as provided in Table \[tab:generic\_model\_definition\]. Defining the critical distance $d_{\rm crit}$ for which the dSph is fully encompassed by the integration region, i.e., $$d_{\rm crit} = \frac{r_\mathrm{s}}{\alpha_{\rm int}}\,.$$ we find $d_{\rm crit}\sim$ 50 kpc and 500 kpc for $r_\mathrm{s}=0.1$ and 1 kpc respectively (the integration range is $\alpha_{\rm int}=0.1^\circ$). If $r_x$ is used instead of $r_\mathrm{s}$, this distance is even smaller. This allows us to test the toy model for the two regimes. The result is shown in Fig. \[fig:toy\_vs\_full\]. The symbols show the full numerical integration while the lines show the toy-model calculations. ![Toy-model calculation (lines) vs full numerical integration (symbols) of $J$ as a function of the distance to the dSph. The integration angle is fixed to $\alpha_{\rm int}=0.1^\circ$ and the $(1,3,\gamma)$ profiles are taken to vary from $\gamma=0$ to $\gamma=1.45$. For each model, $\rho_\mathrm{s}$ is calculated such as to provide $M_{300}=10^7M_\odot$. [**Top:**]{} dSphs for which $r_\mathrm{s}=0.1$ kpc. [**Bottom:**]{} dSphs for which $r_\mathrm{s}=1$ kpc. .[]{data-label="fig:toy_vs_full"}](figB2a.eps "fig:"){width="\columnwidth"} ![Toy-model calculation (lines) vs full numerical integration (symbols) of $J$ as a function of the distance to the dSph. The integration angle is fixed to $\alpha_{\rm int}=0.1^\circ$ and the $(1,3,\gamma)$ profiles are taken to vary from $\gamma=0$ to $\gamma=1.45$. For each model, $\rho_\mathrm{s}$ is calculated such as to provide $M_{300}=10^7M_\odot$. [**Top:**]{} dSphs for which $r_\mathrm{s}=0.1$ kpc. [**Bottom:**]{} dSphs for which $r_\mathrm{s}=1$ kpc. .[]{data-label="fig:toy_vs_full"}](figB2b.eps "fig:"){width="\columnwidth"} For profiles steeper than 0.5, the agreement is better than a factor of 2 for all distances. For flatter profiles, the toy model only gives results within an order of magnitude. However, for $\gamma=0$, the fix applied to the toy-model allows to regain the correct results within a factor of 2. Hence, given the current uncertainties on the profiles, the set of formulae (\[eq:J\_final\_def\]) and (\[eq:J\_final\_def\]) can safely be used for quick inspection of the $J$ value of any profile with an inner slope $\gamma$ of 0, or greater than 0.5. For the sub-clumps {#sec:sub-clumps} ------------------ The influence of DM sub-structures on the $\gamma$-ray production has been widely discussed in the literature. These sub-structures may enhance the detectability by boosting the $\gamma$-ray signal. In this appendix, we give an analytical estimation of the effect of sub-clumps in dSph spheroidal galaxies, in the same spirit as the toy model developed in the previous section for the smooth component. For simplicity, we restrict ourselves to one cored $(\alpha, \beta, \gamma)=(1,3,0)$ and one cusped $(1,3,1)$ profile. To characterise the clump distribution, we use the formalism given in . #### Sub-structure distribution {#sub-structure-distribution .unnumbered} The clump spatial distribution is assumed to follow the dSph DM profile, namely $$\frac{dP(r)}{dV} \propto \left(\frac{r}{r_\mathrm{s}}\right)^{-\gamma} \left[1+\left(\frac{r}{r_\mathrm{s}}\right)^{-\alpha}\right]^{\frac{\gamma-\beta}{\alpha}}\,. \label{eq:distrib_space}$$ The mass distribution of the clumps is taken to be independent of the spatial distribution and takes the usual form, $$\frac{dP}{dM} = A M^{-a}\;, \label{eq:distrib_mass}$$ with $M\in[M_{\rm min},M_{\rm max}]$ and $a\sim 1.9$ from cosmological N-body simulations ($A$ is the normalisation constant for $dP/dM$ to be a probability). #### Clump luminosity {#clump-luminosity .unnumbered} Defining $L_i$ the [*intrinsic luminosity*]{} of the sub-clump $i$ to be $$L_i\equiv\int_{V_{\rm cl}} \rho^2 dV\,,$$ the astrophysical contribution to the $\gamma$-ray flux from the sub-structures of the dSph is $$J_{\rm clumps} = \frac{1}{d^2}\cdot \sum_{i=1}^{N^{\rm cl}} L_i\,, \label{eq:jcl_discreet}$$ where $N^{\rm cl}$ is the number of clumps contained within the integration angle $\alpha$ and $d$ is the distance of dSph. The luminosity depends only on the mass of the clump, once a concentration-mass ($c_{\rm vir}-M_{\rm vir}$) relationship is chosen , so that $ L_i=L (M_i)$. Moving to the continuous limit, Eq. (\[eq:jcl\_discreet\]) reads $$J_{\rm clumps} = \frac{1}{d^2} \cdot N^{\rm cl} \cdot \int_{M_{\rm min}}^{M_{\rm max}} L(M)\frac{dP}{dM} dM\;.$$ Fitting the results from , the intrinsic luminosity[^19] varies almost linearly with the mass of the clump, as $$L^{\rm NFW}(M)=1.17\times 10^{8}\;(M/M_\odot)^{0.91}\;M_\odot^2\;{\rm kpc}^{-3},$$ so we have $$J_{\rm clumps} = \frac{N^{\rm cl}A}{d^2} \left(\frac{1.17\times 10^{8}}{1.91-a}\right)\;\left(M_{\rm max}^{1.91-a}-M_{\rm min}^{1.91-a}\right)\;.$$ #### Number of clumps {#number-of-clumps .unnumbered} The fraction $F$ of clumps in the spherical integration region $r_{\alpha_{\rm int}}\approx\alpha_{\rm int} d$ (cross-hatched region in Fig. \[fig:sketch\_dsph\]) is given by $$F=\frac{N^{\rm cl}}{N_{\rm tot}^{\rm cl}}= \int_0^{r_{\alpha_{\rm int}}} 4\pi r^2 \frac{dP}{dV} dr\;,$$ where $N_{\rm tot}^{\rm cl}$ is the total number of clumps within the dSph. Upon integration and defining $x_{\rm int}=r_{\alpha_{\rm int}}/r_\mathrm{s}$ and $x_{\rm vir}=R_{\rm vir}/r_\mathrm{s}$ this becomes: $$\begin{aligned} \label{eq:frac_core} F_{\rm core}& = &\left[\frac{4x_\alpha+3}{2(x_\alpha+1)^2}+\ln(x_\alpha+1)-\frac{3}{2}\right]\\\nonumber & \times &\left[\frac{4x_{\rm vir}+3}{2(x_{\rm vir}+1)^2}+\ln(x_{\rm vir}+1)-\frac{3}{2}\right]^{-1} \text{for $(1,3,0)$,}\end{aligned}$$ and $$\begin{aligned} \label{eq:frac_cusp} F_{\rm cusp}& = &\left[\frac{1}{(x_\alpha+1)}+\ln(x_\alpha+1)-1\right]\\\nonumber & \times &\left[\frac{1}{(x_{\rm vir}+1)}+\ln(x_{\rm vir}+1)-1\right]^{-1} \text{for NFW.}\end{aligned}$$ Some care is necessary when evaluating the number of clumps $N^{\rm cl}=F\times N_{\rm tot}^{\rm cl}$ in the integration region. Whatever the profile, most of the clumps are located within $r_\mathrm{s}$ so when $r_{\alpha_{\rm int}}>r_\mathrm{s}$, the spherical integration region of our toy model (cross-hatched region in Fig. \[fig:sketch\_dsph\]) is a good enough approximation, and Eq. (\[eq:frac\_core\]) and (\[eq:frac\_cusp\]) hold. However, if $r_{\alpha_{\rm int}}<r_\mathrm{s}$ then the remainder of the intersecting cone (vertically hatched region in Fig. \[fig:sketch\_dsph\]) could amount to a significant contribution to the number of clumps. Cuspy distributions should only be marginally affected given their high central concentration. However, this effect may be important for cored profiles. Whenever $r_{\alpha_{\rm int}} < r_\mathrm{s}$, as for the smooth contribution, Eq. (\[eq:frac\_core\]) is therefore multiplied by the ratio of the intersecting cone volume to the integration sphere volume, in order to account for that effect. If the mass of the dSph is $M_{\rm vir}$ and assuming a fraction $f$ of this mass is in the form of clumps, one gets using Eq. (\[eq:distrib\_mass\]) $$N_{\rm tot}^{\rm cl}=f\; \frac{2-a}{A} M_{\rm vir}\left(M_{\rm max}^{2-a}-M_{\rm min}^{2-a}\right)^{-1}\;.$$ #### Resulting formulae {#resulting-formulae .unnumbered} Adding all ingredients together, the contribution of the sub-structures to the flux is $$\begin{aligned} J_{\rm clumps} & = & 1.17\times 10^{8} \frac{F_{\rm core/cusp}}{d^2} \left(\frac{2-a}{1.91-a}\right)\\\nonumber & \times & \left(\frac{M_{\rm max}^{1.91-a}-M_{\rm min}^{1.91-a}}{M_{\rm max}^{2-a}-M_{\rm min}^{2-a}}\right) f\;M_{\rm vir}\;. \label{eq:toy_cl_final}\end{aligned}$$ #### Toy model [*vs*]{} numerical integration {#toy-model-vs-numerical-integration-1 .unnumbered} The comparison between the two is shown in Fig. \[fig:toy\_vs\_full\_cl\]. The symbols show the full numerical integration while the lines show the toy-model calculations. ![Toy-model calculation (lines) vs full numerical integration (symbols) of $J$ as a function of the distance to the dSph. The integration angle is fixed to $\alpha_{\rm int}=0.1^\circ$ and the two $(1,3,\gamma)$ sub-clump spatial distribution are $\gamma=0$ and $\gamma=1$ (their inner profile is a NFW with a $c_{\rm vir}-M_{\rm vir}$ relation taken from @2001MNRAS.321..559B). The calculations assume the fraction of DM in sub-clumps to be $f=50\%$ of the total mass of the dSphs, where the smooth profile is taken as in Fig. \[fig:toy\_vs\_full\]. [**Top:**]{} $r_\mathrm{s}=0.1$ kpc. [**Bottom:**]{} $r_\mathrm{s}=1$ kpc.[]{data-label="fig:toy_vs_full_cl"}](figB3a.eps "fig:"){width="\columnwidth"} ![Toy-model calculation (lines) vs full numerical integration (symbols) of $J$ as a function of the distance to the dSph. The integration angle is fixed to $\alpha_{\rm int}=0.1^\circ$ and the two $(1,3,\gamma)$ sub-clump spatial distribution are $\gamma=0$ and $\gamma=1$ (their inner profile is a NFW with a $c_{\rm vir}-M_{\rm vir}$ relation taken from @2001MNRAS.321..559B). The calculations assume the fraction of DM in sub-clumps to be $f=50\%$ of the total mass of the dSphs, where the smooth profile is taken as in Fig. \[fig:toy\_vs\_full\]. [**Top:**]{} $r_\mathrm{s}=0.1$ kpc. [**Bottom:**]{} $r_\mathrm{s}=1$ kpc.[]{data-label="fig:toy_vs_full_cl"}](figB3b.eps "fig:"){width="\columnwidth"} For $r_\mathrm{s}=100$ pc, the agreement is better than a factor of 2 for all distances. For $r_\mathrm{s}=1$ kpc, the toy model only gives results correct to within a factor of 4 for $\gamma=1$. Hence, given the current uncertainties on the profiles, Eq. (\[eq:toy\_cl\_final\]) can be used for quick inspection of the $J$ value for the sub-clump contribution. Distance and integration angle dependence on J for generic dSphs {#app:dep_generic} ================================================================ This Appendix completes the study of the $J$-factor dependences started in Section \[sec:gen\_dep\]. All the plots and discussions below rely on the generic profiles given in Table \[tab:generic\_model\_definition\], and the sub-structure reference configuration given in Section \[sec:sub\_reference\]. Distance dependence $J(d)$ -------------------------- Fig. \[fig:generic\_JD\_sm\_and\_subcl\] shows $J_{\rm sm}$ as a function of the distance to the dSph (we assume $\alpha_{\rm int}=0.1^\circ$ here and that we are pointing towards the dSph centre, i.e. $\theta=0$). As we have checked earlier, the sub-clump contribution for the reference model at $\theta=0$ is always sub-dominant, so for clarity only $J_{\rm sm}$ is displayed ($f=0$) in the figure. ![$J_{\rm sm}(\theta=0)$ as a function of the distance to the dSph for three profiles $\gamma$ and three values of $r_\mathrm{s}$. The corresponding values for $\rho_\mathrm{s}$ are given in Table \[tab:generic\_model\_definition\].[]{data-label="fig:generic_JD_sm_and_subcl"}](figC1.eps){width="\linewidth"} If the angular size of the signal is smaller than the integration angle, the distance dependence is expected to be $J_{\rm sm}\propto d^{~-2}$. This is the case for $\gamma=1.5$ for any value of $r_s$ (hollow squares curves). Actually, the three curves follow the point-like source toy formula (\[eq:J\_final\]) appropriate for steep $\gamma$, i.e. $$J(\theta=0) \propto \rho_\mathrm{s}^2 \times \frac{r_\mathrm{s}^3}{d^2}\,.$$ However, when the angular size of the emitting region becomes larger than the integration angle, the above relationship fails. As most of the flux is emitted within $r_\mathrm{s}$, this happens for a critical distance $$d_{\rm crit} \approx \frac{r_\mathrm{s}}{\alpha_{\rm int}}\,. \label{eq:d_crit}$$ For $r_\mathrm{s}=0.1$ kpc, this corresponds to $d_{\rm crit}\approx60$ kpc (see the full circles dashed curve for $\gamma=0$). Having a dSph closer than this critical distance does not increase further the signal (see, e.g., the solid and dotted full circles curves for $\gamma=0$ and $r_s\gtrsim 0.5$ kpc). In the latter case, taking a larger integration region is not always the best strategy as, from an experimental point of view, a larger integration region increases not only the signal but also the background. In this case, the gain in sensitivity from having a dSph close by is not as important as what might naïvely be expected from the point-like approximation (see Section \[sec:detectability\]). Integration angle dependence $J(\alpha_{\rm int})$ -------------------------------------------------- We recall that $ \int_{\Delta\Omega} \mathrm{d}\Omega= \int_{0}^{2\pi} \mathrm{d}\beta_{\rm int} \int_{0}^{\alpha_{\rm int}} \sin(\alpha_{\rm int}) \mathrm{d}\alpha_{\rm int}$, where $\Delta\Omega = 2\pi(1-\cos(\alpha_{\rm int}))$, so that the $J$-factor from Eq. (\[eq:J\]) can be rewritten in the symbolic notation $$J(\psi,\theta, \Delta\Omega) = \int_{0}^{2\pi} F_{[\beta_{\rm int}]} \,\mathrm{d}\beta_{\rm int} \label{eq:J1}$$ with $$F_{[\beta_{\rm int}]} = \int_{0}^{\alpha_{\rm int}} F_{[\beta_{\rm int},\alpha_{\rm int}]} \,\mathrm{d}\alpha_{\rm int}' \label{eq:J2}$$ and $$F_{[\beta_{\rm int},\alpha_{\rm int}]} = \sin(\alpha_{\rm int}) \int_{0}^{l_{max}} {\cal F}\large[r(l,\beta_{\rm int},\alpha_{\rm int})\large] \,\mathrm{d}l. \label{eq:J3}$$ For small integration angles and the case of a flat enough profile, the integrand in Eqs. (\[eq:J2\]) and (\[eq:J3\]) does not vary much with $\alpha_{\rm int}$, so that for the smooth (${\cal F}\equiv\rho^2$) and the mean sub-clumps (${\cal F}\equiv\rho$), we have $$J_{\rm sm}\propto \alpha_{\rm int}^2 \quad {\rm and} \quad \langle J_{\rm subcl}\rangle\propto \alpha_{\rm int}^2\,. \label{eq:alphint2_dep}$$ Fig. \[fig:generic\_Jalphaint\_sm\_and\_subcl\] shows the integration angle dependence for the smooth $(1-f)^2J_{\rm sm}$ (dashed lines) and the sub-clump mean $\langle J_{\rm subcl}\rangle$ (dotted lines) contributions. (The pointing direction is towards the dSph centre.) ![$J \times (d/100$ kpc$)^2$ as a function of $\alpha_{\rm int} \times (d/100$ kpc) for a generic dSph with $r_\mathrm{s}$=0.5 kpc: smooth (thick dashed lines) and sub-clumps (thin dotted lines). With this rescaling, the case $d=10$ kpc (stars) superimposes on the case $d=100$ kpc (empty and full circles).[]{data-label="fig:generic_Jalphaint_sm_and_subcl"}](figC2.eps){width="\linewidth"} For $\gamma=0$ (solid black circles), the $\alpha_{\rm int}^2$ scaling holds up to $\alpha_{\rm int}^{\rm crit} \sim\!3^\circ$ if $d=10$ kpc (as given by Eq. \[eq:d\_crit\]). A plateau is reached when the entire emitting region of the dSph is encompassed (i.e. for a few $r_\mathrm{s}/d$). For $\gamma=1$ (blue empty circles), the curves are slightly more difficult to interpret, as the profile is not steep enough for it to be considered fully point-like (and thus ‘independent’ of $\alpha_{\rm int}$) given the integration angles considered. [^20] Finally, the rescaling used in Fig. \[fig:generic\_Jalphaint\_sm\_and\_subcl\] implies: $$J_{\rm d1} (\alpha_{\rm int}) = J_{\rm d2}\left(\alpha_{\rm int}\frac{d_2}{d_1}\right) \times \left(\frac{d_2}{d_1}\right)^2\;. \label{eq:rescale_Jalpha}$$ Complementary study of the boost factor {#app:boost} ======================================= In Section \[sec:boost\], we concluded that the boost could not be larger than a factor of 2 for all configurations where the sub-clump spatial distribution follows that of the smooth halo in the dSph. The calculations were also made for a ‘reference’ configuration of the sub-clumps. However, the boost can be smaller (or larger) when the latter parameters are varied. In Tab. \[tab:boost\], we systematically vary all the parameters entering the calculation in order to compare with the reference model case. The two quantities of importance are the maximum boost possible (which is obtained when $\alpha_{\rm int}$ fully encompasses the clump), and the transition point $\alpha_{\rm int}d$ for which the boost equals 1 (the minimum value is always given by $(1-f)^2$). The [*reference*]{} results correspond to the numbers obtained from the dotted lines in Fig. \[fig:generic\_boost\_default\], i.e. for $r_\mathrm{s}=1$ kpc. Note that most of the values for $B_{\rm max}$ in the Table would be close to unity if $r_\mathrm{s}=0.1$ kpc were to be selected. Config.$^\dagger$     $\gamma=0$ $\gamma=0.5$ $\gamma=1$ ---------------------------------- ----- --------------------- ------------------- ----------------- --     reference$^{\ddagger}$     $1.9~|~19$ $2.2~|~21$ $2.0~|~30$ [**\[global parameters\]**]{}     $\alpha=1$     $1.0~|~40$ $1.3~|~60$ $~~1.6~|~160$ $\beta=5$     $2.3~|~11$ $2.0~|~18$ $1.3~|~36$ $R_{\rm vir}=6$ kpc     $3.0~|~15$ $3.5~|~20$ $2.9~|~29$ $M_{300}=2\cdot10^7M_\odot$     $1.3~|~66$ $1.4~|~52$ $~1.3~|~64$ [**\[sub-clump parameters\]**]{}     $dP/dV=$Einasto$^\star$     $1.4~|~\dots\!\!$ $1.7~|~\dots\!\!$ $1.7~|~22$ $a=1.7$     $1.3~|~62$ $1.5~|~50$ $1.3~|~61$ $a=2.0$     $2.8~|~0.2\!$ $3.4~|~8~\,$ $2.9~|~16$ $M_{\rm min}=1 M_\odot$     $1.5~|~43$ $1.7~|~37$ $1.5~|~47$ $M_{\rm max}=10^4M_\odot$     $\!\!2.4~|~4~$ $2.8~|~14$ $2.5~|~22$ $f=0.5$     $3.4~|~10$ $4.2~|~16$ $3.5~|~25$ $\rho_{\rm subcl}=$Einasto     $~8.7~|~0.05\!\!\!$ $10.6~|~0.35\!\!$ $9.0~|~4~~$ $c_{\rm vir}\times 2$     $~7.6~|~0.06\!\!\!$ $~\,9.3~|~0.4$ $7.9~|~4.5\!\!$ : Maximum boost and transition regime, i.e. ($\alpha_{\rm int}d)_{B=1}$ in deg kpc, for which $B=1$, for various smooth/sub-clump parameters for three inner slope $\gamma$ (for the smooth).[]{data-label="tab:boost"} $^\dagger$ All parameters are as for [*reference*]{}, except those quoted.\ $^\ddagger$ Reference configuration ($M_{300}=10^7M_\odot$):\ $\cdot$ $\rho_{\rm sm}$ with $(\alpha,\beta,\gamma)=(1,3,\gamma)$ and $dP/dV\propto \rho_{\rm sm}$;\ $\cdot$ $R_{\rm vir}=3$ kpc and $r_\mathrm{s}=1$ kpc (for $\rho_{\rm sm}$ and $dP/dV$);\ $\cdot$ $dP/dM=M^{-a}$ $(a=1.9)$, and $M_{\rm sub} \in [10^{-6}-10^{6}]~M_\odot$;\ $\cdot$ $f=0.2$, $\rho_{\rm subcl}=$NFW, and $c_{\rm vir}-M_{\rm vir}$=B01.\ $^\star$ Einasto parameters taken from @2006AJ....132.2685M. Varying the \[global parameters\] --------------------------------- The four lines under ’\[global parameters\]’ keeps the recipe of $dP/dV \propto\rho_{\rm sm}$, but some previously fixed parameters are now varied. The trend is that a sharper transition zone (larger $\alpha$), a larger radius of the dSph, or a smaller mass imply a larger $B_{\rm max}$. The impact of the outer slope $\beta$ depends on the value of the inner slope $\gamma$. However, the maximum boost factor reached for these parameters is never larger than $\sim3$. The typical transition value lies around $20^\circ$ kpc, which corresponds, for a dSphs located 100 kpc away, to an integration angle of $0.2^{\circ}$. Hence, for all these configuration, large integration angle should be preferred (this is even worse for closer dSph). Varying the \[sub-clump parameters\] ------------------------------------ The remaining lines under \[sub-clump parameters\] show the impact of the choice of the distribution of sub-clumps, the mass distribution parameters (minimal mass and maximal mass of the sub-clumps, slope $a$ of $dP/dM$), and the density profile of the sub-clumps. Relaxing the condition $dP/dV \propto\rho_{\rm sm}$ has no major impact. In @2008MNRAS.391.1685S, a simple Einasto profile with universal parameters was found to fit all halos (from the Aquarius simulation) independently of the halo mass. For that specific case, we use the values found for the Galaxy in @2006AJ....132.2685M. The Einasto profile is steeper than $\gamma=0$ but it decreases logarithmically inwards. Only for $\gamma\gtrsim1$ (for the smooth component) such a model is able to marginally increase the maximum boost w.r.t. the reference model (instead of decreasing it), which is not unexpected.[^21] Varying the mass distribution slope $a$ is understood as follows: for $a\approx1.9$, all decades in mass contribute about the same amount. When $a$ is decreased, the less massive sub-halos dominate, whereas for $a\gtrsim1.9$, the most massive sub-halos dominate the luminosity (e.g. Fig. 4 of ). This has to be balanced by the fact that the fraction of DM going into sub-clumps remain the same ($f=0.2$), regardless of the value of $a$, so that the total number of clumps in a mass decade also changes. The net result is a smaller boost when $a$ is decreased, and a larger boost from the more massive sub-structure when $\alpha$ is increased. In a similar way ($a$ is now fixed to 1.9 again), the mass also impact on $B$, but in a marginal way. The only sizeable impact comes from varying the fraction of mass into clumps, the sub-clump profile or the concentration of sub-clumps. In the first case, when $f$ increases, the smooth signal decreases by $(1-f)^2$ whereas the sub-clump signal increases as $f$. Even if $f$ is increased up to 50%, which is very unlikely (recent simulations such as @2008MNRAS.391.1685S tend to give an upper limit of $f\lesssim 10 \%$) this gives only a mild enhancement. In the second configuration, the NFW profile for the sub-clumps is replaced by an Einasto one. Despite its logarithmic slope decreasing faster than the NFW slope $\gamma=1$ below some critical radius, the latter profile is known to give slightly more signal than the NFW one ($\rho_{\rm Einasto}(r)>\rho_{\rm NFW}(r)$ for a region that matters for the $J$ calculation). This results in a boost close to 10, regardless of the dSph’s smooth profile. Finally, we recall that the B01 $c_{\rm vir}-M_{\rm vir}$ relation is used to calculate the value of the scale parameter for any sub-clump mass. In the last configuration, the concentration parameter is simply multiplied by a factor of 2, which is probably not realistic. Again, the same boost of $\sim10$ is observed. Accordingly, for these last two cases, the transition angle is reduced, and corresponds to $\alpha_{\rm int}< 0.01^\circ$ (for a dSphs at 100 kpc). To conclude, although boosts by as large as a factor of 10 can be obtained through suitable combinations of parameters, most of these combinations are unlikely and require the signal to be integrated on large angles. Impact of the PSF of the instrument {#app:PSF} =================================== Fig. \[fig:PSFapprox\] shows the impact of the instrument angular resolution on the 80% containement radius for $J$ (for the generic dSphs studied in Sec. \[sec:detectability\]). The solid line corresponds to the quadrature approximation given by Eq. (\[eq:PSF\]), whereas the symbols correspond to the convolved PSF$*$halo profile. The PSF is described by the sum of two Gaussians and is a scaled (factor two improved) version of the PSF appropriate for H.E.S.S. at 200 GeV. Calculated halos for a range of $\alpha,\,\beta,\,\gamma$ models consistent with the stellar kinematics of the classical dSphs are shown as gray squares. The quadrature sum approximation used in this work is shown as a solid line. ![80% containment radius ($\theta_{80}$) of PSF-convolved DM annihilation halo models versus $\alpha_{80}$.[]{data-label="fig:PSFapprox"}](figE1.eps){width="\linewidth"} Confidence levels and priors {#app:CLs} ============================ In this Appendix, we describe how confidence intervals for the quantities such as $\rho(r)$ or $J$ are chosen. Sensitivity of the result to the choice of prior ------------------------------------------------ In the Bayesian approach, the PDF of a parameter $x$ is given by the product of the MCMC output PDF ${\cal P}(x)$ and the prior $p(x)$. The resulting PDF is therefore subjective, since it depends on the adoption of a prior. However, whenever the latter are not strongly dependent on $x$, or if ${\cal P}(x)$ falls in a range where $p(x)$ does not strongly varies, the PDF of the parameter becomes insensitive to the prior. This happens for instance if the data give tight constraints on the parameters. In our MCMC analysis, we assumed a flat prior for all our halo parameters, as there is no observationally motivated reason for doing otherwise. Note, however, that flat priors on the model parameters do not necessarily translate into flat priors on quantities derived from those parameters. Specifically, the flat priors on our model parameters imply a non-flat prior on the DM density (and also on its logarithm) at a given radius, and hence a non-flat prior on $J$. In principle it is possible to choose a combination of priors for the parameters that would translate into flat priors on $\rho(r)$, but we have not done so here. The general impact of such choices, and the methodology to study the prior-dependent results, has been discussed in the context of cosmological studies by @2008PhRvD..78f3521V. In this study, we only use a flat prior on the parameters (or on the log for $r_\mathrm{s}$ and $\rho_s$). The test with artificial data demonstrate that our reconstructed $\rho$ and $J$ values are sound. Confidence intervals for $\rho(r)$ and cross-checks --------------------------------------------------- ### Definition Confidence intervals $\Delta_x$ (CI), associated with a confidence level $x\%$ (CL), are constructed from the PDF. The asymmetric interval $\Delta_x\equiv [\theta^-_x, \theta^+_x]$ such as $${\rm CL}(x)\equiv \int_{\Delta_x}{\cal P}(\theta) {\rm d} \theta= 1 - \gamma,$$ defines the $1 - \gamma$ confidence level (CL), along with the CI of the parameter $\theta$. We rely on two standard practises for the CI selection. The first one (used only in this Appendix) is to fix $\theta^-_x$ to be the lowest value of the PDF. The CLs correspond then to quantiles. This is useful for CI selection of $\chi^2$ values, to ensure that the best-fit value of a model (i.e. the lowest $\chi^2$) falls in the CI (see, e.g., Fig. 7 of ). In the second approach, the CI, i.e $\theta^-_x$ (resp. $\theta^+_x$), is found by starting from the median $\theta^{\rm med}$ of the PDF and decrease (resp. increase) $\theta_x$ until we get $x\%/2$ of the integral of the PDF. This approach ensures that the median value of the parameters falls in the CI, any asymmetry in the CI illustrating the departure from a Gaussian PDF: this is the one used thoughout the paper. ### Comparison of several choices for the PDF of $\rho(r)$ Fig. \[fig:best\_CL\_profiles\] shows the projection for each $r$ of the PDF calculated from the output MCMC file. To do so, $\rho(r)$ is calculated for each entry of the thinned chains and then stored as an histogram. This results in ’boxes’: the larger the box, the more likely the value of $\rho(r)$. From this distribution, we can calculate the median (thick solid black line), the most probable value (thick dotted black line). The thick solid red line correspond to the model having the smallest $\chi^2$ value. We see that the latter differ from the median one for this dSph, though they can be close for other dSph in our sample. In this paper, as our analysis is based on the Bayesian approach, we disregard the best-fit model and only retain the median value. ![Projected distribution of $\log_{10}(\rho)$ along with the value of several other estimators for the MCMC analysis of Draco. In this box projection, the larger the box, the most likely the probability of $\log_{10}(\rho)$. For instance, on the top panel, for $\log_{10}(r)=-1.5$ the probability density function of $\log_{10}(\rho)$ is distributed in the range $[8-10]$ and peaks around $9.5$).[]{data-label="fig:best_CL_profiles"}](figF1.eps){width="\linewidth"} In the first approach, the 68% and 95% CLs are calculated from the distribution $\rho_r$ (at each $r$) They are shown as dashed and dotted thick black lines. Note that none of all the above lines corresponds to a [*physical*]{} configuration of $\rho(r)$. A second approach is to construct the 68% CLs from a sampling of the (still) correlated parameters. This is achieved by using all sets of parameters $\{\vec{\theta}\}_{x\% \rm CL}=\{\vec{\theta}_i\}_{i=1\cdots p}$, for which $\chi^2(\vec{\theta}_i)$ falls in the $68\%$ CL of the $\chi^2$ PDF (see above). Once these sets are found, we calculate $\rho(r)$ for each of them, and keep the maximum and minimum values for each position $r$. This defines envelopes of $\rho(r)$ (CIs are found for each $r$). This is shown as dotted and dashed red lines. Such an approach was used in . The CLs obtained from it are larger than the previous one. In the above paper, the uncertainties were small even with that method, so that was not an issue. However, in this study, this makes a huge difference in the resulting value CL of $J$. In order to check which approach was the correct one, we bootstrapped the Draco kinematic data and calculate from the collection of $\rho(r)$ from each bootstrap sample the median value and the uncertainty. The first approach, where the CLs are directly calculated from the full set of MCMC samples was in agreement with the bootstrap approach, meaning that the second one biases the results toward too large uncertainties. The results of the paper rely thus on the first and correct approach. Artificial data sets: validation of the MCMC analysis {#app:fake} ===================================================== In this section, we examine the reliability of the Jeans/MCMC analysis by applying it to artificial data sets of 1000 stellar positions and velocities drawn directly from distribution functions with constant velocity anisotropy. We assume the form $L^{-2\beta_{\rm aniso}} f(\varepsilon)$ for the distribution functions, where the (constant) velocity anisotropy is given by $\beta_{\rm aniso} = 1 - \sigma_{\rm t}^2/\sigma_{\rm r}^2$, with $\sigma_{\rm t}^2$ and $\sigma_{rm r}^2$ being the second moments of the velocity distribution in the radial and tangential directions, respectively. The function $f(\varepsilon)$ is an unspecified function of energy $\varepsilon$ which we determine numerically using an Abel inversion once the halo model and stellar density are specified [@1991MNRAS.253..414C]. We used the same models in @letter_dsph, but we present here a more general study. The set of artificial data covers a grid of models with $\gamma = {0.1,0.5,1.0}, r_{\rm h}/r_{\rm s} = {0.1, 0.5, 1.0}$ and $\beta=3.1$. For each halo model, we assume $\beta_{\rm aniso}$ values of $0$ (isotropic), $0.25$ (radial) or $-0.75$ (tangential): the $\beta_{\rm aniso}$ values for the anisotropic models are chosen to give models with roughly equivalent levels of anisotropy (in terms of the ratios of the velocity dispersions in the radial and tangential directions). We also generate a grid of models with a steeper inner slope $\gamma = 1.5$ and $\beta=4.0$. In all cases, the haloes contain $\sim 10^7$M$_\odot$ within 300pc. We mimic the effects of observational errors by adding Gaussian noise with a dispersion of $2\,$kms$^{-1}$ to each individual stellar velocity generated from the distribution function. The reconstruction depends on the choice of the prior $\gamma_{\rm prior}$, and this effect is explored in the two sections below. Prior: $0\leq\gamma_{\rm prior}\leq1$ versus $0\leq\gamma_{\rm prior}\leq2$ {#app:biases1} --------------------------------------------------------------------------- {width="0.49\linewidth"} {width="0.49\linewidth"} {width="0.49\linewidth"} {width="0.49\linewidth"} We start with the free $\gamma_{\rm prior}$ analysis (see Sec. \[subsec:gamma-free\]) based on two different priors. Top panels of Fig. \[fig:fake\_par6\] show the ratio of the reconstructed median profile to the true profile. There is no significant differences for $\rho(r\gtrsim 1$ kpc) when using the prior $0\leq\gamma_{\rm prior}\leq2$ (top right) or $0\leq\gamma_{\rm prior}\leq1$ (top left): at large radii, the profile does not depend any longer on the $\gamma$ parameter. However, it is striking to see that restricting the prior to $0\leq\gamma_{\rm prior}\leq1$ greatly improves the determination of the inner regions for the profile, regardless of the value of $\gamma_{\rm true}$. Even for $\gamma_{\rm true}=1.5$ (green curves), using an incorrect prior does not degrade to much the reconstruction of the profile. This results is further emphasised when looking at $J$. The bottom panels of Fig. \[fig:fake\_par6\] are plotted with the same scale to emphasise the difference. As $J$ integrates over the inner parts of the profile, the median MCMC value can strongly differ from the true value for cuspy profiles. This difference can reach up to 5 orders of magnitude (over the whole range of $\alpha_{\rm int}$) for $\gamma_{\rm true}\gtrsim 0.5$ when using the prior $0\leq\gamma_{\rm prior}\leq2$ . The prior $0\leq\gamma_{\rm prior}\leq1$ does generally better, and accordingly, the confidence intervals are much smaller than for the other prior (for any integration angle). The behaviour of the $\gamma_{\rm true}=1.5$ case is unexpected. Using the prior $0\leq\gamma_{\rm prior}\leq1$ does better than the other one for any integration angle. Indeed, even if the reconstructed median value is shifted by a factor of 10, its CLs correctly encompass the true value. It does better than the $0\leq\gamma_{\rm prior}\leq1$ prior, which correctly provides CLs (that bracket the true value), but which are completely useless as these CLs can vary on $\sim 8$ orders of magnitude. Strong prior: $\gamma_{\rm prior}$ fixed {#app:biases2} ---------------------------------------- ![Fixed $\gamma_{\rm prior}$ MCMC analysis. [**Top:**]{} $\rho(r)$. [**Bottom:**]{} $J(\alpha_{\rm int})$.[]{data-label="fig:fake_J_par5"}](figG2a.eps "fig:"){width="\linewidth"} ![Fixed $\gamma_{\rm prior}$ MCMC analysis. [**Top:**]{} $\rho(r)$. [**Bottom:**]{} $J(\alpha_{\rm int})$.[]{data-label="fig:fake_J_par5"}](figG2b.eps "fig:"){width="\linewidth"} In Fig. \[fig:fake\_J\_par5\] below, we use a prior $\gamma_{\rm prior}=0$ for models having $\gamma_{\rm true}=0$, a prior $\gamma_{\rm prior}=0.5$ for models having $\gamma_{\rm true}=0.5$, etc. A comparison of Figs. \[fig:fake\_par6\] (using $0\leq\gamma_{\rm prior}\leq1$ or $0\leq\gamma_{\rm prior}\leq2$) and \[fig:fake\_J\_par5\] (fixed $\gamma_{\rm prior}$) shows that the latter prior only slightly improves the precision of the $J$-factor reconstruction for $\gamma_{\rm true}=0$, $\gamma_{\rm true}=0.5$, and $\gamma_{\rm true}=1$. However, if $\gamma_{\rm true}=1.5$ (green curves), although the corresponding $J$-factor is now better reconstructed than when using the prior $0\leq\gamma_{\rm prior}\leq2$ (Figs \[fig:fake\_par6\], top panel), it is surprisingly less reliable than the strongly biased $0\leq\gamma_{\rm prior}\leq1$ prior. The main conclusion is that the knowledge of $\gamma_{\rm true}$ does not help providing tighter constraints on $J$: the uncertainty remains a factor of a few, except when the inner profile is really cuspy ($\gamma_{\rm true}=1.5$), in which case it becomes strongly biased/unreliable. Other reconstruction ’biases’ on the $J$-factor {#app:biases} =============================================== In this Appendix, the MCMC analysis is performed based on the prior $0\leq\gamma_{\rm prior}\leq1$, for which the analysis is found to be the most robust (see previous Appendix). Impact of the binning of the stars {#app:binning} ---------------------------------- Figure \[fig:momentprofiles\_binning\] shows the impact of using different binnings in the MCMC analysis. The left panel shows the reconstructed (median) value of the velocity dispersion as a function of the logarithm of $r$ (to emphasise the differences at small radii), for a binning used in this paper (black; where each of $\sqrt{N}$ bins has $\sqrt{N}$ member stars, where $N$ is the total number of members), a binning with two times (red) and four times (blue) fewer bins. For Fornax and Sculptor, the profiles are insensitive to the binning chosen, so that the reconstruction of the $J$ values median and 68% CLs (right panel) is robust. For other dSphs, either the adjusted velocity dispersion profile is affected at small radii, or at large radii. In the latter case, the $J$ calculation should not be affected, as the outer part does not contribute much to the annihilation signal. In the former case, a deviation even at small radii can affect the associated $J$ by a factor of a few. The exact impact depends on the integration angle, the distance to the dSph (which corresponds to a given radius), and the ’cuspiness’ of the reconstructed profile (the $J$ value of a core profile will be less sensitive to differences in the inner parts than would be a cuspy profile). For instance, Draco and Leo1 both have a 2 km/s uncertainty below 100 pc, but Draco is three times closer than Leo1: their $J$ for a given $\alpha_{\rm int}$ have different behaviours (right panel). The strongest impact is for Leo1 that have the fewest data. The flatness of the $J$ curve seems to indicate a cuspy profile (all the signal in the very inner parts), which we know are the least well reconstructed ones (see Appendix \[app:biases1\]). Leo1 is thus the most sensitive dSph to the binning, for which a balance between a sufficient coverage over $r$ and small error bars cannot be achieved. The ultra-faints dSphs are expected to have even fewer stars, so that their $J$ calculation is expected to be even more uncertain. Overall, the choice of the binning can produce an additional bias of a few on the $J$ reconstruction. This is an extra uncertainty factor that makes Fornax and Sculptor the more robust targets with respect to their annihilation signal. Surveys in the inner parts and outer parts of Carina, Draco, Sextans, Leo I, Leo II and Ursa Minor are desired to get rid of this binning bias. {width="0.495\linewidth"} {width="0.48\linewidth"} Impact of the choice of the light profile {#app:light_profile} ----------------------------------------- Figure \[fig:biases\] shows the various median values and 68% CIs of J when changing the assumptions made on the light profile. The black lines labeled ’physical’ correspond to the Plummer model used for the main analysis (see Eq. \[eq:plummer\]); the red lines labeled ’unphysical’ are also Plummer, but the physical constraint given by Eq. (\[eq:anevans\]) is relaxed; the blue lines and green lines correspond respectively to a light profile modeled with an $(\alpha,\,\beta,\,\gamma)$ profile in order to get a steeper outer slope $(2,6,0)$ or a steeper inner slope $(2,5,1)$ with respect to the Plummer profile. {width="0.7\linewidth"} Regardless of the light profile used, we recover similar critical angles for which $J$ is the most constrained. The impact on the $J$ value is strongest for the least-well measured profiles (Leo 2 and Sextans), but is contained within the 95% CI and marginally within the 68% CL. \[lastpage\] [^1]: E-mails:[email protected] (JH), [email protected] (DM), [email protected] (MGW) [^2]: Technical details are deferred to Appendices. In Appendix \[app:defs\], we comment on the various notations used in similar studies and provide conversion factors to help compare results. In Appendix \[app:toyJ\], we provide a toy model for quick estimates of the $J$-factor. In Appendix \[app:boost\], we calculate in a more systematic fashion the range of the possible ‘boost factor’ (due to DM clumps within the dSphs) for generic dSphs. In Appendix \[app:PSF\], we show that convolving the signal by the PSF of the instrument is equivalent to a cruder quadrature sum approximation. In Appendix \[app:CLs\], we discuss some technical issues related to confidence level determination from the MCMC analysis. In Appendix \[app:fake\], the reconstruction method is validated on simulated dSphs. In Appendix \[app:biases\], we discuss the impact of the choice of the binning of the stars and of the shape of the light profile on the $J$-factor determination. [^3]: This effect depends on the mass and the velocity of the particle; the resulting boost of the signal and the impact on detectability of the dSphs has been discussed, e.g., in @2009MNRAS.399.2033P. [^4]: In Appendix \[app:toyJ\], we provide approximate formulae for quick estimates of the $J$-factor and cross-checks with the numerical results. [^5]: For a different mass for the dSph, the results for $J$ below have to be rescaled by a factor $(M_{300}^{\rm new}/10^7 M_\odot)^2$ since the density is proportional to $M_{300}$, while $J$ goes as the density squared. [^6]: More details about the clump distributions can be found in Appendix \[sec:sub-clumps\]. See also, e.g., Section 2 in and references therein, as we use the same definitions as those given in that paper. [^7]: The difference between the level of boost observed for $r_\mathrm{s}=0.1$ kpc or $r_\mathrm{s}=1$ kpc can be understood if we recall that the total mass of the clump is fixed at 300 pc, regardless of the value of $\gamma$ or $r_\mathrm{s}$. For $r_\mathrm{s}=0.1$ kpc, $\rho_\mathrm{s}\sim {\cal O}(10^9M_\odot$ kpc$^{-3})$, whereas for $r_\mathrm{s}=1$ kpc, $\rho_\mathrm{s}\sim{\cal O}(10^7M_\odot$ kpc$^{-3})$. As $J_{\rm sm} \propto \rho_\mathrm{s}^2$ whereas $J_{\rm sub} \propto \rho_\mathrm{s}$, the relative amount of $J_{\rm sub}$ with respect to $J_{\rm sm}$ is expected to decrease with smaller $r_\mathrm{s}$. This is indeed what we observe in the figure (solid vs dashed lines). [^8]: Kinematic samples are often contaminated by interlopers from the Milky Way foreground. Following W09, we discard all stars for which the algorithm described by @walker09b returns a membership probability less than $0.95$. [^9]: @edo96 and @hargreaves96b conclude that this assumption is valid for the classical dSphs studied here, which have measured velocity dispersions of $\sim 10$ km s$^{-1}$. This conclusion does not necessarily apply to recently-discovered ‘ultra-faint’ Milky Way satellites, which have measured velocity dispersions as small as $\sim 3$ km s$^{-1}$ [@2010ApJ...722L.209M]. [^10]: For consistency with @2009ApJ...704.1274W we define $r_\mathrm{half}$ as the radius of the circle enclosing half of the dSph stellar light as seen in projection. Elsewhere this radius is commonly referred to as the ‘effective radius’. [^11]: available at http://cosmologist.info/cosmomc [^12]: Note that this behaviour is grossly representative of all dSphs, although the integration angle for which the uncertainty is the smallest, and the amplitude of this uncertainty depend, respectively, on the dSph distance (see Section \[sec:gen\_dep\] for the generic dependence), and on the range/precision of the kinematic data (see above). [^13]: The smooth profile is taken to be an Einasto profile, the clump distribution is a core one, whereas their inner profile are Einasto with concentration and parameters [*à la*]{} @2001MNRAS.321..559B Normalising the mass distribution to have 100 clumps more massive than $10^8M_\odot$, and taking $dP/dM\propto M^{-1.9}$ leads to a DM fraction into clumps of $\sim 10\%$ for clumps distributed in the range $10^{-6}-10^{10} M_\odot$ . The local DM distribution is fixed to the fiducial value $\rho_\odot=0.3$ GeV cm$^{-3}$. The exact configuration is unimportant here as this plot is mostly used for illustration purpose. [^14]: This is understood as for the latter, the inner region ($r\ll r_s$) contribute the most to $J$, and even small differences for $\rho(r\sim r_s)$ are bound to translate in sizeable differences for $\rho(r\rightarrow 0)$. Conversely, similar differences on $\rho$ for shallower profiles is not an issue as their inner parts do not contribute to $J$. [^15]: CLs for $J(\alpha_{\rm int})$ are provided along with the paper for readers interested in applying our analysis to existing and future observatories. [^16]: The energy threshold for a ground based instrument is dependent on the zenith angle of observation. This means that the actual energy threshold for a given object will depend on the object’s declination and the latitude of the, yet to be determined FCA site. [^17]: In particular there is a bright GeV emitter 1FGL J0058.4-3235 only $\sim 1.1^{\circ}$ away from Sculptor which significantly worsens the upper limit on that object as discussed by @2010ApJ...712..147A. [^18]: Some authors do however explicitly express the solid angle dependence in their units, e.g. , who express $J$ ($\Phi_{\rm cosmo}$ in their notation) in ${\rm GeV}^2{\rm cm}^{-6} {\rm kpc}\;{\rm sr}$. This is completely equivalent to our $M_\odot^2{\rm kpc}^{-5}$ but for the unit numerical conversion factor. [^19]: In this toy model, we limit ourselves to the NFW profiles for the sub-clumps in the dSph, and a $c_{\rm vir}-M_{\rm vir}$ relation taken from @2001MNRAS.321..559B. [^20]: The dependence can be understood by means of the toy model formulae (\[eq:J\_final\]) and (\[eq:J\_final\_def\]). For $\alpha_{\rm int}<\alpha_{\rm int}^{\rm crit}$, we have $$J_{[\gamma \gtrsim 0.5]} \propto r_\mathrm{s}^{2\gamma} \times (\alpha_{\rm int} d)^{3-2\gamma}\,. \nonumber$$ For $\gamma=1$ (empty blue circles), $J$ is then expected to scale linearly with $\alpha_{\rm int}$, which is observed for the smooth (dashed blue line), and to some extent for the sub-clump contribution (dotted blue line). However, for the latter, the transition region (around $r_\mathrm{s}$) falls from a slope $\alpha=1$ towards an outer slope $\beta=3$ (instead of falling from $\alpha^2=1$ to $\beta^2=6$. Hence, for $\alpha_{\rm int}>\alpha_{\rm int}^{\rm crit}$, the sub-clump contribution continues to build up gradually. [^21]: For smaller $\gamma$, the smooth distribution, in that case, is flatter than the sub-clump one, so that the boost is larger than one for small $\alpha_{\rm int}$ and the transition where $B=1$ is ill-defined. However, such a configuration is highly unlikely as it is exactly the opposite of what is observed in all N-body simulations.

ArXiv

--- abstract: 'In this work, the existence, uniqueness and regularity of solutions to the time-dependent Kohn-Sham equations are investigated. The Kohn-Sham equations are a system of nonlinear coupled Schrödinger equations that describe multi-particle quantum systems in the framework of the time dependent density functional theory. In view of applications with control problems, the presence of a control function and of an inhomogeneity are also taken into account.' author: - 'M. Sprengel[^1]' - 'G. Ciaramella[^2]' - 'A. Borz[ì]{}[^3]' bibliography: - 'Literatur.bib' title: 'A theoretical investigation of time-dependent Kohn-Sham equations[^4]' --- Introduction ============ The time-dependent density functional theory (TDDFT) was introduced to model multi-particle quantum systems avoiding the solution of the full Schrödinger equation (SE) in multi-dimensions [@KohnSham1965; @ParrYang1989; @RuggenthalerPenzVanLeeuwen2015; @RungeGross1984]. The central concept of TDDFT is to describe the configuration of a $N$-particle quantum system using the density function $\rho$ that depends on the $n$-dimensional physical space coordinates and time. This is in contrast to the wave function representation of the full Schrödinger problem where $n\cdot N$ space coordinates are involved. Specifically, in the TDDFT framework a system of $N$ nonlinear SEs is considered that governs the evolution of $N$ single-particle wave functions $\Psi=(\psi_1,\dotsc, \psi_N)$, $\psi_i=\psi_i(x,t)$, $x\in { \mathbb{R} }^n$, $t\in { \mathbb{R} }$. These SEs are coupled through a potential that depends on the density $\rho(x,t)=\sum_{i=1}^N |\psi_i(x,t)|^2$. This time-dependent Kohn-Sham (TDKS) system is given by $$\label{eq:KS} i \partial_t \Psi(x,t) = I_N \otimes \bigl[- \nabla^2 + V_{ext}(x,t;u) + V(x,t;\Psi) \bigr] \Psi(x,t), \quad \Psi(x,0)=\Psi_0(x)$$ where $\nabla^2$ is the Laplacian, $V_{ext}$ is an external potential that includes the confining potential, e.g., the surrounding walls or the Coulomb potential of the nuclei of a molecule, and, possibly, a control potential. $V$ denotes the coupling $KS$ potential. See [@RuggenthalerPenzVanLeeuwen2015] for a review on this model. The purpose of our work is to theoretically investigate , with a given control function $u\in H^1(0,T)$, and an adjoint version of that appears in the following optimal control problem $$\label{Intro:ControlProblem} \begin{split} \min_{(\Psi,u)\in (W,H^1(0,T))} J_1(\Psi)+J_2(\Psi(T)) + \nu\|u\|_{H^1(0,T)}^2 \quad \text{ s.t. } \Psi \text{ solves } \eqref{eq:KS}, \end{split}$$ where $\nu>0$ is a weight parameter and $J_1$ depends on the whole solution $\Psi$, while $J_2$ depends on the wave function at the final time $\Psi(T)$ only. The functionals $J_1, J_2$ are assumed to be lower semicontinuous and Fréchet differentiable with respect to $\Psi$. To characterize the solutions to using the adjoint method [@Borzi2012], the following adjoint equation is considered. $$\label{eq:KSadjoint} \begin{split} i{\frac{{\partial}\Psi}{{\partial}t}}&=I_N \otimes\left( -\nabla^2 +V_{ext}(x,t,u) +V(\Lambda)\right)\Psi\\ &+ I_N \otimes \left(V_H(2{\operatorname{Re}}{\left(\Psi,\,\Lambda\right)_{{ \mathbb{C} }}})+2 {\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda){\operatorname{Re}}{\left(\Psi,\,\Lambda\right)_{{ \mathbb{C} }}}+D_\psi J_1(\Lambda) \right) \Lambda,\\ \Psi(T)&=-D_\Psi J_2 (\Lambda(T)), \end{split}$$ where we again denote by $\Psi$ the adjoint variable while the solution to is denoted with $\Lambda$. We remark that has a similar structure as with an additional inhomogeneity resulting from the Fréchet derivative $D_\psi J_1(\Lambda) \Lambda$ of $J_1$ with respect to the wave function, as well as additional terms resulting from the linearization of the Kohn-Sham potential. On the other hand, $V$ now depends on $\Lambda$ and is no longer a function of the unknown variables. The derivative of $J_2$ gives a terminal condition for that evolves backwards in time. In this paper, we theoretically analyse and as two particular instances of a generalized TDKS equation, proving existence and uniqueness of solutions. At the best of our knowledge, this problem is only addressed in [@Jerome2015] for . In this reference, the Author proves existence and uniqueness of solutions assuming that the Hamiltonian is continuously differentiable in time. We improve these results in such a way that this theory can accommodate TDKS optimal control problems. In particular, existence and uniqueness of solutions with similar regularity as in [@Jerome2015] are proved also in the case when the external potential is only $H^1$ and not $C^1$. These results are achieved in the Galerkin framework. We remark that by this approach, we address the TDKS equation and its adjoint in an unique framework. Notice that the adjoint problem has a different structure that can make it difficult the use of semi-group theory. This paper is organized as follows. In Section \[sec:model\], we discuss the KS potential $V$ and the external potential $V_{ext}$. Further, we formulate our evolution problem in a weak sense that embodies both and . Also in this section, we discuss the initial and boundary conditions, and provide specific assumptions on the potentials and the spatial domain $\Omega$ and the time interval $(0,T)$ where the KS problem is considered. In Section \[sec:PreliminaryEstimates\], we investigate some properties of the KS potential and discuss continuity properties of the bilinear form resulting from the weak formulation. In Section \[sec:Galerkin\], we use the Galerkin framework to obtain a finite dimensional approximation of our weak problem. In Section \[sec:EnergyEstimates\], we present energy estimates for the finite dimensional representation and their extension to the infinite dimensional case. In Section \[sec:ExistenceSolution\] and \[sec:Uniqueness\], we prove existence and uniqueness of solutions to our weak problem. First, we prove convergence of the Galerkin approximation to the infinite dimensional solution and use our results on the Lipschitz properties of the potential to prove uniqueness of this solution. In Section \[sec:ImprovedRegularity\], assuming$\Psi_0\in H^1_0(\Omega)$, we prove that the solution of our problem has higher regularity. The Sections \[sec:ExistenceSolution\], \[sec:Uniqueness\], and \[sec:ImprovedRegularity\] present our main theoretical results. A section of conclusion completes this work. The model description {#sec:model} ===================== In this section, we introduce the weak formulation of our evolution problem, define the potentials and discuss our assumptions. To introduce the weak formulation of the evolution problem, we define the following function spaces. We use $L^2(\Omega;{ \mathbb{C} }^N)$ where ${\left( \cdot ,\, \cdot \right)_{L^2}}$ is the scalar product defined as follows $${\left( \Psi ,\, \Phi \right)_{L^2}} := \int_{\Omega} {\left( \Psi ,\, \Phi \right)_{{ \mathbb{C} }}} dx ,$$ and $\| \cdot \|_{L^2}$ denotes the corresponding norm. Further, we denote by ${\left( \cdot ,\, \cdot \right)_{{ \mathbb{C} }}}$ the scalar product for ${ \mathbb{C} }^N$ and $| \cdot |$ is the corresponding norm. The scalar product of the Sobolev space $H^1(\Omega;{ \mathbb{C} }^N)$ is given by $${\left( \Psi ,\, \Phi \right)_{H^1}} := {\left( \Psi ,\, \Phi \right)_{L^2}} +{\left( \nabla \Psi ,\, \nabla \Phi \right)_{L^2}} ,$$ and $\| \cdot \|_{H^1}$ denotes the corresponding norm. Furthermore, the following spaces of functions of time and space with function values in ${ \mathbb{C} }^N$ are used.$Y := L^2(0,T;L^2(\Omega;{ \mathbb{C} }^N))$, $X:=L^2(0,T;H_0^1(\Omega; { \mathbb{C} }^N))$ with corresponding norms $\|u\|_Y^2=\int_0^T\|u(t)\|_{L^2}^2 {\mathrm{d}}t$ and $\|u\|_X^2=\int_0^T\|u(t)\|_{H^1}^2 {\mathrm{d}}t$ and its dual$X^*=L^2(0,T; H^{-1}(\Omega; { \mathbb{C} }^N))$ and the space of solutions $W := \{u \in X \text{ such that }\linebreak u'\in X^*\}$. We prove the existence of a solution of the controlled Kohn-Sham model and at the same time of on a bounded domain $\Omega \subset \mathbb{R}^n$, $n = 3$, with homogeneous Dirichlet boundary conditions. For this purpose, we denote by $\Psi \in X$ the vector of the wave functions corresponding to $N$ particles $$\Psi := ( \: \psi_1, \dotsc, \psi_N \: ) ,$$ and assume that $\psi_j(x,t) = 0$ for $x\in\partial \Omega$ and consider the initial condition$\psi_j(x,0) = \psi_{0,j}(x)$ with $\psi_{0,j} \in L^2(\Omega;{ \mathbb{C} })$. Moreover, to include a possible inhomogeneity of the PDE, we consider the function $F \in Y$ defined as follows $$F := ( \: f_1, \dotsc , f_N \: ) ,$$ where $f_j \in L^2(0,T;L^2(\Omega;{ \mathbb{C} }))$. The wave function $\Psi$ gives rise to the density $\rho$ defined as follows $$\rho(x,t) := \sum_j |\psi_j(x,t)|^2 ,$$ which is used to characterize the nonlinear potential $V(x,t;\Psi)$. The dependence of $V$ on $\Psi$ is always through the density $\rho$, so we may also write $V(x,t;\rho)$. In the local density approach (LDA) framework, $V$ is given by the sum of the Hartree, the exchange, and the correlation potentials. We have $$\label{def:potential} \begin{split} V(x,t;\Psi) &= V_H+V_{xc}=V_H+V_x+V_c,\\ V_H&=\int_\Omega \frac{\rho(y,t)}{|x-y|} {\mathrm{d}}y, \quad V_x=V_x(\rho(x,t)), \quad V_c= V_c(\rho(x,t)). \end{split}$$ $V_x$ is often derived from an approximation called the homogeneous electron gas [@ParrYang1989] and then given by $V_x= c \rho(x,t)^\beta$, where $c$ is a negative constant and $0<\beta<1$ depends on the dimension $n$. For the correlation potential $V_c$ only numerical approximation exists. In the course of the years, physicists and quantum chemists have developed a collection of different $V_c$ functions. Similar to Jerome [@Jerome2015], who uses a Lipschitz assumption on $V_x+V_c$, we make some general assumptions on the structure of the potentials rather than using an explicit form for one of the approximation used in applications. The external potential is given by $$V_{ext}(x,t;u)=V_0(x)+V_u(x) u(t),$$ where $V_0$ models a confinement potential, e.g., a harmonic trap in a solid state system or a molecule. The control potential $V_u(x) u(t)$ may represent a gate voltage applied to the solid state system or a laser pulse on the molecule. We consider problems and in a unified framework by introducing a parameter $\alpha$ that indicates the case by $\alpha=1$ and by $\alpha=0$. The inhomogeneity $F$ can be zero as in or given as in . The equations are studied in the following weak form: Find a wave function $\Psi\in X$ with $\Psi'\in X^*$, such that $$\label{eq:KSweak} \begin{split} &i {\left( \partial_t \Psi(t),\, \Phi \right)_{L^2}} = B(\Psi(t),\Phi;u(t)) + \alpha{\left( V(\Psi(t)) \Psi(t),\, \Phi \right)_{L^2}} + {\left( F(t) ,\, \Phi \right)_{L^2}} \\ &\text{a.e. in } (0,T) \text{ and } \forall \Phi \in H_0^1(\Omega; { \mathbb{C} }^N) ,\\ &\Psi(0)=\Psi_0 \in L^2(\Omega; { \mathbb{C} }^N), \end{split}$$ where the bilinear form $B(\Psi,\Phi;u)$ is defined as follows $$\begin{split} \label{eq:bilinearForm} B(\Psi,\Phi;u) := &{\left( \nabla \Psi ,\, \nabla \Phi \right)_{L^2}} + {\left( V_{ext}(\cdot,\cdot;u) \Psi ,\, \Phi \right)_{L^2}}\\ &+(1-\alpha){\left(V(\cdot, \cdot, \Lambda)\Psi,\,\Phi\right)_{L^2}}+(1-\alpha) D(\Psi, \Phi) . \end{split}$$ The additional terms of the adjoint equation are given by $$\label{eq:Dadjoint} \begin{split} D(\Psi, \Phi)&=D_H(\Psi, \Phi)+D_{xc}(\Psi, \Phi), \end{split}$$ where $$\begin{aligned} D_H(\Psi, \Phi)&={\left(V_H(2{\operatorname{Re}}{\left(\Psi,\,\Lambda\right)_{{ \mathbb{C} }}})\Lambda,\,\Phi\right)_{L^2}},\\ D_{xc}(\Psi, \Phi)&={\left(2 {\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda){\operatorname{Re}}{\left(\Psi,\,\Lambda\right)_{{ \mathbb{C} }}} \Lambda,\,\Phi\right)_{L^2}}.\end{aligned}$$ We remark that when studying the adjoint equation, the adjoint variable is also denoted with $\Psi$, and $\Lambda=(\lambda_1,\dotsc,\lambda_N)$ corresponds to the solution of the forward equation . As we later prove in Theorem \[thm:ImprovedRegularity\], the solution of the forward equation $\Lambda$ is in $L^2((0,T); H^2(\Omega))$ and the embedding $H^2(\Omega)\hookrightarrow C(\bar{\Omega})$ guarantees that $\Lambda(t)$ is bounded a.e. in $(0,T)$, see, e.g., [@Ciarlet2013 p. 332]. Here, $C(\bar{\Omega})$ is the space of continuous functions with the norm $\|f\|_{C}=\max_{k=1,\dotsc,N}\sup_{x\in \bar{\Omega}}|f_k(x)|$, In a quantum control setting, the inhomogeneity $F$ is zero in the forward equation and contains the derivative of $J_1$ with respect to the wave function in the adjoint equation. However, for generality we allow a nonzero $F$ when studying . As in the argument above, $\Lambda$ and $J_1(\Lambda)$ are continuous functions of $x$ and hence in $L^2(\Omega)$. To incorporate a final condition $\Psi(T)$ instead of an initial condition $\Psi(0)$, we substitute $t\mapsto T-t$. Now, we want to summarize our assumptions that we make throughout our paper. \[assumptions\] We consider the following. 1. A bounded domain $\Omega \subset { \mathbb{R} }^n$ with $n=3$ and a Lipschitz boundary; and, for the improved regularity in Theorem \[thm:ImprovedRegularity\] and \[thm:ImprovedRegularity2\], ${\partial}\Omega\in C^2$. 2. \[assumptionVcBounded\] The correlation potential $V_c$ is uniformly bounded in the sense that$|V_c(\Psi(x,t))|\leq K,$ $\forall x\in\Omega$, $t\in [0,T]$, $\Psi\in Y$; this is the case, e.g. for the Wigner potential [@Wigner1938]. 3. \[assumptionVxLipschitz\] The exchange potential $V_x$ is Lipschitz continuous in the sense$\|V_x(\Psi)\Psi-V_x(\Upsilon)\Upsilon\|_{L^2}\leq L \|\Psi-\Upsilon\|_{L^2}$, for $\Psi$ and $\Upsilon$ being weak solutions of and locally Lipschitz continuous for $\Psi\in L^2(\Omega)$, i.e. $\|V_x(\Psi)\Psi\|_{L^2}\leq L\|\Psi\|_{L^2}$, where $L$ might depend on $\|\Psi\|_{L^2}$; cf. the similar assumption in [@Jerome2015]. This assumption will be motivated further in Remark \[remark:aposteriori\]. 4. \[assumptionVxcDiffable\] $V_x$ and $V_c$ are weakly differentiable as functions of $\rho$ and ${\frac{{\partial}V_{xc}}{{\partial}\rho}}(\rho) \rho$ is bounded for finite values of $\rho$. For $V_x=c\rho^\beta$, this can be shown directly: $$\begin{aligned} {\frac{{\partial}V_x(\rho)}{{\partial}\rho}}\rho=c\beta \rho^{\beta-1} \rho=c\beta\rho^\beta=\beta V_x(\rho). \end{aligned}$$ 5. The confining potential and the spacial dependence of the control potential are bounded, i.e. $V_0, V_u \in L^\infty(\Omega)$, where $\| f \|_{L^\infty}:=\max_{k=1,\dotsc,N}\operatorname*{ess\,sup}_{x\in \Omega} |f_k(x)|$ is the norm for $L^\infty(\Omega;{ \mathbb{C} }^N)$; as we consider a finite domain, this is equivalent to excluding divergent external potentials. 6. The control is $u \in H^1(0,T)$. This a classical assumption in optimal control, see, e.g. [@VonWinckelBorzi2008]. 7. $\Psi_0\in L^2$ for existence and uniqueness of the forward and adjoint equations and $\Psi_0\in H^1$ for the improved regularity. For the adjoint equation, we assume that the solution of the forward problem $\Lambda$ is in $L^2(0,T; H^2(\Omega; { \mathbb{C} }^N))$, this can be shown by applying Theorem \[thm:ImprovedRegularity\] to the forward problem. Preliminary estimates {#sec:PreliminaryEstimates} ===================== In this section, we study continuity properties of the KS potential and of the bilinear form. We begin with a general result on the Coulomb potential $w(x)=\frac{1}{|x|}$. Then we investigate the continuity of the Hartree potential that is defined as the convolution of $w$ with the density $\rho$, and of the KS potential in more detail. Finally, we prove some estimates for the bilinear forms $B$ and $D$. \[lem:CoulombInLp\] Given a bounded domain $\Omega \subset { \mathbb{R} }^n$ containing the origin, it holds that $w\in L^p(\Omega)$ if and only if $n>p$. By $B_R(0):=\{x\in { \mathbb{R} }^m: |x|<R\}$, we denote the open ball of radius $R\in{ \mathbb{R} }^+$ around the origin. Consider now a ball $B_R(0) \subset\Omega$. Then by using spherical coordinates and the fact that $|x|$ does not depend on the orientation of $x$, we get (see e.g. [@Evans2010]) $$\begin{aligned} &\int_{B_R(0)} \frac{1}{|x|^p} {\mathrm{d}}x =\frac{n\pi^{n/2}}{\Gamma(\frac{n}{2}+1)}\int_0^R \frac{1}{r^p} r^{n-1} {\mathrm{d}}r\\ &=\frac{n\pi^{n/2}}{\Gamma(\frac{n}{2}+1)}\left[\frac{r^{n-p}}{n-p} \right]_0^R =\begin{cases} \frac{n\pi^{n/2}}{\Gamma(\frac{n}{2}+1)} \frac{R^{n-p}}{n-p} <\infty & n>p\\ \infty & n\leq p, \end{cases} \end{aligned}$$ where $\Gamma$ is the $\Gamma$-function. Outside this ball, $\frac{1}{|x|}$ is globally bounded. \[CancesLemma\] For $\Phi, \Psi\in H^1(\Omega; { \mathbb{C} }^N)$ there exists a positive constant $C_u$ such that $$\|V_H(\Phi)\Phi-V_H(\Psi)\Psi\|_{L^2} \leq C_u \left( \|\Phi\|_{H^1}^2+\|\Psi\|_{H^1}^2\right) \|\Phi-\Psi\|_{L^2}.$$ We adapt Lemma 5 in [@CancesLeBris1999] to our case of vector valued functions. To this end, we define $g^k(\Phi,\Psi)(x):=\int_\Omega \frac{\phi_k(y){\overline{\psi_k(y)}}}{|x-y|}{\mathrm{d}}y$, $\tilde{g}(\Phi,\Psi)(x):=\sum_{k=1}^N g^k(\Phi,\Psi)(x)$. Then Lemma 3 in [@CancesLeBris1999] gives $$\begin{aligned} |\tilde{g}(\Phi_1,\Phi_2)(x)|\leq \sum_{k=1}^N |g^k(\Phi_1,\Phi_2)(x)|\leq \sum_{k=1}^N \|\phi_{1,k}\|_{L^2}\|\nabla \phi_{2,k}\|_{L^2}. \end{aligned}$$ Using this fact and setting $\Upsilon=(v_1,\dotsc,v_N)$, we have $$\begin{aligned} \|\tilde{g}(\Phi_1,\Phi_2)\Upsilon\|_{L^2}^2&=\sum_{k=1}^N\|\tilde{g}\upsilon_k\|_{L^2}^2=\sum_{k=1}^N \int_\Omega |\tilde{g}(x)\upsilon_k(x)|^2 {\mathrm{d}}x\\ &\leq\sum_{k=1}^N \left(\sum_{l=1}^N \|\phi_{1,l}\|_{L^2} \|\nabla \phi_{2,l}\|_{L^2}\right)^2 \int_\Omega |\upsilon_k(x)|^2{\mathrm{d}}x\\ &=\left(\sum_{l=1}^N \|\phi_{1,l}\|_{L^2} \|\nabla \phi_{2,l}\|_{L^2}\right)^2 \|\Upsilon\|_{L^2}^2\\ &\leq N\sum_{l=1}^N \left(\|\phi_{1,l}\|_{L^2} \|\nabla \phi_{2,l}\|_{L^2}\right)^2 \|\Upsilon\|_{L^2}^2\\ &\leq N\sum_{l=1}^N \|\phi_{1,l}\|_{L^2}^2 \sum_{j=1}^N\|\nabla \phi_{2,j}\|_{L^2}^2 \|\Upsilon\|_{L^2}^2\\ &=N\|\Phi_1\|_{L^2}^2 \|\nabla \Phi_2\|_{L^2}^2\|\Upsilon\|_{L^2}^2. \end{aligned}$$ Now, we apply this to $\Phi_1=\Phi_2=\Phi$, $\Upsilon=\Phi-\Psi$ to obtain $\|(|\Phi|^2\star w)(\Phi-\Psi)\|_{L^2}\leq \sqrt{N}\|\Phi\|_{L^2}\|\nabla\Phi\|_{L^2}\|\Phi-\Psi\|_{L^2}$. Furthermore, using the decomposition $$\begin{aligned} \sum_{k=1}^N|\phi_k|^2-\sum_{k=1}^N|\psi_k|^2=\sum_{k=1}^N\left(\phi_k {\overline{(\phi_k-\psi_k)}}+{\overline{\psi_k}}(\phi_k-\psi_k) \right), \end{aligned}$$ and $\Upsilon=\Psi$, $\Phi_1=\Phi-\Psi$, $\Phi_2=\Phi$ for the first term, and $\Upsilon=\Psi$, $\Phi_1=\Phi-\Psi$,$\Phi_2=\Psi$ for the second term, we find $$\begin{aligned} \|((|\Phi|^2-|\Psi|^2)\star w) \Psi\|_{L^2}\leq \sqrt{N}\|\Psi\|_{L^2}(\|\nabla \Phi\|_{L^2}+\|\nabla \Psi\|_{L^2})\|\Phi-\Psi\|_{L^2}. \end{aligned}$$ With this, the proof of Lemma 5 in [@CancesLeBris1999] extends to the vector case. \[lem:VContinuousFunction\] The nonlinear KS potential $V$ is a continuous function from $L^2(\Omega)$ to $L^2(\Omega)$. First, we show that $\rho(\Psi)$ is a continuous mapping from $L^2(\Omega)$ to $L^1(\Omega)$ in the sense that from $\Psi^n \stackrel{L^2}{\rightarrow} \hat{\Psi}$ follows $ \rho^n \stackrel{L^1}{\rightarrow} \hat{\rho}$. We have $$\begin{aligned} \|\hat{\rho}-\rho_n\|_{L^1}&=\int_{\Omega}\left|\sum_j |\hat{\psi}_j|^2-\sum_j|\psi_j^n+\hat{\psi}_j-\hat{\psi}_j|^2 \right|{\mathrm{d}}x\\ &\leq \int_\Omega \left| \sum_j |\hat{\psi}_j-\psi_j^n|^2+2|\hat{\psi}_j-\psi_j^n|\,|\hat{\psi}_j| \right|{\mathrm{d}}x\\ &\leq\|\hat{\psi}-\psi^n\|_{L^2}^2+2\|\hat{\psi}-\psi^n\|_{L^2}\|\hat{\psi}\|_{L^2}\\ &=3\max\{c,1\}\|\hat{\psi}-\psi^n\|_{L^2}^2, \end{aligned}$$ where $c:=\|\hat{\psi}\|_{L^2}$ and we use Cauchy-Schwarz inequality. Second, $V_H$ is a continuous function of $\rho$ as follows $$\begin{aligned} \|V_H(\rho_1)-V_H(\rho_2) \|_{L^2}&=\|w\star(\rho_1-\rho_2)\|_{L^2}\leq \|w\|_{L^2} \|\rho_1-\rho_2\|_{L^1}, \end{aligned}$$ where Lemma \[lem:CoulombInLp\] and Young’s inequality [@Schilling2005 Theorem 14.6] are used. Finally, since $V_x$ and $V_c$ are pointwise differentiable as functions of $\rho$, they are also continuous. We continue with some estimates for the bilinear form $B$ for arbitrary wave functions. \[lem:bound-weak-form\] There exist positive constants $c_0$, $c_1$, $c_2$ and $c_3$ such that the following estimates hold $$\label{eq:lem3:estD} |D(\Psi, \Phi)|\leq c_0 \|\Psi\|_{L^2} \|\Phi\|_{L^2} ,$$ $$\label{eq:est1} {\operatorname{Re}}B(\Psi,\Phi;u) \leq |B(\Psi,\Phi;u)| \leq c_1 \| \Psi \|_{H^1} \| \Phi \|_{H^1} ,$$ $$\label{eq:est1im} |{\operatorname{Im}}B(\Psi,\Psi;u)| \leq c_0 \| \Psi \|_{L^2}^2 ,$$ and $$\label{eq:est2} \| \Psi \|_{H^1}^2 \leq {\operatorname{Re}}B(\Psi,\Psi;u) + c_3 \| \Psi \|_{L^2}^2 ,$$ for any $\Psi,\Phi \in H^1(\Omega;{ \mathbb{C} }^N)$. For $D(\Psi, \Phi)$ given by , we use the fact that $\Lambda(\cdot, t)\in L^\infty(\Omega)$ and Assumption \[assumptions\] (\[assumptionVxcDiffable\]) to get $$\begin{aligned} |D_{xc}(\Psi, \Phi)| &=\left|\sum_{j=1}^N \sum_{k=1}^N \int_\Omega {\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda) (\psi_j{\overline{\lambda_j}}+{\overline{\psi_j}}\lambda_j)\lambda_k {\overline{\phi_k}} {\mathrm{d}}x\right|\\ &\leq 2\sum_{j=1}^N \sum_{k=1}^N \int_\Omega \left|{\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda) \right| |\psi_j| \left(\sum_{l=1}^N |\lambda_l| \right)^2 |\phi_k| {\mathrm{d}}x\\ &\leq 2N\sum_{j=1}^N \sum_{k=1}^N \int_\Omega \left|{\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda) \right| \left(\sum_{l=1}^N |\lambda_l|^2 \right) |\psi_j| |\phi_k| {\mathrm{d}}x\\ &\leq 2N \sum_{j=1}^N \sum_{k=1}^N \left\|{\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda)\sum_{l=1}^N |\lambda_l|^2 \right\|_{L^\infty} \int_\Omega (|\psi_j|+|{\overline{\psi_j}}|) |{\overline{\phi_k}}|{\mathrm{d}}x\\ &\leq 2N\left\|{\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda)\sum_{l=1}^N |\lambda_l|^2\right\|_{L^\infty} \sum_{j=1}^N \sum_{k=1}^N \|\psi_j\|_{L^2}\|\phi_k\|_{L^2}\\ &\leq 2N^2\left\|{\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda)\sum_{l=1}^N |\lambda_l|^2\right\|_{L^\infty} \|\Psi\|_{L^2}\|\Phi\|_{L^2}\\ &\leq c_0'\| \Psi \|_{L^2} \| \Phi \|_{L^2}.\end{aligned}$$ Similarly, we have $$\begin{aligned} |D_H(\Psi, \Phi)| &=\left| \sum_{k=1}^N \int_\Omega \int_\Omega \frac{\sum_{j=1}^N(\psi_j(y){\overline{\lambda_j(y)}}+{\overline{\psi_j(y)}}\lambda_j(y))}{|x-y|} \lambda_k(x) {\overline{\phi_k(x)}} {\mathrm{d}}y{\mathrm{d}}x\right|\\ &\leq \|\Lambda\|_{L^\infty}^2 {\left(I_N\otimes\sum_{j=1}^N(|{\overline{\psi_j}}|+|\psi_j|)\star w,\,|\Phi|\right)_{L^2}}\\ &\leq\|\Lambda\|_{L^\infty}^2 \sqrt{N}\left\| \sum_{j=1}^N(|{\overline{\psi_j}}|+|\psi_j|)\star w\right\|_{L^2(\Omega;{ \mathbb{C} })} \|\Phi\|_{L^2}\\ &\leq 2N^\frac{3}{2} \left\|w\right\|_{L^1} \|\Lambda\|_{L^\infty}^2 \|\Psi\|_{L^2}\|\Phi\|_{L^2}\\ &\leq c_0''\| \Psi \|_{L^2} \| \Phi \|_{L^2},\end{aligned}$$ where Young’s inequality, see e.g. [@Schilling2005 Theorem 14.6], is used. Together, we have the desired bound on $D$ with $c_0=c_0'+c_0''$. For the second estimate, we first recall that the embedding $H^1(0,T) \hookrightarrow C[0,T]$ is continuous and compact (see, e.g., [@Ciarlet2013]), hence there exists a positive constant $K$ such that $\| u \|_{C[0,T]} \leq K \| u \|_{H^1(0,T)}$ for any $u \in H^1(0,T)$; this is used for the control $u$. Consequently, recalling , we obtain the following estimate $$\label{eq:pppp} \begin{split} | B(\Psi,\Phi;u) | &\leq \| \nabla \Psi \|_{L^2} \| \nabla \Phi \|_{L^2} +(1-\alpha)\bigl(|D(\Psi,\Phi)|+\|V(\Lambda)\Psi\|\|\Phi\|\bigr)\\ &\quad + \Bigl| \int_{\Omega} {\left( V_0(x) \Psi(x) ,\, \Phi(x) \right)_{{ \mathbb{C} }}} {\mathrm{d}}x \Bigr| + \Bigl| \int_{\Omega} {\left( V_u(x) u(t) \Psi(x),\, \Phi(x) \right)_{{ \mathbb{C} }}} {\mathrm{d}}x \Bigr| \\ &\leq \| \Psi \|_{H^1} \|\Phi \|_{H^1} +(c_0+\|V(\Lambda)\|_{L^\infty}) \| \Psi \|_{H^1} \|\Phi \|_{H^1}\\ &\quad + \| V_0 \|_{L^{\infty}} \| \Psi \|_{L^2} \| \Phi \|_{L^2} + K \| u \|_{H^1(0,T)} \| V_u \|_{L^{\infty}} \| \Psi \|_{L^2} \| \Phi \|_{L^2} \\ &= \Bigl( 1 + c_0+\|V(\Lambda)\|_{L^\infty}+\| V_0 \|_{L^{\infty}} + K \| u \|_{H^1(0,T)} \| V_u \|_{L^{\infty}} \Bigr) \| \Psi \|_{H^1} \| \Phi \|_{H^1} , \end{split}$$ hence there exists a constant $c_1$ such that holds. The estimate is easily verified with the above estimates for $D$, as ${\left(\nabla \Psi,\,\nabla \Psi\right)_{L^2}}$ and ${\left(V_{ext}\Psi,\,\Psi\right)_{L^2}}$ are real. To prove the last statement, we recall , and similar to we have the following $${\left( \nabla \Psi ,\, \nabla \Psi \right)_{L^2}} = B( \Psi , \Psi ; u ) - {\left( V_{ext} \Psi ,\, \Psi \right)_{L^2}} - (1-\alpha)D(\Psi,\Psi)-(1-\alpha){\left(V(\Lambda)\Psi,\,\Psi\right)_{L^2}}.\\$$ Taking the real part of this equation results in $$\label{eq:pppp1} \begin{split} {\left( \nabla \Psi ,\, \nabla \Psi \right)_{L^2}} \leq {\operatorname{Re}}B( \Psi , \Psi ; u ) &+ \Bigl( \| V_0 \|_{L^{\infty}} + K \| u \|_{H^1(0,T)} \| V_u \|_{L^{\infty}}\\ &+(1-\alpha)(c_0 +\|V(\Lambda)\|_{L^\infty})\Bigr) \| \Psi \|_{L^2}^2 . \end{split}$$ Adding $\|\Psi\|_{L^2}^2$ on both sides we obtain $$\label{eq:pppp3} \| \Psi \|_{H^1}^2 \leq {\operatorname{Re}}B( \Psi , \Psi ; u ) +c_3 \| \Psi \|_{L^2}^2 ,$$ where $c_3=\Bigl( \| V_0 \|_{L^{\infty}} + K \| u \|_{H^1(0,T)} \| V_u \|_{L^{\infty}} +c_0 +\|V(\Lambda)\|_{L^\infty}+1 \Bigr)$, hence holds. A Galerkin approach {#sec:Galerkin} =================== In this section, we introduce a finite-dimensional subspace $P_m$ of $H^1_0(\Omega; { \mathbb{C} }^N)$, and show existence of a unique solution of in this subspace. To this end, we take smooth functions ($C^\infty_0(\Omega)$) $\phi_k=\phi_k(x)$, for $k=1,2,\dots$, such that $\{\phi_k\}_k$ is an orthogonal basis for $H^1_0(\Omega)$ and an orthonormal basis for $L^2(\Omega)$. Further, we construct a basis $\{\Phi_k\}_k$ that is orthogonal for $H^1_0(\Omega;{ \mathbb{C} }^N)$ and orthonormal for $L^2(\Omega;{ \mathbb{C} }^N)$ by defining $$\Phi_k(x) := \frac{1}{\sqrt{N}} \begin{pmatrix} \phi_{k}(x) \\ \vdots \\ \phi_{k}(x)\\ \vdots \\ \phi_{k}(x) \end{pmatrix} ,$$ where $\frac{1}{\sqrt{N}}$ is a normalization parameter. For a fixed positive integer $m$, we define a function $\Psi_m$ as follows $$\begin{aligned} \label{eq:form} \Psi_m(x,t) &:= \sum_{k=1}^m d^k_m(t) \Phi_k(x)\\ &=\frac{1}{\sqrt{N}}\sum_{k=1}^{m} \begin{pmatrix} d_{m,1}^k(t)&&&&0\\ &\ddots\\ && d_{m,j}^k(t)\\ &&& \ddots\\ 0&&&& d_{m,N}^k(t) \end{pmatrix} \begin{pmatrix} \phi_{k}(x)\\ \vdots \\ \phi_{k}(x)\\\vdots \\ \phi_{k}(x) \end{pmatrix} ,\end{aligned}$$ where the coefficients $d^k_{m,j}:[0,T] \rightarrow { \mathbb{R} }$ are such that $$\label{eq:initial} d^k_{m,j}(0) = {\left( \psi_{0,j} ,\, \phi_{k,j} \right)_{L^2}} ,$$ for $k= 1 , \dots , m$. The space spanned by the first $m$ basis functions is called $$P_m=\operatorname*{span}_{k=1,\dotsc,m}\{\Phi_k\}.$$ Moreover, by testing for $\Phi=\Phi_k$, we obtain the following $$\label{eq:KSweak_test} i {\left( \partial_t \Psi_m ,\, \Phi_k \right)_{L^2}} = B(\Psi_m,\Phi_k;u) + \alpha{\left( V(\Psi_m) \Psi_m ,\, \Phi_k \right)_{L^2}} + {\left( F ,\, \Phi_k \right)_{L^2}} ,$$ for almost all $0 \leq t \leq T$ and all $k = 1 , \dots , m$. Thus we seek a solution $\Psi_m$ in the form that satisfies the projection of problem onto the finite dimensional subspace $W_m=L^2(0,T;P_m)$. \[lem:GlocallyLipschitz\] Recall that $\Psi_m := \sum_{k=1}^m d_m^k \Phi_k$ as in and define $G^k: { \mathbb{C} }^N\rightarrow { \mathbb{C} }^N$ as $d_m\mapsto G^k(d_m):={\left(V(\Psi_m)\Psi_m,\,\Phi_k\right)_{L^2}}$, $d_m=(d_m^1,\dotsc,d_m^m)$. The map $d_m \mapsto G^k(d_m)$ is locally Lipschitz continuous in $d_m$. We want to show local Lipschitz continuity in $d_m$, i.e. that for every $\epsilon>0$ there exists a positive constant $L$ such that $|G^k(d_m)-G^k(b_m)| \leq L |d_m-b_m^2|$ for all $d_m,b_m\in B_\epsilon(0)$. For a wave function in the Galerkin subspace with $d_m\in B_\epsilon(0)$, we have the following bounds $$\begin{aligned} \|\Psi_m\|_{L^2}^2 &= \int_{\Omega} |\sum_{k=1}^m d_m^k \Phi_k(x)|^2{\mathrm{d}}x\\ &\leq m\sum_{k=1}^m |d_m^k|^2 \int_{\Omega} |\Phi_k(x)|^2{\mathrm{d}}x =m\sum_{k=1}^m |d_m^k|^2\leq m^2 \epsilon^2, \end{aligned}$$ $$\begin{aligned} \|\nabla\Psi_m\|_{L^2}^2 &= \int_{\Omega} |\sum_{k=1}^m d_m^k \nabla\Phi_k(x)|^2{\mathrm{d}}x\\ &\leq m\sum_{k=1}^m |d_m^k|^2 \int_{\Omega} |\nabla \Phi_k(x)|^2{\mathrm{d}}x =m\sum_{k=1}^m |d_m^k|^2 C_m \leq m^2 \epsilon^2 C_m, \end{aligned}$$ with $C_m=\max_{k=1,\dotsc, m} \|\nabla \Phi_k\|_{L^2}^2$. From these two bounds, we obtain$\|\Psi_m\|_{H^1}\leq (c_m+1) m^2\epsilon^2$. Now, we prove local Lipschitzianity for the different potentials. Consider $\Psi_m$ and $\Upsilon_m$ with coefficients in $B_\epsilon(0)$. We obtain the following $$\label{eq:GalerkinVH}\begin{split} |{\left(V_H(\Psi_m)\Psi_m-V_H(\Upsilon_m)\Upsilon_m,\,\Phi_k\right)_{L^2}}| &\leq \|V_H(\Psi_m)\Psi_m-V_H(\Upsilon_m)\Upsilon_m\|_{L^2} \|\Phi_k\|_{L^2}\\ &\leq C_u \left( \|\Psi_m\|_{H^1}^2+\|\Upsilon_m\|_{H^1}^2\right)\|\Psi_m-\Upsilon\|_{L^2}\\ &\leq L \|\Psi_m-\Upsilon_m\|_{L^2}, \end{split}$$ where the constant $L$ depends on the dimension of the Galerkin space $m$, the norm of the derivatives of the basis functions $C_m$ and $\epsilon$. For the exchange-correlation potential, we have from the Assumption \[assumptions\] (\[assumptionVcBounded\]) and (\[assumptionVxLipschitz\]) the following estimates $$\label{eq:GalerkinVxc}\begin{split} |{\left(V_c(\Psi_m)\Psi_m-V_c(\Upsilon_m)\Upsilon_m,\,\Phi_k\right)_{L^2}}| &\leq \|V_c\|_{L^\infty}\|\Psi_m-\Upsilon_m\|_{L^2},\\ |{\left(V_x(\Psi_m)\Psi_m-V_x(\Upsilon_m)\Upsilon_m,\,\Phi_k\right)_{L^2}}| &\leq \tilde{L}\|\Psi_m-\Upsilon_m\|_{L^2}. \end{split}$$ Using the estimates and , we have $$\begin{aligned} |G^k(d_m^1)-G^k(d_m^2)| &=|{\left(V_H(\Psi_m)\Psi_m+V_x(\Psi_m)\Psi_m+V_c(\Psi_m)\Psi_m \right.\\ &\left.\quad-V_H(\Upsilon_m)\Upsilon_m-V_x(\Upsilon_m)\Upsilon_m-V_c(\Upsilon_m)\Upsilon_m,\,\Phi_k\right)_{L^2}}|\\ &\leq |{\left(V_H(\Psi_m)\Psi_m-V_H(\Upsilon_m)\Upsilon_m,\,\Phi_k\right)_{L^2}}|\\ &\quad+|{\left(V_c(\Psi_m)+V_x(\Psi_m)-V_c(\Upsilon_m)-V_x(\Upsilon_m),\,\Phi_k\right)_{L^2}}|\\ &\leq L\|\Psi_m-\Upsilon_m\|_{L^2}+L'\|\Psi_m-\Upsilon_m\|_{L^2}. \end{aligned}$$ Further, we have $$\begin{aligned} \|\Psi_m-\Upsilon_m\|_{L^2}^2&=\int_\Omega \left|\sum_{l=1}^m (d_{m,1}^l-d_{m,2}^l)\Phi_l(x)\right|^2 {\mathrm{d}}x\\ &\leq m \int_\Omega \sum_{l=1}^m |d_{m,1}^l-d_{m,2}^l|^2 |\Phi_l(x)|^2 {\mathrm{d}}x\\ &=m\sum_{l=1}^m |d_{m,1}^l-d_{m,2}^l|^2 \int_\Omega|\Phi_l(x)|^2 {\mathrm{d}}x\\ &\leq m |d_{m,1}(t)-d_{m,2}(t)|^2. \end{aligned}$$ All together, we have that $d_m \mapsto G^k(d_m)$ is locally Lipschitz continuous. To show existence of a unique solution in the finite-dimensional Galerkin space, we use the Carathéodory theorem, see, e.g., [@Walter1998], because the time-dependent coefficients satisfy our differential equation only almost everywhere. \[thm:Caratheodory\] Consider the following initial value problem $$\label{eq:ODECaratheodory} {\partial}_t y(t)=f(t,y),\quad y(0)=\eta.$$ Let $S=[0, T]\times { \mathbb{R} }^m$ and assume $f$ satisfies $f(\cdot,y)\in L^1(0,T)$ for fixed $y$ and a generalized Lipschitz condition $$|f(t,y)-f(t,\bar{y})|\leq l(t)|y-\bar{y}| \quad \text{in }S, \text{ where } l(t)\in L^1(0,T).$$ Then there exists a unique absolutely continuous solution satisfying a.e. in $[0,T]$. \[thm:approxS\] For each integer $m=1,2,\dotsc$ there exists a unique function $\Psi_m \in W_m$ of the form satisfying and . Assuming $\Psi_m$ has the structure , we note from the fact that $\Phi_k$ are an orthonormal basis that $${\left({\partial}_t\Psi_m(t),\,\Phi_k\right)_{L^2}}={\partial}_t{d_m^k}(t).$$ Furthermore $$\begin{aligned} B(\Psi_m, \Phi_k; u)-(1-\alpha)D(\Psi_m, \Phi_k)&=\sum_{l=1}^{m} e^{kl}(t) d_m^l(t),\\ D(\Psi_m, \Phi_k)&=\sum_{l=1}^m \tilde{e}^{kl} {\operatorname{Re}}(d_m^l(t)),\end{aligned}$$ for $e^{kl}:=B(\Phi_l,\Phi_k; u)-(1-\alpha)D(\Phi_l,\Phi_k)$, and $\tilde{e}^{kl}:= D(\Phi_l,\Phi_k)$ $k,l=1,\dotsc,m$. The real part comes from the definition of $D$ which already contains ${\operatorname{Re}}\Psi_m$. Define $f^k(t):={\left(F(t),\,\Phi_k\right)_{L^2}}$. Then becomes a nonlinear system of ODEs as follows $$\begin{aligned} \label{eq:ODEm} &i{\partial}_t{d_m^k}(t)=\sum_{l=1}^{m}e^{kl}(t) d_m^l(t)+(1-\alpha)\sum_{l=1}^{m}\tilde{e}^{kl}(t) {\operatorname{Re}}(d_m^l(t))+f^k(t)+\alpha G^k(d_m^l(t)),\end{aligned}$$ for $k=1,\dotsc,m$ with the initial conditions . In , the first term is linear, the second globally Lipschitz continuous with Lipschitz constant 1, and $f$ is constant with respect to $d_m$. By Lemma \[lem:GlocallyLipschitz\], $G^k$ is locally Lipschitz continuous in $d_m$ on every ball $B_r(0)$, so the right hand side is locally Lipschitz in $d_m$. As $G^k(d_m(t))$ depends on $t$ only through $d_m^l(t)$ and $f\in L^2(0,T)$ and $e^{kl}\in H^1(0,T)$ through $u$, the right hand side is also in $L^2(0,T)$ and therefore the required $L^1(0,T)$-bound exists. Hence, we can invoke the Carathéodory theorem to show that has a unique solution in the sense of Theorem \[thm:Caratheodory\]. Energy estimates {#sec:EnergyEstimates} ================ In this section, we discuss energy estimates concerning our evolution problem that are used to prove existence of solutions in $W$. Further, we apply these energy estimates for solutions in $W$ to show uniqueness of the solution. \[thm:estimates\] Let $\Psi\in W_m$ be a solution of $$\label{eq:WeakFormHilbertSpace} i {\left( \partial_t \Psi(t) ,\, \Phi \right)_{L^2}} = B(\Psi(t),\Phi;u(t)) + \alpha {\left( V(\Psi(t)) \Psi(t) ,\, \Phi \right)_{L^2}} + {\left( F(t) ,\, \Phi \right)_{L^2}}, \quad \forall \Phi \in P_m,$$ a.e. in $(0,T)$. Then there exist positive constants $C$, $C_0$, $C_1$, $C'$ and $C''$ such that the following estimates hold $$\begin{aligned} \label{eq:estimate-1} \max_{0\leq t \leq T} \| \Psi(t) \|_{L^2}^2 &\leq C \Bigl( \| \Psi_0 \|_{L^2}^2 + \| F \|_Y^2 \Bigr) ,\\ \label{eq:Bestimate} {\operatorname{Re}}B(\Psi(t), \Psi(t); u)&\leq |B(\Psi(t), \Psi(t); u)| \leq C_0(\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2) \text{ for a.a. }t\in [0,T],\\ \label{eq:H1estimate} \operatorname*{ess\,sup}_{0\leq t \leq T}\|\Psi(t)\|_{H^1}^2 &\leq C_1(\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2),\\ \label{eq:estimate-2} \|\Psi\|_X^2 &\leq C' \Bigl( \| F \|_Y^2 + \| \Psi_0 \|_{L^2}^2 + \bigl( \| F \|_Y^2 + \| \Psi_0 \|_{L^2}^2 \bigr)^2 \Bigr) ,\\ \label{eq:estimate-3} \| \Psi' \|_{X^*}^2 &\leq C'' \bigl(1+ \| F \|_{Y}^2 + \| \Psi_0 \|_{L^2}^2 \bigr)^3.\end{aligned}$$ The same estimates hold for a $\Psi \in W$ solving . $\;$\ \ Testing with $\Psi(\cdot,t)$, we obtain $$\label{eq:KSweak_test2} i {\left( \partial_t \Psi ,\, \Psi \right)_{L^2}} = B(\Psi,\Psi;u) + \alpha {\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} + {\left( F ,\, \Psi \right)_{L^2}} ,$$ a.e. in $(0,T)$. This equation is equivalent to (see e.g. [@Evans2010]) $$\label{eq:KSweak_test3} i \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 = B(\Psi,\Psi;u) + \alpha {\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} + {\left( F ,\, \Psi \right)_{L^2}} .$$ Now, we notice that the left-hand side is purely imaginary, while the terms ${\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}}$ and $B(\Psi,\Psi;u)$ apart from $D(\Psi, \Psi)$ are purely real. Consequently, by splitting into real and imaginary parts, we obtain the following $$\label{eq:KSweak_real} \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 = {\operatorname{Im}}\bigl( {\left( F ,\, \Psi \right)_{L^2}}+(1-\alpha)D(\Psi,\Psi) \bigr) ,$$ and $$\label{eq:KSweak_imag} {\operatorname{Re}}B(\Psi,\Psi;u) + \alpha {\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} + {\operatorname{Re}}\bigl( {\left( F ,\, \Psi \right)_{L^2}} \bigr) = 0 .$$ Now, using Lemma \[lem:bound-weak-form\] and defining $\tilde{c}_0:=(1-\alpha)c_0$, equation becomes as follows $$\label{eq:KSweak_real2} \begin{split} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 &\leq 2 \| F \|_{L^2} \| \Psi \|_{L^2} +2\tilde{c}_0 \|\Psi\|_{L^2}^2\\ &\leq \| F \|_{L^2}^2 + (1+2\tilde{c}_0) \| \Psi \|_{L^2}^2 . \end{split}$$ By defining $\eta(t) := \| \Psi(t) \|_{L^2}^2$ and $\xi(t) := \| F(t) \|_{L^2}^2$ the previous inequality becomes as follows $$\eta'(t) \leq (1+2\tilde{c}_0)\eta(t) + \xi(t) ,$$ a.e. in $(0,T)$. Thus, by applying the Gronwall inequality [@Evans2010] in the differential form, we obtain the following $$\label{eq:afterGronw} \eta(t) \leq e^{(1+2\tilde{c}_0)t} \Bigl( \eta(0) + \int_0^t \xi(s) {\mathrm{d}}s \Bigr) .$$ Notice that by , it holds that $\eta(0) = \| \Psi(0) \|_{L^2}^2 = \| \Psi_0 \|_{L^2}^2$. Consequently, by using , we know that there exists a positive constant $C$ such that the following estimate holds $$\label{eq:estimate-A} \max_{0\leq t \leq T} \| \Psi(t) \|_{L^2}^2 \leq C \Bigl( \| \Psi_0 \|_{L^2}^2 + \| F \|_Y^2 \Bigr) .$$ For $\Psi\in W$, we have the continuous embedding $W \hookrightarrow C([0,T];L^2(\Omega; { \mathbb{C} }^N))$, see, e.g. [@Evans2010 p. 287]. With this, we can evaluate $\Psi$ at time $t$ and find the same estimate if $\Psi\in W$ solves . $ $ \ Taking the real part of , we find $$\begin{aligned} {\operatorname{Re}}B(\Psi, \Psi; u)+\alpha {\left(V(\Psi)\Psi,\,\Psi\right)_{L^2}}+{\operatorname{Re}}{\left(F,\,\Psi\right)_{L^2}}=0. \end{aligned}$$ Using that $V_H\geq 0$ and $V_c\in L^\infty(\Omega)$, we get $$\begin{aligned} {\operatorname{Re}}B(\Psi, \Psi; u)&=\alpha\bigl(-{\left(V_H(\Psi)\Psi,\,\Psi\right)_{L^2}}-{\left(V_x(\Psi)\Psi,\,\Psi\right)_{L^2}}-{\left(V_c(\Psi)\Psi,\,\Psi\right)_{L^2}}\bigr)-{\operatorname{Re}}{\left(F,\,\Psi\right)_{L^2}}\\ &\leq |{\left(V_x(\Psi)\Psi,\,\Psi\right)_{L^2}}|+|{\left(V_c(\Psi)\Psi,\,\Psi\right)_{L^2}}|+|{\operatorname{Re}}{\left(F,\,\Psi\right)_{L^2}}|\\ &\leq |{\left(V_x(\Psi)\Psi,\,\Psi\right)_{L^2}}|+C_{V_c}\|\Psi\|_{L^2}^2+ \|F\|_{L^2}^2+\|\Psi\|_{L^2}^2. \end{aligned}$$ From the assumption \[assumptions\] (\[assumptionVxLipschitz\]) and using , we obtain the following $$\begin{aligned} {\operatorname{Re}}B(\Psi, \Psi; u)\leq C \|\Psi\|_{L^2}^2+\|F\|_{L^2}^2 \leq C_0' (\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2). \end{aligned}$$ By Lemma \[lem:bound-weak-form\], it holds that ${\operatorname{Im}}B(\Psi, \Psi; u)\leq c_0 \|\Psi\|_{L^2}^2$. Combining these two estimates one concludes . As for the first estimate, the same applies in the case when $\Psi\in W$ solves . $ $ \ For the second bound, we simply combine Lemma \[lem:bound-weak-form\] with and . We have $$\begin{aligned} \|\Psi(t)\|_{H^1}^2 &\leq {\operatorname{Re}}B(\Psi(t), \Psi(t); u(t))+c_3\|\Psi(t)\|_{L^2}^2\\ &{\leq}C_0(\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2)+c_3\|\Psi(t)\|_{L^2}^2\\ &{\leq}(C_0+c_3C)(\|\Psi_0\|_{L^2}^2+\|F\|_Y^2). \end{aligned}$$ If $\Psi \in W$, one has to use the fact that given $u_k\rightharpoonup u$ in $L^2(0,T; H^1_0(\Omega))$ with the uniform bound $\operatorname*{ess\,sup}_{0\leq t \leq T} \|u_k(t)\|\leq C$ it follows that $\operatorname*{ess\,sup}_{0\leq t \leq T} \|u(t)\|\leq C$. With this fact, we obtain $$\begin{aligned} \operatorname*{ess\,sup}_{0\leq t \leq T}\|\Psi(t)\|_{H^1}^2 &\leq C_0(\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2)+c_3C(\|\Psi_0\|_{L^2}^2+\|F\|_Y^2). \end{aligned}$$ $ $ \ First, we need an adequate bound for the term ${\left( V(\Psi) \Psi ,\, \Phi \right)_{L^2}}$ for any $\Phi \in L^2(\Omega;{ \mathbb{C} }^N)$. For this reason, we write the following $$\label{eq:bound_pot} \begin{split} {\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} ={\left( V_H(\Psi) \Psi ,\, \Psi \right)_{L^2}}+{\left( V_x(\Psi) \Psi ,\, \Psi \right)_{L^2}}+{\left( V_c(\Psi) \Psi ,\, \Psi \right)_{L^2}}. \end{split}$$ To bound $V_H$, we use the Cauchy-Schwarz inequality, Lemma \[CancesLemma\], and to arrive at $$\begin{aligned} \label{eq:bound_potVH} {\left(V_H(\Psi)\Psi,\,\Psi\right)_{L^2}} &\leq C_u\|\Psi\|_{H^1}^2 \|\Psi\|_{L^2}^2 \leq C_1(\|\Psi_0\|_{L^2}^2+\|F\|_{Y}^2)\|\Psi\|_{L^2}^2.\end{aligned}$$ Next, we recall that $x \mapsto V_c(x,\cdot)$ is bounded (Assumption \[assumptions\] (\[assumptionVcBounded\])) and $V_x$ is Lipschitz continuous (Assumption \[assumptions\] (\[assumptionVxLipschitz\])). Consequently, from , it follows that there exists a positive constant $K'$ such that the following holds $$\label{eq:bound_pot2} {\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} \leq K' \| \Psi \|_{L^2}^2 ,$$ where $K'$ depends on $\|\Psi_0\|_L^2$ and $\|F\|_Y$. By summing term-by-term with , we get the following $$\label{eq:KSweak_new} \begin{split} \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi \|_{L^2}^2 + {\operatorname{Re}}B(\Psi,\Psi;u) = &{\operatorname{Im}}{\left( F ,\, \Psi \right)_{L^2}} +(1-\alpha){\operatorname{Im}}D(\Psi,\Psi) \\ &- \alpha{\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}}- {\operatorname{Re}}{\left( F ,\, \Psi \right)_{L^2}} . \end{split}$$ Adding to both sides the term $c_3 \| \Psi \|_{L^2}^2$, where $c_3$ is the same as in Lemma \[lem:bound-weak-form\], we obtain the following $$\label{eq:KSweak_new2} \begin{split} &\frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 + {\operatorname{Re}}B(\Psi,\Psi;u) + c_3 \| \Psi \|_{L^2}^2\\ &= {\operatorname{Im}}{\left( F ,\, \Psi \right)_{L^2}} +(1-\alpha){\operatorname{Im}}D(\Psi,\Psi) + c_3 \| \Psi \|_{L^2}^2 - \alpha{\left( V(\Psi) \Psi ,\, \Psi \right)_{L^2}} - {\operatorname{Re}}{\left( F ,\, \Psi \right)_{L^2}} . \end{split}$$ Next, by applying Lemma \[lem:bound-weak-form\] and using we get the following $$\label{eq:KSweak_new3} \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 + \| \Psi \|_{H^1}^2 \leq \| F \|_{L^2}^2 + ( 1 + c_0 + c_3 + K' ) \| \Psi \|_{L^2}^2 .$$ By manipulating and integrating over $(0,T)$, we have $$\label{eq:KSweak_new4} \int_0^T \| \Psi \|_{H^1}^2 {\mathrm{d}}t \leq \int_0^T \| F \|_{L^2}^2 + ( 1 + c_0 + c_3 + K' ) \| \Psi \|_{L^2}^2 {\mathrm{d}}t - \int_0^T \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \Psi(t) \|_{L^2}^2 {\mathrm{d}}t ,$$ which implies that $$\label{eq:KSweak_new5} \begin{split} \| \Psi \|_X^2 &\leq \| F \|_Y^2 + ( 1 + c_0 + c_3 + K' ) C T\Bigl(\| \Psi_0 \|_{L^2}^2 + \| F \|_Y^2 \Bigr) + \frac{1}{2} \Bigl( \| \Psi_0 \|_{L^2}^2 - \| \Psi(T) \|_{L^2}^2 \Bigr) \\ &\leq \| F \|_Y^2 + ( 1 + c_0 + c_3 + K' ) C T\Bigl( \| \Psi_0 \|_{L^2}^2 + \| F \|_Y^2 \Bigr) + \frac{1}{2} \| \Psi_0 \|_{L^2}^2 \\ &= \bigl(1+( 1 + c_0 + c_3 + K' ) TC\bigr) \| F \|_Y^2 + \left( ( 1 + c_0 + c_3 + K' ) C T + \frac{1}{2} \right) \| \Psi_0 \|_{L^2}^2 , \end{split}$$ where we used . Using the dependence of $K'$ on the data, the previous implies that there exists a positive constant $C'$ such that $$\label{eq:KSweak_new6} \|\Psi\|_X^2\leq C' \Bigl( \| F \|_Y^2 + \| \Psi_0 \|_{L^2}^2 + \bigl( \| F \|_Y^2 + \| \Psi_0 \|_{L^2}^2 \bigr)^2 \Bigr).$$ The same calculation can be done for $\Psi\in W$ being a solution of . $\;$\ \ Fix any $v \in H^1_0(\Omega;{ \mathbb{C} }^N)$, with $\| v \|_{H^1} \leq 1$. Write $v = v_1 + v_2$, where $v_1 \in {\text{span}}\{\phi_k\}_{k=1}^{m}$ and ${\left(v_2,\,\phi_k\right)_{L^2}}=0$ for $k=1,\dots,m$. Since the functions $\{\phi_k\}_{k=1}^{\infty}$ are orthogonal in $H^1_0(\Omega)$, we have $$\label{eq:norm-leq-1} 1 \ge \| v \|_{H^1}^2 = {\left(v_1+v_2,\,v_1+v_2\right)_{H^1}} = \| v_1 \|_{H^1}^2 + \| v_2 \|_{H^1}^2 \geq \| v_1 \|_{H^1}^2 .$$ Next, utilizing with $\Psi\in W_m$, we obtain $$\label{eq:KSweak_test_boo} i {\left( \partial_t \Psi ,\, v_1 \right)_{L^2}} = B(\Psi,v_1;u) + \alpha{\left( V(\Psi) \Psi ,\, v_1 \right)_{L^2}} + {\left( F ,\, v_1 \right)_{L^2}}$$ a.e. in $[0, T]$. Using the decomposition of $v$, this implies that $$\label{eq:Hm1normsplit} \begin{split} |\langle \Psi' , v \rangle | &= |{\left( \partial_t \Psi ,\, v \right)_{L^2}}| \\ &= |{\left( \partial_t \Psi ,\, v_1 \right)_{L^2}}| \\ &= | B(\Psi,v_1;u) + \alpha{\left( V(\Psi) \Psi ,\, v_1 \right)_{L^2}} + {\left( F ,\, v_1 \right)_{L^2}} | , \end{split}$$ where $\partial_t \Psi \in L^2((0,T);H^1_0(\Omega))$ is the Riesz representative of $\Psi' \in L^2((0,T);H^{-1}(\Omega))$ and ${\langle \cdot ,\, \cdot \rangle} : H^{-1}(\Omega) \times H^{1}_0(\Omega) \rightarrow { \mathbb{C} }$ denotes the dual pairing for $H^1_0(\Omega)$ and its dual $H^{-1}(\Omega)$. By using the Cauchy-Schwarz inequality and Assumptions \[assumptions\] (\[assumptionVcBounded\]) and (\[assumptionVxLipschitz\]) and $\|v_1\|_{H^1}\leq 1$, we have that there exists a positive constant $\tilde{K}$ such that $$\begin{aligned} |{\left( (V_x(\Psi)+V_c(\Psi)) \Psi ,\, v_1 \right)_{L^2}}| \leq \|(V_x(\Psi)+V_c(\Psi)) \Psi\|_{L^2}\|v_1\|_{L^2}\leq \tilde{K}\|\Psi\|_{L^2}.\end{aligned}$$ By recalling Lemma \[lem:bound-weak-form\], and $\| v \|_{H^1} \leq 1$, we obtain that there exists a positive constant $\tilde{C}$ such that $$\label{eq:booo} |\langle \Psi' , v \rangle | \leq \tilde{C}(1+\|\Psi_0\|_{L^2}^2+\|F\|_Y^2) \bigl( \| F \|_{L^2} + \| \Psi \|_{H^1} \bigr) ,$$ and from , we have the following $$\label{eq:booo2} \| \Psi' \|_{H^{-1}}=\sup_{0\neq v\in H_0^1(\Omega)}\frac{|\langle \Psi', v\rangle|}{\|v\|_{H^1}} \leq \tilde{C}(1+\|\Psi_0\|_{L^2}^2+\|F\|_Y^2) \bigl( \| F \|_{L^2} + \| \Psi \|_{H^1} \bigr) .$$ This implies that $$\begin{aligned} \|\Psi'\|_{H^{-1}}^2 \leq \tilde{C}^2(1+\|\Psi_0\|_{L^2}^2+\|F\|_Y^2)^2 \bigl( \| F \|_{L^2} + \| \Psi \|_{H^1} \bigr)^2\\ \leq 2\tilde{C}^2(1+\|\Psi_0\|_{L^2}^2+\|F\|_Y^2)^2\bigl( \| F \|_{L^2}^2 + \| \Psi \|_{H^1}^2 \bigr).\end{aligned}$$ By integrating over $(0,T)$ and using , we obtain that there exists a positive constant $C''$ such that the following estimate holds $$\| \Psi' \|_{X^*}^2 \leq C'' \bigl(1+ \| F \|_{Y}^2 + \| \Psi_0 \|_{L^2}^2 \bigr)^3 ,$$ where $X^* = L^2((0,T);H^{-1}(\Omega))$ and the proof for $\Psi\in W_m$ is completed. For $\Psi\in W$, no decomposition is necessary in , so we can use $v_1=v, v_2=0$ and apply the same estimates to conclude our proof. Existence of a weak solution {#sec:ExistenceSolution} ============================ In the preceding section, we have shown the estimates in Theorem \[thm:estimates\] for solutions $\Psi_m\in W_m$ in the Galerkin subspace. In this section, we use these estimates to show the existence of a solution in the full Sobolev space $W$. To this end, we make use of the following embedding theorem by Lions [@Lions1969 1.5.2]. \[lem:Lions\] Given three Banach spaces $B_0 \Subset B \hookrightarrow B_1$ with $B_0$, $B_1$ reflexive and the embedding $B_0 \hookrightarrow B$ being compact, then the space $$\begin{aligned} V&=\left\{v| v\in L^p((0,T), B_0), v'\in L^q((0,T), B_1) \right\}, \quad 1<p,q<\infty,\\ \|v\|_V&:=\|v\|_{L^p((0,T), B_0)} +\|v'\|_{L^q((0,T), B_1)} \end{aligned}$$ is compactly embedded in $L^p(0,T; B)$. Problem admits a weak solution, i.e. there exists a $\Psi\in W$ such that $$\begin{split} &i {\left( \partial_t \Psi,\, \Phi \right)_{L^2}} = B(\Psi,\Phi;u) + \alpha{\left( V(\rho) \Psi,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} \\ &\text{a.e. in } (0,T), \; \forall \Phi \in H^1_0(\Omega;{ \mathbb{C} }^N),\\ &\Psi_0\in L^2(\Omega;{ \mathbb{C} }^N) . \end{split}$$ Consider a sequence $\{\Psi_m\}_{m=1}^{\infty}$ of solutions of the Galerkin problem , then according to the estimates , , and in Theorem \[thm:estimates\], the sequence is bounded in $X$ and $\{\Psi_m'\}_{m=1}^{\infty}$ is bounded in $X^*$. Consequently, there exists a subsequence $\{ \Psi_{m_l} \}_{l=1}^{\infty}$ and a function $\Psi \in X$ with $\Psi' \in X^*$ such that $\Psi_{m_l} \rightharpoonup \Psi$ in $X$ and $\Psi_{m_l}' \rightharpoonup \Psi'$ in $X^*$; see, e.g., [@Evans2010]. Moreover, by Lions’ theorem (Lemma \[lem:Lions\]) we know that $W$ is compactly embedded in $Y:=L^2(0,T;L^2(\Omega))$, consequently, we have strong convergence of the subsequence $\Psi_{m_l} \rightarrow \Psi$ in $Y$. Next, we fix a positive integer $M$ and construct a test function$\Phi \in C^1([0,T];H^1_0(\Omega;{ \mathbb{C} }^N))$ as follows $$\label{eq:form2} \Phi(x,t) := \sum_{k=1}^M d^k_m(t) \Phi_k(x) ,$$ where $\{ d^k_m \}_{k=1}^{M}$ are given smooth functions. We choose $m \geq M$, multiply by $d^k_m(t)$, sum over $k=1,\dots,M$, and integrate with respect to $t$ to obtain the following $$\label{eq:KSweak_exist} \int_0^T i {\langle \Psi_m' ,\, \Phi \rangle} {\mathrm{d}}t = \int_0^T B(\Psi_m,\Phi;u) + \alpha{\left( V(\Psi_m) \Psi_m ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} {\mathrm{d}}t.$$ By setting now $m=m_l$ and by recalling continuity of $V(\Psi)$ from Lemma \[lem:VContinuousFunction\] and strong convergence $\Psi_{m_l} \rightarrow \Psi$ in $Y$, we can pass to the limit to obtain $$\label{eq:KSweak_exist2} \int_0^T i {\langle \Psi' ,\, \Phi \rangle} {\mathrm{d}}t = \int_0^T B(\Psi,\Phi;u) + \alpha{\left( V(\Psi) \Psi ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} {\mathrm{d}}t .$$ This equality holds for all $\Phi \in X$ as functions of the form are dense in $X$. Hence, in particular $$\label{eq:KSweak_exist3} i {\langle \Psi' ,\, v \rangle} = B(\Psi,\Phi;u) + \alpha{\left( V(\Psi) \Psi ,\, v \right)_{L^2}} + {\left( F ,\, v \right)_{L^2}} ,$$ for any $v \in H^1_0(\Omega;{ \mathbb{C} }^N)$ and a.e. in $[0, T]$. From [@Evans2010 Theorem 3 p. 287], we know also that $\Psi \in C([0,T];L^2(\Omega;{ \mathbb{C} }^N))$. It remains to prove that $\Psi(\cdot,0) = \Psi_0$. For this purpose, we first notice from that the following holds $$\label{eq:KSweak_exist4} \int_0^T - i {\langle \Phi' ,\, \Psi \rangle} {\mathrm{d}}t = \int_0^T B(\Psi,\Phi;u) + \alpha{\left( V(\Psi) \Psi ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} {\mathrm{d}}t + {\left( \Psi(0) ,\, \Phi(0) \right)_{L^2}} ,$$ for any $\Phi \in C^1([0,T];H^1_0(\Omega;{ \mathbb{C} }^N))$ with $\Phi(T)=0$. Similarly, from we get $$\label{eq:KSweak_exist5}\begin{split} \int_0^T - i {\langle \Phi' ,\, \Psi_m \rangle} {\mathrm{d}}t &= \int_0^T B(\Psi_m,\Phi;u) + \alpha{\left( V(\Psi_m) \Psi_m ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} {\mathrm{d}}t\\ &\quad + {\left( \Psi_m(0) ,\, \Phi(0) \right)_{L^2}} . \end{split}$$ We set $m=m_l$ and use again the considered convergences to find $$\label{eq:KSweak_exist6} \int_0^T - i {\langle \Phi' ,\, \Psi \rangle} {\mathrm{d}}t = \int_0^T B(\Psi,\Phi;u) + \alpha{\left( V(\Psi) \Psi ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} {\mathrm{d}}t + {\left( \Psi_0 ,\, \Phi(0) \right)_{L^2}} ,$$ because $\Psi_{m_l}(0) \rightarrow \Psi_0$. As $\Phi(0)$ is arbitrary, by comparing and we conclude that $\Psi(0) = \Psi_0$. Uniqueness of a weak solution ============================= We have shown that there exists at least one solution $\Psi \in W$ of . Now, we can apply the extension of Theorem \[thm:estimates\] to the space $W$ and use the Lipschitz properties of the potentials to show that the solution is indeed unique. \[sec:Uniqueness\] The weak form of the Kohn-Sham equations is uniquely solvable. Seeking a contradiction, we assume that there exists two distinct weak solutions of $\Psi$ and $\Upsilon$ in $W$ with $\| \Psi - \Upsilon \|_X > 0$. Therefore, we have $$\label{eq:KSweakUniq1} \begin{split} &i {\left( \partial_t \Psi ,\, \Phi \right)_{L^2}} = B(\Psi,\Phi;u) + \alpha{\left( V(\Psi) \Psi,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} , \end{split}$$ and $$\label{eq:KSweakUniq2} \begin{split} &i {\left( \partial_t \Upsilon ,\, \Phi \right)_{L^2}} = B(\Upsilon ,\Phi;u) + \alpha{\left( V(\Upsilon ) \Upsilon ,\, \Phi \right)_{L^2}} + {\left( F ,\, \Phi \right)_{L^2}} , \end{split}$$ for all test functions $\Phi \in H^1_0(\Omega; { \mathbb{C} }^N)$. Subtracting term-by-term from and defining $\hat{\Psi} := \Psi - \Upsilon$ we obtain the following $$\label{eq:KSweakUniq3} \begin{split} i {\left( \partial_t \hat{\Psi} ,\, \Phi \right)_{L^2}} &= B(\hat{\Psi} ,\Phi;u) + \alpha{\left( V(\Psi ) \Psi - V(\Upsilon ) \Upsilon ,\, \Phi \right)_{L^2}} . \end{split}$$ By testing the previous with $\Phi = \hat{\Psi}(t)$, we obtain $$\label{eq:KSweakUniq4} i {\left( \partial_t \hat{\Psi} ,\, \hat{\Psi} \right)_{L^2}} = B(\hat{\Psi},\hat{\Psi};u) + \alpha{\left( V(\Psi ) \Psi - V(\Upsilon ) \Upsilon ,\, \hat{\Psi} \right)_{L^2}} .$$ Similarly, as for we have $$\label{eq:KSweak_test4} i \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \hat{\Psi} \|_{L^2}^2 = B(\hat{\Psi},\hat{\Psi};u) + \alpha{\left( V(\Psi ) \Psi - V(\Upsilon ) \Upsilon ,\, \hat{\Psi} \right)_{L^2}} .$$ Now, we notice that the left-hand side is purely imaginary. Consequently, by taking the imaginary part of , we obtain the following $$\label{eq:KSweakU-imag} \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \hat{\Psi} \|_{L^2}^2 =\alpha{\operatorname{Im}}\left( {\left( V(\Psi ) \Psi - V(\Upsilon ) \Upsilon ,\, \hat{\Psi} \right)_{L^2}} +(1-\alpha)D(\hat{\Psi},\hat{\Psi})\right).$$ From and in Lemma \[lem:bound-weak-form\], we get $$\begin{split} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \hat{\Psi} \|_{L^2}^2 &= 2\alpha{\operatorname{Im}}\bigl( {\left( V(\Psi ) \Psi - V(\Upsilon ) \Upsilon ,\, \hat{\Psi} \right)_{L^2}} \bigr)+2(1-\alpha){\operatorname{Im}}D(\hat{\Psi},\hat{\Psi}) \\ &\leq \| V(\Psi) \Psi - V(\Upsilon ) \Upsilon \|_{L^2} \| \hat{\Psi} \|_{L^2} + 2c_0 \| \hat{\Psi} \|_{L^2}^2.\\ \intertext{Using Lemma \ref{CancesLemma}, Assumptions \ref{assumptions} (\ref{assumptionVcBounded}), (\ref{assumptionVxLipschitz}), \eqref{eq:estimate-1}, and Theorem \ref{thm:estimates}, we obtain} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \| \hat{\Psi} \|_{L^2}^2&\leq C_u'(\|\Psi\|_{H^1}^2+\|\Upsilon\|_{H^1}^2) \|\hat{\Psi}\|_{L^2}^2+2c_0\| \hat{\Psi} \|_{L^2}^2\\ &\leq c^\#\bigl(L+K+\| F \|_{L^2}^2 + \| \Psi_0 \|_{L^2}^2+c_0 \bigr) \| \hat{\Psi} \|_{L^2}^2. \end{split}$$ By defining $\eta(t) := \| \hat{\Psi} \|_{L^2} ^2$ and $\vartheta(t) := c^\#\bigl(L+K+\| F \|_{L^2}^2 + \| \Psi_0 \|_{L^2}^2+c_0 \bigr)$, we obtain the following inequality $$\eta'(t) \leq \vartheta(t) \eta(t) .$$ By applying the Gronwall’s inequality, we obtain the following $$\eta(t) \leq \exp \Bigl( \int_0^t \vartheta(s) ds \Bigr) \eta(0) .$$ By noticing that $$\begin{split} \int_0^t \vartheta(s) ds &= \int_0^t c^\#\bigl(L+K+\| F \|_{L^2}^2 + \| \Psi_0 \|_{L^2}^2+c_0 \bigr) {\mathrm{d}}s \\ &\leq \int_0^T c^\#\bigl(L+K+\| F \|_{L^2}^2 + \| \Psi_0 \|_{L^2}^2+c_0 \bigr) {\mathrm{d}}s \\ &\leq c^\#\bigl( \| F \|_{Y}^2 + T \| \Psi_0 \|_{L^2}^2 +T (c_0+L+K) \bigr) , \end{split}$$ and by recalling that $\eta(0) = \| \hat{\Psi}(0) \|_{L^2} = 0$, we obtain that $\| \hat{\Psi} \|_{L^2} \leq 0$ a.e. in $(0,T)$, and the claim follows. Improved regularity {#sec:ImprovedRegularity} =================== We have established the existence and uniqueness of a solution to in $W$. Although our methodology and assumptions on $V$ differ from [@Jerome2015], our result is similar to [@Jerome2015]. Now, we improve these results in the case $\Psi_0\in H_0^1(\Omega)$. With this setting, we prove that the solution to is twice weakly differentiable in space and its first spatial derivative is bounded. \[lem;diffquot\] Assume that for fixed $u\in L^p(V)$, $1<p<\infty$, $V\subset\subset \Omega$, there exists a constant $C$ such that $\|D^h u\|_{L^p(V)} \leq C$ for all $0<|h|<\frac{1}{2} \operatorname{dist}(V, {\partial}\Omega)$ where $$\begin{aligned} D_i^hu(x)=\frac{u(x+e_i h)-u(x)}{h},\quad D^hu=(D_1^hu,\dotsc , D_n^hu).\end{aligned}$$ Then $$\begin{aligned} u\in H^{1,p}(V), \text{ with } \|Du\|_{L^p(V)} \leq C,\end{aligned}$$ where $C$ may depend on $u$, e.g. on $\|u\|_{L^p}(\Omega)$. Furthermore, the statement holds for the case of two half-balls $\Omega=\{|x|<R\}\cap \{x_n> 0\}$ and $V=\{|x|<\frac{R}{2}\}\cap \{x_n> 0\}$. See [@Evans2010 §5.8.2, Theorem 3] and the remark after the proof. Next, we extend the result in [@Evans2010 §6.3.2, Theorem 4] for linear elliptic problems to the case of a specific nonlinear problem. \[lem:EllipticRegularity\] Let $\varphi\in H_0^1(\Omega; { \mathbb{C} }^N)$ be a weak solution of the elliptic boundary value problem $$\begin{aligned} B(\varphi,v;u)+{\left(V(\varphi)\varphi,\,v\right)_{L^2}}={\left(A,\,v\right)_{L^2}}, \forall v\in H_0^1(\Omega; { \mathbb{C} }^N), \; A\in L^2(\Omega; { \mathbb{C} }^N), \end{aligned}$$ such that $\|\varphi\|_{L^2}^2\leq \gamma \|A\|_{L^2}^2$ holds. Furthermore be $\partial \Omega\in C^2$. Then $\varphi\in H^2(\Omega; { \mathbb{C} }^N)$ and $$\begin{aligned} \|\varphi\|_{H^2} &\leq c\left(\|A\|_{L^2}+\|\varphi\|_{L^2} \right),\\ \text{where }c&=\max\left\{1,\|V_0\|_{L^\infty}+\|u\|_{C(0,T)}\|V_u\|_{L^\infty}+\|\varphi\|_{H^1}^2+L+K\right\}. \end{aligned}$$ To extend the results in [@Evans2010 §6.3.2, Theorem 4], two issues have to be treated carefully. First, the nonlinear potential has to be bounded in a suitable way and, second, extra care has to be taken when changing the coordinates. The nonlinear potential has to be bounded in such a way that Lemma \[lem;diffquot\] can be applied. Therefore, we need to find a constant $c$ such that $\|V(\varphi)\varphi\|_{L^2}^2\leq c\|\varphi\|_{L^2}^2$, where $c$ is allowed to depend on $\varphi$. This can be done using Lemma \[CancesLemma\] as follows $$\begin{aligned} \|V_H(\varphi)\varphi\|_{L^2} \leq C_u\|\varphi\|_{H^1}^2 \|\varphi\|_{L^2}, \end{aligned}$$ and using Assumptions \[assumptions\] (\[assumptionVxLipschitz\]) and (\[assumptionVcBounded\]), we obtain $$\begin{aligned} \|V(\varphi)\varphi\|_{L^2}^2 \leq (\|\varphi\|_{H^1}^4+L^2+K^2)\|\varphi\|_{L^2}^2. \end{aligned}$$ Now, we can apply Lemma \[lem;diffquot\] to obtain that the solution is in $H^2(U)$ for a half-ball $U$. Furthermore, in the proof it is necessary to locally flatten out the boundary. This is done by a $C^2$-map that keeps all the coordinates apart from one dimension which is transformed onto a line. This ensures that the determinant of the Jacobian is equal to one. The coordinate transformation of the Laplacian and the linear external potential is as for standard parabolic PDEs. The exchange and correlation potentials do not explicitly depend on space and time but only pointwise on the wave function. Hence a change of coordinates does not change the potential. For the Hartree potential, however, more care is needed. Let the change of coordinates be given by $$\begin{aligned} x&=k(\hat{x}), && \hat{\Psi}(\hat{x})=\Psi(k(\hat{x})). \end{aligned}$$ Regarding the Hartree potential, one has to account for the fact that the transformation $k$ is only locally defined as a $C^2$ map, so the transform to a global integral operator is not well-defined. However, it is possible to evaluate $V_H(\hat{\Psi})(\hat{x})$ as $V_H(\Psi)(x)$ in $x=k(\hat{x})$. With this preparation, let $U$ be the image of a half-ball under $k$. Then we bound $\|\Psi\|_{H^1(U)}\leq C(\|F\|_{L^2(\Omega)}+\|\Psi\|_{L^2(\Omega)})$. As $\Omega$ is compact, it can be covered with finitely many sets $U_i$, so we find $$\begin{aligned} \|\Psi\|_{H^1(\Omega)}\leq \sum_i \|\Psi\|_{H^1(U_i)} \leq \sum_i C\left(\|F\|_{L^2(\Omega)}+\|\Psi\|_{L^2(\Omega)}\right). \end{aligned}$$ Now the standard proof for elliptic equations based on difference quotients can be applied, e.g., [@Evans2010 §6.3.2, Theorem 4]. \[thm:ImprovedRegularity\] Assume $\Psi_0\in H_0^1(\Omega)$, $F\in Y$ and $\partial \Omega \in C^2$. Suppose $\Psi \in W$ is the solution to . Then $$\begin{aligned} \Psi \in L^2(0,T;H^2(\Omega; { \mathbb{C} }^N)) \cap L^\infty(0,T; H_0^1(\Omega; { \mathbb{C} }^N)), \quad \Psi'\in L^2(0,T;L^2(\Omega; { \mathbb{C} }^N)). \end{aligned}$$ Furthermore the following estimate holds $$\begin{aligned} \label{thm:ImprovedRegularity:Estimate} \operatorname*{ess\, sup}_{0\leq t \leq T} \|\Psi(t)\|_{H^1}+\|\Psi \|_{L^2(0,T;H^2(\Omega))}+\|\Psi'\|_Y \leq C\left( \|\Psi_0\|_{H^1}+ \|F\|_Y \right). \end{aligned}$$ We recall , that is, $$\begin{aligned} \label{eq:Improved:Psim} \operatorname*{ess\, sup}_{0\leq t \leq T} \|\Psi(t)\|_{H^1(\Omega)}^2\leq C(\|\Psi_0\|_{L^2(\Omega)}^2+\|F\|_Y^2), \end{aligned}$$ which means that $\Psi\in L^\infty(0,T;H_0^1(\Omega)$). For ${\partial}_t\Psi$, we consider the Galerkin space $W_m$ and take a fixed $m$, multiply with ${\partial}_td_m^k(t)$, and sum for $k=1,\dotsc,m$ to obtain the following $$\label{eq:improved:test}\begin{split} {\left({\partial}_t\Psi_m(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}}&=B(\Psi_m(t), {\partial}_t\Psi_m(t);u(t))+\alpha{\left(V(\Psi(t))\Psi(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}}\\ &\quad+{\left(F(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}} \end{split}$$ a.e. in $(0,T)$. For $D(\Psi_m(t),{\partial}_t\Psi_m(t))$, we have $$\begin{aligned} |D_{xc}(\Psi_m(t),{\partial}_t \Psi_m(t))|&=\left|2{\left({\frac{{\partial}V_{xc}}{{\partial}\rho}}(\Lambda(t)) {\operatorname{Re}}{\left(\Psi_m(t),\,\Lambda(t)\right)_{{ \mathbb{C} }}}\Lambda(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}}\right|. \end{aligned}$$ Because $\Lambda(t)\in L^\infty(\Omega)$, we have $$\label{eq:Imp:Dxc}\begin{split} |D_{xc}(\Psi_m(t),{\partial}_t \Psi_m(t))|&\leq C \int_\Omega \left|\sum_{i=1}^N {\operatorname{Re}}\psi_{i,m}(t) \sum_{j=1}^N {\overline{{\partial}_t \psi_{j,m}(t)}} \right| {\mathrm{d}}x\\ &\leq C \int_\Omega \sum_{i,j=1}^N \frac{1}{\epsilon}|\psi_{i,m}(t)|^2+\epsilon |{\partial}_t\psi_{j,m}(t)|^2\\ &\leq C N (\frac{1}{\epsilon}\|\Psi_m(t)\|_{L^2}^2+\epsilon\|{\partial}_t \Psi_m(t)\|_{L^2}^2), \end{split}$$ where we use Young’s inequality for products. For $D_H$, we use Young’s inequality for convolutions [@Schilling2005 Theorem 14.6] and the fact that $\Lambda \in L^\infty(\Omega)$. We have $$\label{eq:Imp:DH}\begin{split} |D_H(\Psi_m(t),{\partial}_t \Psi_m(t))|&=\left|\sum_{j=1}^N\int_\Omega (2{\operatorname{Re}}{\left(\Psi_m(t),\,\Lambda(t)\right)_{{ \mathbb{C} }}}\star w)(x)\Lambda_j(x,t) {\overline{{\partial}_t\Psi_{m,j}(x,t)}} {\mathrm{d}}x \right|\\ &\leq C'{\left(|\Psi_m(t)|\star w,\,|{\partial}_t\Psi_m(t)|\right)_{L^2}}\\ &\leq C' \| |\Psi_m(t)|\star w\|_{L^2} \|{\partial}_t\Psi_m(t)\|_{L^2}\\ &\leq C' \| \Psi_m(t)\|_{L^2}\| w\|_{L^1} \|{\partial}_t\Psi_m(t)\|_{L^2}, \end{split}$$ where $w$ represents the Coulomb potential. Consequently, by and , we get the following $$\label{eq:Imp:D} |D(\Psi_m(t),{\partial}_t \Psi_m(t))|\leq c (\frac{1}{\epsilon}\|\Psi_m(t)\|_{L^2}^2+\epsilon\|{\partial}_t \Psi_m\|_{L^2}^2) =: \tilde{D}.$$ Estimate is used together with to obtain the following $$\begin{aligned} &\|{\partial}_t\Psi_m(t)\|_{L^2}^2 \\&\leq \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}}{\left(\nabla\Psi_m(t),\,\nabla\Psi_m(t)\right)_{L^2}}+{\left(V_{ext}(t) \Psi_m(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}}+D(\Psi_m(t),{\partial}_t\Psi_m(t))\\ &\quad+{\left(V(\Psi_m(t))\Psi_m(t),\,{\partial}_t\Psi_m\right)_{L^2}}+{\left(F(t),\,{\partial}_t\Psi_m(t)\right)_{L^2}}\\ &\leq \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}}{\left(\nabla\Psi_m(t),\,\nabla\Psi_m(t)\right)_{L^2}}+\|V_{ext}(t)\|_{L^\infty} \|\Psi_m(t)\|_{L^2}\|{\partial}_t\Psi_m(t)\|_{L^2} + \tilde{D}\\ &\quad +c_u\|\Psi_m(t)\|_{L^2}\|\Psi_m(t)\|_{H^1}^2\|{\partial}_t\Psi_m(t)\|_{L^2} +K \|\Psi_m(t)\|_{L^2}\|{\partial}_t\Psi_m(t)\|_{L^2}\\ &\quad +\|F(t)\|_{L^2}\|{\partial}_t\Psi_m(t)\|_{L^2}, \end{aligned}$$ where we use Lemma \[CancesLemma\] and Assumptions \[assumptions\] (\[assumptionVcBounded\]) and (\[assumptionVxLipschitz\]) to estimate $V$. Next, by using Cauchy-Schwarz inequality, and Young’s inequality with an arbitrary positive $\epsilon$, we get $$\begin{aligned} &\|{\partial}_t\Psi_m(t)\|_{L^2}^2\\ &\leq \frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \|\Psi_m(t)\|_{H^1}^2+ \frac{1}{\epsilon}\|F(t)\|_{L^2}^2+\epsilon\|{\partial}_t\Psi_m(t)\|_{L^2}^2 + \tilde{D}\\ &\quad + (\|V_{ext}(t)\|_{L^\infty}+K+C_1(\|\Psi_0\|_{L^2}^2+\|F\|_Y^2))\left(\frac{1}{\epsilon}\|\Psi_m(t)\|_{L^2}^2+\epsilon\|{\partial}_t\Psi_m(t)\|_{L^2}^2 \right)\\ &\leq\frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}}\|\Psi_m(t)\|_{H^1}^2+\frac{\Gamma}{\epsilon}\left( \|\Psi_m(t)\|_{L^2}^2+\|F(t)\|_{L^2}^2\right) +\epsilon \Gamma \|{\partial}_t\Psi_m(t)\|_{L^2}^2, \end{aligned}$$ where $\Gamma$ is a constant depending only on $\|\Psi_0\|_{L^2}$, $\|F\|_Y$, $\max_{t\in [0,T]}\|V_{ext}(t)\|_{L^\infty}$ and $K$. Now, we choose $\epsilon$ small enough, that is $\epsilon < \frac{1}{\Gamma}$ and integrate from $0$ to $T$. We obtain $$\begin{aligned} \int_0^T \|{\partial}_t\Psi_m(t)\|_{L^2}^2{\mathrm{d}}t &\leq \frac{1}{1-\epsilon \Gamma}\left( \operatorname*{ess\,sup}_{0\leq t\leq T} \|\Psi_m(t)\|_{H^1}^2+\frac{\Gamma}{\epsilon}\int_0^T \|\Psi_m(t)\|_{L^2}^2+\|F(t)\|_{L^2}^2 {\mathrm{d}}t \right). \end{aligned}$$ Using and , this gives $$\begin{aligned} \label{eq:Improved:Psimprime} \|{\partial}_t\Psi_m\|_Y^2\leq \Gamma'(\|\Psi_0\|_{L^2}^2+\|F\|_Y^2). \end{aligned}$$ Passing to the limit as $m\rightarrow \infty$ we find $\Psi'\in Y$. Now, we rewrite for a fixed time $t$ as follows $$\begin{aligned} B(\Psi(t), \Phi; u(t))+\alpha{\left(V(\Psi(t))\Psi(t),\,\Phi\right)_{L^2}}={\left(-F(t)+i{\partial}_t \Psi(t),\,\Phi\right)_{L^2}}, \end{aligned}$$ where $\Psi$ is the solution to . Using Theorem \[thm:estimates\] we have that the solution is bounded and, therefore, the estimate in Lemma \[lem:EllipticRegularity\] holds. We have $$\begin{aligned} \label{eq:ellipticH2bound} \|\Psi(t)\|_{H^2} &\leq c (\|A(t)\|_{L^2}+\|\Psi(t)\|_{L^2})\\ &\leq c(\|F(t)\|_{L^2}+\|\Psi'(t)\|_{L^2}+\|\Psi(t)\|_{L^2}),\label{eq:ellipticH2bound2} \end{aligned}$$ where $A(t)=-F(t)+i{\partial}_t \Psi(t)$. Next, we integrate from $0$ to $T$, and use and to obtain the following $$\begin{aligned} \|\Psi\|_{L^2(0,T; H^2(\Omega))}^2 &\leq C(\|\Psi_0\|_{L^2}^2+\|F\|_Y^2). \end{aligned}$$ All together, we have shown the estimate. \[thm:ImprovedRegularity2\] If in addition to the assumptions of Theorem \[thm:ImprovedRegularity\], $\Psi_0\in H^2(\Omega)\cap H_0^1(\Omega)$, and $F\in H^1(0,T;L^2(\Omega))$ hold, then for the solution of , we have $$\begin{aligned} \Psi'\in L^\infty(0,T;L^2(\Omega)) && \text{and} && \Psi\in L^\infty(0,T;H^2(\Omega)). \end{aligned}$$ Take a fixed $m\geq 1$. Differentiate with respect to $t$, multiply this equation with ${\partial}_t d_m^k(t)$, sum over $k$, and integrate over $t$ to obtain $$\label{eq:imp2:weakint} \begin{split} \int_0^T i{\left({\partial}_t^2 \Psi_m,\,{\partial}_t \Psi_m\right)_{L^2}} {\mathrm{d}}t&= \int_0^T B({\partial}_t\Psi_m, {\partial}_t \Psi_m; u)+{\left(V_u {\frac{{\partial}}{{\partial}t}} \left( u\Psi_m\right),\,{\partial}_t\Psi_m\right)_{L^2}}\\+{\left({\partial}_t F,\,{\partial}_t\Psi_m\right)_{L^2}} &+\alpha{\left(V(\Psi_m){\partial}_t \Psi_m+{\frac{{\partial}V(\Psi_m)}{{\partial}t}}\Psi_m,\,{\partial}_t\Psi_m\right)_{L^2}} {\mathrm{d}}t. \end{split}$$ For the left-hand side, we have $$\begin{aligned} \label{eq:imp2:lhsnorm} i{\left({\partial}_t^2 \Psi_m,\,{\partial}_t \Psi_m\right)_{L^2}}=i\frac{1}{2}{\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \|{\partial}_t\Psi_m\|_{L^2}^2. \end{aligned}$$ We remark that for any $f(\Psi, x, t)\in { \mathbb{R} }$, we have $$\begin{aligned} {\left(f(\Psi, \cdot, t) \Psi(t),\,{\partial}_t \Psi(t)\right)_{L^2}}={\left(f(\Psi, \cdot, t),\,{\left(\Psi,\,{\partial}_t\Psi\right)_{{ \mathbb{C} }}}\right)_{L^2}}\\ ={\left(f(\Psi, \cdot, t),\,\frac{1}{2} {\frac{{\mathrm{d}}}{{\mathrm{d}}t}} \|\Psi(t)\|_{ \mathbb{C} }^2\right)_{L^2}} \in { \mathbb{R} }. \end{aligned}$$ Hence, using this result for the product terms in , we get $$\begin{aligned} \label{eq:imp2:realproducts} {\left(V_u {\frac{{\partial}}{{\partial}t}}\left(u\Psi_m(t)\right),\,{\partial}_t\Psi_m\right)_{L^2}}\in { \mathbb{R} }&& \text{and} && {\left( {\frac{{\partial}}{{\partial}t}}\left( V(\Psi_m)\Psi_m \right),\,{\partial}_t\Psi_m\right)_{L^2}} \in { \mathbb{R} }. \end{aligned}$$ Taking the imaginary part of and using and gives the following $$\begin{aligned} \frac{1}{2}( \|{\partial}_t\Psi_m(t)\|_{L^2}^2-\|{\partial}_t\Psi_m(0)\|_{L^2}^2)&=\int_0^t (1-\alpha){\operatorname{Im}}D({\partial}_t\Psi_m, {\partial}_t\Psi_m)+{\operatorname{Im}}{\left({\partial}_t F,\,{\partial}_t\Psi_m\right)_{L^2}}. \end{aligned}$$ From this, using , we obtain the following $$\begin{aligned} \sup_{0\leq t\leq T}\|{\partial}_t \Psi_m(t)\|_{L^2}^2&\leq \|{\partial}_t\Psi_m(0)\|_{L^2}^2+2\int_0^T(1-\alpha)|{\operatorname{Im}}D({\partial}_t\Psi_m, {\partial}_t\Psi_m)|\\ &\quad+|{\operatorname{Im}}{\left({\partial}_t F,\,{\partial}_t\Psi_m\right)_{L^2}}| {\mathrm{d}}t\\ &\leq \|{\partial}_t\Psi_m(0)\|_{L^2}^2+2\int_0^T(1-\alpha)c_0 \|{\partial}_t\Psi_m\|_{L^2}^2+\|F\|_{L^2}^2+\|{\partial}_t \Psi_m\|_{L^2}^2 {\mathrm{d}}t\\ &\leq\|{\partial}_t\Psi_m(0)\|_{L^2}^2+2(c_0+1)\|{\partial}_t\Psi_m\|_Y^2 +2\|F\|_Y^2. \end{aligned}$$ By , $\|{\partial}_t\Psi_m\|_Y$ is bounded by $F$ and $\Psi_0$. Hence, there exists a constant $c_6$ depending only on $T$, $\|\Psi_0\|_{L^2}$, $\|F\|_Y$ and $\|u\|_{H^1(0,T)}$, such that the following holds $$\begin{aligned} \label{eq:Imp2:supwithpsim0} \sup_{0\leq t\leq T}\|{\partial}_t \Psi_m(t)\|_{L^2}^2&\leq \|{\partial}_t\Psi_m(0)\|_{L^2}^2+ c_6. \end{aligned}$$ To bound $\|{\partial}_t\Psi_m(0)\|_{L^2}^2$, we test with ${\partial}_t\Psi_m(0)$ to obtain $$\begin{aligned} i{\left({\partial}_t\Psi_m(0),\,{\partial}_t\Psi_m(0)\right)_{L^2}}&=B(\Psi_m(0),{\partial}_t \Psi_m(0);u)+{\left(V(\Psi_m(0)) \Psi_m(0),\,{\partial}_t \Psi_m(0)\right)_{L^2}}\\ &\quad +{\left(F(0),\,{\partial}_t \Psi_m(0)\right)_{L^2}}, \end{aligned}$$ $$\label{eq:improvedinfty:dtpsizero} \begin{split} \|{\partial}_t\Psi_m(0)\|_{L^2}^2 &\leq |B(\Psi_m(0),{\partial}_t \Psi_m(0);u)|+|{\left(V(\Psi_m(0)) \Psi_m(0),\,{\partial}_t \Psi_m(0)\right)_{L^2}}|\\ &\quad +\|F(0)\|_{L^2}\|{\partial}_t\Psi_m(0)\|_{L^2}\\ &\leq c_1' \|\Psi_m(0)\|_{H^2}\|{\partial}_t\Psi_m(0)\|_{L^2}+K'\|\Psi_m(0)\|_{L^2}\|{\partial}_t \Psi_m(0)\|_{L^2}\\ &\quad +\|F(0)\|_{L^2}\|{\partial}_t\Psi_m(0)\|_{L^2}. \end{split}$$ Here, we used for the nonlinear potential and we use the modified proof of Lemma \[lem:bound-weak-form\] by replacing ${\left(\nabla \Psi,\,\nabla \Phi\right)_{L^2}}$ by ${\left(\nabla^2 \Psi,\,\Phi\right)_{L^2}}$ using integration by parts. Dividing by $\|{\partial}_t\Psi_m(0)\|_{L^2}$ gives $$\begin{aligned} \|{\partial}_t\Psi_m(0)\|_{L^2} &\leq c_1' \|\Psi_m(0)\|_{H^2}+K'\|\Psi_m(0)\|_{L^2}+\|F(0)\|_{L^2}\\ &\leq (c_1'+K) \|\Psi_m(0)\|_{H^2}+\|F(0)\|_{L^2}. \end{aligned}$$ Furthermore, we have $\|\Psi_m(0)\|_{H^2}\leq C\|\Psi_0\|_{H^2}$; see, e.g., [@Evans2010 p. 363]. Using this in gives the following $$\begin{aligned} \label{eq:Imp2:bla} \|{\partial}_t\Psi_m(0)\|_{L^2} &\leq (c_1'+1) C\|\Psi_0\|_{H^2}+\|F(0)\|_{L^2}. \end{aligned}$$ Therefore, using in , we obtain the following $$\begin{aligned} \sup_{0\leq t \leq T} \|{\partial}_t \Psi_m(t)\|_{L^2}^2 \leq c_7 \left( \|\Psi_0\|_{H^2}^2+\|F(0)\|_{L^2}^2 \right)+c_6. \end{aligned}$$ Taking the limit $m\rightarrow \infty$, we find $\Psi'\in L^\infty(0,T; L^2(\Omega))$. Using this result in , we have that $\Psi\in L^\infty(0,T;H^2(\Omega))$ and $\Psi$ is globally bounded by a constant $c_8$ depending on $T$, $\|\Psi_0\|_{L^2}$, $\|F\|_Y$ and $\|u\|_{H^1(0,T)}$ as follows $$\label{eq:globalbound} \operatorname*{ess\,sup}_{0\leq t\leq T} \max_{x\in \Omega} |\Psi(x,t)|\leq c_8.$$ \[remark:aposteriori\] By , the solution of is everywhere and for almost all times bounded by a constant. As $V_x(\Psi)\Psi$ is a convex function of $\Psi$, it is hence Lipschitz continuous for solutions of . Assumption \[assumptions\] (\[assumptionVxLipschitz\]) is hence a reasonable assumption as it holds for all solutions. Conclusion ========== In this paper, the existence, uniqueness and improved regularity of solutions to the time-dependent Kohn-Sham (KS) equations and related equations were proved. These results were proved considering a representative class of KS potentials. This work is instrumental for investigating optimal control problems governed by the KS equations. [^1]: Institut für Mathematik, Universität Würzburg, Emil-Fischer-Strasse 30, 97074 Würzburg, Germany ([[email protected]]{}). [^2]: Section de mathématiques, Université de Genève, 2-4 rue du Lièvre 1211 Genève 4, Switzerland ([[email protected]]{}). [^3]: Institut für Mathematik, Universität Würzburg, Emil-Fischer-Strasse 30, 97074 Würzburg, Germany ([[email protected]]{}). [^4]: Supported in part by the Deutsche Forschungsgemeinschaft (DFG) project “Controllability and Optimal Control of Interacting Quantum Dynamical Systems” (COCIQS).

ArXiv

--- author: - | Yuan Hou$^{a, b}$[^1], An Chang$^b$, Lei Zhang$^b$\ $^a$Department of Computer Engineering, Fuzhou University Zhicheng College,\ Fuzhou, Fujian, P.R. China\ $^b$Center for Discrete Mathematics and Theoretical Computer Science,\ Fuzhou University, Fuzhou, Fujian, P.R. China title: 'A homogeneous polynomial associated with general hypergraphs and its applications [^2]' --- [1.1]{} : In this paper, we define a homogeneous polynomial for a general hypergraph, and establish a remarkable connection between clique number and the homogeneous polynomial of a general hypergraph. For a general hypergraph, we explore some inequality relations among spectral radius, clique number and the homogeneous polynomial. We also give lower and upper bounds on the spectral radius in terms of the clique number. [**AMS**]{}: 15A42; 05C50. [**Keywords**]{}: Spectral radius, general hypergraph, Adjacency tensor, clique number Introduction ============ A general hypergraph is a pair $H=(V, E)$ consisting of a vertex set $V$ and an edge set $E$, where each edge is a subset of $V$. If $H'=(V', E')$ is a hypergraph such that $V'\subseteq V$ and $E'\subseteq E$, then $H'$ is called a subhypergraph of $H$. For a vertex $i$, we denote the family of edges containing $i$ by $E(i)$. Two vertices $i$ and $j$ are said to be adjacent, denoted by $i\sim j$, if there exists an edge $e$ such that $\{ i,j \}\subseteq e$. Otherwise, we call $i$ and $j$ are nonadjacent. The set $R=\{|e|: e\in E \}$ is called the set of edge types of $H$. We also say that $H$ is an $R$-graph. In particular, if $R$ contains only one positive integer $r$, then an $\{r\}$-graph is just an $r$-uniform hypergraph, which is simply written as $r$-graph. Consequently, a $2$-graph is referred to a usual graph. A hypergraph is non-uniform if it has at least two edge types. For a vertex $i$, let $R(i)$ be the multiset of edge types in $E(i)$. For example, assume $E(1)=\{\{1,3\}, \{1,2,3\}, \{1,3,4\}\}$, then $R(1)=\{2,3,3\}$. The rank of $H$, denoted by $rank(H)$, is the maximum cardinality of the edges in the hypergraph. For example, if $R=\{1,4\}$, we say that $H$ is a $\{1,4\}$-graph with $rank(H)=4$. For an integer $n$, let $[n]$ denote the set $\{1,2, \cdots, n\}$. For a set $S$ and integer $i$, let $\left( \begin{array}{ccc} S\\ i \end{array} \right)$ be the family of all $i$-subsets of $S$. An $R$-graph $H$ with vertex set $[n]$ and edge set $\bigcup\limits_{i\in R} \left( \begin{array}{ccc} [n]\\ i \end{array} \right)$ is called a complete $R$-graph. A complete $R$-subgraph in $H$ is called a clique of $H$. A clique is said to be maximal if it is not contained in any other clique, while it is called maximum if it has maximum cardinality. The clique number of a hypergraph $H$, denoted by $\omega(H)$, is defined as the number of vertices of a maximum clique. In other words, the clique number of a hypergraph $H$ is the number of vertices of its maximum complete $R$-subgraph in $H$. In particular, when the edge set of $H$ is empty, we delimit $\omega(H)=1$ for the purpose of complying with mathematical logic. In 1967, Wilf [@Wilf1967] first used spectral graph theory for computing bounds on the chromatic number of graphs. \[Wilf\] Let $G$ be a 2-graph with chromatic number $\chi(G)$ and spectral radius $\rho(G)$. Then $$\chi(G)\leq \rho(G)+1.$$ By above Wilf’s result, it is immediately to obtain that $\omega(G)\leq \rho(G)+1$. Later in 1986, Wilf [@Wilf1986] introduced a lower spectral bound on clique number which was inspired by an elegant result due to Motzkin and Straus [@Motzkin1965]. In 1965, Motzkin and Straus gave an answer to the following problem proposed in [@Macdonald1963]. [*Given a graph $G = (V, E)$ with vertex set $V=\{1,2, \cdots, n\}$. Let $S$ be the simplex in $\mathbb{R}^n$ given by $x_{i}\geq {0}, \sum\limits_{i=1}^{n} {x_{i}}=1$. What is $\max\limits_{x\in S}\sum\limits_{\{i,j\}\in{E}}x_{i}x_{j}$?* ]{} Let $ L(G,x)=\sum\limits_{\{i,j\}\in{E}}x_{i}x_{j}$. The Motzkin and Straus Theorem establishes a link between the problem of finding the clique number of a graph $G$ and the problem of optimizing a homogeneous polynomial $L(G,x)$ of $G$ over the simplex $S=\{ (x_{1},x_{2},\cdots,x_{n})|\sum\limits_{i=1}^{n} {x_{i}}=1, x_{i}\geq {0}\,\, for\,\, i=1,2,\cdots, n. \}$. \[Motzkin-Straus\] Let $G$ be a 2-graph with clique number $\omega(G)$, and $x^{*}$ a maximizer of $L(G, x)$ over $S$. Then $$L(G,x^{*})=\frac{1}{2}(1-\frac{1}{\omega(G)}).$$ \[Wilf1986\] Let $G$ be a 2-graph with spectral radius $\rho(G)$ and principal eigenvector $x$. Then $${\omega(G)}\geq \frac{M^{2}}{M^{2}-\rho(G)}.$$ where $M$ is the sum of the entries of the principal eigenvector $x$. Naturally we want similar comfort and convenience for spectra of hypergraphs. So it is a natural thought to generalize the Wilf’s results to general hypergraphs by using the tool of the spectral hypergraph theory. We first define a homogeneous polynomial of degree $m$ for general hypergraphs. For a general hypergraph $H$ with $rank(H)=m$, and an edge $e=\{{l_1},{l_2}, \cdots, {l_s}\}$ with cardinality $s\leq m$, we define $$\label{edge polynomial}x^{e}_{m}=\sum x_{i_{1}}x_{i_{2}} \cdots x_{i_{m}},$$ where the sum is over $i_{1}, i_{2}, \cdots, i_{m}$ chosen in all possible ways from $\{l_{1}, l_{2}, \cdots, l_{s}\}$ with at least once for each element of the set. \[lagrange x\] Let $S =\{ (x_{1},x_{2},\cdots,x_{n})|\sum\limits_{i=1}^{n} {x_{i}}=1, x_{i}\geq {0}\,\, for\,\, i=1,2,\cdots, n. \}$ and $H$ be a general hypergraph with $rank(H)=m$. For a vector $x$ in $S$, define $$L(H, x)=\sum\limits_{e\in{E}}\frac{1}{\alpha(s)}x_{m} ^{e},$$ where $s$ is the cardinality of the edge $e$ and $\alpha(s)=\sum\limits_{k_1,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m} \frac{m!}{k_1!k_2!\cdots k_s!}$.\ Furthermore, $$\label{L(H)} L(H)=\max\{L(H, x): x\in S \}.$$ A vector $x\in S$ is called an optimal weighting for $H$ if $L(H)=L(H, x)$. Observe that if $H$ is a 2-graph, by Equation (1), we have $\alpha(2)=2$. For an arbitrary vector $x$ in $S$, $$L(H, x)=\sum\limits_{\{i,j\}\in{E}}\frac{1}{2}(x_{i}x_{j}+x_{j}x_{i})=\sum\limits_{\{i,j\}\in{E}}x_{i}x_{j},$$ which is exactly the homogeneous polynomial for 2-graphs in the Motzkin-Straus theorem. In this paper, we apply the homogeneous polynomial methods to study some relations between the largest $H$-eigenvalues of adjacency tensor and clique numbers of general hypergraphs. This work is motivated by the classic results for graphs [@Wilf1967; @Wilf1986] and some recent results [@Frankl1998; @Keevash2013; @Mubayi2006; @Rota; @2009; @Pelillo; @2009; @Talbot2002; @Yuejian; @Peng2016; @peng2016; @chang2013]. Notice that the Graph-Lagrangian of a non-uniform hypergraph is a nonhomogeneous polynomial, which is different from our definition here. In this paper, we also adopt some definitions and methods of previous studies. The rest of this paper is organized as follows. In the next section, we present some definitions and properties on eigenvalues of tensors and hypergraphs. Also we give some useful tools to complete our proof. In Section 3, we attempt to explore the relationships among the homogeneous polynomial, the spectral radius, and the clique number for general hypergraphs. We also bound the spectral radius and clique number for general hypergraphs based on the results obtained. Preliminary =========== In 2005, Qi [@Qi2005_2] and Lim [@Lim2005_3] independently introduced the concept of tensor eigenvalues and the spectra of tensors. An $m$th-order $n$-dimensional real tensor $\mathcal{T}=(\mathcal{T}_{i_{1}\cdots i_{m}})$ consists of $n^{m}$ real entries $\mathcal{T}_{i_{1}\cdots i_{m}}$ for $1\leq{i_{1}, i_{2},\cdots, i_{m}}\leq {n}$. Obviously, a vector of dimension $n$ is a tensor of order $1$ and a matrix is a tensor of order $2$. $\mathcal{T}$ is called symmetric if the value of $\mathcal{T}_{i_{1}\cdots i_{m}}$ is invariant under any permutation of its indices $i_{1}, i_{2},\cdots, i_{m}$. Given a vector $x \in R^{n}$, $\mathcal{T}x^{m}$ is a real number and $\mathcal{T} x^{m-1}$ is an $n$-dimensional vector. $\mathcal{T}x^{m}$ and the $i$th component of $\mathcal{T} x^{m-1}$ are defined as follows: $$\begin{aligned} \mathcal{T}x^{m}&=&\sum_{i_{1},i_{2},\cdots, i_{m} \in [n]} \mathcal{T}_{i_{1}i_{2} \cdots i_{m}}x_{i_{1}}x_{i_{2}} \cdots x_{i_{m}}.\\ (\mathcal{T}x^{m-1})_{i}&=&\sum_{i_{2},\cdots, i_{m}\in [n]} \mathcal{T}_{ii_{2} \cdots i_{m}}x_{i_{2}} \cdots x_{i_{m}}.\end{aligned}$$ Let $\mathcal{T}$ be an $m$th-order $n$-dimensional real tensor. For some $\lambda \in {\mathbb{C}}$, if there exists a nonzero vector $x \in \mathbb{C}^{n}$ satisfying the following eigenequation $$\label{eigenequations} \mathcal{T}x^{m-1} = \lambda x^{[m-1]}.$$ Then $\lambda$ is an eigenvalue of $\mathcal{T}$ and $x$ is its corresponding eigenvector, where $x^{[m-1]}:=({x^{m-1}_{1}},{x^{m-1}_{2}},$ $\cdots,{x^{m-1}_{n}})^{T} \in \mathbb{C}^{n}\setminus\{0\}$. If $x$ is a real eigenvector of $\mathcal{T}$, surely the corresponding eigenvalue $\lambda$ is real. In this case, $\lambda$ is called an $H$-eigenvalue and $x$ is called an $H$-eigenvector associated with $\lambda$. Furthermore, if $x$ is nonnegative and real, we say $\lambda$ is an $H^{+}$-eigenvalue of $\mathcal{T}$. If $x$ is positive and real, $\lambda$ is said to be an $H^{++}$-eigenvalue of $\mathcal{T}$. The maximal absolute value of the eigenvalues of $\mathcal{T}$ is called the spectral radius of $\mathcal{T}$, denoted by $\rho(\mathcal{T})$. For nonnegative tensors, we have the Perron-Frobenius theorem, established as \[Perron-Frobenius theorem for nonnegative tensors\] \(1) (Yang and Yang 2010). If $\mathcal{T}$ is a nonnegative tensor of order $k$ and dimension $n$, then $\rho(\mathcal{T})$ is an $H^{+}$-eigenvalue of $\mathcal{T}$.\ (2) (Friedland Gaubert and Han 2011). If furthermore $\mathcal{T}$ is weakly irreducible, then $\rho(\mathcal{T})$ is the unique $H^{++}$-eigenvalue of $\mathcal{T}$, with the unique eigenvector $x\in R_{++}^{n}$, up to a positive scaling coefficient.\ (3) (Chang Pearson and Zhang 2008). If moreover $\mathcal{T}$ is irreducible, then $\rho(\mathcal{T})$ is the unique $H^{+}$-eigenvalue of $\mathcal{T}$, with the unique eigenvector $x\in R_{+}^{n}$, up to a positive scaling coefficient. In 2012, Cooper and Dutle [@Cooper2012] defined the adjacency tensor of an $r$-graph. Later in 2017, Banerjee et al.[@Banerjee2017] defined the adjacency tensor for general hypergraphs as the following. \[generalhypergraphadjacency\]Let $H=(V,E)$ be a general hypergraph with $rank(H)=m$. The adjacency tensor $\mathcal{A}$ of $H$ is defined as follows $$\mathcal{A}=(a_{i_1i_2\cdots i_m}),1\leq i_1,i_2,\cdots,i_m\leq n.$$ For all edges $e=\{{l_1},{l_2}, \cdots, {l_s}\}\in E$ of cardinality $s\leq m$, $$\label{general adjacency entry}a_{i_1i_2\cdots i_m}=\frac{s}{\alpha(s)}, where \ \alpha(s)=\sum\limits_{k_1,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m} \frac{m!}{k_1!k_2!\cdots k_s!}$$ and $i_1,i_2,\cdots,i_m$ are chosen in all possible ways from $\{l_1,l_2,\cdots,l_s\}$ with at least once for each element of the set. The other positions of the tensor are zeros. Notice that if $H$ is an $r$-graph, it is known from direct calculations that $a_{i_1i_2\cdots i_r}= \frac{1}{(r-1)!}$ for an arbitrary edge $e=\{{i_1},{i_2}, \cdots, {i_r}\}\in E$, which is exactly the definition of Cooper and Dutle in [@Cooper2012]. For an edge $e=\{i,i_{2},\cdots,i_{s}\}$ in a general hypergraph with $rank(H)=m$, we denote an $m$ order $n$ dimensional symmetric tensor $\mathcal{A}(e)$ by $$(\mathcal{A}(e)x)_i=\frac{s}{\alpha(s)} \sum\limits_{k_1\geq 0,k_2,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m-1} \frac{(m-1)!}{k_1!k_2!\cdots k_s!} x_i^{k_1} x_{i_2}^{k_2} \cdots x_{i_s}^{k_s},$$ which indicates that $$x^\top \mathcal{A}(e)x=\frac{s}{\alpha(s)} \sum\limits_{k_1,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m} \frac{m!}{k_1!k_2!\cdots k_s!} x_{i}^{k_1} x_{i_2}^{k_2} \cdots x_{i_s}^{k_s}.$$ Then the adjacency tensor $\mathcal{A}$ of order $m$ and dimension $n$ uniquely defines a homogeneous polynomial in $n$ variables of degree $m$ by: $$F_{\mathcal{A}}(x)=\mathcal{A}x^{m}=\sum\limits_{e\in E} x^\top \mathcal{A}(e)x.$$ Based on the definnitions in [@Banerjee2017], Kang et al.[@Kang2017] obtained the Perron-Frobenius theorem for general hypergraphs. For convenience, let $\rho(H)$ denote the spectral radius of the adjacency tensor of a general hypergraph $H$. \[Perron-Frobenius theorem for general hypergraphs\] \(1) Let $H$ be a general hypergraph, then $\rho(H)$ is an $H^{+}$-eigenvalue of $H$.\ (2) If $H$ is connected, then $\rho(H)$ is the unique $H^{++}$-eigenvalue of $H$, with the unique eigenvector $x\in \mathbb{R}_{++}^{n}$, up to a positive scaling cofficient. By Theorem \[Perron-Frobenius theorem for general hypergraphs\], if $H$ is connected, there is a unique unit positive eigenvector $x$ corresponding to $\rho{(H)}$. Let $\| x \|_{m}=(\sum \limits_{i=1}^{n} x_{i}^{m})^{\frac{1}{m}}$. The positive eigenvector $x$ with $\| x \|_{m}=1$ corresponding to $\rho(H)$ is called the principal eigenvector of $H$. Assume $x$ is the principal eigenvector of an $R$-graph $H$, by the theory of optimization, we have $$\label{spectral radius equation} \rho{(H)}=F_{\mathcal{A}}(x)=\mathcal{A}x^{m}=\sum\limits_{e\in E} x^\top \mathcal{A}(e)x.$$ We give some auxiliary lemmas which will be used in the sequel. \[Maclaurin¡¯s inequality\] Let $x_{1}, x_{2}, \cdots, x_{n}$ be positive real numbers. For any $k\in [n]$, define $S_{k}$ as follows: $$\label{ Maclaurin's inequality} S_{k}=\frac{\sum\limits_{1\leq i_{1}<i_{2}<\cdots<i_{k}\leq n}x_{i_{1}}x_{i_{2}} \cdots x_{i_{k}}}{ \left( \begin{array}{ccc} n\\ k \end{array} \right)}.$$ Then $S_{1}\geq \sqrt{S_{2}} \geq \sqrt[3]{S_{3}}\cdots \geq \sqrt[n]{S_{n}}$ with equality if and only if all the $x_{i}$ are equal. For any $k\in [n]$, by Maclaurin’s inequality, we get $S_{1}\geq \sqrt[k]{S_{k}}$, that is, $$\sum_{1\leq i_{1}<i_{2}<\cdots<i_{k}\leq n}x_{i_{1}}x_{i_{2}} \cdots x_{i_{k}}\leq { \left( \begin{array}{ccc} n\\ k \end{array} \right)}(\frac{x_{1}+x_{2}+\cdots+x_{n}}{n})^{k}.$$ The following lemma is the generalization of the Cauchy-Schwarz inequality for more than two vectors. \[Cauchy¨CSchwarz inequality\] Let $x_{1}=(x_{i}^{(1)}), x_{2}=(x_{i}^{(2)}), \cdots, x_{k}=(x_{i}^{(k)})$ be nonnegative vectors of dimension $n$. Then $$\label{inequality} \sum_{i=1}^{n}\prod_{j=1}^{k}x_{i}^{(j)} \leq \| x_{1} \|_{k}\| x_{2} \|_{k}\cdots| x_{k} \|_{k}.$$ Equality holds if and only if all vectors are collinear to one of them. Main results ============ In this section, we discuss the relations between the clique number and the spectral radius of an $R$-graph. A tight lower bound of the spectral radius of an $R$-graph is presented. Further, we determine the upper bound of the spectral radius based on the clique number for an $\{m,m-1\}$-graph, which derived from a Motzkin-Straus type result due to $L(H)$ for $\{m,m-1\}$-graphs. In the following discussions, without loss of generality, suppose $m\geq3$. \[lower bound for R-hypergraph\] Let $H$ be an $R$-graph with clique number $\omega$. Then $$\rho(H)\geq \sum\limits_{s\in R}\left( \begin{array}{ccc} \omega-1\\ s-1 \end{array} \right)$$ with equality if and only if $H$ is a complete $R$-graph. Assume $H_{0}$ be the maximum complete $R$-subgraph of $H$ on $\omega$ vertices. Suppose $rank(H)=m$. Let $x$ be a nonnegative vector of dimension $n$ with entries $$x_{i}=\begin{cases} \frac{1}{\sqrt[m]{\omega}}&\text{if $i\in H_{0}$},\\ 0&\text{otherwise}. \end{cases}$$ It is easy to verify that $\| x \|^{m} _{m}=1$. For an arbitrary edge $e$ in $H_{0}$, suppose $|e|=s$, we have $$\begin{aligned} \label{clique eq1} x^\top \mathcal{A}(e)x&=&\frac{s}{\alpha(s)} \sum\limits_{k_1,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m} \frac{m!}{k_1!k_2!\cdots k_s!} x_{i}^{k_1} x_{i_2}^{k_2} \cdots x_{i_s}^{k_s}\\&=&\frac{1}{\omega}\frac{s}{\alpha(s)}\sum\limits_{k_1,\cdots,k_s\geq 1,\atop k_1+\cdots+k_s=m} \frac{m!}{k_1!k_2!\cdots k_s!}\\&=&\frac{s}{\omega}.\end{aligned}$$ After using the Equation (\[spectral radius equation\]), we have $$\begin{aligned} \rho(H)&\geq&F_{\mathcal{A}}(x)=\sum\limits_{e\in E(H_{0})} x^\top \mathcal{A}(e)x =\sum\limits_{s\in R}\sum\limits_{e\in E(H_{0}),\atop |e|=s}\frac{s}{\omega} = \sum\limits_{s\in R}\frac{s}{\omega}\left( \begin{array}{ccc} \omega\\ s \end{array} \right) =\sum\limits_{s\in R}\left( \begin{array}{ccc} \omega-1\\ s-1 \end{array} \right).\end{aligned}$$ If $\rho(H)= \sum\limits_{s\in R}\left( \begin{array}{ccc} \omega-1\\ s-1 \end{array} \right)$, that is to say the unit vector $x$ is a maximizer of $F_{\mathcal{A}}(x)$ over $\mathbb{R}_{+}^{n}$. Then by Theorem \[Perron-Frobenius theorem for general hypergraphs\], $x$ must be a positive vector. Thus all vertices in $H$ belong to $H_{0}$, which indicates that $\omega=n$, i.e., $H$ is a complete $R$-graph. On the other hand, if $H$ is a complete $R$-graph, then $\omega=n$ and $x$ is the principal eigenvector. By Equation (\[spectral radius equation\]), we have $\rho(H)= \sum\limits_{s\in R}\left( \begin{array}{ccc} \omega-1\\ s-1 \end{array} \right)$. Therefore the theorem follows. Observe that if $H$ is an $m$-graph, Theorem \[lower bound for R-hypergraph\] indicates that $\rho(H)\geq \left( \begin{array}{ccc} \omega-1\\ m-1 \end{array} \right) $ as proved by Yi and Chang in [@chang2013]. Spectral methods for 2-graphs reside on a solid ground, with traditions settled both in tools and problems, such as number of edges, independence number. Theorem \[lower bound for R-hypergraph\] is also a useful tool for general $R$-graphs. Now, we give the lower bound of clique number for $R$-graphs with $R=\{ m, m-1 \}$. Let $H$ be an $\{m,m-1\}$-graph with clique number $\omega$. Then $$\omega \leq m-2+[(m-1)!\rho(H)]^{\frac{1}{m-1}}.$$ Since $$\left( \begin{array}{ccc} \omega\\ m-1 \end{array} \right)=\frac{\omega(\omega-1)\cdots(\omega-m+2)}{(m-1)!}\geq \frac{(\omega-m+2)^{m-1}}{(m-1)!}$$ By Theorem \[lower bound for R-hypergraph\], if $R=\{m,m-1\}$, we have $$\begin{aligned} \rho(H)&\geq& \left( \begin{array}{ccc} \omega-1\\ m-1 \end{array} \right)+ \left( \begin{array}{ccc} \omega-1\\ m-2 \end{array} \right)= \left( \begin{array}{ccc} \omega\\ m-1 \end{array} \right)\geq \frac{(\omega-m+2)^{m-1}}{(m-1)!}.\end{aligned}$$ Thus it is easy to verify that $\omega \leq m-2+[(m-1)!\rho(H)]^{\frac{1}{m-1}}.$ This completes the proof. \[general M-S\] Let $H$ be an $\{m,m-1\}$-graph with clique number $\omega$. If either $H$ is a complete $\{m,m-1\}$-graph or there exists two nonadjacent vertices $i$ and $j$ such that $R(i)=R(j)$, then $$\begin{aligned} L(H)&=&(\frac{1}{\omega})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right).\end{aligned}$$ Assume $H_{0}$ be the maximum complete $\{m,m-1\}$-subgraph of $H$ on $\omega$ vertices. Let $x$ be a nonnegative vector of dimension $n$ with entries $$x_{i}=\begin{cases} \frac{1}{\omega}&\text{if $i\in H_{0}$},\\ 0&\text{otherwise}. \end{cases}$$ It is easy to verify that $\| x \|_{1}=\sum \limits_{i=1}^{n} x_{i}=1$, i.e., $x\in S$. After using the Definition \[lagrange x\], we have $$\begin{aligned} L(H)&\geq& L(H,x)\\&=& \sum\limits_{e\in{E(H_{0})}}\frac{1}{\alpha(s)}x_{m} ^{e}\\ &=& \sum\limits_{s\in \{m,m-1\}}\frac{1}{\alpha(s)}\sum\limits_{ e\in E(H_{0}),\atop |e|=s}x^{e}_{m}\\ &=&\sum\limits_{s\in \{m,m-1\}}\frac{1}{\alpha(s)}\sum\limits_{e\in E(H_{0}),\atop |e|=s}\alpha(s)({\frac{1}{\omega}})^{m}\\ &=&({\frac{1}{\omega}})^{m}\sum\limits_{s\in \{m,m-1\}}\left( \begin{array}{ccc} \omega\\ s \end{array} \right)\\ &=&(\frac{1}{\omega})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right).\end{aligned}$$ To prove the opposite inequality, we proceed by induction on $n$. For $n\leq m-2$, we have $\omega =1$ and $L(H)=0$. Without loss of generality, suppose $n$ is sufficiently large. Assume the theorem true for $\{m,m-1\}$-graphs with fewer than $n$ vertices. Suppose $x$ is an optimal weighting for $H$, where $x_{1}\geq x_{2}\geq \cdots \geq x_{k} > x_{k+1}=x_{k+2}=\cdots=x_{n}=0$. If $k<n$, there exists one of the entries $x_{i}=0$. Let $H'$ be obtained from $H$ by deleting the corresponding vertex $i$ and the edges containing $i$. Since the theorem holds for $H'$, we have $$\begin{aligned} L(H)&=& L(H')= ({\frac{1}{\omega'}})^{m}\left( \begin{array}{ccc} \omega'+1\\ m \end{array} \right),\end{aligned}$$ where $\omega'$ is the clique number of $H'$. It is clearly that $\omega' \leq \omega$. For $\omega\geq 1$, $$\begin{aligned} ({\frac{1}{\omega}})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right)&=&\frac{(1+\frac{1}{\omega})(1-\frac{1}{\omega})(1-\frac{2}{\omega})\cdots(1-\frac{m-2}{\omega})}{m!} =\frac{(1-\frac{1}{\omega^{2}})(1-\frac{2}{\omega})\cdots(1-\frac{m-2}{\omega})}{m!},\end{aligned}$$ which is an monotonically increasing function of $\omega$ . Thus $$\begin{aligned} L(H)&=& L(H')\leq ({\frac{1}{\omega}})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right).\end{aligned}$$ If $k=n$, that is $x_{1}\geq x_{2}\geq \cdots \geq x_{n}>0$. To proceed our proof, we consider the following two cases. [**[Case 1.]{}**]{} There exists two nonadjacent vertices $i$ and $j$ such that $R(i)=R(j)$. Notice that $R(i)$and $R(j)$ are multisets. For a vector $x$ in $R_{+}^{n}$, we write $$L^{i}(H,x)=\sum\limits_{e\in E(i)}\frac{1}{\alpha(s)}x_{m}^{e},$$ and $$L^{j}(H,x)=\sum\limits_{e\in E(j)}\frac{1}{\alpha(s)}x_{m}^{e}.$$ Define two vectors $y=(y_{k})$ and $z=(z_{k})$ in $R_{+}^{n}$ as follows. Let $y_{k} = x_{k}$ for $k\neq i,j$, $y_{i}=x_{i}+x_{j}$ and $y_{j}=0$. Clearly, $y\in S$. Further, $z_{i} = x_{j}$ and $z_{k} = x_{k}$ for $k\neq i$. Without loss of generality, assume $L^{i}(H,z)\geq L^{j}(H,x)$. By definition \[lagrange x\] and $R(i)=R(j)$, we have $$\begin{aligned} L(H,y)-L(H,x)&=& L^{i}(H,y)-L^{i}(H,x)-L^{j}(H,x)\\ &=& \sum\limits_{e\in E(i)}\frac{1}{\alpha(s)}y_{m}^{e}-\sum\limits_{e\in E(i)}\frac{1}{\alpha(s)}x_{m}^{e} -\sum\limits_{e\in E(j)}\frac{1}{\alpha(s)}x_{m}^{e}.\\\end{aligned}$$ For an arbitrary positive interger $k$, it is clearly that $(x_{i}+x_{j})^{k}> x_{i}^{k}+x_{j}^{k}$. Consequently, $L^{i}(H,y)>L^{i}(H,x)+L^{i}(H,z).$ Thus $$L(H,y)-L(H,x)= L^{i}(H,y)-L^{i}(H,x)-L^{j}(H,x)> L^{i}(H,z)-L^{j}(H,x)\geq {0},$$ So that the maximum is attained for the subgraph $H'$ obtained from $H$ by deleting the vertex $j$ and the corresponding edges containing $j$. Similarly, the theorem is again true by the induction hypothesis. [**[Case 2.]{}**]{} If $H$ is a complete $\{m,m-1\}$-graph, it is obvious that $\omega=n$ and the maximizer of $L(H, x)$ over $S$ must be a positive vector. Otherwise, we can also use the induction hypothesis. Then we have $$\begin{aligned} L(H)&=&\max\limits_{x\in S}L(H,x)=\max\limits_{x\in S}\sum\limits_{e\in E}\frac{1}{\alpha(s)}x_{m}^{e}=\max\limits_{x\in S}\sum\limits_{s\in{\{m-1,m\}}}\frac{1}{\alpha(s)}\sum\limits_{e\in{E}\atop {|e|=s}}x_{m} ^{e} \\ &\leq& \max\limits_{x\in S}\frac{1}{\alpha(m)}\sum\limits_{e\in{E}\atop {|e|=m}}x_{m} ^{e}+\max\limits_{x\in S}\frac{1}{\alpha(m-1)}\sum\limits_{e\in{E}\atop {|e|=m-1}}x_{m} ^{e}.\\\end{aligned}$$ It is easy to verify that $\alpha(m)=m!$ and $$\sum\limits_{e\in{E}\atop {|e|=m}}x_{m} ^{e}=\sum\limits_{\{i_{1},i_{2},\cdots,i_{m}\}\in E,\atop {1\leq i_{1}<i_{2}<\cdots<i_{k}\leq n}}m!x_{i_{1}}x_{i_{2}}\cdots x_{i_{m}},$$ where the sum is over all edges with cardinality $m$. By Maclaurin’s inequality, we have $$\begin{aligned} \frac{1}{\alpha(m)}\sum\limits_{e\in{E}\atop {|e|=m}}x_{m} ^{e}&=& \sum\limits_{\{i_{1},i_{2},\cdots,i_{m}\}\in E,\atop {1\leq i_{1}<i_{2}<\cdots<i_{k}\leq n}}x_{i_{1}}x_{i_{2}}\cdots x_{i_{m}}\\&\leq& \left( \begin{array}{ccc} n\\ m \end{array} \right)(\frac{\sum\limits_{i=1}^{n}x_{i}}{n})^{m}\\ &=&(\frac{1}{n})^{m}\left( \begin{array}{ccc} n\\ m \end{array} \right).\end{aligned}$$ The equality holds if and only if $x_{1}=x_{2}=\cdots=x_{n}=\frac{1}{n}.$ Similarly, $\alpha(m-1)=\sum\limits_{k_1,\cdots,k_{m-1}\geq 1,\atop k_1+\cdots+k_{m-1}=m} \frac{m!}{k_1!k_2!\cdots k_{m-1}!}=\frac{(m-1)m!}{2}$, and $$\begin{aligned} \sum\limits_{e\in{E}\atop {|e|=m-1}}x_{m} ^{e}&=&\sum\limits_{\{i_{1},i_{2},\cdots,i_{m-1}\}\in E}\sum\limits_{k_1,\cdots,k_{m-1}\geq 1,\atop k_1+\cdots+k_{m-1}=m} \frac{m!}{k_1!k_2!\cdots k_{m-1}!}x_{i_{1}}^{k_{1}}x_{i_{2}}^{k_{2}}\cdots x_{i_{m-1}}^{k_{m-1}}\\ &=&\frac{m!}{2}\sum\limits_{\{i_{1},i_{2},\cdots,i_{m-1}\}\in E}x_{i_{1}}x_{i_{2}}\cdots x_{i_{m-1}}(x_{i_{1}}+x_{i_{2}}+\cdots +x_{i_{m-1}})\\ &=&\frac{m!}{2}[x_{1}^{2}\sum\limits_{\{1,i_{2},\cdots,i_{m-1}\}\in E, \atop {2\leq i_{2}<\cdots<i_{m-1}\leq n}}x_{i_{2}}x_{i_{3}}\cdots x_{i_{m-1}} +\cdots\\ &+&x_{n}^{2}\sum\limits_{\{i_{2},\cdots,i_{m-1},n\}\in E,\atop {1\leq i_{2}<\cdots<i_{k}\leq n-1}} x_{i_{2}}\cdots x_{i_{m-1}}]\end{aligned}$$ By Maclaurin’s inequality, we have $$\begin{aligned} \frac{1}{\alpha(m-1)}\sum\limits_{e\in{E}\atop {|e|=m-1}}x_{m} ^{e}&=& \frac{1}{m-1}[x_{1}^{2}\sum\limits_{\{1,i_{2},i_{3},\cdots,i_{m-1}\}\in E,\atop {2\leq i_{2}<\cdots<i_{m-1}\leq n}}x_{i_{2}}x_{i_{3}}\cdots x_{i_{m-1}}+\cdots\\ &+&x_{n}^{2}\sum\limits_{\{i_{2},\cdots,i_{m-1},n \}\in E,\atop {1\leq i_{2}<\cdots<i_{k}\leq n-1}} x_{i_{2}}\cdots x_{i_{m-1}}]\\ &\leq&\frac{1}{m-1}\left( \begin{array}{ccc} n-1\\ m-2 \end{array} \right)[x_{1}^{2}(\frac{\sum\limits_{i=2}^{n}x_{i}}{n-1})^{m-2}+\cdots+x_{n}^{2}(\frac{\sum\limits_{i=1}^{n-1}x_{i}}{n-1})^{m-2}].\\\end{aligned}$$ The equality holds if and only if $x_{1}=x_{2}=\cdots=x_{n}=\frac{1}{n}.$ Hence, the maximum is attained on $x_{1}=x_{2}=\cdots=x_{n}=\frac{1}{n}$. After setting $x_{1}=x_{2}=\cdots=x_{n}=\frac{1}{n}$, we get $$\begin{aligned} \max_{x\in S}\frac{1}{\alpha(m-1)}\sum\limits_{e\in{E}\atop {|e|=m-1}}x_{m} ^{e}&=& (\frac{1}{n})^{m-1}\frac{1}{m-1}\left( \begin{array}{ccc} n-1\\ m-2 \end{array} \right)\\ &=&(\frac{1}{n})^{m}\frac{n}{m-1}\left( \begin{array}{ccc} n-1\\ m-2 \end{array} \right)\\&=&(\frac{1}{n})^{m}\left( \begin{array}{ccc} n\\ m-1 \end{array} \right)\end{aligned}$$ Combining the above inequalities, if $H$ is a complete $\{m,m-1\}$-graph, then $\omega=n$ and $$\begin{aligned} L(H)&\leq&(\frac{1}{n})^{m}\left( \begin{array}{ccc} n\\ m \end{array} \right)+(\frac{1}{n})^{m}\left( \begin{array}{ccc} n\\ m-1 \end{array} \right)=(\frac{1}{n})^{m}\left( \begin{array}{ccc} n+1\\ m \end{array} \right).\end{aligned}$$ The equality holds if and only if $x_{1}=x_{2}=\cdots=x_{n}=\frac{1}{n}$. This completes the proof. \[upper bound on clique and spectral radius\] Let $H$ be an $\{m,m-1\}$-graph with clique number $\omega$. If either $H$ is a complete $\{m,m-1\}$-graph or there exists two nonadjacent vertices $i$ and $j$ such that $R(i)=R(j)$, then $$\rho(H)\leq m({\frac{U}{\omega}})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right)$$ where $U$ is the sum of the entries of the principal eigenvector. Let $x$ be the principal eigenvector of $H$. By Equation (\[spectral radius equation\]), we have $$\begin{aligned} \rho(H)&=&\sum\limits_{e\in E} x^\top \mathcal{A}(e)x=\sum\limits_{s\in \{m,m-1\}}\frac{s}{\alpha(s)}\sum\limits_{e\in E, \atop |e|=s}x^{e}_{m}.\end{aligned}$$ Set $y=\frac{x}{U}$, where $U$ is the sum of the entries of $x$. It is clearly that $y\in S$. Apply the Theorem \[general M-S\], then $$\begin{aligned} \frac{\rho(H)}{U^{m}}&=&\sum\limits_{s\in \{m,m-1\}}\frac{s}{\alpha(s)}\sum\limits_{e\in E, \atop |e|=s}y^{e}_{m}\\ &\leq&\sum\limits_{s\in \{m,m-1\}}\frac{m}{\alpha(s)}\sum\limits_{e\in E, \atop |e|=s}y^{e}_{m}\\ &\leq& m({\frac{1}{\omega}})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right)\end{aligned}$$ Thus $$\rho(H)\leq m({\frac{U}{\omega}})^{m}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right)$$ Therefore the theorem follows. And further applying Lemma \[Cauchy¨CSchwarz inequality\], it is easy to check the following result. Let $H$ be an $\{m,m-1\}$-graph with clique number $\omega$. If either $H$ is a complete $\{m,m-1\}$-graph or there exists two nonadjacent vertices $i$ and $j$ such that $R(i)=R(j)$, then $$\rho(H)\leq {\frac{mn^{m-1}}{\omega^{m}}}\left( \begin{array}{ccc} \omega+1\\ m \end{array} \right)$$ Let $x$ be the principal eigenvector of $H$ and $U$ is the sum of the entries of $x$. By Lemma \[Cauchy¨CSchwarz inequality\], we have $$\begin{aligned} U&=&\sum\limits_{i=1}^{n}x_{i}=\sum\limits_{i=1}^{n}1\cdot 1\cdots x_{i}\leq (\sum\limits_{i=1}^{n}1)^{\frac{m-1}{m}}(\sum\limits_{i=1}^{n}{x_{i}^{m}})^{\frac{1}{m}}=n^{\frac{m-1}{m}}.\end{aligned}$$ Obviously, $U^{m}\leq n^{m-1}.$ Combining with Theorem \[upper bound on clique and spectral radius\], the result follows. [99]{} C. Berge, Hypergraph: Combinatorics of Finite Sets, third edition, North-Holland, Amsterdam, 1973. A. Bretto, Hypergraph Theory: An Introduction, Springer, 2013. A. Banerjee, A. Char, B. Mondal, Spectral of general hypergraphs, Linear Algebra Appl. 518 (2017) 14-30. J. Cooper, A. Dutle, Spectra of uniform hypergraphs, Linear Algebra Appl. 436 (2012) 3268-3292. Y. Fan, Y. Tan, X. Peng, A. Liu, Maximizing spectral radii of uniform hypergraphs with few edges, Discussiones Math. Graph Theory 36(2016) 845-856. P. Frankl, Z. Füredi, Extremal problems and the Lagrange function of hypergraphs, Bulletin Institute Math. Academia Sinica 16(1988) 305-313. R. Gu, X. Li, Y. Peng, Y. Shi, Some Motzkin-Straus type results for non-uniform hypergraphs, J. Comb. Optim. 31(2016) 223-238. G. Hardy, J. Littlewood, G. Pólya, Inequalities, 2nd edition, Cambridge University Press, 1988. D. Hefetz, P. Keevash, A hypergraph Turán theorem via lagrangians of intersecting families, J. Combin. Theory Ser. A 120 (2013) 2020-2038. J. MacDonald Jr., Problem E1643, Amer. Math. Monthly 70 (1963) 1099. T. Motzkin, E. Straus, Maxima for graphs and a new proof of a theorem of Turán, Canad. J. Math. 17(1965) 533-540. D. Mubayi, A hypergraph extension of Turans theorem, J. Combin. Theory Ser. B 96 (2006) 122-134. V. Nikiforov, Analytic methods for uniform hypergraphs, Linear Algebra Appl. 457 (2014) 455-535. B. Papendieck, P. Recht, On maximal entries in the principal eigenvector of graphs, Linear Algebra Appl. 310 (2000) 129-138. Y. Peng, H. Peng, Q. Tang, C. Zhao, An extension of Motzkin-Straus Thorem to non-uniform hypergraphs and its applications, Discrete Appl. Math. 200 (2016) 170-175. L. Qi, Symmetric nonnegative tensors and copositive tensors, Linear Algebra Appl. 439(2013) 228-238. L. Lim, Singular values and eigenvalues of tensors: a variational approach, in: Proceedings of the IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP 05) 1(2005) 129-132. L. Liu, L. Kang, X. Yuan, On the principal eigenvectors of uniform hypergraphs, Linear Algebra Appl. 511(2016) 430-446. L. Qi, Eigenvalues of a real supersymmetric tensor, J. Symb. Comput. 40(2005) 1302-1324. S. Rota Bulò, M. Pelillo, A generalization of the Motzkin-Straus theorem to hypergraphs, Optim. Lett. 3 (2009) 187-295. S. Rota Bulò, M. Pelillo, New bounds on the clique number of graphs based on spectral hypergraph theory, Learning and Intelligent Optim. 5851(2009) 45-58. J. Talbot, Lagrangians of hypergraphs, Combin. Prob. Comput. 11 (2002) 199-216. H. Wilf., The eigenvalues of a graph and its chromatic number, J. London Math. Soc. 42(1967) 330-332. H. Wilf., Spectral bounds for the clique and independence numbers of graphs, J. Comb. Theory Series B 40(1986) 113-117. G. Yi, A. Chang, The spectral bounds for the clique numbers of $r$-uniform hypergraphs, Manuscript, Fuzhou University, 2013. W. Zhang, L. Liu, L. Kang, Y. Bai, Some properties of the Spectral radius for general hypergraphs, Linear Algebra Appl. 513(2017) 103-119. [^1]: E-mail addresses: [email protected](Y. Hou), [email protected](A. Chang). [^2]: This research is supported by the National Natural Science Foundation of China (Grant No. 11471077).

ArXiv

--- author: - Daniel Hug and Rolf Schneider title: | Approximation properties of random polytopes\ associated with Poisson hyperplane processes --- > [**Abstract**]{} > > We consider a stationary Poisson hyperplane process with given directional distribution and intensity in $d$-dimensional Euclidean space. Generalizing the zero cell of such a process, we fix a convex body $K$ and consider the intersection of all closed halfspaces bounded by hyperplanes of the process and containing $K$. We study how well these random polytopes approximate $K$ (measured by the Hausdorff distance) if the intensity increases, and how this approximation depends on the directional distribution in relation to properties of $K$.\ > [*Keywords:*]{} Poisson hyperplane process; zero polytope; approximation of convex bodies; directional distribution\ > 2010 Mathematics Subject Classification: Primary 60D05 Introduction ============ Asymptotic properties of the convex hull of $n$ independent, identically distributed random points in ${\mathbb{R}}^d$, as $n$ tends to infinity, are an actively studied topic of stochastic geometry; see, for example, Subsection 8.2.4 of the book [@SW08] and the more recent survey by Reitzner [@Rei10]. Very often, one studies uniform random points in a given convex body and measures the rate of approximation by the volume difference, or the difference of other global functionals, or one investigates the asymptotic behaviour of combinatorial quantities such as face numbers. In contrast, approximation by random polytopes, measured in terms of the Hausdorff metric $\delta$, has been investigated less frequently. We recall that the Hausdorff distance of two nonempty compact sets $K,L\subset {\mathbb R}^d$ is defined by $$\delta(K,L)=\max\left\{\max_{{\mbox{\boldmath$\scriptstyle x$}}\in K}\min_{{\mbox{\boldmath$\scriptstyle y$}}\in L} \|{\mbox{\boldmath$x$}}-{\mbox{\boldmath$y$}}\|,\max_{{\mbox{\boldmath$\scriptstyle x$}}\in L}\min_{{\mbox{\boldmath$\scriptstyle y$}}\in K} \|{\mbox{\boldmath$x$}}-{\mbox{\boldmath$y$}}\|\right\}.$$ For results on Hausdorff distances of random polytopes we refer to Note 5 for Subsection 8.2.4 in [@SW08] and mention here only the following. For a convex body $K$ of class $C^2_+$ (that is, with a twice continuously differentiable boundary with positive Gauss curvature), Bárány [@Bar89] (Theorem 6) showed that the Hausdorff distance from $K$ to the convex hull $K_n$ of $n$ i.i.d. uniform random points in $K$ satisfies $${{\rm E} }\,\delta(K,K_n)\sim\left(\frac{\log n}{n}\right)^{2/(d+1)}$$ as $n\to\infty$ (here $f(n)\sim g(n)$ means that there are constants $c_1,c_2$ such that $c_1 g(n)<f(n)< c_2g(n)$). A result of Dümbgen and Walther [@DW96] (Corollary 1) says that, for an arbitrary convex body $K$, $$\delta(K,K_n)= {\rm O}\left(\left(\frac{\log n}{n}\right)^{1/d}\right)\quad\mbox{almost surely}.$$ The second standard approach to convex polytopes, generating them as intersections of closed halfspaces instead of convex hulls of points, was, for the case of random polygons in the plane, already considered in the third of the seminal papers by Rényi and Sulanke [@RS63; @RS64; @RS68], which initiated this subject. Nevertheless, this approach has later not found equal attention in the study of random polytopes. About the role that duality, either in an exact or a heuristic sense, can play here, we refer to the introduction of [@BS10]. This alternative approach has to offer some new aspects, in particular since random hyperplanes naturally come with some directional distribution, which influences the random polytopes that they generate. This aspect is emphasized in the present article, where we consider random polytopes generated by a stationary Poisson hyperplane process, with an arbitrary directional distribution. Let $X$ be a stationary nondegenerate (see [@SW08 p.486]) Poisson hyperplane process in Euclidean space ${\mathbb{R}}^d$, $d\ge 2$ (with scalar product $\langle\cdot,\cdot\rangle$ and norm $\|\cdot\|$). The reader is referred to Chapters 3 and 4 of [@SW08] for an introduction, and also for some notational conventions used here. In particular, we recall the convention that a simple point process $X$, which is by definition a simple random counting measure, is often identified with its support, which is a locally finite random set. For a hyperplane $H$ in ${\mathbb{R}}^d$, not passing through the origin ${\mbox{\boldmath$o$}}$, we denote by $H^-_{{\mbox{\boldmath$\scriptstyle o$}}}$ the closed halfspace bounded by $H$ that contains ${\mbox{\boldmath$o$}}$. The random polytope $$Z_0 := \bigcap_{H\in X} H^-_{{\mbox{\boldmath$\scriptstyle o$}}}$$ is called the [*zero cell*]{} of $X$ (it is also known as the [*Crofton polytope*]{} of $X$). A generalization of this notion is obtained as follows. Let $K\subset {\mathbb{R}}^d$ be a convex body, by which we understand, in the following, a compact convex subset with interior points. For a hyperplane $H$ not intersecting $K$ we denote by $H^-_K$ the closed halfspace bounded by $H$ that contains $K$. Then we define the $K$[*-cell*]{} of $X$ as the random polytope $$Z_K:= \bigcap_{H\in X,\,H\cap K=\emptyset} H^-_K.$$ The almost sure boundedness of $Z_K$ follows as in the proof of [@SW08 Theorem 10.3.2]. In the following we are interested in the question how well $K$ is approximated by $Z_K$, if the intensity of the process $X$ tends to infinity. Since the intensity is a constant multiple of the expected number of hyperplanes in the process that hit $K$, the analogy to convex hulls of an increasing number of points is evident. We consider approximation in sense of the Hausdorff metric $\delta$ on the space ${\mathcal K}^d$ of convex bodies in ${\mathbb{R}}^d$. Of course, in order that approximation of $K$ by $Z_K$ be possible at all, the convex body $K$ and the directional distribution of the hyperplane process $X$ must somehow be adapted to each other. For example, a ball $K$ cannot be approximated arbitrarily closely by $Z_K$ if the hyperplane process $X$ has only hyperplanes of finitely many directions. To make this more precise, let $N$ be a closed subset of the unit sphere ${\mathbb{S}^{d-1}}$, not contained in a closed halfsphere. For a given convex body $K$, we denote by ${\mathcal P}(K,N)$ the set of all polytopes which are finite intersections of closed halfspaces containing $K$ and with outer unit normal vectors in $N$. [**Proposition 1.**]{} [*The convex body $K$ can be approximated arbitrarily closely, with respect to the Hausdorff metric, by polytopes from ${\mathcal P}(K,N)$ if and only if ${\rm supp}\,S_{d-1}(K,\cdot)\subset N$.*]{} Here [supp]{} denotes the support of a measure, and $S_{d-1}(K,\cdot)$ is the surface area measure of $K$ (see [@Sch14], Section 4.2, for example). We shall give a proof of Proposition 1 in the next section. It serves here only to motivate the assumption (\[n1\]) made below. The intensity measure $\Theta= {{\rm E} }X(\cdot)$ of $X$ is assumed, as usual, to be locally finite. It can then be represented in the form (see [@SW08], (4.33)) $$\label{2.0} \Theta(A) = 2\gamma\int_{{\mathbb S}^{d-1}}\int_0^\infty {{\bf 1}}_A(H({\mbox{\boldmath$u$}},t))\,{{\rm d}}t\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})$$ for $A\in{\cal B}({\mathcal H}^d)$, where $\gamma>0$ is the intensity and $\varphi$ is the spherical directional distribution of $X$; the latter is an even Borel probability measure on the unit sphere ${\mathbb{S}^{d-1}}$ which is not concentrated on a great subsphere. Later, when $\varphi$ is fixed and $\gamma$ varies, we write $\Theta_\gamma$ instead of $\Theta$. By ${\mathcal H}^d$ we denote the space of hyperplanes in ${\mathbb{R}}^d$, and ${\mathcal B}(T)$ is the $\sigma$-algebra of Borel sets of a topological space $T$. Further, $$H({\mbox{\boldmath$u$}},t)=\{{\mbox{\boldmath$x$}}\in{\mathbb{R}}^d: \langle{\mbox{\boldmath$x$}},{\mbox{\boldmath$u$}}\rangle=t\}$$ for ${\mbox{\boldmath$u$}}\in {\mathbb S}^{d-1}$ and $t>0$ is the standard parametrization of a hyperplane not passing through the origin ${\mbox{\boldmath$o$}}$. For convenience (in view of some later estimations of constants), we also assume that $\gamma\ge 1$. For $K\in{\mathcal K}^d$, the Hausdorff distance $\delta(K,P)$ of $K$ from a polytope $P$ containing it is the smallest number $\varepsilon\ge 0$ such that $P\subset K(\varepsilon)$, where $K(\varepsilon)=K+\varepsilon B^d$ ($B^d$ is the unit ball) denotes the outer parallel body of $K$ at distance $\varepsilon$. Thus, for given $\varepsilon>0$ the probability ${{\rm P}}\{\delta(K,Z_K)>\varepsilon \}$, in which we are interested, is equal to ${{\rm P}}\{Z_K\not\subset K(\varepsilon)\}$. First we give a necessary and sufficient condition that this probability tends to zero if the intensity of the process $X$ tends to infinity; if the condition is satisfied, we obtain that the decay is exponential. Under a slightly stronger assumption, this can then be used to derive our main results, concerning the rate of convergence. We assume in the following that the surface area measure of the given convex body $K$ satisfies $$\label{n1} {\rm supp}\,S_{d-1}(K,\cdot) \subset {\rm supp}\,\varphi.$$ By Proposition 1, this assumption is necessary for arbitrarily good approximation of $K$ by $Z_K$. Theorem 1 shows, in a stronger form, that it is also sufficient. For ${\mbox{\boldmath$y$}}\in {\mathbb{R}}^d\setminus K$, let $K^{{\mbox{\boldmath$\scriptstyle y$}}}:= {\rm conv}(K\cup\{{\mbox{\boldmath$y$}}\})$. For $\varepsilon>0$ we define $$\label{n4} \mu(K,\varphi,\varepsilon):= \min_{{\mbox{\boldmath$\scriptstyle y$}}\in {\rm bd}\,K(\varepsilon)} \int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}}),$$ where $h$ denotes the support function. Lemma 1, to be proved in the next section, shows that condition (\[n1\]) implies $\mu(K,\varphi,\varepsilon)>0$. [**Theorem 1.**]{} [*Let $K\in{\mathcal K}^d$ be a convex body. Let $X$ be a stationary Poisson hyperplane process in ${\mathbb R}^d$ with intensity $\gamma$ and with a directional distribution $\varphi$ satisfying $(\ref{n1})$. There are positive constants $C_1(\varepsilon), C_2$ (both depending on $K$, $\varphi$, $d$) such that the following holds. If $0<\varepsilon\le 1$, then*]{} $$\label{5.0} {{\rm P}}\left\{\delta(K,Z_K)>\varepsilon \right\} \le C_1(\varepsilon)\exp\left[-C_2 \mu(K,\varphi,\varepsilon)\gamma\right],$$ where $\mu(K,\varphi,\varepsilon)>0$. In order to be able to deal with convergence for increasing intensities, we consider an embedding of the stationary Poisson hyperplane processes $X_\gamma$ with intensity $\gamma>0$, directional distribution $\varphi$ and intensity measure $${{\rm E} }X_\gamma(\cdot) = 2\gamma \int_{{\mathbb S}^{d-1}}\int_0^\infty {{\bf 1}}\{H({\mbox{\boldmath$u$}},t)\in\cdot\}\,{{\rm d}}t\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})=: \Theta_\gamma$$ into a Poisson process $\xi$ in $[0,\infty)\times\mathcal{H}^d$ (on a suitable probability space) with intensity measure $\lambda\otimes \Theta_1$, where $\lambda$ denotes Lebesgue measure on $[0,\infty)$. Then $\xi([0,\gamma]\times\cdot)$ is a Poisson hyperplane process in ${\mathbb R}^d$ with intensity measure $\Theta_\gamma$, thus $X_\gamma$ is stochastically equivalent to $\xi([0,\gamma]\times\cdot)$ (e.g., by [@SW08], Theorem 3.2.1). In the following, we can identify $X_\gamma$ with $\xi([0,\gamma]\times\cdot)$. Let $Z_K^{(\gamma)}$ denote the $K$-cell associated with $\xi([0,\gamma]\times\cdot)$. Then we have $K\subset Z_K^{(\tau)}\subset Z_K^{(\gamma)}$ for $\tau\ge \gamma>0$, and therefore $\delta(K,Z_K^{(\tau)})\le \delta(K,Z_K^{(\gamma)})$. This shows that $${{\rm P}}\left\{\sup_{\tau\ge \gamma}\delta(K,Z_K^{(\tau)})\ge \varepsilon\right\}={{\rm P}}\left\{\delta(K,Z_K^{(\gamma)})\ge \varepsilon\right\} \le C_1(\varepsilon)\exp\left[-C_2 \mu(K,\varphi,\varepsilon)\gamma\right]$$ for all $\varepsilon>0$, and thus $$\lim_{\gamma\to \infty} \delta(K,Z_K^{(\gamma)}) = 0$$ holds almost surely. We state this as a corollary. [**Corollary.**]{} [*If the Poisson hyperplane processes $X_\gamma$, $\gamma\ge 1$, are defined as above on a common probability space and if $Z_K^{(\gamma)}$ denotes the $K$-cell of $X_\gamma$ for a convex body $K\in {\mathcal K}^d$, then condition $\rm (\ref{n1})$ is necessary and sufficient in order that*]{} $$\label{n6} \lim_{\gamma\to \infty} \delta(K,Z_K^{(\gamma)}) = 0 \quad\mbox{almost surely}.$$ In the following, we will be interested in rates of convergence. For this, we consider the sequence $X_1,X_2,\dots$ of Poisson hyperplane processes defined as above, with spherical directional distribution $\varphi$, where $X_n$ has intensity $n$. Under the sole assumption (\[n1\]), no statement stronger than (\[n6\]), involving also a rate of convergence, is possible. In fact, if any decreasing sequence $(\varepsilon_n)_{n\in{\mathbb N}}$ with $\varepsilon_n\to 0$ for $n\to\infty$ is given and if $K$ is a convex body, then the directional distribution $\varphi$ of the hyperplane processes $X_n$ can be chosen in such a way that (\[n1\]) is satisfied but $$\label{C1} {{\rm P}}\{\delta(K,Z_K^{(n)})\ge\varepsilon_n \mbox{ for almost all } n\}=1.$$ We prove this at the end of the paper. Therefore, no assumption on the convex body $K$ alone allows us to estimate the rate of convergence of $\delta(K,Z_K^{(n)})$ for arbitrary directional distributions $\varphi$. On the other hand, suitable assumptions on the directional distribution, for example $$\label{n7} \varphi \ge b\sigma$$ with a constant $b>0$, where $\sigma$ denotes spherical Lebesgue measure, permit to estimate the rate of convergence for arbitrary convex bodies. This is shown by the first assertion of Theorem 2. If the directional distribution does not satisfy such a strong assumption, then rates of convergence can only be estimated if this distribution is suitably adapted to the given convex body. In this sense, we assume that $$\label{n3} \varphi \ge bS_{d-1}(K,\cdot)$$ with some constant $b$. If $(Y_n)_{n\in\mathbb{N}}$ is a sequence of real random variables and $f(n)_{n\in\mathbb{N}}$ is a sequence of nonnegative real numbers, we write $Y_n={\rm O}(f(n))$ [*almost surely*]{} if there is a constant $C<\infty$ such that with probability one we have $Y_n\le Cf(n)$ for sufficiently large $n$. Moreover, we write $Y_n\sim f(n)$ [*almost surely*]{} if there are constants $0<c\le C<\infty$ such that with probability one we have $cf(n)\le Y_n\le Cf(n)$ for all sufficiently large $n$. A ‘ball’ in the following is a Euclidean ball of positive radius. One says that a convex body $M$ [*slides freely*]{} inside a convex body $K$ if $K$ is the union of all translates of $M$ that are contained in $K$. [**Theorem 2.**]{} *Let $K\in{\mathcal K}^d$ be a convex body. Let $X$ be a stationary Poisson hyperplane process in ${\mathbb R}^d$ with intensity $\gamma$ and with a directional distribution $\varphi$ satisfying $\rm (\ref{n7})$ or $\rm (\ref{n3})$. Then $$\label{5.0z} \delta(K,Z_K^{(n)})={\rm O}\left(\left(\frac{\log n}{n}\right)^{1/d}\right)\quad\mbox{almost surely,}$$ as $n\to\infty$.* Suppose that $\rm (\ref{n3})$ holds. If a ball slides freely inside $K$, then the exponent $1/d$ in $(\ref{5.0z})$ can be replaced by $2/(d+1)$, and if $K$ is a polytope, then it can be replaced by $1$. Under stronger assumptions on $K$ and $\varphi$, we can determine the exact asymptotic order of approximation. [**Theorem 3.**]{} [*Let the convex body $K\in{\mathcal K}^d$ be such that a ball slides freely inside $K$ and that $K$ slides freely inside a ball. Suppose that the directional distribution $\varphi$ of the stationary Poisson hyperplane processes $X_n$ satisfies $$\label{n9} a\sigma \ge \varphi\ge b\sigma$$ with some positive constants $a,b$. Then $$\label{n8} \delta(K,Z_K^{(n)})\sim \left(\frac{\log n}{n}\right)^{2/(d+1)}\quad\mbox{almost surely,}$$ as $n\to\infty$.* ]{} Note that Theorem 3 covers, in particular, the case where $K$ is of class $C^2_+$ and the hyperplane processes $X_n$ are isotropic, that is, their directional distribution $\varphi$ is invariant under rotations and thus is equal to the normalized spherical Lebesgue measure. If $K$ is of class $C^2_+$, then the assumptions on $K$ are satisfied by Blaschke’s rolling theorem (Corollary 3.2.13 in [@Sch14]). In the next section, we prove some auxiliary results. Theorem 1 is proved in Section 3, and the proofs of Theorems 2 and 3 follow in Section 4. Auxiliary results ================= [*Proof of Proposition $1$.*]{} By [@Sch14], Theorem 4.5.3, the support of the area measure $S_{d-1}(K,\cdot)$ is equal to ${\rm cl\;extn}\, K$, the closure of the set of extreme (unit) normal vectors of $K$. Suppose now that $K$ can be approximated arbitrarily closely by polytopes from ${\mathcal P}(K,N)$. Let ${\mbox{\boldmath$x$}}$ be a regular boundary point of $K$, and let $({\mbox{\boldmath$x$}}_i)_{i\in{\mathbb N}}$ be a sequence of points in ${\mathbb{R}}^d\setminus K$ converging to ${\mbox{\boldmath$x$}}$. To each $i$, there exists a polytope $P_i\in {\mathcal P}(K,N)$ not containing ${\mbox{\boldmath$x$}}_i$, hence there is a closed halfspace $H^-_i$ with outer normal vector ${\mbox{\boldmath$u$}}_i\in N$ containing $K$ but not ${\mbox{\boldmath$x$}}_i$. For $i\to\infty$, the sequence of hyperplanes $H_i$ bounding $H^-_i$ has a convergent subsequence; its limit is the unique supporting hyperplane of $K$ at ${\mbox{\boldmath$x$}}$. It follows that the outer unit normal vector of $K$ at ${\mbox{\boldmath$x$}}$ belongs to the closed set $N$. A normal vector at a regular boundary point of $K$ is a $0$-exposed normal vector. Since ${\mbox{\boldmath$x$}}$ was an arbitrary regular boundary point of $K$, the set $N$ contains the set of $0$-exposed normal vectors of $K$. The closure of the $0$-exposed normal vectors is equal to the closure of the extreme normal vectors (see Theorem 2.2.9 of [@Sch14], also for the terminology used here). Hence, ${\rm cl\;extn}\, K\subset N$. Conversely, suppose that ${\rm cl\;extn}\, K\subset N$. The body $K$ is the intersection of its supporting halfspaces with a regular point of $K$ in the boundary (see [@Sch14], Theorem 2.2.5). The outer unit normal vector of such a halfspace is extreme and hence belongs to $N$. Thus, denoting by $H^-(K,{\mbox{\boldmath$u$}})$ the supporting halfspace of $K$ with outer unit normal vector ${\mbox{\boldmath$u$}}$, we have $K=\bigcap_{{\mbox{\boldmath$\scriptstyle u$}}\in N}H^-(K,{\mbox{\boldmath$u$}})$. Therefore, if $\varepsilon>0$, then $$\bigcap_{{\mbox{\boldmath$\scriptstyle u$}}\in N} {\rm bd}(K+\varepsilon B^d)\cap H^-(K,{\mbox{\boldmath$u$}})=\emptyset.$$ By compactness, there is a finite subset $F\subset N$ such that the corresponding intersection is empty, which implies that $$P:= \bigcap_{{\mbox{\boldmath$\scriptstyle u$}}\in F}H^-(K,{\mbox{\boldmath$u$}})\subset{\rm int}(K+\varepsilon B^d ).$$ Thus, $P$ is a polytope in ${\mathcal P}(K,N)$ with $\delta(K,P)<\varepsilon$. Since $\varepsilon>0$ was arbitrary, this shows that $K$ can be approximated arbitrarily closely by polytopes from ${\mathcal P}(K,N)$. In the rest of this paper, $c_1,c_2,\dots$ denote positive constants that depend only on $K$, $\varphi$ and the dimension $d$. [**Lemma 1.**]{} *Let $K\in{\mathcal K}^d$ and let $\varphi$ be a probability measure on ${\mathbb{S}^{d-1}}$. Let $0<\varepsilon\le 1$.* $\rm (a)$ If $(\ref{n1})$ holds, then $ \mu(K,\varphi,\varepsilon)>0$. $\rm (b)$ If $\rm (\ref{n7})$ holds, then there exists a constant $c_1$ such that $$\label{5.0a} \mu(K,\varphi,\varepsilon) \ge c_1\varepsilon^d.$$ In $\rm (c),\,(d),\,(e)$ it is assumed that $(\ref{n3})$ is satisfied. $\rm (c)$ For $\varepsilon\le D(K)$, where $D(K)$ denotes the diameter of $K$, there exists a constant $c_2$ such that $$\label{5.1} \mu(K,\varphi,\varepsilon) \ge c_2\varepsilon^d.$$ $\rm (d)$ If a ball slides freely inside $K$, then there exists a constant $c_3$ such that $$\label{5.2} \mu(K,\varphi,\varepsilon)\ge c_3\varepsilon^{(d+1)/2}.$$ $\rm (e)$ If $K$ is a polytope, then there exists a constant $c_4$ such that $$\label{5.3} \mu(K,\varphi,\varepsilon) \ge c_4\varepsilon.$$ [*Proof.*]{} (a) Let $(\ref{n1})$ be satisfied. Let ${\mbox{\boldmath$y$}}\in{\mathbb{R}}^d\setminus K$. Let $V_d$ denote the volume and $V$ the mixed volume in ${\mathbb{R}}^d$. Using a formula for mixed volumes ([@Sch14], (5.19)) and Minkowski’s inequality (e.g., [@Sch14], (7.18)), we get $$\begin{aligned} & & \frac{1}{d} \int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,S_{d-1}(K,{{\rm d}}{\mbox{\boldmath$u$}})\\ && =V(K^{{\mbox{\boldmath$\scriptstyle y$}}},K,\dots,K)-V_d(K)\\ &&\ge V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})^{\frac{1}{d}}V_d(K)^{\frac{d-1}{d}}-V_d(K)\\ &&=V_d(K)^{\frac{d-1}{d}}\left[V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})^{\frac{1}{d}}-V_d(K)^{\frac{1}{d}}\right]\\ &&>0.\end{aligned}$$ The integrand is nonnegative and continuous as a function of ${\mbox{\boldmath$u$}}$. Since the integral is positive, there exists a neighbourhood (in ${\mathbb{S}^{d-1}}$) of some point ${\mbox{\boldmath$u$}}_0\in{\rm supp}\,S_{d-1}(K,\cdot)$ on which the integrand is positive. By (\[n1\]), ${\mbox{\boldmath$u$}}_0\in {\rm supp}\,\varphi$, and hence $$g({\mbox{\boldmath$y$}}):=\int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})>0.$$ The function $g$ is continuous, hence on each compact subset of ${\mathbb{R}}^d\setminus K$ it attains a minimum. This proves that $\mu(K,\varphi,\varepsilon)>0$. \(b) Suppose that (\[n7\]) holds. For the proof of (\[5.0a\]), let $K\in{\mathcal K}^d$ be given. Let ${\mbox{\boldmath$y$}}\in {\rm bd}\, K(\varepsilon)$ and let ${\mbox{\boldmath$x$}}$ be the point in $K$ nearest to ${\mbox{\boldmath$y$}}$. Then $N({\mbox{\boldmath$x$}}):=({\mbox{\boldmath$y$}}-{\mbox{\boldmath$x$}})/\varepsilon$ is an outer unit normal vector of $K$ at ${\mbox{\boldmath$x$}}$. We denote by $H^-$ the closed halfspace bounded by the hyperplane through ${\mbox{\boldmath$x$}}$ and orthogonal to $N({\mbox{\boldmath$x$}})$ and containing $K$. If $D(K)$ denotes the diameter of $K$, then $K\subset H^-\cap ({\mbox{\boldmath$x$}}+D(K)B^d)$. Define $\beta=\beta(\varepsilon)\in [0,\pi/2)$ by $\cos\beta=D(K)/\sqrt{D(K)^2+\varepsilon^2}$ and let $S({\mbox{\boldmath$y$}},\varepsilon)$ be the set of all ${{\mbox{\boldmath$u$}}}\in{\mathbb{S}^{d-1}}$ such that $\angle ({{\mbox{\boldmath$u$}}},N({\mbox{\boldmath$x$}}) )\le\beta/2$. Then $$\label{eqle1} \sigma(S({\mbox{\boldmath$y$}},\varepsilon))\ge c_5\sin^{d-1}(\beta/2)\ge c_6\varepsilon^{d-1}.$$ For ${\mbox{\boldmath$u$}}\in S({\mbox{\boldmath$y$}},\varepsilon)\setminus \{N({\mbox{\boldmath$x$}})\}$ there is a unique unit vector ${\mbox{\boldmath$e$}}$ orthogonal to $N({\mbox{\boldmath$x$}})$ such that ${\mbox{\boldmath$u$}}=\tau N({\mbox{\boldmath$x$}})+\sqrt{1-\tau^2}\,{\mbox{\boldmath$e$}}$ with $0<\tau<1$. With ${\mbox{\boldmath$z$}}:=D(K){\mbox{\boldmath$e$}}$ we then obtain $$\begin{aligned} h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})&\ge\langle {\mbox{\boldmath$y$}},{\mbox{\boldmath$u$}}\rangle-\langle {\mbox{\boldmath$z$}},{\mbox{\boldmath$u$}}\rangle =\langle {\mbox{\boldmath$y$}}-{\mbox{\boldmath$z$}},{\mbox{\boldmath$u$}}\rangle\nonumber\\ &\ge D(K)\left\langle \frac{{\mbox{\boldmath$y$}}-{\mbox{\boldmath$z$}}}{\|{\mbox{\boldmath$y$}}-{\mbox{\boldmath$z$}}\|},{\mbox{\boldmath$u$}}\right\rangle\ge D(K)\sin(\beta/2)\nonumber\\ &\ge c_7\varepsilon,\label{le1eq2}\end{aligned}$$ for all ${\mbox{\boldmath$u$}}\in S({\mbox{\boldmath$y$}},\varepsilon)$. Combining (\[n7\]), and , we obtain $$\begin{aligned} & & \frac{1}{b}\int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\, \varphi({{\rm d}}{\mbox{\boldmath$u$}})\\ & & \ge\int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\, \sigma({{\rm d}}{\mbox{\boldmath$u$}})\ge \sigma(S({\mbox{\boldmath$y$}},\varepsilon)) c_7\varepsilon\ge c_8\varepsilon^d,\end{aligned}$$ which completes the proof of (b). Now suppose that $(\ref{n3})$ holds. From the estimate in the proof of (a) we get $$\begin{aligned} \frac{1}{bd} \int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}}) &\ge& \frac{1}{d} \int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,S_{d-1}(K,{{\rm d}}{\mbox{\boldmath$u$}})\\ & \ge&V_d(K)^{\frac{d-1}{d}}\left[V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})^{\frac{1}{d}}-V_d(K)^{\frac{1}{d}}\right]\\ &\ge& c_9\left[V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})-V_d(K)\right].\end{aligned}$$ \(c) For the proof of (\[5.1\]), let ${\mbox{\boldmath$y$}}\in{\rm bd}\,K(\varepsilon)$ and let $C$ be the cone with apex ${\mbox{\boldmath$y$}}$ spanned by $K$. Let ${\mbox{\boldmath$y$}}'$ be the point in $K$ nearest to ${\mbox{\boldmath$y$}}$. The vector ${\mbox{\boldmath$y$}}-{\mbox{\boldmath$y$}}'$ has length $\varepsilon$, and the hyperplane $H'$ orthogonal to it and passing through ${\mbox{\boldmath$y$}}'$ supports $K$. Let $H$ be the other supporting hyperplane of $K$ parallel to $H'$. Let $\Delta$ be the convex hull of ${\mbox{\boldmath$y$}}$ and $H\cap C$ and $\Delta'$ the convex hull of ${\mbox{\boldmath$y$}}$ and $H'\cap C$. Denoting by $D(K)$ the diameter of $K$ and assuming that $\varepsilon\le D(K)$, we have $$V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})-V_d(K) \ge V_d(\Delta') \ge \left(\frac{\varepsilon}{D(K)+\varepsilon}\right)^d V_d(\Delta) \ge \left(\frac{\varepsilon}{2D(K)}\right)^dV_d(K).$$ This gives (\[5.1\]). \(d) Suppose that a ball of radius $r>0$ slides freely inside $K$. Since $\mu(\cdot,\varphi,\varepsilon)$ is translation invariant, we can assume that $K$ contains the ball $B({\mbox{\boldmath$o$}},r)$ of radius $r$ centred at ${\mbox{\boldmath$o$}}$. Let $R>0$ be such that $K\subset B({\mbox{\boldmath$o$}},R)$. For $s>0$, the convex body $$K^s:=\{{\mbox{\boldmath$x$}}\in{\mathbb{R}}^d:V_d(K^{{\mbox{\boldmath$\scriptstyle x$}}})-V_d(K)\le s\}$$ is known as an illumination body of $K$ (cf. [@Werner1994 p. 258]; the convexity follows from Satz 4 in Fáry and Rédei [@FR50]). Now let ${\mbox{\boldmath$y$}}\in \text{bd}\, K(\varepsilon)$ and put $\nu:=V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})-V_d(K)$, then ${\mbox{\boldmath$y$}}\in\text{bd}\, K^\nu$. Let ${\mbox{\boldmath$x$}}\in\text{bd}\, K$ be determined by $\{{\mbox{\boldmath$x$}}\}=[{\mbox{\boldmath$o$}},{\mbox{\boldmath$y$}}]\cap\text{bd}\,K$, and denote by $N({\mbox{\boldmath$x$}})$ the unique exterior unit normal vector of $K$ at ${\mbox{\boldmath$x$}}$ (the normal vector is unique since by assumption there is a ball $B'$ of radius $r>0$ with ${\mbox{\boldmath$x$}}\in B'\subset K$). Since $B({\mbox{\boldmath$o$}},r)\subset K$, we have $$\langle {\mbox{\boldmath$x$}},N({\mbox{\boldmath$x$}})\rangle\ge r,\qquad \langle {\mbox{\boldmath$x$}}/\|{\mbox{\boldmath$x$}}\|,N({\mbox{\boldmath$x$}})\rangle\ge r/R.$$ From $\|{\mbox{\boldmath$y$}}\|-\|{\mbox{\boldmath$x$}}\| \ge\varepsilon$ we get $\|{\mbox{\boldmath$y$}}\|^d-\|{\mbox{\boldmath$x$}}\|^d \ge dr^{d-1}\varepsilon$. Therefore, Lemma 2 in [@Werner1994] yields $$\nu^{2/(d+1)} \ge c_{10} rr^{(d-1)/(d+1)}\left(\left(\frac{\|{\mbox{\boldmath$y$}}\|}{\|{\mbox{\boldmath$x$}}\|}\right)^d-1\right) \ge c_{11}R^{-d}\left(\|{\mbox{\boldmath$y$}}\|^d-\|{\mbox{\boldmath$x$}}\|^d\right) \ge c_{12}\varepsilon,$$ hence $$V_d(K^{{\mbox{\boldmath$\scriptstyle y$}}})-V_d(K)\ge c_{13}\,\varepsilon^{(d+1)/2},$$ which gives (\[5.2\]). \(e) Now suppose that $K$ is a polytope. Let ${\mbox{\boldmath$y$}}\in{\rm bd}\,K(\varepsilon)$ and let ${\mbox{\boldmath$y$}}'$ be the point in $K$ nearest to ${\mbox{\boldmath$y$}}$. Put ${\mbox{\boldmath$v$}}:=({\mbox{\boldmath$y$}}-{\mbox{\boldmath$y$}}')/\|{\mbox{\boldmath$y$}}-{\mbox{\boldmath$y$}}'\|$, and let $F$ denote the unique (proper) face of $K$ which contains ${\mbox{\boldmath$y$}}'$ in its relative interior. Let $F_1,\dots,F_m$ be the facets of $K$ that contain $F$, and let ${\mbox{\boldmath$u$}}_1,\dots,{\mbox{\boldmath$u$}}_m$ be their outer unit normal vectors. By [@Sch14 p. 85 and Theorem 2.4.9], we have $${\mbox{\boldmath$v$}}\in N(K,F)=N(K,{\mbox{\boldmath$y$}}')=\text{pos}\{{\mbox{\boldmath$u$}}_i:i=1,\ldots,m\},$$ where $N(K,F)$ and $N(K,{\mbox{\boldmath$y$}}')$ are the normal cones of $K$ at $F$ and ${\mbox{\boldmath$y$}}'$, respectively, and $\text{pos}$ denotes the positive hull. For any unit vector ${\mbox{\boldmath$w$}}\in N(K,F)$ there is some $i\in\{1,\ldots,m\}$ such that $\langle {\mbox{\boldmath$w$}}, {\mbox{\boldmath$u$}}_i\rangle>0$; in particular, $$a(F,{\mbox{\boldmath$w$}}):=\max\{\langle {\mbox{\boldmath$w$}},{\mbox{\boldmath$u$}}_i\rangle:i=1,\dots,m\}>0$$ and $a(F,{\mbox{\boldmath$v$}})=\langle {\mbox{\boldmath$v$}},{\mbox{\boldmath$u$}}_{i_0}\rangle>0$ for some $i_0\in\{1,\dots,m\}$. Since $N(K,F)\cap {\mathbb{S}^{d-1}}$ is compact, we have $$a(F):=\min\{a(F,{\mbox{\boldmath$w$}}):{\mbox{\boldmath$w$}}\in N(K,F)\cap {\mathbb{S}^{d-1}}\}>0$$ and thus $$c_{14}:=\min\{a(F):F\text{ is a proper face of }K\}>0.$$ Therefore, with $c_{15}:=\min\{V_{d-1}(F):F\text{ is a facet of }K\}>0$, where $V_{d-1}$ denotes the $(d-1)$-dimensional volume, we get $$\begin{aligned} \int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,S_{d-1}(K,{{\rm d}}{\mbox{\boldmath$u$}})\ge& \langle {\mbox{\boldmath$y$}}-{\mbox{\boldmath$y$}}', {\mbox{\boldmath$u$}}_{i_0}\rangle V_{d-1}(F_{i_0})\\ \ge& \|{\mbox{\boldmath$y$}}-{\mbox{\boldmath$y$}}'\|\cdot c_{14} c_{15}=c_{16}\varepsilon .\end{aligned}$$ This yields (\[5.3\]). [**Remark.**]{} Although in the case of a general convex body $K$, the derivation of the estimate (\[5.1\]) may seem rather crude, the order of $\varepsilon^d$ cannot be improved. In fact, if (\[5.1\]) would be replaced by $\mu(K,\varphi,\varepsilon) \ge c_2\varepsilon^\alpha$ with $1<\alpha< d$, then a counterexample would be provided by a body $K$ which in a neighbourhood of some boundary point is congruent to a suitable part of a body of revolution with meridian curve given by $\mu(t)= |t|^r$ with $1< r< \frac{d-1}{\alpha-1}$. [**Lemma 2.**]{} [*Let the convex body $K\in{\mathcal K}^d$ be such that a ball slides freely inside $K$. Assume further that $$\label{n9a} a\sigma\ge\varphi$$ with some positive constant $a$. Then $$\int_{{\mathbb{S}^{d-1}}} [h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})\le c_{17} \varepsilon^{(d+1)/2}$$ for $\varepsilon>0$ and ${\mbox{\boldmath$y$}}\in{\rm bd}\,K(\varepsilon)$.*]{} [*Proof.*]{} Let ${\mbox{\boldmath$y$}}\in {\rm bd}\,K(\varepsilon)$. From (\[n9a\]) we get $$\int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}}) \le c_{18}\int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\sigma({{\rm d}}{\mbox{\boldmath$u$}}).$$ Let ${\mbox{\boldmath$x$}}$ be the point in $K$ nearest to ${\mbox{\boldmath$y$}}$; then ${\mbox{\boldmath$y$}}={\mbox{\boldmath$x$}}+\varepsilon N({\mbox{\boldmath$x$}})$, where $N({\mbox{\boldmath$x$}})$ is the outer unit normal vector of $K$ at ${\mbox{\boldmath$x$}}$. By assumption, a ball, say of radius $r>0$, slides freely inside $K$. In particular, some ball $B$ of radius $r$ satisfies ${\mbox{\boldmath$x$}}\in B\subset K$. Let $${\rm Cap}\,({\mbox{\boldmath$y$}},\varepsilon):=\left\{{\mbox{\boldmath$u$}}\in{\mathbb{S}^{d-1}}:\langle {\mbox{\boldmath$u$}},N({\mbox{\boldmath$x$}}) \rangle\ge \frac{r}{r+\varepsilon}\right\}.$$ For ${\mbox{\boldmath$u$}}\in {\mathbb S}^{d-1}\setminus {\rm Cap}\,({\mbox{\boldmath$y$}},\varepsilon)$ we have $h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}}) -h(K,{\mbox{\boldmath$u$}})=0$. If $h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}}) -h(K,{\mbox{\boldmath$u$}})\not=0$, then $$h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}}) -h(K,{\mbox{\boldmath$u$}}) \le \langle {\mbox{\boldmath$y$}}-{\mbox{\boldmath$x$}},{\mbox{\boldmath$u$}}\rangle \le\varepsilon.$$ With $\alpha(\varepsilon):=\arccos r/(r+\varepsilon)$ this gives $$\begin{aligned} \int_{{\mathbb{S}^{d-1}}}[h(K^{{\mbox{\boldmath$\scriptstyle y$}}},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\sigma({{\rm d}}{\mbox{\boldmath$u$}}) &\le& \int_{{\rm Cap}\,({\mbox{\boldmath$\scriptstyle y$}},\varepsilon)}\varepsilon\,\sigma({{\rm d}}{\mbox{\boldmath$u$}})\\ &\le& c_{18} \varepsilon \sin^{d-1}\alpha(\varepsilon)= c_{18}\varepsilon\sqrt{1-(r/(r+\varepsilon))^2}^{\,d-1}\\ &\le& c_{19}\varepsilon^{(d+1)/2}.\end{aligned}$$ This yields the assertion. The following lemma is sufficient for our purpose; it does not aim at an optimal order. [**Lemma 3.**]{} [*Let $K\in{\mathcal K}^d$ be a convex body which slides freely in some ball. There are constants $c_{20}, c_{21} >0$ such that the following holds. For $0< \varepsilon< c_{20}$, let $m(\varepsilon)$ be the largest number $m$ such that there are $m$ points in ${\rm bd}\,K(\varepsilon)$ with the property that each segment connecting any two of them intersects the interior of $K$. Then* ]{} $$m(\varepsilon)\ge c_{21}\varepsilon^{-1/2}.$$ [*Proof.*]{} The convex body $K$ (which has interior points, by our general assumption) contains some ball, without loss of generality the ball $rB^d$. Let $R$ be such that $K$ slides freely in a ball of radius $R$. We put $c_{20}:= \min\{2R,(\pi r)^2/64R\}$ and assume that $0< \varepsilon< c_{20}$. For points ${\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}\in{\rm bd}\,K(\varepsilon)$, we assert that $$\label{2.10} \|{\mbox{\boldmath$x$}}-{\mbox{\boldmath$y$}}\|\ge 4\sqrt{R\varepsilon}\enspace \Rightarrow\enspace [{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}]\cap{\rm int}\,K\not=\emptyset.$$ For the proof, let ${\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}\in{\rm bd}\,K(\varepsilon)$ and suppose that $ [{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}]\cap{\rm int}\,K=\emptyset$. Let ${\mbox{\boldmath$p$}}\in K$ and ${\mbox{\boldmath$q$}}\in{\rm aff}\,\{{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}\}$ be points of smallest distance. If ${\mbox{\boldmath$p$}}\not={\mbox{\boldmath$q$}}$, then the hyperplane $H$ through ${\mbox{\boldmath$p$}}$ orthogonal to ${\mbox{\boldmath$q$}}-{\mbox{\boldmath$p$}}$ supports $K$. If ${\mbox{\boldmath$p$}}={\mbox{\boldmath$q$}}$, then the line through ${\mbox{\boldmath$x$}}$ and ${\mbox{\boldmath$y$}}$ touches $K$, and we choose $H$ as a supporting hyperplane of $K$ containing that line. The body $K$ slides freely in a ball, say $B$, of radius $R$, hence $K$ is a summand of $B$ ([@Sch14], Theorem 3.2.2). This means that there exists a compact convex set $M\subset{\mathbb R}^d$ such that $K+M=B$. Let ${\mbox{\boldmath$u$}}$ denote the outer unit normal vector of the supporting hyperplane $H$ of $K$ at ${\mbox{\boldmath$p$}}$, so that $h(K,{\mbox{\boldmath$u$}})=\langle {\mbox{\boldmath$p$}},{\mbox{\boldmath$u$}}\rangle$. There is a point ${\mbox{\boldmath$t$}}\in M$ with $h(M,{\mbox{\boldmath$u$}})=\langle {\mbox{\boldmath$t$}},{\mbox{\boldmath$u$}}\rangle$, and the point ${\mbox{\boldmath$z$}}:= {\mbox{\boldmath$p$}}+{\mbox{\boldmath$t$}}$ satisfies ${\mbox{\boldmath$z$}}\in B$ and $h(B,{\mbox{\boldmath$u$}})=\langle{\mbox{\boldmath$z$}},{\mbox{\boldmath$u$}}\rangle$. It follows that $K\subset B-{\mbox{\boldmath$t$}}$ and that $H$ is a supporting hyperplane of $B-{\mbox{\boldmath$t$}}$ at ${\mbox{\boldmath$p$}}$. The ball $(B-{\mbox{\boldmath$t$}})+\varepsilon B^d$ contains $K(\varepsilon)$ and hence the segment $[{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}]$. The line parallel to $[{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}]$ through ${\mbox{\boldmath$p$}}$ lies in $H$ and intersects the ball $(B-{\mbox{\boldmath$t$}})+\varepsilon B^d$ in a segment $S$, which is not shorter than $[{\mbox{\boldmath$x$}},{\mbox{\boldmath$y$}}]$. Thus, $\|{\mbox{\boldmath$x$}}-{\mbox{\boldmath$y$}}\|\le{\rm length}(S)= 2\sqrt{2R\varepsilon+\varepsilon^2} <4\sqrt{R\varepsilon}$, since $\varepsilon< 2R$. This proves (\[2.10\]). Let $m$ be the largest integer with $$m\le \frac{\pi r}{4\sqrt{R}}\,\varepsilon^{-1/2}.$$ Then $m\ge 2$ (by the choice of $c_{20}$), and there is a constant $c_{21}$ with $m\ge c_{21}/\sqrt{\varepsilon}$. Let $C$ be an arbitrary great circle of the ball $rB^d$. On $C$, we choose $m$ equidistant points ${\mbox{\boldmath$y$}}_1,\dots,{\mbox{\boldmath$y$}}_m$. For $i\not=j$ we have $\|{\mbox{\boldmath$y$}}_i-{\mbox{\boldmath$y$}}_j\|\ge 2r\sin(\pi/m)>r\pi/m$. Let ${\mbox{\boldmath$x$}}_i= \lambda_i{\mbox{\boldmath$y$}}_i\in{\rm bd}\,K(\varepsilon)$ with $\lambda_i>0$, then $\lambda_i>1$ for $i=1,\dots,m$ and hence $\|{\mbox{\boldmath$x$}}_i-{\mbox{\boldmath$x$}}_j\|>r\pi/m\ge 4\sqrt{R\varepsilon}$ for $i\not=j$. By (\[2.10\]), this completes the proof. Proof of Theorem 1 ================== We assume that $X$ and $K$ are as in Theorem 1 and satisfy the assumptions mentioned above, that is, $\varphi$ is not concentrated on a great subsphere, $\gamma\ge 1$, and the inclusion (\[n1\]) holds. Without loss of generality, we may assume that ${\mbox{\boldmath$o$}}\in{\rm int}\,K$. Recalling that $Z_0$ denotes the zero cell of $X$, we note that by the independence properties of the Poisson process we have $${{\rm P}}\{Z_K\not\subset K(\varepsilon)\}= {{\rm P}}\left\{Z_0\not\subset K(\varepsilon)\mid K\subset Z_0\right\}.$$ The conditional probability involving the zero cell is slightly more convenient to handle. For a compact convex set $L\subset{\mathbb{R}}^d$ we define $${\cal H}_L:=\{H\in {\mathcal H}^d:H\cap L\not=\emptyset\}$$ and $$\Phi(L):= \Theta({\cal H}_L).$$ By (\[2.0\]) we have $$\label{n2} \Phi(L) = 2\gamma\int_{{\mathbb{S}^{d-1}}} h(L,{\mbox{\boldmath$u$}})\,\varphi({{\rm d}}{\mbox{\boldmath$u$}}).$$ The following two lemmas use ideas from the proofs of Lemmas 3 and 5 in [@HS07], but the present situation is simpler. As there, we use the abbreviation $$H_1^-\cap\dots\cap H_n^-=:P(H_{(n)}),$$ where $H_1,\dots,H_n$ are hyperplanes not passing through ${\mbox{\boldmath$o$}}$ and $H_i^-$ is the closed halfspace bounded by $H_i$ that contains ${\mbox{\boldmath$o$}}$. Let $\|{\mbox{\boldmath$x$}}\|_K= \min\{\lambda\ge 0: {\mbox{\boldmath$x$}}\in\lambda K\}$ for ${\mbox{\boldmath$x$}}\in {\mathbb R}^d$. For a nonempty compact convex set $L$, we define $\|L\|_K:=\max\{\|{\mbox{\boldmath$x$}}\|_K:{\mbox{\boldmath$x$}}\in L\}$. For $\varepsilon\ge 0$ and $m\in {\mathbb N}$, let $${\cal K}^d_\varepsilon(m):=\{L\in{\cal K}^d: K\subset L \not\subset K(\varepsilon),\,\|L\|_K\in (m,m+1]\}$$ and $$q_\varepsilon(m):= {{\rm P}}\{Z_0\in{\cal K}^d_\varepsilon(m)\}.$$ We abbreviate $$(m+1)K=:K_m.$$ We have $$\label{1} q_\varepsilon(m) = \sum_{N=d+1}^\infty {{\rm P}}\{X({\cal H}_{K_m})=N\}p(N,m,\varepsilon)$$ with $$\begin{aligned} p(N,m,\varepsilon) &:=& {{\rm P}}\{Z_0\in{\cal K}^d_\varepsilon(m)\mid X({\cal H}_{K_m})=N\}\\ &=& \Phi(K_m)^{-N} \int_{{\cal H}_{K_m}^N}{\bf 1}\{P(H_{(N)})\in{\cal K}^d_\varepsilon(m)\}\, \Theta^N({{\rm d}}(H_1,\dots,H_N)),\end{aligned}$$ the latter by a well-known property of Poisson processes (e.g., [@SW08 Th. 3.2.2(b)]), and $$\label{2} {{\rm P}}\{X({\cal H}_{K_m})=N\} = \frac{\Phi(K_m)^{N}}{N!} \exp\left[-\Phi(K_m)\right].$$ [**Lemma 4.**]{} [*There exists a number $m_0$, depending only on $K$, $\varphi$ and $d$, such that $$q_0(m) \le c_{22}\exp[-\Phi(K)-c_{23}\gamma m]$$ for $m\ge m_0$.* ]{} [*Proof.*]{} We modify and adapt the proof of Lemma 3 in [@HS07]. If $H_1,\dots,H_N\in {\cal H}_{K_m}$ and if $P:= P(H_{(N)})\in {\cal K}^d_0(m)$, then $P$ has a vertex ${\mbox{\boldmath$v$}}$ with $m<\|{\mbox{\boldmath$v$}}\|_K\le m+1$. Since ${\mbox{\boldmath$v$}}$ is the intersection of some $d$ facets of $P$, there exists a $d$-element set $J\subset \{1,\dots,N\}$ with $$\{{\mbox{\boldmath$v$}}\}= \bigcap_{j\in J}H_j.$$ We denote the segment $[{\mbox{\boldmath$o$}},{\mbox{\boldmath$v$}}]$ by $S=S(H_i,\, i\in J)$ (where it is assumed that the hyperplanes $H_i$, $i\in J$, have linearly independent normal vectors) and note that $$H_{i} \cap {\rm relint}\, S = \emptyset \qquad\mbox{for}\qquad i=1,\dots,N.$$ For any segment $S=[{\mbox{\boldmath$o$}},{\mbox{\boldmath$v$}}]$ with $\|{\mbox{\boldmath$v$}}\|_K\ge m$ we have (writing $a^+:=\max\{a,0\}$) $$\Phi(S) = 2\gamma \int_{{\mathbb S}^{d-1}} \langle {\mbox{\boldmath$v$}},{\mbox{\boldmath$u$}}\rangle^+ \varphi({{\rm d}}{\mbox{\boldmath$u$}}) \ge 2c_{24}\gamma m$$ with a positive constant $c_{24}$. This follows from the fact that the function $${\mbox{\boldmath$v$}}_1\mapsto \int_{{\mathbb S}^{d-1}} \langle {\mbox{\boldmath$v$}}_1, {\mbox{\boldmath$u$}}\rangle^+ \varphi({{\rm d}}{\mbox{\boldmath$u$}}),\qquad {\mbox{\boldmath$v$}}_1\in{\mathbb S}^{d-1},$$ is positive (since $\varphi$ is not concentrated on a great subsphere) and continuous. Let $m_0$ be the smallest integer $\ge (2/c_{24})\int_{{\mathbb S}^{d-1}}h(K,{\mbox{\boldmath$u$}})\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})$. For $m\ge m_0$ we then have $$\Phi(S) \ge \Phi(K) +c_{24}\gamma m,$$ and hence $$\int_{{\cal H}_{K_m}} {\bf 1} \{H \cap S = \emptyset\} \,\Theta({{\rm d}}H) = \Phi(K_m) - \Phi(S)\le \Phi(K_m) - \Phi(K) -c_{24}\gamma m,$$ where we used that $S\subset K_m$, since $\|{\mbox{\boldmath$v$}}\|_K\le m+1$. Now we obtain $$\begin{aligned} p(N,m,\varepsilon) & \le & \binom{N}{d}\Phi(K_m)^{-N} \int_{{\cal H}_{K_m}^{d}} {\bf 1} \left\{ \|S(H_j,\,j\in\{1,\dots,d\})\|_K \ge m\right\} \\ & & \int_{{\cal H}_{K_m}^{N-d}} {\bf 1} \left\{H_i \cap S(H_j,\,j\in\{1,\dots,d\})= \emptyset \mbox{ for } i=d+1,\dots,N\right\} \\ & & \times \;\Theta^{N-d} ({{\rm d}}(H_{d+1},\dots,H_{N})) \,\Theta^{d}({{\rm d}}(H_{1},\dots,H_{d})) \\ & \le & \binom{N}{d} \Phi(K_m)^{-N} \int_{{\cal H}_{K_m}^{d}} [\Phi(K_m) - \Phi(K)- c_{24}\gamma m ]^{N-d}\, \Theta^{d}({{\rm d}}(H_{1},\dots,H_{d})) \\ & = & \binom{N}{d} \Phi(K_m)^{d-N} \left[\Phi(K_m) - \Phi(K)- c_{24}\gamma m\right]^{N-d}.\end{aligned}$$ With (\[1\]) (for $\varepsilon=0$) and (\[2\]) this gives $$\begin{aligned} & & q_0(m)\\ & & \le \sum_{N=d+1}^{\infty} \frac{\Phi(K_m)^{N}}{N!} \exp\left[-\Phi(K_m)\right] \binom{N}{d} \Phi(K_m)^{d-N} \left[\Phi(K_m) - \Phi(K)-c_{24}\gamma m\right]^{N-d} \\ & & = \frac{1}{d!} \Phi(K_m)^{d} \exp[-\Phi(K_m)] \sum_{N=d+1}^{\infty} \frac{1}{(N-d)!} \left[\Phi(K_m)- \Phi(K)- c_{24}\gamma m \right]^{N-d} \\ & & \le \frac{1}{d!} \Phi(K_m)^{d} \exp\left[- \Phi(K)-c_{24}\gamma m\right] \\ & & = \frac{1}{d!} \left(2\gamma (m+1) \int_{{\mathbb{S}^{d-1}}} h(K,{\mbox{\boldmath$u$}})\,\varphi({{\rm d}}{\mbox{\boldmath$u$}}) \right)^d \exp\left[- \Phi(K)-c_{24}\gamma m\right]\\ & & \le c_{22} \exp\left[- \Phi(K)-c_{23} \gamma m\right]\end{aligned}$$ with $c_{23}=c_{24}/2$, say. [**Lemma 5.**]{} [*Let $0<\varepsilon\le 1$. Then, for $m\in\mathbb{N}$,*]{} $$q_\varepsilon(m) \le c_{25}(\gamma m)^d\exp\left[-\Phi(K)- 2\gamma \mu(K,\varphi,\varepsilon)\right].$$ [*Proof.*]{} With $H_1,\dots,H_N \in {\cal H}_{K_m}$ and $P = P(H_{(N)})\in {\cal K}^d_0(m)$ as in the previous proof, the polytope $P$ has a vertex ${\mbox{\boldmath$x$}}\in K_m\setminus K(\varepsilon)$. This vertex is the intersection of $d$ facets of $P$. Hence, there exists an index set $J \subset \{1,\dots,N\}$ with $d$ elements such that $$\{{\mbox{\boldmath$x$}}\} = \bigcap_{j \in J} H_j.$$ There exists a point ${\mbox{\boldmath$y$}}\in {\rm bd}\,K(\varepsilon)$ such that $$\Phi({\rm conv}(K\cup\{{\mbox{\boldmath$x$}}\})) \ge \Phi({\rm conv}(K\cup\{{\mbox{\boldmath$y$}}\})) = \Phi(K^{{\mbox{\boldmath$\scriptstyle y$}}}) \ge \Phi(K) + 2\gamma \mu(K,\varphi,\varepsilon),$$ where the last inequality follows from (\[n2\]) and (\[n4\]), together with the monotonicity of $\Phi$. This gives $$\begin{aligned} \int_{{\cal H}_{K_m}} {\bf 1} \{H\cap{\rm conv}(K\cup\{{\mbox{\boldmath$x$}}\})=\emptyset\}\,\Theta({{\rm d}}H) &=& \Phi(K_m)- \Phi({\rm conv}(K\cup\{{\mbox{\boldmath$x$}}\}))\\ &\le& \Phi(K_m)- \Phi(K) - 2\gamma \mu(K,\varphi,\varepsilon).\end{aligned}$$ We write ${\mbox{\boldmath$x$}}={\mbox{\boldmath$x$}}(H_1,\dots,H_d)$ for the intersection point of the hyperplanes $H_1,\dots,H_d$ (supposed in general position) and obtain $$\begin{aligned} p(N,m,\varepsilon) &\le & \binom{N}{d} \Phi(K_m)^{-N} \int_{{\cal H}_{K_m}^d} {\bf 1} \{ {\mbox{\boldmath$x$}}(H_1,\dots,H_d) \in K_m\setminus K(\varepsilon)\}\\ & & \int_{{\cal H}_{K_m}^{N-d}} {\bf 1}\{ H_i \cap {\rm conv}(K \cup \{{\mbox{\boldmath$x$}}(H_1,\dots,H_d)\}) = \emptyset \mbox{ for } i=d+1,\dots,N \}\\ & & \times \; \Theta^{N-d}({{\rm d}}(H_{d+1},\dots,H_N))\, \Theta^d({{\rm d}}(H_1\dots,H_d))\\ & \le & \binom{N}{d}\Phi(K_m)^{d-N}\left[\Phi(K_m) - \Phi(K)- 2\gamma \mu(K,\varphi,\varepsilon) \right]^{N-d}.\end{aligned}$$ Similarly as in the proof of Lemma 4, summation over $N$ gives $$\begin{aligned} & & q_\varepsilon(m)\\ & & \le \sum_{N=d+1}^{\infty} \frac{\Phi(K_m)^{N}}{N!} \exp\left[-\Phi(K_m)\right] \binom{N}{d} \Phi(K_m)^{d-N} \left[\Phi(K_m) - \Phi(K)-2\gamma \mu(K,\varphi,\varepsilon)\right]^{N-d} \\ & & \le \frac{1}{d!}\Phi(K_m)^d \exp \left[-\Phi(K)- 2\gamma \mu(K,\varphi,\varepsilon)\right]\\ & & \le c_{25} (\gamma m)^d\exp \left[-\Phi(K)- 2\gamma \mu(K,\varphi,\varepsilon)\right].\end{aligned}$$ [*Proof of Theorem $1$.*]{} We have $$\begin{aligned} {{\rm P}}\left\{\delta(K,Z_K)>\varepsilon\right\} &=& {{\rm P}}\left\{ Z_0 \not\subset K(\varepsilon) \mid K \subset Z_0\right\}\\ &=& \frac{{{\rm P}}\left\{K \subset Z_0,\; Z_0 \not\subset K(\varepsilon) \right\}}{{{\rm P}}\left\{K \subset Z_0\right\}}= \frac{\sum_{m=1}^\infty q_{\varepsilon}(m)}{\exp\left[-\Phi(K)\right]}.\end{aligned}$$ To estimate the last numerator, we choose $m_0$ according to Lemma 4 and use Lemma 5 for $m\le m_0$ and Lemma 4 together with $q_{\varepsilon}(m) \le q_0(m)$ for $m>m_0$. By the assumptions of Theorem 1, relation $(\ref{n1})$ is satisfied. We obtain $${{\rm P}}\left\{ Z_0 \not\subset K(\varepsilon) \mid K \subset Z_0\right\} \le \sum_{m=1}^{m_0} c_{25}(\gamma m)^d \exp[-2\gamma \mu(K,\varphi,\varepsilon)] + \sum_{m>m_0} c_{22}\exp[-c_{23}\gamma m].$$ The first sum can be estimated by $$\begin{aligned} \label{4.1} & & \sum_{m=1}^{m_0} c_{25}(\gamma m)^d \exp[-2\gamma \mu(K,\varphi,\varepsilon)] \\ & & \le c_{25} m_0^{d+1} \gamma^d \exp\left[ -\gamma \mu(K,\varphi,\varepsilon)\right]\exp\left[ -\gamma \mu(K,\varphi,\varepsilon)\right]\nonumber\\ & & \le c_{26}(\varepsilon)\exp\left[-\gamma \mu(K,\varphi,\varepsilon)\right],\nonumber\end{aligned}$$ since $\mu(K,\varphi,\varepsilon)>0$ by condition (\[n1\]) and Lemma 1. The second sum can be estimated by $$\sum_{m>m_0} c_{22}\exp[-c_{23}\gamma m] \le c_{22}\exp[-c_{23}\gamma]\sum_{m>m_0} \exp[-c_{23}(m-1)] \le c_{27}\exp\left[-c_{23}\gamma\right],$$ where we have used that $\gamma\ge 1$ (by assumption) and that the last sum converges. Both estimates together yield (\[5.0\]). Proofs of Theorems 2 and 3 ========================== Under the assumptions (\[n7\]) or (\[n3\]), we can conclude from Lemma 1 that $\mu(K,\varphi,\varepsilon) \ge c_{28} \varepsilon^\alpha$ with suitable $\alpha\le d$. Therefore, in estimating (\[4.1\]) we can use that $$\gamma^d \exp\left[ -\gamma \mu(K,\varphi,\varepsilon)\right] \le \gamma^d \exp\left( -\gamma c_{28}\varepsilon^\alpha\right)\le c_{29}\varepsilon^{-d\alpha}.$$ This gives $$\sum_{m=1}^{m_0} c_{25}(\gamma m)^d \exp[-2\gamma \mu(K,\varphi,\varepsilon)] \le c_{30}\varepsilon^{-d\alpha}\exp\left(-c_{31}\gamma \varepsilon^\alpha\right).$$ The estimation of the second sum above remains unchanged. Hence, under the assumptions of Theorem 2 and with $\gamma=n$, we can conclude that $${{\rm P}}\left\{\delta(K,Z_K^{(n)})>\varepsilon\right\} \le c_{32} \varepsilon^{-d\alpha} \exp \left(-c_{33}n \varepsilon^\alpha \right).$$ We choose $$C>\frac{d+1}{c_{33}}$$ and put $$\varepsilon_n:= \left(\frac{C\log n}{n}\right)^{1/\alpha}.$$ Then $$\begin{aligned} \label{n10} \sum_{n=1}^\infty {{\rm P}}\left\{\delta(K,Z_K^{(n)})>\varepsilon_n\right\} &\le& \sum_{n=1}^\infty c_{32}\left(\frac{n}{C\log n} \right)^d \exp\left(-c_{33} C\log n\right) \nonumber\\ &=& c_{34}\sum_{n=1}^\infty (\log n)^{-d} n^{d-c_{33}C} <\infty.\end{aligned}$$ The Borel–Cantelli lemma gives $${{\rm P}}\left\{\delta(K,Z_K^{(n)})>\varepsilon_n\mbox{ for infinitely many }n\right\}=0,$$ hence $${{\rm P}}\left\{\delta(K,Z_K^{(n)})\le\left(\frac{C\log n}{n}\right)^{1/\alpha} \mbox{ for sufficiently large }n\right\}=1.$$ This completes the proof of Theorem 2. [*Proof of Theorem $3$.*]{} Since $K$ slides freely in some ball, say of radius $R$, there is a convex body $L$ with $K+L=RB^d$ ([@Sch14 Theorem 3.2.2]). From the polynomial expansion of $S_{d-1}(K+L,\cdot)$ ([@Sch14 (5.18)]) it follows that $S_{d-1}(K,\cdot) \le S_{d-1}(RB^d,\cdot)=R^{d-1}\sigma$. Together with the assumption (\[n9\]) this shows that (\[n3\]) is satisfied. Therefore, Theorem 2 yields that $$\label{5.0y} \delta(K,Z_K^{(n)})={\rm O}\left(\left(\frac{\log n}{n}\right)^{\frac{2}{d+1}}\right)\quad\mbox{almost surely,}$$ as $n\to\infty$. Let $0<\varepsilon < c_{20}$ (with $c_{20}$ as in Lemma 3). According to Lemma 3, we can choose $$m=m(\varepsilon) \ge c_{21}\varepsilon^{-1/2}$$ points ${\mbox{\boldmath$x$}}_1,\dots,{\mbox{\boldmath$x$}}_m\in{\rm bd}\,K(\varepsilon)$ such that the segment joining any two of them intersects the interior of $K$. Let $n\in{\mathbb N}$. Suppose that $\delta(K,Z_K^{(n)})<\varepsilon$. Then each point ${\mbox{\boldmath$x$}}_i$ is strictly separated from $K$ by some hyperplane from $X_n$. Let ${\mathcal A}_i\subset {\mathcal H}^d$ be the set of hyperplanes strictly separating ${\mbox{\boldmath$x$}}_i$ and $K$. By the choice of the points ${\mbox{\boldmath$x$}}_1,\dots,{\mbox{\boldmath$x$}}_m$, the sets ${\mathcal A}_1,\dots,{\mathcal A}_m$ are pairwise disjoint. Since $X_n$ is a Poisson process, the processes $X_n{\,\rule{.1mm}{.26cm}\rule{.24cm}{.1mm}\,}{\mathcal A}_1,\dots,X_n{\,\rule{.1mm}{.26cm}\rule{.24cm}{.1mm}\,}{\mathcal A}_m$ are stochastically independent (e.g., [@SW08 Theorem 3.2.2]). It follows that $$\begin{aligned} {{\rm P}}\{ \delta(K,Z_K^{(n)}) < \varepsilon\} &\le& {{\rm P}}\{ X_n({\mathcal A}_i)\ge 1\mbox{ for }i=1,\dots,m\}\\ &=& \prod_{i=1}^m {{\rm P}}\{ X_n({\mathcal A}_i)\ge 1\}= \prod_{i=1}^m \left[ 1-{{\rm P}}\{ X_n({\mathcal A}_i) =0\}\right]\\ &=& \prod_{i=1}^m \left( 1-\exp[-\Theta_n({\mathcal A}_i)] \right),\end{aligned}$$ where $\Theta_n$ is the intensity measure of $X_n$. Since the assumptions on $K$ in Lemma 2 are satisfied, we can conclude that $$\begin{aligned} \Theta_n({\mathcal A}_i) &=& \Theta_n({\mathcal H}_{K^{{\mbox{\boldmath$\scriptstyle x$}}_i}}) - \Theta_n({\mathcal H}_K)\\ &=& 2n\int_{{\mathbb S}^{d-1}} [h(K^{{\mbox{\boldmath$\scriptstyle x$}}_i},{\mbox{\boldmath$u$}})-h(K,{\mbox{\boldmath$u$}})]\,\varphi({{\rm d}}{\mbox{\boldmath$u$}})\\ &\le& 2nc_{17}\varepsilon^{(d+1)/2}.\end{aligned}$$ This gives $${{\rm P}}\left\{ \delta(K,Z_K^{(n)}) < \varepsilon\right \} \le \left[1-\exp\left(-2c_{17}n \varepsilon^{(d+1)/2}\right) \right]^{m(\varepsilon)}.$$ Now we choose $$\varepsilon_n^{(d+1)/2} = \frac{c\log n}{n}$$ with $$0<c<\frac{1}{4c_{17}(d+1)}.$$ Then $${{\rm P}}\left\{ \delta(K,Z_K^{(n)}) < \varepsilon_n\right \} \le \left(1- n^{-2c_{17}c} \right)^{m(\varepsilon_n)}$$ with $$m(\varepsilon_n) \ge c_{21}\varepsilon_n^{-1/2} = c_{21}\left(\frac{n}{c\log n}\right)^{1/(d+1)}> c_{35} n^{1/(2d+2)}$$ for sufficiently large $n$. With $p:= 2c_{17}c$ and $q:= 1/(2d+2)$ we have $q>p$ and $$\left(1-n^{-2c_{17}c}\right)^{m(\varepsilon_n)} <\left(1-\frac{1}{n^p}\right)^{c_{35}n^q} =\left[\left(1-\frac{1}{n^p}\right)^{n^p\cdot n^{q-p}}\right]^{c_{35}} \le ({\rm e}^{-c_{35}})^{n^{q-p}}.$$ It follows that $$\sum_{n=1}^\infty {{\rm P}}\left\{\delta(K,Z_K^{(n)}) <\left(\frac{c\log n}{n}\right)^{\frac{2}{d+1}}\right\} <\infty.$$ From the Borel–Cantelli lemma we conclude that $${{\rm P}}\left\{\delta(K,Z_K^{(n)}) <\left(\frac{c\log n}{n}\right)^{\frac{2}{d+1}}\mbox{ for infinitely many }n\right\} =0$$ and hence $${{\rm P}}\left\{\delta(K,Z_K^{(n)}) \ge \left( \frac{c\log n}{n}\right)^{\frac{2}{d+1}} \mbox{ for almost all }n \right\} =1.$$ Together with (\[5.0y\]), this completes the proof of Theorem 3. Finally, we construct a directional distribution exhibiting the property (\[C1\]) for a given convex body $K$. We do that at this stage, since arguments appearing in the previous proofs are employed. As explained before (\[C1\]), we assume that a decreasing sequence $(\varepsilon_n)_{n\in{\mathbb N}}$ with $\lim_{n\to\infty}\varepsilon_n=0$ is given. For $n\in{\mathbb N}$, let $X_n$ be a stationary Poisson hyperplane process with intensity $n$ and directional distribution $\varphi$, to be constructed. The $d$-dimensional convex body $K$ contains some ball touching the boundary, hence there are a number $r>0$ and a point ${\mbox{\boldmath$x$}}\in{\rm bd}\,K$ such that ${\mbox{\boldmath$x$}}$ is contained in a ball of radius $r$ that is contained in $K$. Let $N({\mbox{\boldmath$x$}})$ be the unique outer unit normal vector of $K$ at ${\mbox{\boldmath$x$}}$ and let ${\mbox{\boldmath$y$}}={\mbox{\boldmath$x$}}+\varepsilon N({\mbox{\boldmath$x$}})$. Let $n\in{\mathbb N}$, and suppose that $\delta(K,Z_K^{(n)})<\varepsilon_n$. Then the point ${\mbox{\boldmath$y$}}$ is strictly separated from $K$ by some hyperplane of $X_n$. Similarly as in the proof of Theorem 3, this yields $${{\rm P}}\left\{\delta(K,Z_K^{(n)})<\varepsilon_n\right\}\le 1-\exp[-2n\varepsilon_n\varphi(S_n)]$$ with $$S_n:=\left\{{\mbox{\boldmath$u$}}\in{\mathbb{S}^{d-1}}:\langle {\mbox{\boldmath$u$}},N({\mbox{\boldmath$x$}}) \rangle\ge \frac{r}{r+\varepsilon_n}\right\}.$$ It is easy to construct an even positive measurable function $g$ on ${\mathbb{S}^{d-1}}$ such that the measure $\varphi$ defined by ${{\rm d}}\varphi=g \,{{\rm d}}\sigma$ is a probability measure and that $$2n\varepsilon_n\varphi(S_n) < |\log(1-n^{-2})|$$ for all $n\in{\mathbb N}$ (for example, $g$ can be a suitable constant on $S_n\setminus S_{n+1}$). The directional distribution $\varphi$ then satisfies $$1-\exp[-2n\varepsilon_n\varphi(S_n)]<\frac{1}{n^2}$$ and hence $$\sum_{n=1}^\infty {{\rm P}}\left\{\delta(K,Z_K^{(n)})<\varepsilon_n\right\}<\infty.$$ As in the proof of Theorem 3, this yields (\[C1\]). [99]{} <span style="font-variant:small-caps;">Bárány, I.,</span> Intrinsic volumes and $f$-vectors of random polytopes. *Math. Ann.* [**285**]{} (1989), 671–699. <span style="font-variant:small-caps;">Böröczky, K. J. and Schneider, R.,</span> The mean width of circumscribed random polytopes. *Canad. Math. Bull.* **53** (2010), 614–628. <span style="font-variant:small-caps;">Dümbgen, L. and Walther, G.,</span> Rates of convergence for random approximations of convex sets. *Adv. Appl. Prob. (SGSA)* [**28**]{} (1996), 384–393. <span style="font-variant:small-caps;">Fáry, I. and Rédei, L.,</span> Der zentralsymmetrische Kern und die zentralsymmetrische Hülle von konvexen Körpern. *Math. Ann.* [**122**]{} (1950), 205–220. <span style="font-variant:small-caps;">Hug, D. and Schneider, R.,</span> Asymptotic shapes of large cells in random tessellations. *Geom. Funct. Anal.* **17** (2007), 156–191. <span style="font-variant:small-caps;">Reitzner, M.,</span> Random polytopes. In [*New Perspectives in Stochastic Geometry*]{} (eds. W. S. Kendall, I. Molchanov), pp. 45–76, Oxford University Press 2010. <span style="font-variant:small-caps;">Schneider, R.,</span> *Convex Bodies – The Brunn–Minkowski Theory.* Second ed., Cambridge University Press, Cambridge 2014. <span style="font-variant:small-caps;">Schneider, R. and Weil, W.,</span> *Stochastic and Integral Geometry.* Springer, Berlin 2008. <span style="font-variant:small-caps;">Rényi, A. and Sulanke, R.,</span> Über die konvexe Hülle von $n$ zufällig gewählten Punkten. *Z. Wahrscheinlichkeitsth. verw. Geb.* **2** (1963), 75–84. <span style="font-variant:small-caps;">Rényi, A. and Sulanke, R.,</span> Über die konvexe Hülle von $n$ zufällig gewählten Punkten. II. *Z. Wahrscheinlichkeitsth. verw. Geb.* **3** (1964), 138–147. <span style="font-variant:small-caps;">Rényi, A. and Sulanke, R.,</span> Zufällige konvexe Polygone in einem Ringgebiet. *Z. Wahrscheinlichkeitsth. verw. Geb.* **9** (1968), 146–157. <span style="font-variant:small-caps;">Werner, E.,</span> Illumination bodies and affine surface area. *Studia Math.* **110** (1994), 257–269. Authors’ addresses:\ Daniel Hug\ Karlsruhe Institute of Technology, Department of Mathematics\ D-76128 Karlsruhe, Germany\ E-mail: [email protected]\ Rolf Schneider\ Mathematisches Institut, Albert-Ludwigs-Universit[ä]{}t Freiburg\ D-79104 Freiburg i. Br., Germany\ E-mail: [email protected]

ArXiv

--- abstract: 'Rapid variability in the radio flux density of the BL Lac object PKSB1144$-$379 has been observed at four frequencies, ranging from 1.5 to 15 GHz, with the VLA and the University of Tasmania’s Ceduna antenna. Intrinsic and line of sight effects were examined as possible causes of this variability, with interstellar scintillation best explaining the frequency dependence of the variability timescales and modulation indices. This scintillation is consistent with a compact source 20–40$~{\mu}$as, or 0.15–0.3 pc in size. The inferred brightness temperature for PKSB1144$-$379 (assuming that the observed variations are due to scintillation) is $6.2 \times 10^{12}$ K at 4.9 GHz, with approximately 10 percent of the total flux in the scintillating component. We show that scintillation surveys aimed at identifying variability timescales of days to weeks are an effective way to identify the AGN with the highest brightness temperatures.' author: - 'R. J. Turner, S. P. Ellingsen, S. S. Shabala, J. Blanchard, J. E. J. Lovell, and J. N. McCallum' - 'G. Cimò' title: 'BL LAC PKSB1144$-$379 an extreme scintillator' --- Introduction ============ The BL Lac object PKSB1144$-$379 [@Nicolson; @et; @al.+1979] ($11^{\rm h} 47^{\rm m} 01^{\rm s} .4$, $-38^{\circ} 12^{\prime} 11^{\prime\prime}$ in J2000 coordinates), located at $z=1.048$ [@Stickel; @et; @al.+1989] is known for both its long and short-term flux density variability at centimetre wavelengths. Some of the first observations of PKSB1144$-$379 showed variability at a frequency of 5 GHz. The flux density increased from 0.9 Jy in 1970 December to 1.6 Jy in 1971 February and again to 2.22 Jy in 1971 September [@Shimmins; @Bolton+1972; @Bolton; @Shimmins+1973; @Gardner; @et; @al.+1975]. Between May and August 1994, the flux density of PKSB1144$-$379 at 4.8 GHz dropped 17%, and 9% at 8.6 GHz [@Kedziora-Chudczer; @et; @al.+2001b]. Long-term (months to years) variability is an indicator of intrinsic variations in the source. For example, VLBI imaging of PKSB1144$-$379 shows changes in the extended jet-like components on milliarcsecond scales over these timescales, in addition to variations in the flux density of the quasar core [@Ojha; @et; @al.+2010]. This potentially degrades the utility of this object as an international celestial reference frame (ICRF) defining source [@Ma; @et; @al.+2009]. Intraday variability in PKSB1144$-$379 was first identified by @Kedziora-Chudczer [@et; @al.+2001a] and on this basis it was included in the long term 6.7 GHz monitoring which commenced 2003 April 3 using the University of Tasmania’s 30 m Ceduna antenna as part of the Continuous Single-dish Monitoring of Intraday variability at Ceduna (COSMIC) project [@McCulloch; @et; @al.+2005; @Carter; @et; @al.+2009]. Observations between 2003 and 2010 measured the mean 6.7 GHz flux density to range from 0.8 to 3.2 Jy, with the largest amplitude variations occurring on timescales of hundreds of days (the data from this longer term monitoring will be presented as part of a future publication). Quasars are known to exhibit short timescale variability and in particular, intraday variability [@Wagner; @Witzel+1995]. Sources PKSB0405$-$385, PKSB1257$-$326, and J1819+3845 [@Kedziora-Chudczer; @et; @al.+1997; @Bignall; @et; @al.+2006; @Dennett-Thorpe; @de; @Bruyn+2000] have been observed to vary on timescales of less than 2 hours at centimeter wavelengths. These variations can either be due to intrinsic or line of sight effects. The assumption of rapid intrinsic variability typically requires source brightness temperatures well in excess of the inverse Compton limit ($\sim 10^{12}$ K), or unrealistically high Doppler beaming factors [see for example @Kedziora-Chudczer; @et; @al.+1997]. Extrinsic effects include jet precession [e.g. @Camenzind; @Krockenberger+1992], gravitational microlensing [e.g. @Paczynski+1986] and interstellar scintillation (ISS) [e.g. @Walker+1998]. Of these, ISS best fits the observed frequency dependence, annual cycles and time-delay for the variations observed in objects such as PKSB0405$-$385, PKSB1257$-$326, and J1819+3845. Limited studies of the short timescale variability of PKSB1144$-$379 exist. In 1999 May, @Kedziora-Chudczer [@et; @al.+2001a] observed PKSB1144$-$379 at four frequencies over a period of six days. Intra day variability (IDV) in the flux density on timescales of 3 to 4 days was observed and @Kedziora-Chudczer [@et; @al.+2001a] suggested interstellar scintillation as the likely cause. The MASIV survey of 443 compact radio sources demonstrated that where variations on timescales of days are observed in AGN at centimetre wavelengths, scintillation is the predominant cause [@Lovell; @et; @al.+2003; @Lovell; @et; @al.+2008]. Interstellar scintillation is the interference phenomenon seen when radiation from a distant source passes through the turbulent, ionised interstellar medium (ISM) of our Galaxy. Interference arises due to small variations in the medium’s refractive index for different paths through the ISM. As this interstellar material moves with respect to the observer, intensity variations will be observed as a function of time. Scintillation behaviour is characterised as either strong or weak scattering. In weak scattering only small phase changes are introduced by the ISM over the first Fresnel zone. In strong scattering the wavefront is rapidly varying on scales smaller than the first Fresnel zone. Two types of variability are expected in the strong scattering regime, slow broadband variability due to refractive scattering and fast narrowband variability due to diffractive scattering. Unlike jet precession or microlensing, the measured ISS variability timescales are expected to be frequency dependent [@Walker+1998]. Observations of PKSB1144$-$379 during 2004 with the Ceduna antenna at 6.7 GHz showed both the flux density and the amplitude of variability to be decreasing. However, lower amplitude quasi-periodic variations with a period in excess of a week persisted throughout this “quiet” phase and the purpose of the observations presented here is to use multi-frequency radio data to determine the mechanism for these variations. In this paper, we report 15 days of observations of PKSB1144$-$379 at frequencies of 1.5, 4.9 and 15 GHz using the Very Large Array (VLA), and 6.7 GHz observations using the University of Tasmania’s Ceduna antenna. Observations and data reduction {#sec:obs} =============================== The source PKSB1144$-$379 was observed over a 15 day period between 2005 January 10 and 2005 January 24 using the VLA. Observations were made at three frequencies centered at 1.5, 4.9 and 15 GHz, each with a 50 MHz bandwidth. Primary flux density calibration was obtained through observations of 1331+305 (3C286) for each of the frequency bands. Daily observations were made for each frequency for a duration of approximately 3 minutes for 1331+305 and approximately 10 minutes for PKSB1144$-$379. These observations were conducted during the reconfiguration of the VLA (from the A array to the B array), and consequently on some days not all antennas were available (typically no more than 2 or 3 antennas were missing from the array). The data collected using the VLA was processed using the Astronomical Image Processing System (AIPS)[^1]. Standard VLA calibration and data reduction procedures were used for each frequency and each day, with only minor differences (e.g. changes to the averaging time for self-calibration). Antennas or time ranges for which the data were clearly erroneous were removed prior to calibration. The data for PKSB1144$-$379 were calibrated against 1331+305 for which the assumed flux densities were 1.16, 0.87 and 0.54 Jy at the frequencies 1.5, 4.9 and 15 GHz respectively [@Baars; @et; @al.+1977]. The data for PKSB1144$-$379 were then self-calibrated using an initial model of a point source of peak intensity 1 Jy. The resulting image was cleaned with clean boxes placed around PKSB1144$-$379 and other visible sources within the primary beam. This image was self-calibrated using only phases, then cleaned, with this procedure repeated a further time correcting both the amplitudes and phases. The flux densities of all sources were measured from the final image using the AIPS task SAD. Typical noise levels in the final images were of the order of 1 mJy, and this is taken as representative of the uncertainty in our flux density measurements. The University of Tasmania’s Ceduna antenna in South Australia was also used to observe PKSB1144$-$379 over a longer period as part of the COSMIC project [@McCulloch; @et; @al.+2005], with observations commencing 2003 April 3 and still being undertaken at the time of writing. Observations of PKSB1144$-$379 were made on just over half the days in this time range in contiguous blocks, typically ranging from 5 to 15 days. The observing frequency band has a center frequency of 6.7 GHz and a 500 MHz bandwidth for each of two orthogonal circular polarizations. Around 50 independent flux density measurements of PKSB1144$-$379 are made each day as part of COSMIC [see @Carter; @et; @al.+2009]. To improve the signal to noise ratio here we present the daily mean flux density measured with the Ceduna antenna. Data from the Ceduna antenna were only available for 11 of the 15 days for which observations were made with the VLA. Results ======= The light curves of PKSB1144$-$379 observed with both the VLA and the Ceduna antenna are plotted in Figure  \[fig:lightcurve\]. To assess the variability timescale we used the approach of @Lovell [@et; @al.+2008] to produce structure functions for each frequency. The characteristic timescale is defined to be the time for the structure function to reach half the value of its maximum [@Lovell; @et; @al.+2008]. At both 4.9 and 6.7 GHz, a characteristic timescale of between 1 and 1.5 days was estimated. At 15 GHz there is evidence of two timescales of approximately 0.4 and 2 days. A complete cycle was not observed at 1.5 GHz in the 15 day sampling interval, so conservatively, the characteristic timescale is greater than 4.8 days. The peak-to-peak variability timescales evident in Figure \[fig:lightcurve\] are $2\pi$ times greater than these characteristic timescales. Due to the low number of data points, we sought to validate these timescale estimates by crudely fitting sinusoids to the data points. The estimated peak-to-peak timescale for the 4.9 GHz frequency is 7.7 days, and for the 6.7 GHz frequency it is 6.6 days. Ordinary least-squares regression analysis[^2] yields $R^2$ values of 0.76 and 0.83, respectively, for these fits. For the 15 GHz frequency, least-squares analysis shows a timescale of 8.6 days ($R^2=0.57$), and a weaker fit ($R^2=0.26$) at 1.7 days. In Figure \[fig:timescales\] we plot the resultant characteristic timescales. These estimates are consistent with the characteristic timescales estimated from the structure functions. Peak-to-peak timescales of approximately 2 days are below the Nyquist frequency with the daily sampling regime. Observations of PKSB1144$-$379 taken in 1999 by @Kedziora-Chudczer [@et; @al.+2001a] clearly show a timescale at 8.6 GHz 3 – 4 times shorter than at 4.8 GHz. However, the source is likely to have undergone significant intrinsic evolution since then, as evidenced by the 4.8 GHz flux falling from 2Jy mean in May 1999 to 1Jy in January 2005. ![The 1.5, 4.9 and 15 GHz light curves (orange, black and blue respectively) of PKSB1144$-$379 observed with the VLA and the 6.7 GHz light curve (red) observed with the University of Tasmania Ceduna antenna over the period MJD 53375 – 53400 (2005 January 5–30).[]{data-label="fig:lightcurve"}](lightcurves_5e.eps){width="8cm"} The Discrete Correlation Function [@Edelson; @Krolik+1988] was used to determine if the light curves of any of the observed frequencies are correlated. The correlation between the 4.9 and 6.7 GHz observations is greatest ($\sim 1$) for a time lag of between 0.5 and 1 days (4.9 GHz lagging 6.7 GHz). None of the other frequency pairs showed a correlation coefficient above 0.6. The modulation index, $m$, was calculated as $m=\sigma/\langle S \rangle$ with $\sigma$ the standard deviation of the flux density and $\langle S \rangle$ the mean flux density. The modulation indices calculated from the observations for the 1.5, 4.9, 6.7 and 15 GHz frequencies are 0.014, 0.081, 0.080 and 0.059 respectively. These are plotted in Figure \[fig:modind\]. Since only a fraction of a cycle is sampled over the 15 day observation period at 1.5 GHz, the derived modulation index is a lower limit. Comparisons with the MASIV sample [@Lovell; @et; @al.+2008] show that the high 4.9 GHz modulation index of PKSB1144-379 places it in the top 1–2 % (in terms of fractional variability) of bright ($>1$ Jy) quasars. ![The modulation index as a function of frequency. The uncertainty in the modulation index $m$ is estimated as $\sigma_{\rm m}=0.9\,m(t_{\rm char}/{\rm 15\,days})^{1/2}$, where 15 days is the observing interval [@Stinebring; @et; @al.+2000]. Also plotted is a one-parameter ($f_{\rm c}$) theoretical fit for the expected scintillation behaviour. The rising part of each curve corresponds to point-like scintillation. The falling part of each curve corresponds to quenched scintillation. The coloured lines represent models where source size exceeds the refractive scale at 4.9 (blue) and 6.7 (red) GHz. Solid lines are for models where source size $\propto \nu^{-1}$. Dashed lines are models where source size is independent of frequency. Unquenched scintillation at all frequencies (dotted line) clearly cannot explain the data.[]{data-label="fig:modind"}](mod_freq_28e.eps){width="8cm"} At 1.5 and 4.9 GHz, additional sources are visible within the primary beam of the VLA for the observations of PKSB1144$-$379. At both these frequencies, a source was observed with declination and right ascension offset from PKSB1144$-$379 by 107 and -99 arcseconds on the sky respectively, while a second source was observed at only 1.5 GHz offset in right ascension and declination by -201 and 189 arcseconds on the sky respectively. These sources were found to have a constant flux density (within the measurement uncertainty) over the 15 day period of our observations. This, combined with the close agreement between the pattern of the Ceduna 6.7 and VLA 4.9 GHz variations demonstrates that the observed variability has not been produced by errors in the calibration or data reduction process. Figure 2 of @Carter [@et; @al.+2009] shows that there are systematic variations in the Ceduna flux densities of around 0.15 Jy on timescales less than a day. However, the close agreement between the pattern of variations at 4.9 GHz from the VLA and 6.7 GHz from Ceduna demonstrates that the daily average flux densities produced from the Ceduna 30 m (which have been used here) are very reliable. ![Variability timescales as a function of frequency. Only lower limits are available at 1.5 GHz; these are shown numerically at the top of each curve. 15 GHz variability suggests two contributing timescales. The smaller formal uncertainty at 6.7 GHz is due a longer time series for Ceduna compared with the VLA data. The break in model curves corresponds to a frequency above which source size exceeds the refractive scale. Colours and line styles are as in Figure \[fig:modind\].[]{data-label="fig:timescales"}](time_freq_22e.eps){width="8cm"} Discussion ========== The observed variability timescales of approximately 7 days can potentially be explained by either line of sight effects or intrinsic mechanisms. The former includes interstellar scintillation, jet precession or microlensing, and the latter rapid intrinsic variability. Brightness temperature due to intrinsic variability --------------------------------------------------- If the observed variability in the source is intrinsic, then using causality to limit the linear scale we can infer the brightness temperature of the source. Using Planck’s law we can determine the brightness temperature $T_{\rm B}$ (in K) from the variation in flux density $\triangle S$ (in Jy), the frequency $\nu$ (in GHz), the peak-to-peak variability timescale $\tau$ (in days) and the angular diameter distance $D_{\rm A}$ (in Gpc) $$T_{\rm B}=5.9 \times 10^{15}\left[\frac{\triangle S}{\mathrm{1\:Jy}}\right]\left[\frac{\nu}{\mathrm{GHz}}\right]^{-2}\left[\frac{\tau}{\mathrm{100\:days}}\right]^{-2}\left[\frac{D_{\rm A}}{\mathrm{1\:Gpc}}\right]^2\ \mbox{K} \label{brightness}$$ For our 4.9 GHz VLA observations, the variation in flux density is $\triangle S \sim \pm 0.14$ Jy and the variability timescale is $\tau \sim 7.7$ days. PKSB1144$-$379 has a redshift of $z=1.048$ and an angular diameter distance of $D_{\rm A} = 1.677$ Gpc (using standard $\Lambda$CDM with $H_0$ = 71kms$^{-1}$Mpc$^{-1}$ ; $\Omega_m = 0.27$ ; $\Omega_\Lambda = 0.73$). Thus the implied brightness temperature is $T_{\rm B}=1.7 \times 10^{16}$ K. When mapped into the rest frame of the source the brightness temperature is reduced by a factor of $\mathcal{D}^3(1+z)^{-3}$, under the hypothesis of intrinsic variation. At high brightness temperatures, $T_{\rm B} \geqslant 10^{12}$ K, energy losses of radiating electrons due to inverse Compton scattering become extremely large, resulting in a rapid cooling of the system and thereby bringing the brightness temperature quickly below this value [@Kellermann; @Pauliny-Toth+1969; @Readhead+1994]. So assuming a source brightness temperature of $T_{\rm B} \leqslant 10^{12}$ K, a Doppler factor of $\mathcal{D} > 53$ is required. This is a large Doppler factor. Although it is in principle possible for the observed variations to be intrinsic, as we show in § \[sec:scintill\], the frequency dependence does not favor an intrinsic interpretation. Precession & microlensing ------------------------- Quasi-periodic variation in the flux density can be caused by jet precession close to the line of sight. Using the model of @Camenzind [@Krockenberger+1992] we can replicate the characteristics of light curves observed at 4.9 GHz for PKSB1144$-$379, namely the 7.7 day period and 0.081 modulation index, with realistic parameters for beaming angle, black hole mass and jet opening angle. However, just like intrinsic variability, precession is expected to produce light curves whose period is independent of frequency. This is clearly in disagreement with our data (Figure \[fig:lightcurve\]). Thus, jet precession does not provide a viable explanation for the observed variability. Microlensing is ruled out on two counts, since it is expected to be both achromatic and not quasi-periodic. Scintillation {#sec:scintill} ------------- The Galactic coordinates of PKSB1144$-$379 are 289.24 degrees longitude and 22.95 degrees latitude. For this line of sight, the transition frequency estimated by the @Cordes [@Lazio+2001] model of the Galactic electron distribution is $\nu_0 \sim 14.4$ GHz. Thus, all our observing frequencies are expected to be in the refractive scattering regime ($\nu < \nu_0$), with the exception of 15GHz which is close to the transition frequency. We can use our multi-frequency timescales and modulation indices to infer physical characteristics of the source. The refractive angular scale is given by $$\theta_{\rm r} = \theta_{\rm r0} (\nu/\nu_0)^{-11/5} \label{eqn:refScale}$$ where $\theta_{\rm r0} = 2.3$ $\mu$as at the Galactic coordinates of interest [@Walker+2001]. For a point source, the frequency dependence of the scintillation timescale is thus $t_{\rm char} \propto \nu^{-11/5}$. If the source size evolves with frequency slower than $\nu^{-11/5}$, it is possible that at some $\nu > \nu_{\rm crit}$ the source size exceeds the refractive scale. At that point, the scintillation is quenched, and the timescale will take on the same frequency dependence as source size. There are a number of possible models for source size evolution with frequency. Here we consider the limiting cases $\theta_{\rm s} \propto \nu^{-1}$ (corresponding to a self-absorbed synchrotron source), and $\theta_{\rm s}=$const. While in principle the flux density of the compact component (and, hence, the scintillating fraction of total flux) may change with frequency, we refrain from introducing this extra, highly uncertain free parameter into our models. Comparison of the 4.9 and 6.7 GHz timescales (Figure \[fig:timescales\]) places an upper limit on source size. These characteristic timescales are [@Walker+1998] $$t_{\rm char} = % \begin{cases} t_{\rm r0}(\nu_0/\nu)^{11/5} & \text{if } \theta_{\rm s} < \theta_{\rm r}\\ t_{\rm r0}(\theta_{\rm s}/\theta_{\rm r0}) = \theta_{\rm s} \left( D_{\rm screen} / V_{\rm screen} \right) & \text{if } \theta_{\rm s} \geq \theta_{\rm r} . \end{cases} \label{time}$$ where $t_{\rm r0}$ is the light crossing time of the refractive scale at the transition frequency $\nu_0=14.4$ GHz; and $D_{\rm screen}$ and $V_{\rm screen}$ are the distance and transverse velocity of the screen causing the scintillation. From Equation \[time\], if the source still appears point-like at 6.7 GHz (i.e. if $\theta_{\rm s}(6.7)<\theta_{\rm r}(6.7)$), we expect the 4.9 GHz scintillation timescale to be longer than the 6.7 GHz timescale by a factor of $(6.7/4.9)^{11/5}=2.0$. This does not appear to be consistent with the data (Figure \[fig:lightcurve\]). Formal uncertainties on the characteristic timescales are given by $\sigma_{\rm t\,char}=2.2\,t_{\rm char}(t_{\rm char}/{\rm 15\,days})^{1/2}$, where 15 days is the observing interval [@Stinebring; @et; @al.+2000]. This yields quite large formal uncertainties of $0.7$ and $0.6$ days for the 4.9 and 6.7 GHz timescales, respectively. This is largely due to not sampling enough scintles over the 15-day observing period. Using the available 6.7GHz Ceduna data for an additional 70 days, the same $t_{\rm char}$ is recovered, but with a much smaller uncertainty of 0.2 days. Furthermore, as reported in Section \[sec:obs\] we find a strong correlation between the 6.7 and 4.9 GHz light curves over the 15-day observing period. This makes our timescales highly inconsistent with a $\nu^{-11/5}$ evolution over the range 4.9 – 6.7 GHz. Thus, $\theta_{\rm s}(6.7) \geq \theta_{\rm r}(6.7)$. The frequency evolution of the modulation index (Figure \[fig:modind\]) provides a further constraint. This is given by [@Walker+1998] $$m_{\rm p} = % \begin{cases} f_{\rm c} (\nu/\nu_0)^{17/30} & \text{if } \theta_{\rm s} < \theta_{\rm r}\\ f_{\rm c} (\nu/\nu_0)^{-2} \left( \theta_{\rm s} / \theta_{\rm r0} \right)^{-7/6} & \text{if } \theta_{\rm s} \geq \theta_{\rm r} . \end{cases} \label{eqn:modIndex}$$ where $f_{\rm c}$ is the fraction of total flux that scintillates; this may in general be a function of frequency. The observed modulation index (Figure \[fig:modind\]) decreases between 6.7 and 15 GHz. This is clearly inconsistent with a $\nu^{17/30}$ evolution expected for a point source, and provides further confirmation that we begin to see source structure rather than the refractive scale at some $\nu_{\rm crit}<6.7$GHz. By fitting a series ($\nu_{\rm crit}=$ 4.9, 6.7, 20GHz) of simple one-parameter curves (for scintillating fraction $f_{\rm c}$), we find that models in which source size is independent of frequency consistently underpredict the modulation index at 15 GHz. The $\theta_{\rm s} \propto \nu^{-1}$ model provides a better fit to the data, requiring source size to exceed the refractive scale at some $\nu_{\rm crit} > 4.9$ GHz. A final consistency check comes from our lower limit on the 1.5GHz timescale, $t_{\rm char}>4.8$ days. Fitting a transition from $\nu^{-11/5}$ to $\nu^{-1}$ dependence at $\nu=4.9$GHz through our timescale data yields a 1.5 GHz timescale of 20 days, well in excess of the expected lower limit of 4.8 days. Knowledge of the transition frequency at which scintillation becomes quenched allows us to estimate the size of the scintillating core. Equation \[eqn:refScale\] gives the refractive scales at 4.9 and 6.7 GHz as 25 and 12$\mu$as respectively. The discussion above suggests the scintillation becomes quenched in this frequency range, and we therefore estimate the source size (Gaussian FWHM) as between 20 and 41$\mu$as at these frequencies[^3]. This is approximately a factor of 10 – 100 smaller than the implied angular size for rapidly scintillating sources such as PKSB1257$-$326 and J1819+3845 [@Bignall; @et; @al.+2006; @Dennett-Thorpe; @de; @Bruyn+2003]. Interestingly, in 1999 @Kedziora-Chudczer [@et; @al.+2001a] found that the scintillation timescale changed by a factor of 3 – 4 between 8.6 and 4.8GHz, exactly as expected for a point source (Equation \[time\]). This implies a source angular size less than $12$$\mu$as, and suggests significant source evolution between the May 1999 epoch and our observations in January 2005. The long-term COSMIC data [@Carter; @et; @al.+2009] indeed show a significant monotonic rise in flux density for the source, beginning some 5 – 6 months before our epoch of observations. This is suggestive of a new outburst, with a $\mu$as jet component expected to move away from the compact core over time. At a redshift of $1.048$, the angular diameter distance is 1.65 Gpc, and a source size of 20$\mu$as corresponds to a physical size of 0.15 pc. Expansion at $0.6c$ is required for such a source evolution. Thus, it is a plausible scenario. The distance to the phase screen is $$D_{\rm screen}=\frac{c}{2\pi{\theta_{\rm r0}}^2 \nu_0} \label{dist}$$ Using the @Cordes [@Lazio+2001] model, we get $D_{\rm screen}=0.86$kpc. This is a factor of 100 more distant than the screen distances typically inferred from time delay and annual cycle fitting for the rapidly scintillating sources. For unquenched scintillation, the transverse speed of the phase screen relative to the line of sight to the source is related to this distance, the angular scale of the source $\theta_{\rm s}$ and the corresponding characteristic timescale $t_{\rm char}$ by $$V_{\rm screen}= D_{\rm screen}(\theta_{\rm s}/t_{\rm char})= D_{\rm screen}(\theta_{\rm r0}/ t_{\rm r0}) \label{speed}$$ As discussed above, we observe unquenched scintillation at 4.9GHz ($\theta_{\rm s}(4.9)=25$$\mu$as) with a timescale $t_{\rm char}(4.9)=1.2$ days. This yields a transverse speed of the phase screen is $V_{\rm screen}=30$ kms$^{-1}$. As a final sanity check, we also know that the scintillation is quenched at $6.7$ GHz, and so the screen velocity must exceed $D_{\rm screen} \theta_{\rm s}(6.7)/t_{\rm char}(6.7)=18$ kms$^{-1}$. We can also use the source size inferred from scintillation to estimate the brightness temperature of the most compact emission. Using the approach outlined in @Walker+1998 we find that at a frequency of 4.9 GHz the implied angular size of the scintillating component (41$\mu$as) corresponds to a brightness temperature of $6.2 \times 10^{12}$ K, a factor of 2500 less than that implied by an intrinsic interpretation of the variability. At a redshift of 1.048 a modest [@Hovatta; @et; @al.+2009] Doppler factor of 3.8 is required to reduce the brightness temperature in the source frame to beneath the inverse Compton limit. It is also worth noting that the inferred angular size of the region containing the varying flux density if we assume the variations are intrinsic is $<$ 1 $\mu$as, and hence we would expect to see scintillation from this source in addition to any intrinsic changes. Conclusions =========== Scintillation has been conclusively identified as the cause of short timescale variability at radio frequencies in a number of sources. Some of these (e.g. PKSB1257$-$326 and J1819+3845) show variability with characteristic timescales as short as a few hours, but sources with longer timescales such as PKSB1519$-$273 and PKSB1622$-$253 [e.g. @Carter; @et; @al.+2009] are much more common [@Lovell; @et; @al.+2008]. PKSB1144$-$379 is a member of a small class of extreme scintillators which are strong ($>$ 1 Jy), have high modulation index ($m > 0.05$) and vary on timescales longer than a day. Such sources can be monitored with single-dish radio telescopes [as has been demonstrated by @Carter; @et; @al.+2009]. That short-timescale scintillators are much rarer than longer-timescale ones implies that suitable nearby screens (within a few 10’s of pc) are relatively rare, since if such a screen is present the requirements for a background source to scintillate should be met by a large fraction of AGN. We thus expect a large number of more distant screens (the Cordes & Lazio (2001) model predicts screen distances of 0.5 – 3 kpc), but these will produced high modulation index scintillation (i.e. not heavily quenched) only for the most compact AGN. So when seeking the most compact AGN either to test physical models, or as targets for space Very Long Baseline Interferometry missions, those sources with high modulation indices and the longest scintillation timescales would appear to be the best targets. The 1.2 day characteristic timescale (corresponding to a peak-to-peak period of 7.7 days) and high modulation index observed in PKSB1144$-$379 make it one of the most extreme bright scintillators identified to date. However, there are likely to be other sources with still longer timescales yet to be identified. The high-cadence, long timescale monitoring of AGN being undertaken by the COSMIC project is ideally suited to identifying such sources, and other rare but brief phenomena such as extreme-scattering events [e.g. @Senkbeil; @et; @al.+2008]. Our detection of a distant ($\sim 0.86$ kpc, according to the Cordes & Lazio model) screen was only possible due to our 15 day monitoring campaign. It is entirely possible that such screens are ubiquitous, and simply not picked up in the short term (2 to 3 days) intensive, or longer term low-cadence observing campaigns which are more commonly undertaken. We thank the anonymous referee for a number of thoughtful comments that helped improve the paper significantly. RT is grateful to the University of Tasmania for a Dean’s Summer Research Scholarship. SS and JM thank the ARC for Super Science Fellowships. The National Radio Astronomy Observatory is a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc. This research has made use of the NASA/IPAC Extragalactic Database (NED) which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. Baars, J. W. M., Genzel, R., Pauliny-Toth, I. I. K., & Witzel, A. 1977, A&A, 61, 99 Bignall, H. E., Macquart, J.-P., Jauncey, D. L., et al. 2006, ApJ, 652, 1050 Bolton, J. G., & Shimmins, A. J. 1973, Aust J Phys Astrophys Suppl, No. 30 Camenzind, M., & Krockenberger, M. 1992, A&A, 255, 59 Carter, S. J. B., Ellingsen, S. P., Macquart, J.-P., & Lovell, J. E. J. 2009, MNRAS, 396, 1222 Cordes, J. M., & Lazio, T. J. W. 2001, ApJ, 549, 997 Dennett-Thorpe, J., & de Bruyn, A. G. 2000, ApJ, 529, 65 Dennett-Thorpe, J., & de Bruyn, A. G. 2003, A&A, 404, 113 Edelson, R. A., & Krolik, J. H. 1988, ApJ, 333, 646 Gardner, F. F., Whiteoak, J. B., & Morris, D. 1975, Aust J Phys Astrophys Suppl, No. 35 Hovatta, T., Valtaoja, E., Tornikoski, M. & L[ä]{}hteenm[ä]{}ki, A. 2009, A&A 494, 527 Kedziora-Chudczer, L. L., Jauncey, D. L., Wieringa, M. H., Walker, M. A., Nicolson, G. D., Reynolds, J. E., Tzioumis, A. K. 1997, ApJ, 490, L9 Kedziora-Chudczer, L. L., Jauncey, D. L., Wieringa, M. H., Tzioumis, A. K., & Bignall H. E. 2001, Ap&SS, 278, 113 Kedziora-Chudczer, L. L., Jauncey, D. L., Wieringa, M. H., Tzioumis, A. K., & Reynolds J. E. 2001, MNRAS, 325, 1411 Kellermann, K. I., & Pauliny-Toth, I. I. K. 1969, ApJ, 155, L71 Lovell, J. E. J., Jauncey, D.L., Bignall, H.E., Kedziora-Chudczer, L., Macquart, J.-P., Rickett, B. J. & Tzioumis, A. K., 2003, AJ, 126, 1699 Lovell, J. E. J., Rickett, B. J., Macquart, J.-P., et al. 2008, ApJ, 689, 108 Ma, C., et al. 2009, in IERS Technical Note. No. 35, The Second Realization of the International Celestial Reference Frame by Very Long Baseline Interferometry, ed. A.L. Fey, D. Gordon, C.S. Jacobs (Frankfurt am Main, Germany: IERS) McCulloch, P. M., Ellingsen, S. P., Jauncey, D. L., et al. 2005, AJ, 129, 2034 Narayan, R., 1992, Phil. Trans. Roy. Soc., 341, 151 Nicolson, G. D., Glass, I. S., Feast, M. W., & Andrews, P. J. 1979, MNRAS, 189, 29 Ojha, R., Kadler, M., Böck, M., et al. 2010, A&A, 519, 45 Paczyński, B. 1986, ApJ, 301, 503 Readhead, A.C.S. 1994, ApJ, 426, 51 Rickett, B. J., Quirrenbach, A., Wegner, R., Krichbaum, T. P., & Witzel, A., 1995, A&A, 293, 479 Rickett, B. J., Lazio, T. J. W. & Ghigo, F. D., 2006, ApJS, 165, 439 Senkbeil, C.E., Ellingsen, S.P., Lovell, J.E.J., Macquart, J.-P., Cimò, G., Jauncey, D.L., 2008, ApJ, 672, L95 Shimmins, A. J., & Bolton, J. G., 1972, Aust J Phys Astrophys Suppl, No. 23 Stickel, M., Fried, J. W., & Kühr, H., 1989, A&AS, 80, 103 Stinebring, D. R., Smirnova, T. V., Hankins, T. H., et al. 2000, ApJ, 539, 300 Wagner, S. J., & Witzel, A., 1995, ARA&A, 33, 163 Walker, M. A., 1998, MNRAS, 294, 307 Walker, M. A., 2001, MNRAS, 321, 176 [^1]: http://www.aips.nrao.edu/index.shtml [^2]: In ordinary least-squares analysis, $R^2$ value represents the proportion of variability in the data that can be accounted for by a given model. Values of $R^2$ close to one thus represent a good fit, in our case suggesting that the observed light curves can be approximated by sinusoidal functions. [^3]: A factor of $2 \sqrt{\ln 2}$ between refractive scale and source size comes about because observationally the source size is typically defined in terms of the Full Width at Half Maximum (FWHM) of a Gaussian intensity profile fitted to the source; while the angular scales (e.g. $\theta_{\rm r}$, $\theta_{\rm s}$ etc) used in the scintillation literature are the standard deviation of the fitted Gaussian [@Narayan+1992].

ArXiv

HU-EP 00/40,\ hep-th/0010086 [**Rank 48 gauge group in heterotic string**]{} $^\dag$\ \ \ [*D-10115 Berlin, Germany*]{}\ \ We discuss the existence of a non-perturbative gauge sector that can raise the rank of the gauge group of the ${\cal N}_4=2$ heterotic string up to 48. These gauge bosons, that don’t exist in six dimensions, co-exist with those originating from small instantons shrinking to zero size. ------------------------------------------------------------------------ width 6.7cm $^\dag$  Research supported by the “Marie Curie” fellowship HPMF-CT-1999-00396.\ \ $^1$e-mail: [email protected] {#section .unnumbered} It is well known that, under certain conditions, the gauge group of the heterotic string, whose maximal rank is fixed, in the perturbative construction, to be sixteen, can be non-perturbatively enhanced. This has been shown to happen, for instance, in ${\cal N}_6=1$ compactifications of the $SO(32)$ string, when small instantons shrink to zero size [@w] [^1]. This extension of the heterotic gauge group, whose rank can be raised to 32, can be explicitly observed in the type I dual construction [@gp], where it appears perturbatively. Indeed, through the map to F-theory, configurations have been analyzed, for which the gauge group can receive a bigger enhancement [@am], but here we will consider only situations that we can study through string-string duality, without passing through higher dimensional theories. Intrinsically, the existence of such non-perturbative states is related to the nature of the space on which the string is compactified, so that it exists for any value of the coupling constant. Indeed, from a geometrical point of view, in the case of the heterotic/type I dual pair mentioned above, the spaces on which the heterotic and the dual type I string are compactified are the same. It is natural therefore to ask whether in certain cases a non-perturbative phenomenon such as that at work on the small instantons of the heterotic string, can provide a further enhancement of the type I gauge group, that would eventually raised to a group of maximal rank $3 \times 16$. By duality, such an extension should exist also on the heterotic side. The aim of this note is to discuss this issue, and to provide evidence for the existence of such a further non-perturbative enhancement, in compactifications to four dimensions. As we will see, the further compactification of the ${\cal N}_6=1$ theory is essential for the appearance of new gauge bosons, whose coupling does not depend on a volume but rather on a complex structure modulus. We will consider the problem both from the heterotic and the type I point of view. Finally, we will discuss also the type IIA dual point of view, commenting on a possible higher dimensional (M-theory) interpretation. Our analysis ultimately provides a step toward the investigation of type II/heterotic duality for ${\cal N}_4=2$ compactifications of the type IIA string which are not realized on K3 fibrations. 0.3cm {#section-1 .unnumbered} We start by reviewing some facts about compactifications of the heterotic string, obtained by toroidal compactifications on $T^2$ of the ${\cal N}_6=1$ theory in six dimensions. This theory has ${\cal N}_4=2$ supersymmetry. The perturbative corrections to the couplings of the $F_{\mu \nu} F^{\mu \nu}$ or $R^2$ terms of the effective action can be shown to have a general form of the type: S + (T) + (U) + … , where $T$ and $U$ are the moduli associated to the Kähler class and complex structure of the two-torus respectively [@dkl]–[@greek]. The $\Im S$-term is the tree-level contribution, while the functions $\Delta(T)$, $\Delta(U)$ appear at the one loop [^2]. When a type I dual orbifold exists, it can be shown [@ap] that the correction to the $F^2$ terms does not depend on $T$, which on the other hand is mapped, under duality, into the modulus $S^{\prime}$. This field parameterizes the tree level effective coupling of the D5-branes sector [@s], namely the sector dual to the heterotic small instantons [@w]. The non-dependence of the effective coupling of the heterotic $F^2$ terms on this modulus is then a necessary requirement in order for the duality to work, because on the type I side there is no perturbative mixing of the two moduli in the effective gauge couplings. On the other hand, the analysis of this duality tells us that the heterotic field $T$, entering generically into the corrections of various terms of the effective action, is not just a geometric modulus, but indeed the coupling of a non-perturbative sector. For generic heterotic ${\cal N}_4=2$ compactifications, the dependence of the corrections on this modulus can be therefore interpreted as the signal of the running of states charged under both the perturbative and non-perturbative sectors. We may now ask how should we interpret the dependence of the corrections on the other modulus, $U$. Can this be seen in some way as the coupling of another, non-perturbative sector? And if yes, what is this sector? For sure, if $U$ has to be interpreted as the modulus parameterizing the gauge coupling of another sector, this latter cannot exist in six dimensions: whatever could in fact be the coupling of this sector in six dimensions, toroidal compactification would then give him a dependence on the volume of the torus, namely on the field $T$, and not on $U$. It must therefore necessarily be a sector that appears only *after compactification* from six to lower dimensions. By duality, this must be true, whenever it exists, also on the type I dual of the heterotic construction. Indeed, the $F^2$ and $R^2$ corrections of the type I ${\cal N}_4=2$ effective action depend on this modulus [@ap; @ms], and we could ask the same question, namely what is the interpretation we must give to the modulus $U$, also in the type I framework. If we indicate respectively by $g_{(a)}$ and $g_{(b)}$ the couplings of the small instantons and of this new, unknown, sector, we would have: & \~& T ;\ [1 g\^2\_[(b)]{}]{} & \~& U . When the heterotic torus $T^2$ is described by a product of two circles, of radii $R_1$ and $R_2$, we would have: & \~& R\_1 R\_2 ;\ [1 g\^2\_[(b)]{}]{} & \~& R\_1 / R\_2 . The two couplings are therefore exchanged under T-duality along the second circle. If we start from the $SO(32)$ heterotic string, we can put on the second circle a Wilson line that breaks $SO(32)$ to $SO(16) \times SO(16)$, chosen in order to act as W in Ref. [@w]. T-duality along this circle exchanges then the “$SO(32)$” with the “$E_8 \times E_8$” theory, making clear that, while the modulus $T$ has to be interpreted as the coupling of the small instantons of the $SO(32)$ theory, $U$ should be interpreted as the coupling of the small instantons of the $E_8 \times E_8$ theory. Our claim is that indeed, in four dimensions, both these sectors are *present at the same time*. What makes this possible is the fact that, unlike in field theory, in the toroidally compactified heterotic string there is T-duality, with the presence, at the same time, of momentum and winding states. The ordinary small instantons correspond to non-perturbative objects of the “field theory” part of the heterotic string, i.e. the one built on the Kaluza–Klein, momentum states, while the other non-perturbative states are pure stringy non-perturbative states. Indeed, as discussed in Ref. [@gh; @sw], zero-size $E_8 \times E_8$ instantons don’t give rise to vector multiplets in six dimensions, but rather to tensor multiplets. This fits with our interpretation of the field $U$: only after compactification to lower dimensions the tensor multiplets give rise to vector multiplets. If we decompactify the four dimensional theory to six dimensions $(V_{(2)} \equiv R_1 R_2 \to \infty )$, we indeed observe the disappearance of this gauge sector from the effective action: & & [ T V\_[(2)]{}]{} = 1 ;\ [1 g\^2\_[(b)]{}]{} & & [ U V\_[(2)]{}]{} 0 . \[4-6\] On the type I side, this is related to the fact that, as explained in Refs. [@6danom], the fixed points of the $T^4 \big/ Z_2 \sim K3$ compact space of the six dimensional ${\cal N}_6=1$ theory are associated to a gauge bundle without vector structure, so that the non-perturbative states arising from small type I instantons in six dimensions contain tensor multiplets instead of vector multiplets. The appearance of new massless vector multiplets after circle compactification from six dimensions was interpreted, in Ref. [@gh], from the heterotic point of view, as due to the appearance of tensionless strings. Here we want to stress that the appearance of these states in the lower dimensional theory is not related to the actual existence of a perturbative (subgroup of the) $E_8 \times E_8$ gauge group: they are indeed present also in the $SO(32)$ theory. The simultaneous presence of both the small instantons of $SO(32)$ and $E_8 \times E_8$ is due to the T-duality of the heterotic string, and matches with the simultaneous presence of these states on the type I dual theory, in which however only one of these two sectors is non-perturbative. Here however we are faced with a puzzle. If we go to six dimensions, by decompactifying the two-torus of the heterotic string, and reinterpret the vector multiplets of four dimensions in terms of vector and tensor multiplets of six dimensions, we see that the number of these states does not in general satisfies the constraint imposed by the vanishing of the six dimensional anomaly: N\_H-N\_V+29 N\_T = 273 . \[6d\] The solution to this puzzle comes from the analysis of the type II dual point of view. 0.3cm {#section-2 .unnumbered} Since heterotic/type IIA duality in four dimensions maps the heterotic dilaton field into a perturbative, volume-form modulus of the type IIA string, we expect that inspection of the latter can give a hint in understanding what is happening. On the type IIA string side, the Horava–Witten orbifold of the M-theory is realized as an ordinary K3 compactification, which admits an orbifold realization as $T^4 \big/ Z_2$. This projection produces a twisted sector dual to the $E_8 \times E_8$ gauge sector of the heterotic string. A further $Z_2$ orbifold projection, besides the breaking of supersymmetry to ${\cal N}_4=2$, produces also a new twisted sector, corresponding to the $SO(32)$ sector. Indeed, the two orbifold projections are equivalent and interchangeable. Moreover, the $SO(32)$ and $E_8 \times E_8$ points are connected in the K3 moduli space, and in type IIA only the the Cartan subgroup of the gauge group appears perturbatively. It has therefore no meaning to distinguish between $SO(32)$ and $E_8 \times E_8$: in the following we will simply refer to “rank 16” factors of the whole gauge group. Only one of these factors appears perturbatively on the heterotic side. The second one appears perturbatively only on the type I side. There is however a third twisted sector, corresponding to the fixed points of the product of the two $Z_2$ orbifold projections. This gives rise to another rank 16 sector, which is non-perturbative on both the heterotic and type I sides. On the type IIA side, this $Z_2 \times Z_2$ orbifold has in total 48 fixed points, and it was repeatedly considered in the literature, both in the framework of string [@gkr] or of F-theory compactifications [@mvII; @gm]. It corresponds to the orbifold limit of the compactification on a Calabi–Yau manifold with Hodge numbers $(h^{1,1},h^{2,1})=(51,3)$. Its perturbative spectrum contains 3+48 vector multiplets and 4 hyper multiplets. The perturbative corrections to the effective coupling of the $R^2$ term were computed in Ref. [@gkr], and read [^3]: = - T\^1 | ( T\^1 ) |\^4 - T\^2 | ( T\^2 ) |\^4 - T\^3 | ( T\^3 ) |\^4 , where $T^1$, $T^2$, $T^3$ are the moduli associated to the Kähler classes of the three tori of the six dimensional compact space. Since the compact manifold is not self-mirror, the above correction is most probably modified by non-perturbative corrections. However, the above expression already indicates us that, under type II/heterotic duality, these moduli should be mapped into the three moduli $S$, $T$, $U$ [^4]. From the type IIA analysis, it appears clearly that these moduli parameterize the couplings of three *equivalent* gauge sectors: only one of them shares the “bare” coupling with the perturbative heterotic string, namely the one parameterized by the field $S$, and an entirely perturbative heterotic dual exists only when the other two sectors are not present. This happens if one of the two orbifold projections acts freely, in such a way that neither the corresponding twisted sector nor the twisted sector corresponding to the product $Z_2 \times Z_2$ possesses fixed points. This freely acting orbifold corresponds to the $CY^{11,11}$ manifold, which is a K3 fibration. This construction, together with its heterotic dual, has been considered in Refs. [@mvII]–[@gkp]. In that case, the corrections to the $R^2$ coupling read, on the type IIA side: = - T\^1 | ( T\^1 ) |\^4 - T\^2 | \_4 ( T\^2 ) |\^4 - T\^3 | \_4 ( T\^3 ) |\^4 . The $\vartheta$ functions in the second and third term signal that the corresponding sectors are massive. We remark that, in general, it is not possible to construct an heterotic model with a behavior of the coupling like: \~ S - T | \_4 ( T ) |\^4 - U | \_4 ( U ) |\^4 , namely, a model in which only the vectors corresponding to the “SO(32)” small instantons are present. Again, the substantial “equivalence” of the moduli $T$ and $U$ in the heterotic perturbative construction is a consequence of T-duality of the heterotic string [^5]. The example we provided was just intended to give the flavor of what is happening, without any aim to provide a concrete dual for a specific situation. In general, the four dimensional ${\cal N}_4=2$ heterotic string is dual to a type IIA string compactified on a K3 fibration ${\cal M}$. One can easily realize that, under heterotic/type II duality, the heterotic torus $T^2$ is mapped inside the fiber on the type IIA side. In this space, there are always at least three two-cycles, that must correspond to the three vector moduli present in any ${\cal N}_4=2$ heterotic compactification, namely $S$ (the base of the fibration) and the two moduli of the torus, $T$ and $U$, in such a way that $(\Im S \times \Im T \times \Im U) \approx {\rm Volume}({\cal M})$. The $T \leftrightarrow U$ symmetry of the perturbative heterotic string implies a corresponding symmetry also for their images in ${\cal M}$, at least when the volume of the base is large. When on the heterotic side extra, small-instanton-like states appear, and the type IIA fibration degenerates, the singular points must appear in a symmetric way with respect to the two cycles corresponding to the image of $T$ and $U$. For what matters the six dimensional anomaly constraint, we have to keep in mind that the decompactification of the image of the heterotic two-torus is a singular limit on the type II side, that leads in general to what resembles a “non-compact orbifold”. The “six dimensional” theory obtained via this decompactification is not therefore a true, smooth six dimensional theory, and there is no reason to expect that the massless states satisfy the six dimensional anomaly constraint, Eq. (\[6d\]) [^6]. 0.3cm {#section-3 .unnumbered} It is natural to ask why these new states never appeared in the various analysis of small instanton-like phenomena of the heterotic string, and whether there can exist a description of them in terms of some effective theory of membranes or, in general, solitonic solutions of some effective theory. In order to answer to these questions, we must go back and see *how* the “ordinary” small instantons have been detected. These states are always described in terms either of D5-branes of the type I string, therefore using string-string duality, or in terms of solitonic, or membranes, solutions of “field theory” configurations that are supposed to contain the heterotic string, such as those of M-theory. However, in both these cases one always “projects” the underlying theory onto a subspace spanned by the only Kaluza–Klein states. Namely, in order to give a description in terms of field theory solutions, one looks for a geometric picture, that necessarily discards a part of the states. In other words, one never has, at the same time, one manifold and its T-dual. If the states of a certain kind of small instantons (those of the $E_8 \times E_8$ or those of the $SO(32)$ string as well) are described in terms of a field theory built on geometric objects, like membranes, it is not possible to describe, at the same time, also the T-dual states. As we said in the previous section, the new states correspond instead to “true” small instantons of the type I string, something very different from what one usually calls the “type I small instantons”, namely the duals of the heterotic small instantons, the D5 branes. Their existence can be inferred from the fact that, like the heterotic string, also the type I string is compactified on a K3, with a bundle with tensor structure. However, the only way of explicitly observing all these states is through the direct type II construction in four dimensions, as we did in the previous section. There, T-duality of the heterotic string appears as a symmetry in the moduli space, that maps, in the specific example we considered, a twisted sector into another one. We argue that the place where these phenomena can be properly investigated is the type IIA string construction in four dimensions [^7]: these new states should appear at singularities of K3 fibrations, and their coupling would be derived from the intersection of the corresponding cycles with those associated to the moduli $T$ and $U$ [@al]. From this point of view, it is clear that, as we said, the type II fibration must possess a symmetry corresponding to the heterotic $T \leftrightarrow U$ symmetry, at least when the volume of the base is large. This implies a symmetry in the fiber, so that the appearance of only cycles that “intersect” $T$ and not, at the same time, cycles that “intersect” $U$ (or vice-versa), would correspond to a breaking of the heterotic T-duality much stronger than the one expected [^8], which is due to instanton phenomena, suppressed at the weak coupling (large $S$). In this case, the breaking of T-duality would be “explicit”, for instance in the effective action corresponding to the massless degrees of freedom, and persist even in the weak coupling limit. We consider this phenomenon quite unlikely to happen, because it would contradict a perturbative symmetry of the heterotic string even in the weak coupling regime. To this regard, a comment is in order about certain results of Ref. [@am], for which it seems that the number of tensor multiplets and the rank of the extra, non-perturbative gauge group one obtains at special points of the six dimensional heterotic theory, give, after compactification to four dimensions, a result in contrast with T-duality. More precisely, as discussed in section 6 of Ref. [@am], the theory is only “covariant”, but not “invariant” under T-duality (see discussion around Eq. (47) of that reference). The analysis is based on the duality between the heterotic string compactified on a K3 and F-theory compactified on a Calabi–Yau three-fold. The existence of such a duality was conjectured in Ref. [@mvI], as a consequence of the decompactification of both the four dimensional heterotic string compactified on $T^2 \times K3$ and its type II dual compactified on a K3 fibration. At the base of the argument, there was the conjecture of Ref. [@vafa], according to which the type II theory is essentially F-theory compactified on a torus. However, the chain of arguments leading from a string/string duality in four dimensions to a string/F-theory duality in six dimensions presented in Ref. [@mvI] is in general not correct: the passage from four to six dimensions is in fact very singular. In order to get a well defined six dimensional limit, one has to assume in fact that the heterotic torus $T^2$ is mapped to a torus, or a ${\bf P}^1$, on the type II side. However, in the examples of heterotic/type II dual pairs discussed in Refs. [@gkp; @gkp2], where the map of the moduli of the two-torus has been carefully taken into account, it happens that the moduli $T$ and $U$, associated respectively to the Kähler class and to the complex structure of the heterotic torus, are mapped to two volume moduli of the fiber on the type II side, in such a way that $\Im T \times \Im U = V_{(4)}$, the volume of the fiber. Something similar happens also more in general: on the heterotic side both the Kähler class and the complex structure of a two-torus belong to the vector multiplets. On the type II side, instead, they belong one to the vector and the other to the hyper multiplets (which one, it depends on whether we consider type IIA or type IIB). It is therefore clear that the heterotic torus must map into something in the fiber, but it cannot be a torus. In general, it is not a ${\bf P}^1$ either: both $T$ and $U$ must correspond to vector multiplets, represented by two independent cycles in the fiber. More precisely, they must map into two of the three cycles that are always present in any K3 fibration (see the lower bound of Eq. (30) of Ref. [@al]). In fact, in the case of duality with type IIA, together with the divisor dual to the heterotic dilaton, they must span the “volume” of the Calabi–Yau manifold. The decompactification of the heterotic torus, i.e. the limit $\Im T \to \infty$, is therefore something ill defined on the type II side: if this decompactification is achieved by passing through a five-dimensional theory, then also $\Im U \to \infty$, or $\Im U \to 0$. In any case, even if one trades, via a bi-rational transformation, a ${\bf P}^1$, or a torus, in the fiber with one in the base, there remains a mismatch regarding what happens to the other cycle. The decompactification of the heterotic string is a very singular operation on the type II side, that “destroys” the initial K3 fibration of the type II string. Conversely, let’s consider the F-theory compactified to six dimensions on that K3 fibration, ${\cal M}$. Now, either ${\cal M}$ is also fibered over a torus, or a ${\bf P}^1$, that can be considered to correspond to the type II coupling, so that, in a certain limit, this vacuum can be seen to correspond to a (weakly coupled) type II vacuum, or is not. In the first case, further compactification on a two-torus adds (at least) one vector multiplet and one hypermultiplet. In the second case, compactification on a torus must add at least one hypermultiplet, in order to provide the coupling of the type II string. In any case, this compactification adds at least one hypermultiplet. This means that, away from special points, it adds one vector and one hypermultiplet. On the heterotic side, instead, compactification of the six dimensional theory on a two-torus adds, away from special points, two vector multiplets. There is therefore in general a mismatch between the two theories. The loss of track of the heterotic T-duality on the type II side has its origin in the ill-definiteness of the decompactification. These problems arise when we want to use an analysis based on six dimensional anomaly in order to derive, via toroidal compactification, the non-perturbative spectrum of the four dimensional theory. In this case, the (singular) Calabi–Yau space of the type II string cannot be used as a good compactification space: the F-theory dual of the so derived six dimensional heterotic string may well be compactified on a very different space, that just “projects” on ${\cal M}$. Moreover, there is in general no consistent field theory description accounting for all the massless states at the same time. In the cases in which one considers only the perturbative moduli space of the heterotic theory, and translates its properties into geometrical properties of smooth manifolds, implementing then the analysis with a set of rules based essentially on field theory, this duality [@mvI] works, and can be transferred from six to four dimensions without problems: in six dimensions, the perturbative theory is in fact highly constrained, and the four dimensional theory is simply a dimensional reduction of that. 0.3cm {#section-4 .unnumbered} In this note we discussed the existence, in heterotic and type I $T^2 \times T^4 \big/ Z_2$ compactifications to four dimensions, of a non-perturbative gauge sector, that extends the whole gauge group to a maximal rank 48. This sector originates from wrapped tensor multiplets associated, on the type I side, to the $Z_2$ fixed points. On the heterotic side, these vector multiplets are generically present together with those appearing when small instantons shrink to zero size. The simultaneous presence of both these non-perturbative sectors in the heterotic string is essentially due to T-duality, that exchanges momentum states with winding states. The coupling of these extra gauge bosons, that don’t exist in six dimensions, is parameterized by the field $U$, associated to the complex structure of $T^2$. The analysis we carried out is only qualitative, and we leave for future investigation interesting issues like what really are the allowed gauge groups and what is the massless spectrum, including hypermultiplets. From the investigation of the type IIA string we learn that, from the M-theory point of view, the three gauge sectors, namely the one corresponding to the heterotic perturbative gauge group and the two non-perturbative ones, should appear essentially on the same footing. By this we mean that physics cannot “prefer” one sector with respect to the other. Only the identification of the heterotic dual chooses a preferred direction. This fact may have interesting consequences for the string phenomenology: there is in fact no a-priori reason to prefer one of the three gauge sectors to be the one that should contain the Standard Model [^9], and the relation between gauge couplings and string parameters may be rather different from what expected from a perturbative heterotic string analysis. In this case, we have in fact three gauge sectors, in which the coupling is parameterized respectively by the fields $S$, the axion–dilaton field of the heterotic string, $T$, the modulus associated to the volume of a two-torus, and $U$, associated to its complex structure. In particular, in the third sector the coupling does dependent neither on the string scale, nor on the size of internal dimensions, but only on a “shape” modulus, both for the heterotic and the type I string. [99]{} E. Witten, . M. Faux, D. Lust and B. A. Ovrut, . E. G. Gimon and J. Polchinski, . P. S. Aspinwall and D. R. Morrison, . L. J. Dixon, V. Kaplunovsky and J. Louis, . J. A. Harvey and G. Moore, . E. Kiritsis, C. Kounnas, P.M. Petropoulos and J. Rizos, . See also hep-th/9605011, published in the [*proceedings of 5th Hellenic School and Workshops on Elementary Particle Physics*]{}, Corfu, Greece, 3-24 Sep 1995. I. Antoniadis, C. Bachas, C. Fabre, H. Partouche and T.R. Taylor, ;\ I. Antoniadis, H. Partouche and T.R. Taylor, ; Nucl. Phys. Proc. Suppl. [**61A**]{}, 1998, 58; Nucl. Phys. Proc. Suppl. [**67**]{}, 1998, 3. A. Sagnotti, . J. F. Morales and M. Serone, . O. J. Ganor and A. Hanany, . N. Seiberg and E. Witten, . M. Berkooz, R. G. Leigh, J. Polchinski, N. Seiberg and E. Witten, . A. Gregori, C. Kounnas and J. Rizos, . E. Kiritsis and C. Kounnas, . D. R. Morrison and C. Vafa, ; R. Gopakumar and S. Mukhi, . S. Ferrara, J. A. Harvey, A. Strominger and C. Vafa, . J. A. Harvey and G. Moore, . A. Gregori, C. Kounnas and P. M. Petropoulos, ;\ A. Gregori, hep-th/9811096, in the [*proceedings of the 6th Hellenic School and Workshop on Elementary Particle Physics*]{}, Corfu, Greece, 6-26 Sep 1998. P. S. Aspinwall and J. Louis, . P. S. Aspinwall and M.R. Plesser, JHEP [**9908**]{} (1999) 001. D. R. Morrison and C. Vafa, . C. Vafa, . A. Gregori, C. Kounnas and P.M. Petropoulos, . K. Benakli and Y. Oz, . A. Gregori, hep-th/0005198. A. Gregori, hep-th/0110025. A. Gregori, hep-th/0110201. [^1]: Extensions of the gauge group, that appear as non-perturbative from the heterotic point of view, have also been observed in the M-theory compactified on $S^1 \times T^4 \big/ Z_2$ in Ref. [@flo]. [^2]: In the above formula, we give only the dominant behavior, omitting any contribution of Wilson lines and terms mixing the contribution of these moduli. [^3]: For simplicity, here and in the following we will omit all the normalization coefficients and the cut-off dependent term accounting for the infrared running. For more details about this, we refer the reader to Ref. [@infra]. [^4]: We recall that, for large $\Im X$, $-\log \Im X | \eta (X)|^4 \to 3 \Im X$, $-\log \Im X | \vartheta_4 (X)|^4 \to \log \Im X$. [^5]: This duality can be broken by Wilson lines, but never in a way to lift the mass of all the states of only one of the two sectors. [^6]: For a further discussion, see Ref. [@poster]. [^7]: For a detailed analysis of certain specific cases, see Ref. [@stria]. [^8]: See for instance Ref. [@apl]. [^9]: For a discussion of this issue, see Refs. [@bo; @gshort].

ArXiv

--- abstract: | India like many other developing countries is characterized by huge proportion of informal labour in its total workforce. The percentage of Informal Workforce is close to 92% of total as computed from NSSO 68th round on Employment and Unemployment, 2011-12^[@nsso68]^. There are many traditional and geographical factors which might have been responsible for this staggering proportion of Informality in our country.\ As a part of this study, we focus mainly on finding out how Informality varies with Region, Sector, Gender, Social Group, and Working Age Groups. Further we look at how Total Inequality[^1] is contributed by Formal and Informal Labour, and how much do occupations/industries contribute to inequality within each of formal and informal labour groups separately. For the purposes of our study we use NSSO rounds 61 (2004-05) and 68 (2011-12) on employment and unemployment. The study intends to look at an overall picture of Informality, and based on the data highlight any inferences which are visible from the data. author: - | Vinay Reddy Venumuddala\ PhD Student in Public Policy\ Indian Institute of Management, Bangalore. bibliography: - 'reflibrary.bib' title: Informal Labour in India --- Introduction ============ The term *Informal Sector* in Indian context is more or less synonymous to *Unorganized sector*, which signifies those enterprises that are outside purview of a majority of labour laws^[@nceus_stats]^. Therefore workforce in these enterprises lack the protection of fair standards that are essential for enabling decent forms of work. These fair labour standards may include protection in terms of Social Security, Employment Security, Occupational Safety, freedom of collective bargaining, absence of any forms of discrimination at work and so on. In addition to workforce in Informal Sector, there are workforce in Formal Sector and also in households who often lack similar protection. Since majority of workforce in the country are informal in nature, in a way it reflects the extent to which labour laws are ensuring protection to the total workforce.\ \ From a legal standpoint, an important notion that is widely used in countries around the world to assess the scope and extent of labour standards applicable between employee and employer is the *Employment Relationship*. It is the legal interpretation attributed to Employment Relationship that guides the level of enforcement of labour standards. Following is the formal definition as described by an ILO report on Employment Relationship^[@EmpRel]^.\ *“The employment relationship is a legal notion widely used in countries around the world to refer to the relationship between a person called an employee (frequently referred to as a worker) and an employer for whom the employee performs work under certain conditions in return for remuneration.^[@EmpRel]^”*\ \ With globalization and growing use of Information Technology, the nature of working arrangements has become enormously diverse, and hence it poses a challenge in terms of the applicability of an Employment Relationship. More often, certain work arrangements like self-employed, triangular work relationships, and disguised employment restricts or completely nullifies the applicability of an employment relationship. To improve the extent of labour protection offered under law, it is therefore necessary to clarify the legal nature of employment relationship to cover most of the Occupations and Industries.\ \ In order to approach the problem of Informality from a legal perspective, it is essential to look at overall nature to understand the complexity involved. With this as the motivation we try to look at the variation of informality across different categories (like Occupation, Industries, Gender, Soical Group, Sector, Region, etc..,) so that it can give a picture of the complexity we are dealing with. We also look at the contribution of formal and informal employment to Total Inequality and further within formal and informal groups, how different occupations or industries are contributing to these within group (formal/informal) inequalities.\ \ In Section-I we briefly give the definitional aspect of Informal labour, In Section-II we describe the classification within each of the categories in our study, In Sections-III and IV, we deal with variation of informality within categories, and the inequality analysis as described above. In Section-V we conclude by highlighting the limitations of this study and try to suggest possible areas of future work in this area. Section-I: Definition of Informal Employment ============================================ In its report on Definitional and Statistical Issues relating to Informal Economy^[@nceus_stats]^, NCEUS has given clear and concise definitions for *Informal Sector* and *Informal Employment* in the context of India. In the following sub-section we give a brief overview of the definitional aspects of these terms. Informal Sector and Informal Employment --------------------------------------- In order to arrive at the definition of Informal Sector, the commission has investigated three relevant major characteristics in line with international recommendations. 1. *Enterprises owned by individuals or households that are not constituted as separate legal entities independent of their owners:* Committee has carefully investigated into data sources on unorganized enterprises (56th and 57th rounds of NSSO on unorganized manufacturing and service sector enterprises, 3rd Census of Small Scale Industries and 5th Economic Census on unorganized sector) and found out that more than 95% of enterprises not covered under Organized Sector Surveys and Statistics (Like Annual Survey of Industries for organized manufacturing, National Accounts Statistics, etc..,) are proprietary and partnership in nature. 2. *Employment Size of the enterprise has to be below a certain threshold to be determined according to national circumstances:*The committee in its report has studied various labour legislations offering protection in terms of social security, and income security to its workers, and has shown that the enterprises where number of workers are less than 10 are essentially not covered by any of the existing labour laws. 3. *Non-maintenance of complete accounts that would permit financial separation of production activities:* Committee identified that all the proprietary and partnership enterprises employing less than 10 total workers are not under any legal obligation to maintain separate accounts. Based on observing the characteristics of enterprises compatible with international recommendations, the committee has given the following definition of informal sector,\ *“The informal sector consists of all unincorporated private enterprises owned by individuals or households engaged in the sale and production of goods and services operated on a proprietary or partnership basis[^2] and with less than ten total workers”*\ In addition to employees working in Informal Enterprises, there is informal employment also present in households and also in some formal sector enterprises. In order to cover these, the committee adopted the conceptual framework on Informal Employment, recommended in 17th ICLS, and has defined informal workers as follows,\ *“**Informal Employment:** Informal workers consist of those working in the informal sector or households, excluding regular workers with social security benefits provided by the employers and the workers in the formal sector without any employment and social security benefits provided by the employers”.*\ In our analysis in subsequent sections we therefore categorize the workforce into formal and informal groups, using the above definition of Informal employment given by NCEUS^[@nceus_stats]^. Section-II: Classification within categories ============================================ In order to analyze the variation of Informality, we adopt the following major classifications so as to look at how different is the level of informality within each of these categories. We use Occupation, Industry, Sector (Rural/Urban), Region, Gender (Male/Female), Social Group (ST,SC,OBC and Others), and Working Age Group as our categories of interest. Following tables show the classification in some of the major categories we use during our data analysis in the further sections.\ {width="6.0in"} \[fig:PovOcc\] {width="7.0in"} \[fig:IndCodes\] Section-III: Analysis of Informality ==================================== Informality across Occupations and Industries --------------------------------------------- In the following tables below, we give an insight into the percentage of Informal labour within and across each of the various industries and occupations in the country computed from NSSO 68th round data on Employment and Unemployment^[@nsso68]^. {width="7.0in"} \[fig:Ind68\] {width="6.0in"} \[fig:Occ68\] From the tables[^3] above we can see that Industry codes ..,9,6,19,20,7,12,1 contribute to higher proportion of informal labour within them, compared to 15,4,16,11,17,... Former include sectors like ‘Accommodation and Food Service Activities’, ‘Construction’, ‘Wholesale and retail trade and repair of motorcycles’, ‘Real estate activities’ and primarily ‘Agriculture, forestry and fishing’. While the latter include sectors like ‘Public Administration, defense and social security’, ‘Electricity, Gas, and Steam Supply’, ‘Education’, ‘Financial and Insurance activities’, ‘Human health and social work’, and so on.\ Similarly, occupational classes ..,7,9,1,6 have higher proportion of informal labour within them, than 4,3,2,... Former include classes like ‘Craft and related trades workers’, ‘Skilled agriculture and fishery occupations’, ‘Elementary occupations’, and more importantly it also includes ‘Senior officials and managers’. Latter include classes like ‘Technicians’, ‘Professionals’, ‘Clerks’, and so on.\ \ Above data clearly shows that the proportion of workforce beyond the coverage of labour laws are varying with occupation and industries. As we mentioned previously about the legal perspective of looking at fair labour standards, we can say that the employment relationship clarified by various labour laws in the country is somehow leaving out varying proportion of workers in different occupations and industries. The point we try to convey here is that, while determining the factors and indicators[^4] in describing employment relationship under labour legislations, it is essential to look at the differences in terms of work relationships that might arise in various industrial and occupational categories. Only when a comprehensive analysis of this sort is carried out, one can look into the applicability of fair labour standards across the country covering a majority of its workforce. Variation of Informality within categories ------------------------------------------ In this sub-section we consider occupations as a way of looking at total informal workforce. And, we show how across occupations proportion of informality is varying with Working Age Groups, Regions, Sector, Gender, and Social Groups. We use NSSO 68th round on Employment and Unemployment^[@nsso68]^ to show the variation of Informality within categories. [2]{} {width="2.5in"} \[fig:OccOverall\] {width="2.5in"} \[fig:OccAgeGroups\] {width="3.0in"} \[fig:OccSector\] {width="3.0in"} \[fig:OccRegionCat\] {width="3.0in"} \[fig:OccGender\] {width="3.0in"} \[fig:OccSocialGroups\] In the above figures, proportion of informal labour is shown in green while proportion of formal labour is shown in black. And, thickness of each bar represents the number of workforce in respective occupations. Following are some inferences we can get from looking at the above plots. - From Figure-\[fig:OccAgeGroups\], we can see that Working Age Groups 1 (15yrs-24yrs) and 4 (above 65yrs) account for higher proportion of informality in almost all the occupational classes. - From Figure-\[fig:OccSector\], we can see that the proportion of informal labour working in Occupational codes 6 *(Skilled Agricultural and fishery workers)*, and 9 *Elementary Occupations* are differing significantly between Rural and Urban areas. This is obvious because of the predominance of agriculture workforce in villages. - Apart from number of labour working in different occupational classes, there is no clear visible variation (from just looking at the plots) in proportion of informality within categories of regions, Gender and Social Group. Deeper analysis is required in this context to identify if there are any trends in terms of variation of informality in certain occupational classes within each of these categories. Section-IV: Inequality Decomposition on the basis of Informality ================================================================ Inequality Decomposition - Theory --------------------------------- We use decomposition by population subgroups as explained by Cowell et.al~[@cowell2011inequality]~. Following portion is a summary of the method taken from the paper\ \ Consider a Data Generating Process (DGP) represented in the following linear form: $$\begin{aligned} Y &= \beta_0+\sum_{k=1}^{K}\beta_k X_k + U \\ Where, \ \textbf{X}: &= (X_1,..,X_k) \ are \ explanatory \ random \ variables\end{aligned}$$ Now consider $X_1$ is a discrete random variable that take only the values ${X_{1,j}:j = 1,...,t_1}$. Data Generating Process (or population regression equation) for Sub-Group $j$ can be represented as: $$\begin{aligned} Y_j &= \beta_{0,j}+\beta_{1,j} X_{1,j} + \sum_{k=2}^{K}\beta_{k,j} X_{k,j} + U_j \\\end{aligned}$$ Define $P_j = Pr(X_1 = X_{1,j})$, the proportion of population for which $X_1 = X_{1,j}$. Then within group inequality can be written as: $$\begin{aligned} I_w(Y) = \sum_{j=1}^{t_1}W_j I(Y_j)\end{aligned}$$ where $t_1$ is the number of groups considered, $W_j$ is a weight that is a function of the $P_j$. The decomposition by population subgroups can be written as: $$\begin{aligned} I(Y) = I_b(Y) + I_w(Y)\end{aligned}$$ where $I_b$ is the between-group inequality. In the case of Generalized Entropy (GE) Indices we have, for any $\alpha \in (-\infty,\infty)$, $$\begin{aligned} W_j = P_j\left[ \frac{\mu(Y_j)}{\mu(Y)} \right]^{\alpha} = R_j^{\alpha} P_j^{1-\alpha},\end{aligned}$$ where $R_j := P_j \mu(Y_j)/\mu(Y)$ is the income share of group $j$, $\mu(Y_j)$ is mean income for subgroup $j$, $\mu(Y)$ is the mean income for the whole population.\ Empirical Application --------------------- Our empirical application of this inequality decomposition proceeds as follows: 1. We divide the individual data into subgroups by the variable Informality. Relating to the above theory, the random variable representing formal and informal groups can be given as\ \ **Inequality Decomposition of Total Employed Workforce by Formal or Informal Groups:** $$\begin{aligned} X_{1,j}:j=0,1 \;\ \ 0 = Formal, \ 1=Informal \end{aligned}$$ $$\begin{aligned} Y_j = \beta_{0,j}+\beta_{1,j} X_{1,j} + \sum_{k=2}^{K}\beta_{k,j} X_{k,j} + U_j \end{aligned}$$ We then compute the within group and between group inequality using $\alpha = 1.3$. Following are the equations of the total inequality in employed workforce, and decomposition. $$\begin{aligned} I(Y) &= \frac{1}{\alpha^2 - \alpha}\left[\int \left[ \frac{Y}{\mu(Y)} \right]^\alpha dF(Y) - 1\right] \\ I(Y) &= I_b(Y) + I_w(Y)\\ I_w(Y) &= W^fI(Y^f) + W^iI(Y^i)\\ I_b(Y) &= I(Y) - I_w(Y) \\ W^f &= P^f\left[ \frac{\mu(Y^f)}{\mu(Y)} \right]^{\alpha} = (R^f)^{\alpha} (P^f)^{1-\alpha} \\ W^i &= P^i\left[ \frac{\mu(Y^i)}{\mu(Y)} \right]^{\alpha} = (R^i)^{\alpha} (P^i)^{1-\alpha} \end{aligned}$$ Contributions to Within are given by $$\begin{aligned} C_w(Formal) &= W^fI(Y^f) \\ C_w(Informal) &= W^iI(Y^i) \\ \end{aligned}$$ Contributions to Total is given by $$\begin{aligned} C_t(Formal) &= 100\left[\frac{W^fI(Y^f)}{I(Y)} \right] \\ C_t(Informal) &= 100\left[\frac{W^iI(Y^i)}{I(Y)} \right] \\ \end{aligned}$$ 2. In this step, we subset the individual data on the basis of informality. We have therefore two sub-sets, a. Individuals who are informally employed, and b. Individuals who are formally employed. Now, we separately compute the within group and between group inequality on each of these data subsets by decomposing on the basis of occupation codes. The random variable(s) and the sub-group DGP pertaining to these sub-sets can be given by\ **Inequality Decomposition of Formal Workforce by Occupational Codes:** $$\begin{aligned} X^f_{1,j} :j=1,2,..9; \ (Occupational \ Codes) \end{aligned}$$ $$\begin{aligned} Y^f_j = \beta^f_{0,j}+\beta^f_{1,j} X^f_{1,j} + \sum_{k=2}^{K}\beta^f_{k,j} X^f_{k,j} + U^f_j \end{aligned}$$ Compute the within group and between group inequality using $\alpha = 1.3$. Following are the equations of the inequality in formal workforce, and further decomposition. $$\begin{aligned} I(Y^f) &= I_b(Y^f) + I_w(Y^f)\\ I_w(Y^f) &= \sum_{j=1}^{9}W^f_jI(Y^f_j) \\ I_b(Y^f) &= I(Y^f) - I_w(Y^f)\\ W^f_j &= P^f_j\left[ \frac{\mu(Y^f_j)}{\mu(Y^f)} \right]^{\alpha} = (R^f_j)^{\alpha} (P^f_j)^{1-\alpha} \end{aligned}$$ Contributions to Within are given by $$\begin{aligned} C_w(Each \ Occupation) &= W^f_jI(Y^f_j) \\ \end{aligned}$$ Contributions to Total is given by $$\begin{aligned} C_t(Each \ Occupation) &= 100\left[\frac{W^fW^f_jI(Y^f_j)}{I(Y)} \right] \\ \end{aligned}$$ **Inequality Decomposition of Informal Workforce by Occupational Codes:** $$\begin{aligned} X^i_{1,j} :j=1,2,..9; \ (Occupational \ Codes) \end{aligned}$$ $$\begin{aligned} Y^i_j = \beta^i_{0,j}+\beta^i_{1,j} X^i_{1,j} + \sum_{k=2}^{K}\beta^i_{k,j} X^i_{k,j} + U^i_j \end{aligned}$$ Compute the within group and between group inequality using $\alpha = 1.3$. Following are the equations of the inequality in informal workforce, and further decomposition. $$\begin{aligned} I(Y^i) &= I_b(Y^i) + I_w(Y^i)\\ I_w(Y^i) &= \sum_{j=1}^{9}W_jI(Y^i_j) \\ I_b(Y^i) &= I(Y^i) - I_w(Y^i) \\ W^i_j &= P^i_j\left[ \frac{\mu(Y^i_j)}{\mu(Y^i)} \right]^{\alpha} = (R^i_j)^{\alpha} (P^i_j)^{1-\alpha} \end{aligned}$$ Contributions to Within are given by $$\begin{aligned} C_w(Each \ Occupation) &= W^i_jI(Y^i_j) \\ \end{aligned}$$ Contributions to Total is given by $$\begin{aligned} C_t(Each \ Occupation) &= 100\left[\frac{W^iW^i_jI(Y^i_j)}{I(Y)} \right] \\ \end{aligned}$$ [llllllll]{} & $ C_w $ & $ GEI $ & $ P $ & $ R $ & W/B & Index & $ C_t $(%)\ \ $ I(Y) $ & - & - & - & - & - & 0.282 & 100\ $ I_w(Y) $ & - & - & - & - & 0.241 & - & 85.29\ $ I_b(Y) $ & - & - & - & - & 0.041 & - & 14.71\ Formal & 0.056 & 0.275 ($ I(Y^f) $) & 0.08 & 0.165 & - & - & 20\ Informal & 0.184 & 0.227 ($ I(Y^i) $) & 0.92 & 0.835 & - & - & 65.29\ \ $ I(Y^f) $ & - & - & - & - & - & 0.275 & -\ $ I_w(Y^f) $ & - & - & - & - & 0.229 & - & 16.65\ $ I_b(Y^f) $ & - & - & - & - & 0.046 & - & 3.35\ (F)Mgrs & 0.039 & 0.251 ($ I(Y^f_1) $) & 0.078 & 0.134 & - & - & 2.86\ (F)Professionals & 0.064 & 0.223 ($ I(Y^f_2) $) & 0.195 & 0.264 & - & - & 4.69\ (F)Technicians & 0.048 & 0.252 ($ I(Y^f_3) $) & 0.201 & 0.192 & - & - & 3.47\ (F)Clerks & 0.025 & 0.189 ($ I(Y^f_4) $) & 0.149 & 0.136 & - & - & 1.83\ (F)Service and Sales & 0.017 & 0.201 ($ I(Y^f_5) $) & 0.112 & 0.09 & - & - & 1.23\ (F)Agri and Allied & 0 & 0.106 ($ I(Y^f_6) $) & 0.005 & 0.003 & - & - & 0.02\ (F)Craft related & 0.009 & 0.14 ($ I(Y^f_7) $) & 0.101 & 0.073 & - & - & 0.68\ (F)Machine Op. & 0.011 & 0.198 ($ I(Y^f_8) $) & 0.083 & 0.062 & - & - & 0.81\ (F)Elem. Occ & 0.014 & 0.361 ($ I(Y^f_9) $) & 0.078 & 0.047 & - & - & 1.05\ \ $ I(Y^i) $ & - & - & - & - & - & 0.223 & -\ $ I_w(Y^i) $ & - & - & - & - & 0.194 & - & 55.68\ $ I_b(Y^i) $ & - & - & - & - & 0.033 & - & 9.61\ (I)Mgrs & 0.037 & 0.283 ($ I(Y^i_1) $) & 0.071 & 0.113 & - & - & 10.59\ (I)Professionals & 0.018 & 0.35 ($ I(Y^i_2) $) & 0.024 & 0.043 & - & - & 5.23\ (I)Technicians & 0.009 & 0.247 ($ I(Y^i_3) $) & 0.019 & 0.03 & - & - & 2.45\ (I)Clerks & 0.003 & 0.178 ($ I(Y^i_4) $) & 0.009 & 0.015 & - & - & 0.87\ (I)Service and Sales & 0.018 & 0.191 ($ I(Y^i_5) $) & 0.075 & 0.089 & - & - & 5.13\ (I)Agri and Allied & 0.05 & 0.177 ($ I(Y^i_6) $) & 0.322 & 0.29 & - & - & 14.28\ (I)Craft related & 0.02 & 0.16 ($ I(Y^i_7) $) & 0.13 & 0.127 & - & - & 5.77\ (I)Machine Op. & 0.008 & 0.14 ($ I(Y^i_8) $) & 0.047 & 0.056 & - & - & 2.35\ (I)Elem. Occ & 0.031 & 0.142 ($ I(Y^i_9) $) & 0.302 & 0.237 & - & - & 9\ Section-V: Discussion ===================== Informal Labour is a major policy issue, which particularly the developing countries are facing. To tackle the problem of informality, it is crucial to look at it holistically, before formulating broad policy solutions. Above study is just a small step towards understanding the problem of Informality in India based on recent NSSO rounds on Employment and Unemployment. It is limited by the extent to which it highlights the nature of informality in the country, and we see that there is a lot that needs to be researched in this area. Further work can be done to look within each of the industry and occupational classes and see how categories like Gender, Social Group, Sector, Working Age Groups etc.., are interacting with them to impact the extent of informality. It is essential to understand the problem of informality from the view-point of all the relevant stakeholders involved, so that formulating policies can be streamlined properly.\ Further there is a need to look from a legal perspective into each of the industrial and occupational classes and find if the scope of employment relationship can be improved in our labour laws. We conclude by saying that the above analysis is only one of the many perspectives of looking at how a country can realize fair labour standards for its workforce. Even after expanding the scope and extent of employment relationship, many forms of work arrangements where the workforce is Self-Employed or Casual Labour are still going to be outside this purview. Therefore it is important to carry out further research into how a country can accommodate labour of this nature within the definition of its employment relationship, or may be find out how it can legally enable fair labour standards to these workers outside the traditional framework of an employment relationship. Appendix ======== [^1]: For the purpose of measuring inequality, we use Monthly Per-capita Consumption Expenditure [^2]: An enterprise is classified as proprietary if an individual is its sole owner and as partnership if there are two or more owners on a partnership basis with or without formal registration. Partnerships also include SHGs and associations of individuals which are not recognized as separate legal entities. It excludes all corporate entities, cooperatives, trusts and other legal entities which are independent of their owners. [^3]: First two columns in Figure\[fig:Ind68\] shows the percentage of formal labour and informal labour within each industries, while the second set of two columns indicate the overall percentage of formal or informal labour across all the industries (Note that the total percentages in Column3 and in Column4 each add up to 100%). Interpretation is similar for Figure \[fig:Occ68\] [^4]: Factors are the criteria used in determining the employment relationship, and indicators are used to identify whether or not relevant factors are present which determine the existence of an employment relationship.^[@EmpRel]^

ArXiv

--- abstract: 'We introduce the Neural Collaborative Subspace Clustering, a neural model that discovers clusters of data points drawn from a union of low-dimensional subspaces. In contrast to previous attempts, our model runs without the aid of spectral clustering. This makes our algorithm one of the kinds that can gracefully scale to large datasets. At its heart, our neural model benefits from a classifier which determines whether a pair of points lies on the same subspace or not. Essential to our model is the construction of two affinity matrices, one from the classifier and the other from a notion of subspace self-expressiveness, to supervise training in a collaborative scheme. We thoroughly assess and contrast the performance of our model against various state-of-the-art clustering algorithms including deep subspace-based ones.' bibliography: - 'deepsubspace.bib' nocite: '[@langley00]' --- Introduction ============ In this paper, we tackle the problem of subspace clustering, where we aim to cluster data points drawn from a union of low-dimensional subspaces in an unsupervised manner. Subspace Clustering (SC) has achieved great success in various computer vision tasks, such as motion segmentation [@kanatani2001motion; @elhamifar2009sparse; @ji2014robust; @ji2016robust], face clustering [@ho2003clustering; @elhamifar2013sparse] and image segmentation [@yang2008unsupervised; @ma2007segmentation]. Majority of the SC algorithms [@yan2006general; @chen2009spectral; @elhamifar2013sparse; @liu2013robust; @wang2013provable; @lu2012robust; @ji2015shape; @you2016oracle] rely on the linear subspace assumption to construct the affinity matrix for spectral clustering. However, in many cases data do not naturally conform to linear models, which in turns results in the development of non-linear SC techniques. Kernel methods [@chen2009kernel; @patel2013latent; @patel2014kernel; @yin2016kernel; @xiao2016robust; @Ji2017AdaptiveLK] can be employed to implicitly map data to higher dimensional spaces, hoping that data conform better to linear models in the resulting spaces. However and aside from the difficulties of choosing the right kernel function (and its parameters), there is no theoretical guarantee that such a kernel exists. The use of deep neural networks as non-linear mapping functions to determine subspace friendly latent spaces has formed the latest developments in the field with promising results [@ji2017deep; @peng2016deep]. Despite significant improvements, SC algorithms still resort to spectral clustering which in hindsight requires constructing an affinity matrix. This step, albeit effective, hampers the scalability as it takes $O(n^2)$ memory and $O(kn^2)$ computation for storing and decomposing the affinity matrix for $n$ data points and $k$ clusters. There are several attempts to resolve the scalability issue. For example, [@you2016oracle] accelerate the construction of the affinity matrix using orthogonal matching pursuit; [@DBLP:journals/corr/abs-1811-01045] resort to $k$-subspace clustering to avoid generating the affinity matrix. However, the scalability issue either remains due to the use of spectral clustering [@you2016oracle], or mitigates but at the cost of performance [@DBLP:journals/corr/abs-1811-01045]. In this paper, we propose a neural structure to improve the performance of subspace clustering while being mindful to the scalablity issue. To this end, we first formulate subspace clustering as a classification problem, which in turn removes the spectral clustering step from the computations. Our neural model is comprised of two modules, one for classification and one for affinity learning. Both modules collaborate during learning, hence the name “Neural Collaborative Subspace Clustering”. During training and in each iteration, we use the affinity matrix generated by the subspace self-expressiveness to supervise the affinity matrix computed from the classification part. Concurrently, we make use of the classification part to improve self-expressiveness to build a better affinity matrix through collaborative optimization. We evaluate our algorithm on three datasets , namely MNIST [@lecun1998gradient], Fashion-MNIST [@xiao2017/online], and hardest one Stanford Online Products datasets [@oh2016deep] which exhibit different levels of difficulty. Our empirical study shows the superiority of the proposed algorithm over several state-of-the-art baselines including deep subspace clustering techniques. Related Work ============ {width="1.7\columnwidth"} -0.3in Subspace Clustering ------------------- Linear subspace clustering encompasses a vast set of techniques, among them, spectral clustering algorithms are more favored to cluster high-dimensional data [@vidal2011subspace]. One of the crucial challenges in employing spectral clustering on subspaces is the construction of an appropriate affinity matrix. We can categorize the algorithms based on the way the affinity matrix is constructed into three main groups: factorization based methods [@gruber2004multibody; @mo2012semi], model based methods [@chen2009spectral; @ochs2014segmentation; @purkait2014clustering], self-expressiveness based methods [@elhamifar2009sparse; @ji2014efficient; @liu2013robust; @vidal2014low]. The latter, [*i*.*e*.]{}, self-expressiveness based methods have become dominant due to their elegant convex formulations and existence of theoretical analysis. The basic idea of subspace self-expressiveness is that one point can be represented in terms of a linear combination of other points from the same subspace. This leads to several advantages over other methods: (i) it is more robust to noise and outliers; (ii) the computational complexity of the self-expressiveness affinity does not grow exponentially with the number of subspaces and their dimensions; (iii) it also exploits the non-local information without the need of specifying the size of the neighborhood ([*i*.*e*.]{}, the number of nearest neighbors as usually used for identifying locally linear subspaces [@yan2006general; @zhang2012hybrid]). The assumption of having linear subspaces does not necessarily hold in practical problems. Several works are proposed to tackle the situation where data points do not form linear subspaces but nonlinear ones. Kernel sparse subspace clustering (KSSC) [@patel2014kernel] and Kernel Low-rank representation [@xiao2016robust] benefit from pre-defined kernel functions, such as polynomial or Radial Basis Functions (RBF), to cast the problem in high-dimensional (possibly infinite) reproducing kernel Hilbert spaces. However, it is still not clear how to choose proper kernel functions for different datasets and there is no guarantee that the feature spaces generated by kernel tricks are well-suited to linear subspace clustering. Recently, Deep Subspace Clustering Networks (DSC-Net) [@ji2017deep] are introduced to tackle the non-linearity arising in subspace clustering, where data is non-linearly mapped to a latent space with convolutional auto-encoders and a new [*self-expressive layer*]{} is introduced between the encoder and decoder to facilitate an end-to-end learning of the affinity matrix. Although DSC-Net outperforms traditional subspace clustering methods by large, their computational cost and memory footprint can become overwhelming even for mid-size problems. There are a few attempts to tackle the scalability of subspace clustering. The SSC-Orthogonal Matching Pursuit (SSC-OMP) [@you2016scalable] replaces the large scale convex optimization procedure with the OMP algorithm to represent the affinity matrix. However, SSC-OMP sacrifices the clustering performance in favor of speeding up the computations, and it still may fail when the number of data points is very large. $k$-Subspace Clustering Networks ($k$-SCN) [@DBLP:journals/corr/abs-1811-01045] is proposed to make subspace clustering applicable to large datasets. This is achieved via bypassing the construction of affinity matrix and consequently avoiding spectral clustering, and introducing the iterative method of $k$-subspace clustering [@tseng2000nearest; @bradley2000k] into a deep structure. Although $k$-SCN develops two approaches to update the subspace and networks, it still shares the same drawbacks as iterative methods, for instance, it requires a good initialization, and seems fragile to outliers. Model fitting ------------- In learning theory, distinguishing outliers and noisy samples from clean ones to facilitate training is an active research topic. For example, Random Sample Consensus (RANSAC) [@fischler1981random] is a classical and well-received algorithm for fitting a model to a cloud of points corrupted by noise. Employing RANSAC on subspaces [@yang2006robust] in large-scale problems does not seem to be the right practice, as RANSAC requires a large number of iterations to achieve an acceptable fit. Curriculum Learning [@bengio2009curriculum] begins learning a model from easy samples and gradually adapting the model to more complex ones, mimicking the cognitive process of humans. Ensemble Learning [@dietterich2000ensemble] tries to improve the performance of machine learning algorithms by training different models and then to aggregate their predictions. Furthermore, distilling the knowledge learned from large deep learning models can be used to supervise a smaller model [@hinton2015distilling]. Although Curriculum Learning, Ensemble Learning and distilling knowledge are notable methods, adopting them to work on problems with limited annotations, yet aside the unlabeled scenario, is far-from clear. Deep Clustering --------------- Many research papers have explored clustering with deep neural networks. Deep Embedded clustering (DEC) [@xie2016unsupervised] is one of the pioneers in this area, where the authors propose to pre-train a stacked auto-encoder (SAE) [@bengio2007greedy] and fine-tune the encoder with a regularizer based on the student-t distribution to achieve cluster-friendly embeddings. On the downside, DEC is sensitive to the network structure and initialization. Various forms of Generative Adversarial Network (GAN) are employed for clustering such as Info-GAN [@chen2016infogan] and ClusterGAN [@DBLP:journals/corr/abs-1809-03627], both of which intend to enforce the discriminative feature in the latent space to simultaneously generate and cluster images. The Deep Adaptive image Clustering (DAC) [@chang2017deep] uses fully convolutional neural nets [@springenberg2014striving] as initialization to perform self-supervised learning, and achieves remarkable results on various clustering benchmarks. However, sensitivity to the network structure seems again to be a concern for DAC. In this paper, we formulate subspace clustering as a binary classification problem through collaborative learning of two modules, one for image classification and the other for subspace affinity learning. Instead of performing spectral clustering on the whole dataset, we train our model in a stochastic manner, leading to a scalable paradigm for subspace clustering. Proposed Method =============== To design a scalable SC algorithm, our idea is to identify whether a pair of points lies on the same subspace or not. Upon attaining such knowledge (for a large-enough set of pairs), a deep model can optimize its weights to maximize such relationships (lying on subspaces or not). This can be nicely cast as a binary classification problem. However, since ground-truth labels are not available to us, it is not obvious how such a classifier should be built and trained. In this work, we propose to make use of two confidence maps (see Fig. \[structure\] for a conceptual visualization) as a supervision signal for SC. To be more specific, we make use of the concept of self-expressiveness to identify positive pairs, [*i*.*e*.]{}, pairs that lie on the same subspaces. To identify negative pairs, pairs that do not belong to same subspaces, we benefit from a negative confidence map. This, as we will show later, is due to the fact that the former can confidently mine positive pairs (with affinity close to 1) while the latter is good at localizing negative pairs (with affinity close to 0). The two confidence maps, not only provide the supervision signal to optimize a deep model, but act collaboratively as partial supervisions for each other. Binary Classification --------------------- Given a dataset with $n$ points $\mathcal { X } = \left\{ \mathbf { x } _ { i } \right\} _ { i = 1 } ^ { n }$ from $k$ clusters, we aim to train a classifier to predict class labels for data points without using the groundtruth labels. To this end, we propose to use a multi-class classifier which consists of a few convolutional layers (with non-linear rectifiers) and a softmax output layer. We then convert it to an affinity-based binary classifier by $$\mathbf{A}_{c}(i,j) = \mathbf{\nu}_i\mathbf{\nu}_j^T,$$ where $\mathbf{\nu}_i\in \mathbb{R}^{k}$ is a $k$ dimensional prediction vector after $\ell_2$ normalization. Ideally, when $\mathbf{\nu}_i$ is one-hot, $\mathbf{A}_{c}$ is a binary matrix encoding the confidence of data points belonging to the same cluster. So if we supervise the classifier using $\mathbf{A}_{c}$, we will end up with a binary classification problem. Also note that $\mathbf{A}_{c}(i,j)$ can be interpreted as the cosine similarity between softmax prediction vectors of $\mathbf { x } _ { i }$ and $\mathbf { x } _ { j }$, which has been widely used in different contexts [@nguyen2010cosine]. However, unlike the cosine similarity which lies in $[-1,1]$, $\mathbf{A}_{c}(i,j)$ lies within $[0,1]$, since the vectors are normalized by softmax and $\ell_2$ norm. We illustrate this in Fig. \[classifier\]. 0.2in ![By normalizing the feature vectors after softmax function and computing their inner product, an affinity matrix can be generated to encode the clustering information. []{data-label="classifier"}](illustration.JPG){width="\columnwidth"} -0.2in Self-Expressiveness Affinity ---------------------------- Subspace self-expressiveness can be worded as: one data point $\mathbf{x}_i$ drawn from linear subspaces $ {\left\{ \mathcal { S } _ { i } \right\} _ { i = 1}^{k}}$ can be represented by a linear combination of other points from the same subspace. Stacking all the points into columns of a data matrix $\mathbf{X}$, the self-expressiveness can be simply described as $\mathbf{X} = \mathbf{X}\mathbf{C}$, where $\mathbf{C}$ is the coefficient matrix. It has been shown ([*e*.*g*.]{}, [@ji2014efficient]) that by minimizing certain norms of coefficient matrix $\mathbf{C}$, a block-diagonal structure (up to certain permutations) on $\mathbf{C}$ can be achieved. This translates into $c_{ji} \neq 0$ only if data points coming from the same subspace. Therefore, the loss function of learning the affinity matrix can be written as: $$\label{eq:self-expressive} \min\limits_{\mathbf{C}} \|{\mathbf{C} }\|_p \quad \rm s.t. \quad {\bf X} = {\bf X}{\bf C} ,\; {\rm diag}({\bf C}) = {\bf 0} , $$ where $\|\cdot\|_p$ denotes a matrix norm. For example, Sparse Subspace Clustering (SSC) [@elhamifar2009sparse] sticks to the $\ell_1$ norm, Low Rank Representation (LRR) models [@liu2011latent; @vidal2014low] pick the nuclear norm, and Efficient Dense Subspace Clustering [@ji2014efficient] uses the $\ell_2$ norm. To handle data corruption, a relaxed version can be derived as: $$\label{eq:self-expressive_relax} \begin{split} {\bf C}^* = &\operatorname*{arg\,min}_ { \mathbf { C } } \| \mathbf { C } \| _ { p } + \frac { \lambda } { 2 } \| \mathbf { X } - \mathbf { X } \mathbf { C } \| _ { F } ^ { 2 } \quad \\ & {\rm s.t.} \quad \operatorname {\rm diag } ( \mathbf { C } ) = \mathbf { 0 } . \end{split}$$ Here, $\lambda$ is a weighting parameter balancing the regularization term and the data fidelity term. To handle subspace non-linearity, one can employ convolutional auto-encoders to non-linearly map input data $\mathbf{X}$ to a latent space $\mathbf{Z}$, and transfer the self-expressiveness into a linear layer (without non-linear activation and bias parameters) named [*self-expressive layer*]{} [@ji2017deep] (see the bottom part of Fig. \[structure\]). This enables us to learn the subspace affinity $\mathbf{A}_{s}$ in an end-to-end manner using the weight parameters ${\bf C}^*$ in the self-expressive layer: $$\label{eq:definition-As} \mathbf{A}_{s}(i,j) = \begin{cases} (|c^*_{ij}|+|c^*_{ji}|) / 2c_{\rm max} & \text{if $i\neq j$,} \\ 1 & \text{if $i=j$,} \end{cases}$$ where $c_{\rm max}$ is the maximum absolute value of off-diagonal entries of the current row. Note that $\mathbf{A}_{s}(i,j)$ then lies within $[0,1]$. Collaborative Learning ---------------------- The purpose of collaborative learning is to find a principled way to exploit the advantages of different modules. The classification module and self-expressive module distill different information in the sense that the former tends to extract more abstract and discriminative features while the latter focuses more on capturing the pairwise correlation between data samples. From our previous discussion, ideally, the subspace affinity $\mathbf{A}_{s}(i,j)$ is nonzero only if $\mathbf{x}_i$ and $\mathbf{x}_j$ are from the same subspace, which means that $\mathbf{A}_{s}$ can be used to mine similar pairs ([*i*.*e*.]{}, positive samples). On the other hand, if the classification affinity $\mathbf{A}_{c}(i,j)$ is close to zero, it indicates strongly that $\mathbf{x}_i$ and $\mathbf{x}_j$ are dissimilar ([*i*.*e*.]{}, negative sample). Therefore, we carefully design a mechanism to let both modules collaboratively supervise each other. Given $\mathbf{A}_{c}$ and $\mathbf{A}_{s}$, we pick up the high-confidence affinities as supervision for training. We illustrate this process in Fig. \[structure\]. The “positive confidence” in Fig. \[structure\] denotes the ones from the same class, and the “negative confidence” represents the ones from different classes. As such, we select high affinities from $\mathbf{A}_{s}$ and small affinities from $\mathbf{A}_{c}$, and formulate the collaborative learning problem as: $$\label{Collaborative} \begin{split} & \min_{\mathbf{A}_{s},\mathbf{A}_{c}} \mathbf{\Omega}(\mathbf{A}_{s},\mathbf{A}_{c},l,u) = \\ L_{pos}&(\mathbf{A}_{s}, \mathbf{A}_{c},u) + \alpha L_{neg}(\mathbf{A}_{c},\mathbf{A}_{s},l), \end{split}$$ where the $L_{pos}(\mathbf{A}_{s},\mathbf{A}_{c},u)$ and $L_{neg}(\mathbf{A}_{c},\mathbf{A}_{s},l)$ denote the cross-entropy function with sample selection process, which can be defined as follows: $$\begin{split}\label{positive} L_{pos}(&\mathbf{A}_{s}, \mathbf{A}_{c},u) = H(\mathbf{M}_{s} || \mathbf{A}_{c} ) \\ & \mathrm{s.t} \quad \mathbf{M}_{s} = \mathds{1}(\mathbf{A}_{s} > u), \end{split}$$ and $$\begin{split}\label{negative} L_{neg}(\mathbf{A}_{c}&, \mathbf{A}_{s},l ) = H(\mathbf{M}_{c} || (\mathbf{1} - \mathbf{A}_{s} ) ) \\ & \mathrm{s.t} \quad \mathbf{M}_{c} = \mathds{1}(\mathbf{A}_{c} < l), \end{split}$$ where $\mathds{1}(\cdot)$ is the indicator function returning $1$ or $0$, $\{u,l\}$ are thresholding parameters, and $H$ is the entropy function, defined as $H(\mathbf{p}||\mathbf{q}) = \sum_j p_j \log(q_j)$. Note that the cross-entropy loss is a non-symmetric metric function, where the former probability serves a supervisor to the latter. Therefore, in Eqn. , the subspace affinity matrix $\mathbf{A}_{s}$ is used as the “teacher” to supervise the classification part (the “student”). Conversely, in Eqn. , the classification affinity matrix $\mathbf{A}_{c}$ works as the “teacher” to help the subspace affinity learning module to correct negative samples. However, to better facilitate gradient back-propagation between two modules, we can approximate indicator function by replacing $\mathbf{M}_{s}$ with $\mathbf{M}_{s} \mathbf{A}_{s}$ in Eqn.  and $\mathbf{M}_{c}$ with $ \mathbf{M}_{c} (1-\mathbf{A}_{c})$ in Eqn. . The weight parameter $\alpha$ in Eqn. \[Collaborative\], called collaboration rate, controls the contributions of $L_{pos}$ and $L_{neg}$. It can be set as the ratio of the number of positive confident pairs and the negative confident pairs, or slightly tuned for better performance. Loss Function ------------- After introducing all the building blocks of this work, we now explain how to jointly organize them in a network and train it with a carefully defined loss function. As shown in Fig. \[structure\], our network is composed of four main parts: (i) a convolutional encoder that maps input data ${\bf X}$ to a latent representation ${\bf Z}$; (ii) a linear self-expressive layer which learns the subspace affinity through weights ${\bf C}$; (iii) a convolutional decoder that maps the data after self-expressive layer, [*i*.*e*.]{}, ${\bf Z}{\bf C}$, back to the input space $\hat{\bf X}$; (iv) a multi-class classifier that outputs $k$ dimensional prediction vectors, with which a classification affinity matrix can be constructed. Our loss function consists of two parts, [*i*.*e*.]{}, collaborative learning loss and subspace learning loss, which can be written as: $$\label{eq:cost_function} \begin{split} \mathds{L}(\mathbf{X};\Theta) & = L_{sub}(\mathbf{X};\Theta,\mathbf{A}_{s}) \\ & + \lambda_{cl} \mathbf{\Omega}(\mathbf{X},u,l;\Theta, \mathbf{A}_{s},\mathbf{A}_{c}), \end{split}$$ where $\Theta$ denotes the neural network parameters and $\lambda_{cl}$ a weight parameter for the collaborative learning loss. The $L_{sub}(\mathbf{X};\Theta,\mathbf{A}_{s})$ is the loss to train the affinity matrix through self-expressive layer. Combining Eqn.  and the reconstruction loss of the convolutional auto-encoder, we arrive at: $$\begin{split} \label{eq:loss_subs} L_{sub}(\mathbf{X};\Theta,\mathbf{A}_{s}) & = \|\mathbf{C}\|_F^2 + \frac{\lambda_1}{2}\|{\bf Z} - {\bf Z}\mathbf{C}\|_F^2 \\ & + \frac{1}{2}\|{\bf X} - \hat{\bf X}\|_F^2 \quad \\ &{\rm s.t.}\quad {\rm diag}(\mathbf{C}) = {\bf 0} , \end{split}$$ where $\mathbf{A}_{s}$ is a function of ${\bf C}$ as defined in . After the training stage, we no longer need to run the decoder and self-expressive layer to infer the labels. We can directly infer the cluster labels through the classifier output $\mathbf{\nu}$: $$\label{eq:cluster} s_i = \operatorname*{arg\,max}_h \mathbf{\nu}_{ih},\;\; h = 1, \dots, k,$$ where $s_i$ is the cluster label of image $\mathbf{x}_i$. Optimization and Training ========================= dataset $\mathcal { X } = \left\{ \mathbf{ x } _ { i } \right\} _ { i = 1 } ^ { N }$, number of clusters $k$, sample selection threshold $u$ and $l$ , learning rate of auto-encoder $\eta_{ae}$, and learning rate of other parts $\eta$ Pre-train the Convolutional Auto-encoder by minimizing the reconstruction error. For every mini-batch data Train auto-encoder with [*self-expressive layer*]{} to minimize loss function in Eqn.  to update $\mathbf{A}_{s}$. Forward the batch data through the classifier to get $\mathbf{A}_{c}$. Do sample selection and collaborative learning through minimizing Eqn.  to update the classifier. Jointly update all the parameters by minimizing Eqn. . Get the cluster $s_i$ for all samples by Eqn.  In this section, we provide more details about how training will be done. Similarly to other auto-encoder based clustering methods, we pretrain the auto-encoder by minimizing the reconstruction error to get a good initialization of latent space for subspace clustering. According to [@elhamifar2009sparse], the solution to formulation  is guaranteed to have block-diagonal structure (up to certain permutations) under the assumption that the subspaces are independent. To account for this, we make sure that the dimensionality of the latent space (${\bf Z}$) is greater than (the subspace intrinsic dimension) $\times$ (number of clusters) [^1]. In doing so, we make use of the stride convolution to down-sample the images while increasing the number of channels over layers to keep the latent space dimension large. Since we have pretrained the auto-encoder, we use a smaller learning rate in the auto-encoder when the collaborative learning is performed. Furthermore, compared to DSC-Net or other spectral clustering based methods which require to perform sophisticated techniques to post process the affinity matrix, we only need to compute $\mathbf{A}_{s} = (|\mathbf{C}^*| + |\mathbf{C}^{*T}|)/2$ and normalize it (divided by the largest value in each row and assign $1.0$ to the diagonal entries) to ensure the subspace affinity matrix lie in the same range with the classification affinity matrix. We adopt a three-stage training strategy: first, we train the auto-encoder together with the self-expressive layer using the loss in  to update the subspace affinity $\mathbf{A}_{s}$; second, we train the classifier to minimize Eqn. ; third, we jointly train the whole network to minimize the loss in . All these details are summarized in Algorithm \[algorithm\]. Experiments =========== We implemented our framework with Tensorflow-1.6 [@abadi2016tensorflow] on a Nvidia TITAN X GPU. We mainly evaluate our method on three standard datasets, [*i*.*e*.]{}, MNIST, Fashion-MNIST and the subset of Stanford Online Products dataset. All of these datasets are considered challenging for subspace clustering as it is hard to perform spectral clustering on datasets of this scale, and the linearity assumption is not valid. The number of clusters $k$ is set to 10 as input to all competing algorithms. For all the experiments, we pre-train the convolutional auto-encoder for 60 epochs with a learning rate $1.0 \times 10^{-3}$ and use it as initialization, then decrease the learning rate to $1.0 \times 10^{-5}$ in training stage. The hyper parameters in our loss function are easy to tune. $\lambda_1$ in Eqn.  controls self-expressiveness, and it also affects the choice of $u$ and $l$ in Eqn. . If $\lambda_1$ set larger, the coefficient in affinity matrix will be larger, and in that case the $l$ should be higher. The other parameter $\lambda_{cl}$ balances the cost of subspace clustering and collaborative learning, and we usually set it to keep these two terms in the same scale to treat them equally. We keep the $\lambda_1 = 10$ in all experiments, and slightly change the $l$ and $u$ for each dataset. Our method is robust to different network design choices. We test different structures in our framework and get similar results on the same datasets. For MNIST, we use a three-layer convolutional encoder; for Fashion-MNIST and Stanford online Product, we use a deeper network consisting of three residual blocks [@he2016deep]. We do not use batch normalization in our network because it will corrupt the subspace structure that we want to learn in latent space. We use the Rectified Linear Unit (ReLu) as the non-linear activation in our all experiments. Since there are no ground truth labels, we choose to use a larger batch size compared with supervised learning to make the training stable and robust. Specifically, we set the batch size to 5000, and use Adam [@kingma2014adam], an adaptive momentum based gradient descent method to minimize the loss for all our experiments. We set the learning rate to $1.0 \times 10^{-5}$ the auto-encoder and $1.0 \times 10^{-3}$ for other parts in all training stages. **Baseline Methods.** We use various clustering methods as our baseline methods including the classic clustering methods, subspace clustering methods, deep clustering methods, and GAN based methods. Specifically, we have the following baselines: - classic methods: $K$-Means [@lloyd1982least] (KM), $K$-Means with our CAE-feature (CAE-KM) and SAE-feature (SAE-KM); - subspace clustering algorithms: sparse subspace clustering (SSC) [@elhamifar2013sparse], Low Rank Representation (LRR) [@liu2013robust], Kernel Sparse Subspace Clustering (KSSC) [@patel2014kernel], Deep Subspace Clustering Network (DSC-Net) [@ji2017deep], and $k$-Subspace Clustering Network ($k$-SCN) [@DBLP:journals/corr/abs-1811-01045]; - deep clustering methods: Deep Embedded Clustering (DEC) [@xie2016unsupervised], Deep Clustering Network (DCN) [@pmlr-v70-yang17b], and Deep Adaptive image Clustering (DAC) [@chang2017deep]; - GAN based clustering methods: Info-GAN [@chen2016infogan] and ClusterGAN [@DBLP:journals/corr/abs-1809-03627]. **Evaluation Metric.** For all quantitative evaluations, we make use of the unsupervised clustering accuracy rate, defined as $${\rm ACC }\;\% =\max_M \frac{\sum_{i=1}^n \mathds{1}(y_i = M(c_i))}{n}\times 100\%\;.$$ where $y_i$ is the ground-truth label, $c_i$ is the subspace assignment produced by the algorithm, and $M$ ranges over all possible one-to-one mappings between subspaces and labels. The mappings can be efficiently computed by the Hungarian algorithm. We also use normalized mutual information (NMI) as the additional quantitative standard. NMI scales from 0 to 1, where a smaller value means less correlation between predict label and ground truth label. Another quantitative metric is the adjusted Rand index (ARI), which is scaled between -1 and 1. It computes a similarity between two clusters by considering all pairs of samples and counting pairs that are assigned in the same or different clusters in ground truth and predicted clusters. The larger the ARI, the better the clustering performance. MNIST ----- MNIST consists of $70000$ hand-written digit images of size $28\times 28$. Subspace non-linearity arises naturally for MNIST due to the variance of scale, thickness and orientation among all the images of each digit. We thus apply our method on this dataset to see how well it can handle this type of subspace non-linearity. In this experiment, we use a three-layer convolutional auto-encoder and a self-expressive layer in between the auto-encoder for the subspace affinity learning module. The convolution kernel sizes are $5-3-3-3-3-5$ and channels are $10-20-30-30-20-10$. For the classification module, we connect three more convolutional layers after the encoder layers with kernel size 2, and one convolutional layer with kernel size 1 to output the feature vector. For the threshold parameters $u$ and $l$, we set them to $0.1$ and $0.7$ respectively in the first epoch of training, and increase $l$ to $0.9$ afterwards. Our algorithm took around 15 mins to finish training on a normal PC with one TITAN X GPU. We report the clustering results of all competing methods in Table  \[tab:MNIST\_results\_test\]. Since spectral clustering based methods ([*i*.*e*.]{}, SSC-CAE, LRR-CAE, KSSC-CAE, DSC-Net) can not apply on the whole dateset (due to memory and computation issue), we only use the 10000 samples to show how they perform. As shown in Table \[tab:MNIST\_results\_test\], subspace algorithms do not perform very well even on 10000 samples. Although the DSC-Net is trapped by training the self-expressive layer, it outperforms other subspace clustering algorithm, which shows the potential of learning subspace structure using neural networks. On the other hand, DEC, DCN, $k$-SCN and our algorithm are all based on auto-encoder, which learn embeddings with different metrics to help clustering. However, our classification module boost our performance through making the latent space of auto-encoder more discriminative. Therefore, our algorithm incorporates the advantage of different classes, [*e*.*g*.]{}, self-expressivess, nonlinear mapping and discriminativeness, and achieves the best results among all the algorithms thanks to the collaborative learning paradigm. ACC(%) NMI(%) ARI(%) ---------- ----------- ----------- ----------- CAE-KM 51.00 44.87 33.52 SAE-KM 81.29 73.78 67.00 KM 53.00 50.00 37.00 DEC 84.30 80.00 75.00 DCN 83.31 80.86 74.87 SSC-CAE 43.03 56.81 28.58 LRR-CAE 55.18 66.54 40.57 KSSC-CAE 58.48 67.74 49.38 DSC-Net 65.92 73.00 57.09 $k$-SCN 87.14 78.15 75.81 Ours **94.09** **86.12** **87.52** : Clustering results of different methods on MNIST. For all quantitative metrics, the larger the better. The best results are shown in bold.[]{data-label="tab:MNIST_results_test"} Fashion-MNIST ------------- Same as in MINIST, Fashion-MNIST also has $70000$ images of size $28\times 28$. It consists of various types of fashion products. Unlike MNIST, every class in Fashion-MNIST has different styles with different gender groups ([*e*.*g*.]{}, men, women, kids and neutral). As shown in Fig. \[samples\_fashion\], the high similarity between several classes (such as { Pullover, Coat, Shirt}, { T-shirt, Dress }) makes the clustering more difficult. Compared to MNIST, the Fashion-MNIST clearly poses more challenges for unsupervised clustering. On Fashion-MNIST, we employ a network structure with one convolutional layer and three following residual blocks without batch normalization in the encoder, and with a symmetric structure in the decoder. As the complexity of dataset increases, we also raise the dimensionality of ambient space to better suit self-expressiveness, and increase capacity for the classification module. For all convolutional layers, we keep kernel size as 3 and set the number of channels to 10-20-30-40. We report the clustering results of all methods in Table \[tab:Fashon-MNIST\_results\], where we can clearly see that our framework outperforms all the baselines by a large margin including the the best-performing baseline $k$-SCN. Specifically, our method improves over the second one by $8.36\%$, $4.2 \%$ and $9\%$ in terms of accuracy, NMI and ARI. We can clearly observe from Fig. \[fashion\_visualization\] that the latent space of our framework, which is collaboratively learned by subspace and classification modules, has strong subspace structure and also keeps each subspace discriminative. For subspace clustering methods we follow the way as on MNIST to use only 10000 samples. DSC-Net does not drop a lot while the performance of other subspace clustering algorithms decline sharply campared with their performance on MNIST. Since the code of ClusterGAN is not available currently, we can only provide results from their paper (without reporting ARI). ![The data samples of the Fashion-Mnist Dataset []{data-label="samples_fashion"}](Fashion_Mnist_display.png){width="1.0\columnwidth"} -0.2in ![The visualization of our latent space through dimension reduction by PCA.[]{data-label="fashion_visualization"}](Fashion_display.png){width="0.5\columnwidth"} -0.2in ACC(%) NMI(%) ARI(%) ------------ ----------- ----------- ----------- SAE-KM 54.35 58.53 41.86 CAE-KM 39.84 39.80 25.93 KM 47.58 51.24 34.86 DEC 59.00 60.10 44.60 DCN 58.67 59.4 43.04 DAC 61.50 63.20 50.20 ClusterGAN 63.00 64.00 - InfoGAN 61.00 59.00 44.20 SSC-CAE 35.87 18.10 13.46 LRR-CAE 34.48 25.41 10.33 KSSC-CAE 38.17 19.73 14.74 DSC-Net 60.62 61.71 48.20 $k$-SCN 63.78 62.04 48.04 Ours **72.14** **68.60** **59.17** : Clustering results of different methods on Fashion-MNIST. For all quantity metrics, the larger the better. The best results are shown in bold.[]{data-label="tab:Fashon-MNIST_results"} Stanford Online Products ------------------------ The Stanford Online Products dataset is designed for supervised metric learning, and it is thus considered to be difficult for unsupervised clustering. Compared to the previous two, the challenging aspects of this dataset include: (i) the product images contain various backgrounds, from pure white to real world environments; (ii) each product has different shapes, colors, scales and view angles; (iii) products across different classes may look similar to each other. To create a manageable dataset for clustering, we manually pick 10 classes out of 12 classes, with around 1000 images per class (10056 images in total), and then re-scale them to $32 \times 32$ gray images, as shown Fig. \[samples\_stanford\]. Our networks for this dataset start from one layer convolutional kernel with 10 channels, and follow with three pre-activation residual blocks without batch normalization, which have 20, 30 and 10 channels respectively. Table \[stanford\_table\] shows the performance of all algorithms on this dataset. Due to the high difficulty of this dataset, most deep learning based methods fail to generate reasonable results. For example, DEC and DCN perform even worse than their initialization, and DAC can not self-supervise their model to achieve a better result. Similarly, infoGAN also fails to find enough clustering pattern. In contrast, our algorithm achieves better results compared to other algorithms, especially the deep learning based algorithms. Our algorithm along with KSSC and DSC-Net achieve top results, due to the handling of non-linearity. Constrained by the size of dataset, our algorithm does not greatly surpass the KSSC and DSC-Net. We can easily observe that subspace based clustering algorithms perform better than clustering methods. This illustrates how effective the underlying subspace assumption is in high dimension data space, and it should be considered to be a general tool to help clustering in large scale datasets. In summary, compared to other deep learning methods, our framework is not sensitive to the architecture of neural networks, as long as the dimensionality meets the requirement of subspace self-expressiveness. Furthermore, the two modules in our network progressively improve the performance in a collaborative way, which is both effective and efficient. ACC (%) NMI (%) ARI (%) ---------- ---------- ----------- ---------- -- DEC 22.89 12.10 3.62 DCN 21.30 8.40 3.14 DAC 23.10 9.80 6.15 InfoGAN 19.76 8.15 3.79 SSC-CAE 12.66 0.73 0.19 LRR-CAE 22.35 **17.36** 4.04 KSSC-CAE 26.84 15.17 7.48 DSC-Net 26.87 14.56 **8.75** $k$-SCN 22.91 16.57 7.27 Ours **27.5** 13.78 7.69 : The clustering results of different algorithms on subset of Stanford Online Products. The best results are in bold.[]{data-label="stanford_table"} 0.2in ![The data samples of the Stanford Online Products Dataset []{data-label="samples_stanford"}](stanford_samples.png){width="1.1\columnwidth"} -0.2in Conclusion ========== In this work, we have introduced a novel learning paradigm, dubbed collaborative learning, for unsupervised subspace clustering. To this end, we have analyzed the complementary property of the classifier-induced affinities and the subspace-based affinities, and have further proposed a collaborative learning framework to train the network. Our network can be trained in a batch-by-batch manner and can directly predict the clustering labels (once trained) without performing spectral clustering. The experiments in our paper have shown that the proposed method outperforms the-state-of-art algorithms by a large margin on image clustering tasks, which validates the effectiveness of our framework. [^1]: Note that our algorithm does not require specifying subspace intrinsic dimensions explicitly. Empirically, we found a rough guess of the subspace intrinsic dimension would suffice, [*e*.*g*.]{}, in most cases, we can set it to 9.

ArXiv

--- abstract: 'We present a comprehensive theoretical analysis and computational study of four-wave mixing (FWM) of optical pulses co-propagating in one-dimensional silicon photonic crystal waveguides (Si-PhCWGs). Our theoretical analysis describes a very general set-up of the interacting optical pulses, namely we consider nondegenerate FWM in a configuration in which at each frequency there exists a superposition of guiding modes. We incorporate in our theoretical model all relevant linear optical effects, including waveguide loss, free-carrier (FC) dispersion and FC absorption, nonlinear optical effects such as self- and cross-phase modulation (SPM, XPM), two-photon absorption (TPA), and cross-absorption modulation (XAM), as well as the coupled dynamics of FCs and optical field. In particular, our theoretical analysis based on the coupled-mode theory provides rigorously derived formulae for linear dispersion coefficients of the guiding modes, linear coupling coefficients between these modes, as well as the nonlinear waveguide coefficients describing SPM, XPM, TPA, XAM, and FWM. In addition, our theoretical analysis and numerical simulations reveal key differences between the characteristics of FWM in the slow- and fast-light regimes, which could potentially have important implications to the design of ultra-compact active photonic devices.' author: - Spyros Lavdas - 'Nicolae C. Panoiu' title: 'Theory of Pulsed Four-Wave-Mixing in One-dimensional Silicon Photonic Crystal Slab Waveguides' --- Introduction {#sIntr} ============ One of the most promising applications of photonics is the development of ultra-compact optical interconnects for chip-to-chip and even intra-chip communications. The driving forces behind research in this area are the perceived limitations at high frequency of currently used copper interconnects [@hmh01pieee], combined with a rapidly increasing demand to move huge amounts of data within increasingly more confined yet increasingly intricate communication architectures. An approach showing great potential towards developing optical interconnects at chip scale is based on high-index contrast optical waveguides, such as silicon photonic waveguides (Si-PhWGs) implemented on the silicon-on-insulator material platform [@lll00apl; @apc02ptl]. Among key advantages provided by this platform are the increased potential for device integration facilitated by the enhanced confinement of the optical field achievable in high-index contrast photonic structures, as well as the particularly large optical nonlinearity of silicon, which makes it an ideal material for active photonic devices. Many of the basic device functionalities required in networks-on-chip have in fact already been demonstrated using Si-PhWGs, including parametric amplification [@cdr03oe; @edo04oe; @rjl05n; @fts06n; @lov10np], optical modulation [@cir95ptl; @ljl04n; @xsp05n], pulse compression [@cph06ptl; @m08ptl], supercontinuum generation [@bkr04oe; @hcl07oe], pulse self-steepening [@plo09ol], modulational instability [@pco06ol], and four-wave mixing (FWM) [@fys05oe; @edo05oe; @fts07oe; @zpm10np; @dog12oe]; for a review of optical properties of Si-PhWGs see [@opd09aop]. However, since the parameter space of Si-PhWGs is rather limited, there is little room to engineer their optical properties. A promising solution to this problem has its roots in the advent of photonic crystals (PhCs) in the late 80’s [@y87; @j87]. Thus, by patterning an optical medium in a periodic manner, with the spatial periods of the pattern being comparable to the operating optical wavelength, the optical properties of the resulting medium can be modified and engineered to a remarkable extent. Following this approach, a series of photonic devices have been demonstrated using PhCs, including optical waveguides and bends [@mck96prl; @lch98s; @bfy99el; @mck99prb; @cn00prb], optical micro-cavities [@fvf97nat; @pls99s; @aas03nat; @rsl04nat; @ysh04nat], and optical filters [@fvj99prb; @cmi01apl]. One of the most effective approaches to affect the optical properties of PhCs is to modify the group-velocity (GV), $v_{g}$, of the propagating modes. Unlike the case of waves propagating in regular optical media, whose GV can hardly be altered, by varying the geometrical parameters of PhCs one can tune the corresponding GV over many orders of magnitude. Perhaps the most noteworthy implication of the existence of optical modes with significantly reduced GV, the so-called slow-light [@sj04nm; @k08nphot; @b08nphot], is that both linear and nonlinear optical effects can be dramatically enhanced in the slow-light regime [@nys01prl; @sjf02josab; @pbo03ol; @pbo04oe; @vbh05n; @myp06ol; @cmg09nphot; @rls11prb]. One of the most important nonlinear optical process, as far as nonlinear optics applications are concerned, is FWM. In the generic case, it consists of the combination of two photons with frequencies, $\omega_{1}$ and $\omega_{2}$, belonging to two pump continuous-waves (CWs) or pulses, followed by the generation of a pair of photons with frequencies $\omega_{3}$ and $\omega_{4}$. The energy conservation requires that $\omega_{1}+\omega_{2}=\omega_{3}+\omega_{4}$. In practice, however, an easier to implement FWM configuration is usually employed, namely degenerate FWM. In this case one uses just one pump with frequency, $\omega_{p}$, the generated photons belonging to a signal ($\omega_{s}$) and an idler ($\omega_{i}$) beam; in this case the conservation of the optical energy is expressed as: $2\omega_{p}=\omega_{s}+\omega_{i}$. Among the most important applications of degenerate FWM it is noteworthy to mention optical amplification, wavelength generation and conversion, phase conjugation, generation of squeezed states, and supercontinuum generation. While FWM has been investigated theoretically and experimentally in PhC waveguides [@myk10oe; @ssv10oe; @meg10oe; @jli11oe; @csl12oe] and long-period Bragg waveguides [@dog12oe; @lzd14ol], a comprehensive theory of FWM in silicon PhC waveguides (Si-PhCWGs), which rigorously incorporates in a unitary way all relevant linear and nonlinear optical effects as well as the influence of photogenerated free-carriers (FCs) on the pulse dynamics is not available yet. In this article we introduce a rigorous theoretical model that describes FWM in Si-PhCWGs. Our model captures the influence on the FWM process of linear optical effects, including waveguide loss, FC dispersion (FCD) and FC absorption (FCA), nonlinear optical effects such as self- and cross-phase modulation (SPM, XPM), two-photon absorption (TPA), and cross-absorption modulation (XAM), as well as the mutual interaction between FCs and optical field. We also illustrate how our model can be applied to investigate the characteristics of FWM in the slow- and fast-light regimes, showing among other things that by incorporating the effects of FCs on the optical pulse dynamics new physics emerge. One noteworthy example in this context is that the well-known linear dependence of FCA on $v_{g}^{-1}$ is replaced in the slow-light regime by a $v_{g}^{-3}$ power-law dependence. ![(a) Geometry of the 1D Si-PhC slab waveguide. The height of the slab is $h=0.6a$ and the radius of the holes is $r=0.22a$. The primed coordinate system shows the principal axes of the Si crystal with the input facet of the waveguide in the (1$\bar{1}$0) plane of the Si crystal lattice. (b) Projected band structure. Dark yellow and brown areas correspond to slab leaky and guiding modes, respectively. The red and blue curves represent the guiding modes of the 1D waveguides.[]{data-label="fig:geom"}](Figure1.eps){width="8.0cm"} The remaining of the paper is organized as follows. In the next section we present the optical properties of the PhC waveguide considered in this work. Then, in Sec. \[sMath\], we develop the theory of pulsed FWM in Si-PhCWGs whereas the particular case of degenerate FWM is analyzed in Sec. \[sDFWM\]. Then, in Sec. \[sPMC\], we apply these theoretical tools to explore the physical conditions in which efficient FWM can be achieved. The results are subsequently used, in Sec. \[sResults\], to study via numerical simulations the main properties of pulsed FWM in Si-PhCWGs. We conclude our paper by summarizing in the last section the main findings of our article and discussing some of their implications to future developments in this research area. Finally, an averaged model that can be used in the case of broad optical pulses is presented in an Appendix.  by the circles *A*, *B*, *C*, *D*, and *E*, respectively.[]{data-label="fig:modes"}](Figure2.eps){width="8.0cm"} Description of the photonic crystal waveguide {#sGeometry} ============================================= In this section we present the geometrical and material properties of the PhC waveguide considered in this work, as well as the physical properties of its optical modes. Thus, our Si-PhCWG consists of a one-dimensional (1D) waveguide formed by introducing a line defect in a two-dimensional (2D) honeycomb-type periodic lattice of air holes in a homogeneous slab made of silicon (a so-called W1 PhC waveguide). The line defect is oriented along the $z$-axis, which is chosen to coincide with one of the $\Gamma K$ symmetry axes of the crystal, and is created by filling in a row of holes \[see Fig. \[fig:geom\](a)\]. The slab height is $h=0.6a$ and the radius of the holes is $r=0.22a$, where $a=\SI{412}{\nano\meter}$ is the lattice constant, whereas the index of refraction of silicon is $n_{\mathrm{Si}}\equiv n=3.48$. The defect line breaks the discrete translational symmetry of the photonic system along the $y$-axis, so that the optical modes of the waveguide are invariant only to discrete translation along the $z$-axis [@j08book]. Moreover, based on experimental considerations, we restrict our analysis to in-plane wave propagation, namely the wave vector, $\mathbf{k}$, lies in the $x=0$ plane. The $k_{z}$ component, on the other hand, can be restricted to the first Brillouin zone, $k_{z}\in[-\pi/a,\pi/a]$, which is an immediate consequence of the Bloch theorem. Under these circumstances, we determined numerically the photonic band structure of the system and the guiding optical modes of the waveguide using MPB, a freely available code based on the plane-wave expansion (PWE) method [@jj01oe]. To be more specific, we used a supercell with size of $6a \times 19 \sqrt{3}/2 a \times a$ along the $x$-, $y$-, and $z$-axis, respectively, the corresponding step size of the computational grid being $a/60$, $a\sqrt{3}/120$, and $a/60$, respectively. Figure \[fig:geom\](b) summarizes the results of these calculations. Thus, the waveguide has two fundamental TE-like optical guiding modes located in the band-gap of the unperturbed PhC, one $y$-even and the other one $y$-odd. ![(a), (b), (c), and (d) Frequency dependence of waveguide dispersion coefficients $n_g$, $\beta_2$, $\beta_3$, and $\beta_4$, respectively, determined for the even and odd modes. Light green, blue, and red shaded regions correspond to slow-light regime, defined as $n_g>20$. The dashed vertical line in panel (b) indicates the zero-GVD wavelength.[]{data-label="fig:disp"}](Figure3.eps){width="8.0cm"} In order to better understand the physical properties of the optical guiding modes, we plot in Fig. \[fig:modes\] the profile of the magnetic field $H_{x}$, which is its only nonzero component in the $x=0$ symmetry plane. These field profiles, calculated for several values of $k_{z}$, show that although the optical field is primarily confined at the location of the defect (waveguide), for some values of $k_{z}$ it is rather delocalized in the transverse direction. This field delocalization effect is particularly strong in the spectral domains where the modal dispersion curves are relatively flat, namely in the so-called slow-light regime, and increases when the group index of the mode, defined as $n_g=c/v_g$, increases. The dispersion effects upon pulse propagation in the waveguide are characterized by the waveguide dispersion coefficients, defined as $\beta_n=d^nk_{z}/d\omega^n$. In particular, the first-order dispersion coefficient is related to the pulse GV via $\beta_1=1/v_g$, whereas the second-order dispersion coefficient, $\beta_2$, quantifies the GV dispersion (GVD) as well as pulse broadening effects. The wavelength dependence of the first four dispersion coefficients, determined for both guided modes, is presented in Fig. \[fig:disp\], the shaded areas indicating the spectral regions of slow-light. For the sake of clarity, we set the corresponding threshold to $c/v_{g}=20$, that is the slow-light regime is defined by $n_g>20$. As it can be seen in Fig. \[fig:disp\], the even mode possesses two slow-light regions, one located at the band-edge ($\lambda\approx\SI{1.6}{\micro\meter}$) and the other one at $k_z\approx 0.3(2\pi/a)$, i.e. $\lambda\approx\SI{1.52}{\micro\meter}$, whereas the odd mode contains only one such spectral domain located at the band-edge ($\lambda\approx\SI{1.67}{\micro\meter}$). Moreover, the even mode can have both positive and negative GVD, the zero-GVD point being at $\lambda=\SI{1.56}{\micro\meter}$, whereas the odd mode has normal GVD ($\beta_{2}>0$) throughout. Since usually efficient FWM can only be achieved in the anomalous GVD regime ($\beta_{2}<0$), we will assume that the interacting pulses propagate in the even mode unless otherwise is specified. Derivation of the Mathematical Model {#sMath} ==================================== This section is devoted to the derivation of a system of coupled-mode equations describing the co-propagation of a set of mutually interacting optical pulses in a Si-PhCWG, as well the influence of photogenerated FCs on the pulse evolution. We will derive these coupled-mode equations in the most general setting, namely the nondegenerate FWM, then show how they can be applied to a particular case most used in practice, the so-called degenerate FWM configuration. Our derivation follows the general approach used to develop a theoretical model for pulse propagation in silicon waveguides with uniform cross-section [@cpo06jqe] and Si-PhCWGs [@pmw10jstqe]. Optical modes of photonic crystal waveguides -------------------------------------------- In the presence of an external perturbation described by the polarization, $\mathbf{P}_{\mathrm{pert}}(\mathbf{r},\omega)$, the electromagnetic field of guiding modes with frequency, $\omega$, is described by the Maxwell equations, which in the frequency domain can be written in the following form: \[Me\] $$\begin{aligned} &\nabla\times \mathbf{E}(\mathbf{r},\omega)=i\omega\mu\mathbf{H}(\mathbf{r},\omega),\label{ee} \\ &\nabla\times \mathbf{H}(\mathbf{r},\omega)=-i\omega[\epsilon_{c}(\mathbf{r},\omega)\mathbf{E}(\mathbf{r},\omega)+\mathbf{P}_{\mathrm{pert}}(\mathbf{r},\omega)],\label{he}\end{aligned}$$ where $\mu$ is the magnetic permeability, which in the case of silicon and other nonmagnetic materials can be set to $\mu=\mu_{0}$, $\epsilon_{c}(\mathbf{r},\omega)$ is the dielectric constant of the PhC, and $\mathbf{E}$ and $\mathbf{H}$ are the electric and magnetic fields, respectively. In our case, $\mathbf{P}_{\mathrm{pert}}$ is the sum of polarizations describing the refraction index change induced by photogenerated FCs and nonlinear (Kerr) effects. In order to understand how the modes of the PhC waveguide are affected by external perturbations, let us consider first the unperturbed system, that is $\mathbf{P}_{\mathrm{pert}}=0$. Thus, let us assume that, at the frequency $\omega$, the unperturbed PhC waveguide has $M$ guiding modes. It follows then from the Bloch theorem that the fields of these modes can be written as: \[modedef\] $$\begin{aligned} &\mathbf{E}_{m\sigma}(\mathbf{r},\omega)=\mathbf{e}_{m\sigma}(\mathbf{r},\omega)e^{i\sigma\beta_{m}z},~~~m=1,2,\ldots,M, \\ &\mathbf{H}_{m\sigma}(\mathbf{r},\omega)=\mathbf{h}_{m\sigma}(\mathbf{r},\omega)e^{i\sigma\beta_{m}z},~~~m=1,2,\ldots,M,\end{aligned}$$ where $\beta_{m}$ is the $m$th mode propagation constant and $\sigma=+$ ($\sigma=-$) denotes forward (backward) propagating modes. Here, we consider that the harmonic time dependence of the fields was chosen as $e^{-i\omega t}$. The mode amplitudes $\mathbf{e}_{m\sigma}$ and $\mathbf{h}_{m\sigma}$ are periodic along the $z$-axis, with period $a$. Moreover, the forward and backward propagating modes obey the following symmetry relations: \[modesymm\] $$\begin{aligned} &\mathbf{e}_{m-}(\mathbf{r},\omega)=\mathbf{e}_{m+}^{*}(\mathbf{r},\omega), \\ &\mathbf{h}_{m-}(\mathbf{r},\omega)=-\mathbf{h}_{m+}^{*}(\mathbf{r},\omega),\end{aligned}$$ where the symbol “$^{*}$” denotes complex conjugation. As such, one only has to determine either the forward or the backward propagating modes. The guiding modes can be orthogonalized, the most commonly used normalization convention being $$\label{modeorth} \frac{1}{4}\int_{S}\left(\mathbf{e}_{m\sigma}\times\mathbf{h}_{m'\sigma'}^{*}+\mathbf{e}_{m'\sigma'}^{*}\times\mathbf{h}_{m\sigma}\right)\cdot\hat{\mathbf{z}}dS=\sigma P_{m}\delta_{\sigma\sigma'}\delta_{mm'},$$ where $P_{m}$ is the power carried by the $m$th mode. This mode power is related to the mode energy contained in one unit cell of the PhC waveguide, $W_{m}$, via the relation: $$\label{PEnDens} P_{m}=\frac{W_{m}^{\mathrm{el}}+W_{m}^{\mathrm{mag}}}{a}v_{g} = \frac{2W_{m}^{\mathrm{el}}}{a}v_{g} = \frac{2W_{m}^{\mathrm{mag}}}{a}v_{g},$$ where \[energEH\] $$\begin{aligned} \label{energE} &W_{m}^{\mathrm{el}} = \frac{1}{4}\int_{V_{\mathrm{cell}}}\frac{\partial}{\partial\omega}(\omega\epsilon_{c})\vert\mathbf{e}_{m\sigma}(\mathbf{r},\omega)\vert^{2}dV, \\ \label{energH} &W_{m}^{\mathrm{mag}} = \frac{1}{4}\int_{V_{\mathrm{cell}}}\mu_{0}\vert\mathbf{h}_{m\sigma}(\mathbf{r},\omega)\vert^{2}dV,\end{aligned}$$ are the electric and magnetic energy of the mode, respectively, and $V_{\mathrm{cell}}$ is the volume of the unit cell. Note that in Eq.  we used the fact that the mode contains equal amounts of electric and magnetic energy. It should be stressed that the waveguide modes defined by Eqs.  are exact solutions of the Maxwell equations with $\mathbf{P}_{\mathrm{pert}}=0$, and thus they should not be confused with the so-called local modes of the waveguide. The latter modes correspond to waveguides whose optical properties vary adiabatically with $z$, on a scale comparable to the wavelength and have been used to describe, *e.g.*, wave propagation in tapered waveguides [@s70tmtt] or pulse propagation in 1D long-period Bragg gratings [@sse02jmo]. Perturbations of the photonic crystal waveguide ----------------------------------------------- Due to the photogeneration of FCs and nonlinear optical effects, the dielectric constant of Si-PhCWGs undergoes a certain local variation, $\delta\epsilon(\mathbf{r})$, upon the propagation of optical pulses in the waveguide. The corresponding perturbation polarization, $\mathbf{P}_{\mathrm{pert}}$ in Eq. , can be divided in two components according to the physical effects they describe: the linear change of the dielectric constant *via* generation of FCs and the nonlinearly induced variation of the index of refraction. Assuming an instantaneous response of the medium, the linear contribution to $\mathbf{P}_{\mathrm{pert}}$, $\delta \mathbf{P}_{\mathrm{lin}}(\mathbf{r},t)$, is written as: $$\label{linp} \delta \mathbf{P}_{\mathrm{lin}}(\mathbf{r},t)=\left[\delta\epsilon_{\mathrm{fc}}(\mathbf{r})+\delta\epsilon_{\mathrm{loss}}(\mathbf{r})\right]\mathbf{E}(\mathbf{r},t),$$ where [@cpo06jqe]: \[epslin\] $$\begin{aligned} \label{epslinfc}&\delta \epsilon_{\mathrm{fc}}(\mathbf{r}) = \left(2\epsilon_{0}n\delta n_{\mathrm{fc}} + i\frac{\displaystyle \epsilon_{0}cn}{\displaystyle \omega}\alpha_{\mathrm{fc}}\right)\Sigma(\mathbf{r}),\\ \label{epslinfc}&\delta\epsilon_{\mathrm{loss}}(\mathbf{r}) = i\frac{\displaystyle \epsilon_{0}cn}{\displaystyle \omega}\alpha_{\mathrm{in}}\Sigma(\mathbf{r}).\end{aligned}$$ Here, $\alpha_{\mathrm{in}}$ is the intrinsic loss coefficient of the waveguide and $\Sigma(\mathbf{r})$ is the characteristic function of the domain where FCs can be generated, namely $\Sigma=1$ in the domain occupied by Si and $\Sigma=0$ otherwise. Based on the Drude model, the FC-induced change of the index of refraction, $\delta n_{\mathrm{fc}}$, and FC losses, $\alpha_{\mathrm{fc}}$, are given by [@sb87jqe]: \[nepsFC\] $$\begin{aligned} \label{nFC}& \delta n_{\mathrm{fc}} = -\frac{\displaystyle e^{2}}{\displaystyle 2\epsilon_{0}n\omega^{2}} \left ( \frac{N_{e}}{m_{ce}} + \frac{N_{h}^{0.8}}{m_{ch}} \right ), \\ \label{epsFC}& \alpha_{\mathrm{fc}} = \frac{\displaystyle e^{3}}{\displaystyle \epsilon_{0}cn\omega^{2}} \left ( \frac{N_{e}}{\mu_{e}m_{ce}^{2}} + \frac{N_{h}}{\mu_{h}m_{ch}^{2}} \right ).\end{aligned}$$ Here, $e$ is the charge of the electron, $\mu_{e}$ ($\mu_{h}$) is the electron (hole) mobility, $m_{ce}=0.26m_{0}$ ($m_{ch}=0.39m_{0}$) is the conductivity effective mass of the electrons (holes), with $m_{0}$ the mass of the electron, and $N_{e}$ ($N_{h}$) is the induced variation of the electrons (holes) density (in what follows, we assume that $N_{e}=N_{h}\equiv N$). The nonlinear contribution to $\mathbf{P}_{\mathrm{pert}}$, $\delta \mathbf{P}_{\mathrm{nl}}(\mathbf{r},t)$, is described by a third-order nonlinear susceptibility, $\hat{\chi}^{(3)}(\mathbf{r})$, and can be written as: $$\label{Kerr} \delta \mathbf{P}_{\mathrm{nl}}(\mathbf{r},t)=\epsilon_{0}\hat{\chi}^{(3)}(\mathbf{r}) \vdots\mathbf{E}(\mathbf{r},t)\mathbf{E}(\mathbf{r},t)\mathbf{E}(\mathbf{r},t).$$ The real part of the susceptibility $\hat{\chi}^{(3)}$ describes parametric optical processes such as SPM, XPM, and FWM, while the imaginary part of $\hat{\chi}^{(3)}$ corresponds to TPA and XAM. Note that in this study we neglect the stimulated Raman scattering effect as it is assumed that the frequencies of the interacting pulses do not satisfy the condition required for an efficient, resonant Raman interaction. Since silicon belongs to the crystallographic point group $m3m$ the susceptibility tensor $\mathbb{\hat{\chi}}^{(3)}$ has 21 nonzero elements, of which only 4 are independent, namely, $\mathbb{\chi}_{1111}$, $\mathbb{\chi}_{1122}$, $\mathbb{\chi}_{1212}$, and $\mathbb{\chi}_{1221}$ [@b08book]. In addition, the frequency dispersion of the nonlinear susceptibility can be neglected as we consider optical pulses with duration of just a few picoseconds or larger. As a consequence, the Kleinman symmetry relations imply that $\mathbb{\chi}_{1122}=\mathbb{\chi}_{1212}=\mathbb{\chi}_{1221}$. Moreover, experimental studies have shown that $\hat{\chi}^{(3)}_{1111} = 2.36\hat{\chi}^{(3)}_{1122}$ [@zlp07apl] within a broad frequency range. Therefore, the nonlinear optical effects considered here can be described by only one element of the tensor $\hat{\chi}^{(3)}$. Because of fabrication considerations, in many instances the waveguide is not aligned with any of the crystal principal axes and as such these axes are different from the coordinate axes in which the optical modes are calculated. Therefore, one has to transform the tensor $\hat{\chi}^{(3)}$ from the crystal principal axes into the coordinate system in which the optical modes are calculated [@cpo06jqe], $$\label{rotChi} \hat{\chi}^{(3)}_{ijkl} = \hat{R}_{i\alpha}\hat{R}_{j\beta}\hat{R}_{k\gamma}\hat{R}_{l\delta}\hat{\chi}^{\prime(3)}_{\alpha\beta\gamma\delta},$$ where $\hat{\chi}^{\prime(3)}$ is the nonlinear susceptibility in the crystal principal axes and $\hat{R}$ is the rotation matrix that transforms one coordinate system into the other. In our case, $\hat{R}$ is the matrix describing a rotation with $\pi/4$ around the $x$-axis (see Fig. \[fig:geom\]). Coupled-mode equations for the optical field -------------------------------------------- In order to derive the system of coupled-mode equations describing pulsed FWM in Si-PhCWGs we employ the conjugated form of the Lorentz reciprocity theorem [@cpo06jqe; @mpw03pre; @ks07jqe; @s83book]. To this end, let us consider two solutions of the Maxwell equations , $[\mathbf{E}_{a}(\mathbf{r},\omega_{a}),\mathbf{H}_{a}(\mathbf{r},\omega_{a})]$ and $[\mathbf{E}_{b}(\mathbf{r},\omega_{b}),\mathbf{H}_{b}(\mathbf{r},\omega_{b})]$, which correspond to two different spatial distribution of the dielectric constant, $\epsilon_{a}(\mathbf{r},\omega_{a})$ and $\epsilon_{b}(\mathbf{r},\omega_{b})$, respectively. If we insert the vector $\mathbf{F}$, defined as $\mathbf{F} = \mathbf{E}_{b} \times \mathbf{H}_{a}^{*} + \mathbf{E}_{a}^{*} \times \mathbf{H}_{b}$, in the integral identity: $$\label{Lorentz} \int_{S} \nabla \cdot \mathbf{F} dS = \frac{\displaystyle \partial}{\displaystyle \partial z} \int_{S} \mathbf{F} \cdot \hat{\mathbf{z}} dS + \oint_{\partial S}\mathbf{F}\cdot\mathbf{n} dl,$$ where $S$ is the transverse section at position, $z$, and $\partial S$ is the boundary of $S$, and use the Maxwell equations, we arrive at the following relation: $$\begin{aligned} \label{Lorentzred} \frac{\displaystyle \partial}{\displaystyle \partial z} &\int_{S} \mathbf{F} \cdot \hat{\mathbf{z}}dS = i\mu_{0}(\omega_{b}-\omega_{a})\int_{S} \mathbf{H}_{a}^{*} \cdot \mathbf{H}_{b} dS \notag \\ &+i\int_{S}(\omega_{b}\epsilon_{b}-\omega_{a}\epsilon_{a}) \mathbf{E}_{a}^{*} \cdot \mathbf{E}_{b} dS-\oint_{\partial S}\mathbf{F}\cdot\mathbf{n} dl.\end{aligned}$$ Let us consider now a nondegenerate FWM process in which two pulses at carrier frequencies $\bar{\omega}_{1}$ and $\bar{\omega}_{2}$ interact and generate two optical pulses at carrier frequencies $\bar{\omega}_{3}$ and $\bar{\omega}_{4}$, with the energy conservation expressed as $\bar{\omega}_{1}+\bar{\omega}_{2}=\bar{\omega}_{3}+\bar{\omega}_{4}$. Then, in the Lorentz reciprocity theorem given by Eq.  we choose as the first set of fields a mode of the unperturbed waveguide ($\mathbf{P}_{\mathrm{pert}}=0$), which corresponds to the frequency $\omega_{a}=\bar{\omega}_{i}$, where $\bar{\omega}_{i}$ is one of the carrier frequencies $\bar{\omega}_{1}$, $\bar{\omega}_{2}$, $\bar{\omega}_{3}$, or $\bar{\omega}_{4}$: \[field1\] $$\begin{aligned} &\mathbf{E}_{a}(\mathbf{r},\bar{\omega}_{i})=\frac{\mathbf{e}_{n_{i}\rho_{i}}(\mathbf{r},\bar{\omega}_{i})}{\sqrt{\bar{P}_{n_{i}}}}e^{i\rho_{i}\bar{\beta}_{n_{i}}z}, \\ &\mathbf{H}_{a}(\mathbf{r},\bar{\omega}_{i})=\frac{\mathbf{h}_{n_{i}\rho_{i}}(\mathbf{r},\bar{\omega}_{i})}{\sqrt{\bar{P}_{n_{i}}}}e^{i\rho_{i}\bar{\beta}_{n_{i}}z},\end{aligned}$$ where $\rho_{i}=\pm1$ and $n_{i}$ is an integer, $1\leq n_{i}\leq N_{i}$, $i=1,\ldots,4$, with $N_{i}$ being the number of guiding modes at the frequency $\bar{\omega}_{i}$. In Eqs. , and in what follows, a bar over a symbol means that the corresponding quantity is evaluated at one of the carrier frequencies. As the second set of fields we take those that propagate in the perturbed waveguide, at the frequency $\omega_{b}=\omega$. These fields are written as a series expansion of the guiding modes at frequencies $\bar{\omega}_{i}$, $i=1,\ldots,4$, thus neglecting the frequency dispersion of the guiding modes and the radiative modes that might exist at the frequency $\omega$. This approximation is valid as long as all interacting optical pulses have narrow spectra centered at the corresponding carrier frequencies, that is the physical situation considered in this work. In particular, this modal expansion becomes less accurate when any of the pulses propagates in the slow-light regime, as generally the smaller the GV of a mode is the larger its frequency dispersion is. Thus, the second set of fields are expanded as: \[field2\] $$\begin{aligned} &\mathbf{E}_{b}(\mathbf{r},\omega)=\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}a_{m_{j}\sigma_{j}}^{(j)}(z,\omega)\frac{\mathbf{e}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}}e^{i\sigma_{j}\bar{\beta}_{m_{j}}z}, \\ &\mathbf{H}_{b}(\mathbf{r},\omega)=\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}a_{m_{j}\sigma_{j}}^{(j)}(z,\omega)\frac{\mathbf{h}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}}e^{i\sigma_{j}\bar{\beta}_{m_{j}}z}.\end{aligned}$$ With the fields normalization used in Eqs. , the mode amplitudes $a_{m_{i}\sigma_{i}}^{(i)}(z,\omega)$, $i=1,\ldots,4$, are measured in units of $\sqrt{\mathrm{W}}$. Note that since the optical pulses are assumed to be spectrally narrow, the mode amplitudes $a_{m_{i}\sigma_{i}}^{(i)}(z,\omega)$ have negligible values except when the frequency $\omega$ lies in a narrow spectral domain centered at the carrier frequency, $\bar{\omega}_{i}$. The dielectric constant in the two cases is $\epsilon_{a}=\bar{\epsilon}_{c}(\mathbf{r},\bar{\omega}_{i})$ and $\epsilon_{b}=\epsilon_{c}(\mathbf{r},\omega)+\delta\epsilon(\mathbf{r},\omega)$, where $\epsilon_{c}(\mathbf{r},\omega)$ is the dielectric constant of the unperturbed PhC. If the material dispersion is neglected, $\epsilon_{c}(\mathbf{r},\omega)=\epsilon_{c}(\mathbf{r},\bar{\omega}_{i})=\bar{\epsilon}_{c}(\mathbf{r})$. Inserting the fields given by Eqs.  and Eqs.  in Eq. , and neglecting the line integral in Eq. , which cancels for exponentially decaying guiding modes, one obtains the following set of coupled equations: $$\begin{aligned} \label{cmefreq} \rho_{i}&\frac{\displaystyle \partial a_{n_{i}\rho_{i}}^{(i)}(z)}{\displaystyle \partial z} + \sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum_{m_{j}\sigma_{j}}C_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}(z) \notag \\& \times\left[\frac{\displaystyle \partial a_{m_{j}\sigma_{j}}^{(j)}(z)}{\displaystyle \partial z} +i(\sigma_{j}\bar{\beta}_{m_{j}}-\rho_{i}\bar{\beta}_{n_{i}})a_{m_{j}\sigma_{j}}^{(j)}(z)\right] \notag \\ &= B_{n_{i}\rho_{i}}^{i}a_{n_{i}\rho_{i}}^{(i)}(z) + {\sum_{jm_{j}\sigma_{j}}}^{\prime}D_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}a_{m_{j}\sigma_{j}}^{(j)}(z) \notag \\ &+\frac{i\omega e^{-i\rho_{i}\bar{\beta}_{n_{i}}z}}{4\sqrt{\bar{P}_{n_{i}}}}\int_{S}\bar{\mathbf{e}}_{n_{i}\rho_{i}}^{*} \cdot \mathbf{P}_{\mathrm{pert}}(\mathbf{r},\omega) dS,~~~i=1,\ldots,4,\end{aligned}$$ where \[dispcoeffgen\] $$\begin{aligned} \label{dispcoeffgenC} C_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}&(z)=\frac{e^{i(\sigma_{j}\bar{\beta}_{m_{j}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{4\sqrt{\bar{P}_{n_{i}}\bar{P}_{m_{j}}}}\notag\\ &\times\int_{S}\left(\bar{\mathbf{e}}_{m_{j}\sigma_{j}}\times\bar{\mathbf{h}}_{n_{i}\rho_{i}}^{*}+\bar{\mathbf{e}}_{n_{i}\rho_{i}}^{*}\times\bar{\mathbf{h}}_{m_{j}\sigma_{j}}\right)\cdot\hat{\mathbf{z}}dS, \\ \label{dispcoeffgenB} B_{n_{i}\rho_{i}}^{i}&=\frac{i}{4\bar{P}_{n_{i}}}\int_{S}\left[\mu_{0}(\omega-\bar{\omega}_{i})\vert\bar{\mathbf{h}}_{n_{i}\rho_{i}}\vert^{2}\right. \notag\\&\left.+(\omega\epsilon_{c}-\bar{\epsilon}_{c}\bar{\omega}_{i})\vert\bar{\mathbf{e}}_{n_{i}\rho_{i}}\vert^{2}\right]dS, \\ \label{dispcoeffgenD} D_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}&(z)=\frac{ie^{i(\sigma_{j}\bar{\beta}_{m_{j}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{4\sqrt{\bar{P}_{n_{i}}\bar{P}_{m_{j}}}} \int_{S}\left[\mu_{0}(\omega-\bar{\omega}_{i})\right.\notag\\&\left.\times\bar{\mathbf{h}}_{m_{j}\sigma_{j}}\cdot\bar{\mathbf{h}}_{n_{i}\rho_{i}}^{*} +(\omega\epsilon_{c}-\bar{\epsilon}_{c}\bar{\omega}_{i})\bar{\mathbf{e}}_{m_{j}\sigma_{j}}\cdot\bar{\mathbf{e}}_{n_{i}\rho_{i}}^{*}\right]dS.\end{aligned}$$ In Eq.  and what follows a prime symbol to a sum means that the summation is taken over all modes, except that with $j=i$, $m_{j}=n_{i}$, and $\sigma_{j}=\rho_{i}$. Moreover, in deriving the l.h.s. of Eq.  we used the orthogonality relation given by Eq. . The time-dependent fields are obtained by integrating over all frequency components contained in the spectra of the system of interacting optical pulses: \[fieldtime\] $$\begin{aligned} \label{fieldtimeE} \mathbf{E}&(\mathbf{r},t)=\frac{1}{2}\int_{0}^{\infty}\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}a_{m_{j}\sigma_{j}}^{(j)}(z,\omega)\frac{\mathbf{e}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}} \notag \\ &\times e^{i(\sigma_{j}\bar{\beta}_{m_{j}}z-\omega t)} d\omega+c.c.\equiv\frac{1}{2}\left[\mathbf{E}^{(+)}(\mathbf{r},t)+\mathbf{E}^{(-)}(\mathbf{r},t)\right], \\ \label{fieldtimeH} \mathbf{H}&(\mathbf{r},t)=\frac{1}{2}\int_{0}^{\infty}\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}a_{m_{j}\sigma_{j}}^{(j)}(z,\omega)\frac{\mathbf{h}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}} \notag \\ &\times e^{i(\sigma_{j}\bar{\beta}_{m_{j}}z-\omega t)}d\omega+c.c.\equiv\frac{1}{2}\left[\mathbf{H}^{(+)}(\mathbf{r},t)+\mathbf{H}^{(-)}(\mathbf{r},t)\right],\end{aligned}$$ where $\mathbf{E}^{(+)}(\mathbf{r},t)$, $\mathbf{H}^{(+)}(\mathbf{r},t)$ and $\mathbf{E}^{(-)}(\mathbf{r},t)$, $\mathbf{H}^{(-)}(\mathbf{r},t)$ are the positive and negative frequency parts of the spectrum, respectively. Let us now introduce the envelopes of the interacting pulses in the time domain, $A_{n_{i}\rho_{i}}^{(i)}(z,t)$, defined as the integral of the mode amplitudes taken over the part of the spectrum that contains only positive frequencies, $$\begin{aligned} \label{env} A_{n_{i}\rho_{i}}^{(i)}(z,t) = \int_{0}^{\infty}a_{n_{i}\rho_{i}}^{(i)}(z,\omega)e^{-i(\omega-\bar{\omega}_{i})t}d\omega.\end{aligned}$$ With this definition, the time-dependent fields given in Eqs.  become: \[fieldtimesimpl\] $$\begin{aligned} \mathbf{E}(\mathbf{r},t)=&\frac{1}{2}\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}A_{m_{j}\sigma_{j}}^{(j)}(z,t)\notag \\ &\times\frac{\bar{\mathbf{e}}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}}e^{i(\sigma_{j}\bar{\beta}_{m_{j}}z-\bar{\omega}_{j} t)}+c.c., \\ \mathbf{H}(\mathbf{r},t)=&\frac{1}{2}\sum_{j=1}^{4}\sum_{m_{j}\sigma_{j}}A_{m_{j}\sigma_{j}}^{(j)}(z,t)\notag \\ &\times\frac{\bar{\mathbf{h}}_{m_{j}\sigma_{j}}(\mathbf{r},\bar{\omega}_{j})}{\sqrt{\bar{P}_{m_{j}}}}e^{i(\sigma_{j}\bar{\beta}_{m_{j}}z-\bar{\omega}_{j} t)}+c.c..\end{aligned}$$ Following the same approach, the time-dependent polarization, too, can be decomposed in two components, which contain positive and negative frequencies, that is, it can be written as: $$\begin{aligned} \label{poltime} \mathbf{P}_{\mathrm{pert}}(\mathbf{r},t)&=\frac{1}{2}\int_{0}^{\infty}\mathbf{P}_{\mathrm{pert}}(\mathbf{r},\omega)e^{-i\omega t}d\omega +c.c.\notag \\ &\equiv\frac{1}{2}\left[\mathbf{P}_{\mathrm{pert}}^{(+)}(\mathbf{r},t)+\mathbf{P}_{\mathrm{pert}}^{(-)}(\mathbf{r},t)\right].\end{aligned}$$ The next step of our derivation is to Fourier transform Eq.  in the time domain. To this end, we first expand the coefficients $B_{n_{i}\rho_{i}}^{i}$ and $D_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}$ in Taylor series, around the carrier frequency $\bar{\omega}_{i}$ \[note that according to Eq. , $C_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}$ is frequency independent\]: \[dispcoeffTs\] $$\begin{aligned} \label{dispcoeffTsB} B_{n_{i}\rho_{i}}^{i}=\sum_{q\geq1}&\frac{(\Delta\omega_{i})^{q}}{q!}\left.\frac{\partial^{q}B_{n_{i}\rho_{i}}^{i}}{\partial\omega^{q}}\right|_{\omega=\bar{\omega}_{i}} \equiv \sum_{q\geq1}\frac{i\beta^{(q)i}_{n_{i}\rho_{i}}}{q!}(\Delta\omega_{i})^{q}, \\ \label{dispcoeffTsD} D_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}&=\sum_{q\geq1}\frac{(\Delta\omega_{i})^{q}}{q!}\left.\frac{\partial^{q}D_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}}{\partial\omega^{q}}\right|_{\omega=\bar{\omega}_{i}}\nonumber \\ &\equiv \sum_{q\geq1}\frac{i\beta^{(q)ij}_{n_{i}\rho_{i},m_{j}\sigma_{j}}}{q!}(\Delta\omega_{i})^{q},\end{aligned}$$ where $\Delta\omega_{i}=\omega-\bar{\omega}_{i}$, $i=1,\ldots,4$. Combining Eqs. , , , and leads to the following expression for the dispersion coefficients, $\beta^{(q)i}_{n_{i}\rho_{i}}$: \[dispcoeffZdep\] $$\begin{aligned} &\beta^{(1)i}_{n_{i}\rho_{i}}(z) = \frac{\delta^{i}_{n_{i}}(z)}{v_{g,n_{i}}^{i}}, \\ &\beta^{(n)i}_{n_{i}\rho_{i}}(z) = \delta^{i}_{n_{i}}(z)\frac{\partial^{n-1}}{\partial\omega^{n-1}}\left(\frac{1}{v_{g,n_{i}}^{i}}\right),~n\geq2,\end{aligned}$$ where $$\begin{aligned} \label{delDef} \delta^{i}_{n_{i}}(z)=&\frac{a}{4W_{m_{i}}}\int_{S}\left[\mu_{0}\vert\mathbf{h}_{n_{i}\rho_{i}}(\mathbf{r},\bar{\omega}_{i})\vert^{2} \right. \notag \\ &+\left.\frac{\partial}{\partial\omega}(\omega\epsilon_{c})\vert\mathbf{e}_{n_{i}\rho_{i}}(\mathbf{r},\bar{\omega}_{i})\vert^{2}\right]dS.\end{aligned}$$ It can be easily seen from this equation that the average of $\delta^{i}_{n_{i}}(z)$ over one lattice cell of the PhC waveguide is equal to 1, *i.e.*, $$\label{delAv} \tilde{\delta}^{i}_{n_{i}}\equiv\frac{1}{a}\int_{z}^{z+a}\delta^{i}_{n_{i}}(z')dz'=1.$$ Here and in what follows the tilde symbol indicates that the corresponding physical quantity has been averaged over a lattice cell of the waveguide. With this notation, Eqs.  become: \[dispcoeffZdepAv\] $$\begin{aligned} &\tilde{\beta}^{(1)i}_{n_{i}\rho_{i}} \equiv \beta^{i}_{1,n_{i}} = \frac{1}{v_{g,n_{i}}^{i}}, \\ &\tilde{\beta}^{(n)i}_{n_{i}\rho_{i}} \equiv \beta^{i}_{n,n_{i}} = \frac{\partial^{n-1} \beta^{i}_{1,n_{i}}}{\partial\omega^{n-1}},~n\geq2.\end{aligned}$$ These relations show that $\tilde{\beta}^{(n)i}_{n_{i}\rho_{i}} = \beta^{i}_{n,n_{i}}$ is the $n$th order dispersion coefficient of the waveguide mode characterized by the parameters $\{n_{i},\rho_{i}\}$, evaluated at $\omega=\bar{\omega}_{i}$. We now multiply Eq.  by $e^{-i(\omega-\bar{\omega}_{i})t}$ and integrate over the positive-frequency domain. These simple calculations lead to the time-domain coupled-mode equations for the field envelopes, $A_{n_{i}\rho_{i}}^{(i)}(z,t)$: $$\begin{aligned} \label{cmetime} &\rho_{i}\frac{\displaystyle \partial A_{n_{i}\rho_{i}}^{(i)}}{\displaystyle \partial z} + \sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum_{m_{j}\sigma_{j}}C_{n_{i}\rho_{i},m_{j}\sigma_{j}}^{ij}e^{-i(\bar{\omega}_{j}-\bar{\omega}_{i})t}\left[\frac{\displaystyle \partial A_{m_{j}\sigma_{j}}^{(j)}}{\displaystyle \partial z}\right. \notag \\ &\left.+i(\sigma_{j}\bar{\beta}_{m_{j}}-\rho_{i}\bar{\beta}_{n_{i}})A_{m_{j}\sigma_{j}}^{(j)}\right]= i\sum_{q\geq1}\frac{\beta^{(q)i}_{n_{i}\rho_{i}}}{q!}\left(i\frac{\partial}{\partial t}\right)^{q}A_{n_{i}\rho_{i}}^{(i)} \notag \\ &+i\sum_{q\geq1}{\sum_{jm_{j}\sigma_{j}}}^{\prime}\frac{\beta^{(q)ij}_{n_{i}\rho_{i},m_{j}\sigma_{j}}}{q!}e^{-i(\bar{\omega}_{j}-\bar{\omega}_{i})t}\left(i\frac{\partial}{\partial t}\right)^{q}A_{n_{j}\rho_{j}}^{(j)}\notag\\ &+\frac{i\bar{\omega}_{i}e^{-i(\rho_{i}\bar{\beta}_{n_{i}}z-\bar{\omega}_{i}t)}}{4\sqrt{\bar{P}_{n_{i}}}}\int_{S}\bar{\mathbf{e}}_{n_{i}\rho_{i}}^{*} \cdot \mathbf{P}_{\mathrm{pert}}^{(+)}(\mathbf{r},t) dS,~i=1,\ldots,4,\end{aligned}$$ The temporal width of the pulses considered in this analysis is much smaller as compared to the nonlinear electronic response time of silicon and therefore the latter can be approximated to be instantaneous. In addition, we assume that the spectra of the interacting pulses are narrow and do not overlap. Under these circumstances, the optical pulses can be viewed as quasi-monochromatic waves and their nonlinear interactions can be treated in the adiabatic limit. Separating the nonlinear optical effects contributing to the nonlinear polarization, one can express in the time domain this polarization as [@bc91book]: $$\begin{aligned} \label{polnltime} &\delta\mathbf{P}_{\mathrm{nl},\bar{\omega}_{i}}(\mathbf{r},t)=\frac{3}{4}\sum_{m_{i}\sigma_{i}}\epsilon_{0}\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots \bar{\mathbf{e}}_{m_{i}\sigma_{i}}(\mathbf{r},\bar{\omega}_{j})\bar{\mathbf{e}}_{m_{i}\sigma_{i}}^{*}(\mathbf{r},\bar{\omega}_{j}) \bar{\mathbf{e}}_{m_{i}\sigma_{i}}(\mathbf{r},\bar{\omega}_{j})\vert A_{m_{i}\sigma_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}\frac{e^{i\sigma_{i}\bar{\beta}_{m_{i}}z}}{\bar{P}_{m_{i}}\sqrt{\bar{P}_{m_{i}}}} \notag \\ &~+\frac{3}{2}\sum_{m_{i}\sigma_{i}}\sum\limits_{\scriptsize{\begin{array}{c} (p_{i}\varrho_{i})\neq(m_{i}\sigma_{i}) \\ p_{i}>m_{i} \end{array}}}\epsilon_{0}\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots \bar{\mathbf{e}}_{p_{i}\varrho_{i}}(\mathbf{r},\bar{\omega}_{i})\bar{\mathbf{e}}_{p_{i}\varrho_{i}}^{*}(\mathbf{r},\bar{\omega}_{i}) \bar{\mathbf{e}}_{m_{i}\sigma_{i}}(\mathbf{r},\bar{\omega}_{i})\vert A_{p_{i}\varrho_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}\frac{e^{i\sigma_{i}\bar{\beta}_{m_{i}}z}}{\bar{P}_{p_{i}}\sqrt{\bar{P}_{m_{i}}}} \notag \\ &~+\frac{3}{2}\sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum\limits_{\scriptsize{\begin{array}{c} m_{i}\sigma_{i} \\ p_{j}\varrho_{j} \end{array}}}\epsilon_{0}\hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{j},\bar{\omega}_{i})\vdots \bar{\mathbf{e}}_{p_{j}\varrho_{j}}(\mathbf{r},\bar{\omega}_{j})\bar{\mathbf{e}}_{p_{j}\varrho_{j}}^{*}(\mathbf{r},\bar{\omega}_{j}) \bar{\mathbf{e}}_{m_{i}\sigma_{i}}(\mathbf{r},\bar{\omega}_{i})\vert A_{p_{j}\varrho_{j}}^{(j)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}\frac{e^{i\sigma_{i}\bar{\beta}_{m_{i}}z}}{\bar{P}_{p_{j}}\sqrt{\bar{P}_{m_{i}}}} \notag \\ &~+\frac{3}{2}\sum\limits_{\scriptsize{\begin{array}{c} p_{j}q_{k}m_{l} \\ \varrho_{j}\tau_{k}\sigma_{l} \end{array}}}\epsilon_{0}\hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{k},\bar{\omega}_{l})\vdots \bar{\mathbf{e}}_{p_{j}\varrho_{j}}(\mathbf{r},\bar{\omega}_{j})\bar{\mathbf{e}}_{q_{k}\tau_{k}}^{*}(\mathbf{r},\bar{\omega}_{k}) \bar{\mathbf{e}}_{m_{l}\sigma_{l}}(\mathbf{r},\bar{\omega}_{l}) A_{p_{j}\varrho_{j}}^{(j)}A_{q_{k}\tau_{k}}^{(k)*}A_{m_{l}\sigma_{l}}^{(l)} \notag\\ &~~~~\left.\times\frac{e^{i\left[(\varrho_{j}\bar{\beta}_{p_{j}}-\tau_{k}\bar{\beta}_{q_{k}}+\sigma_{l}\bar{\beta}_{m_{l}})z -(\bar{\omega}_{j}-\bar{\omega}_{k}+\bar{\omega}_{l})t\right]}}{\sqrt{\bar{P}_{p_{j}}\bar{P}_{q_{k}}\bar{P}_{m_{l}}}} \right\vert_{\scriptsize{\begin{array}{lc} j\neq k\neq l\neq i \\ \bar{\omega}_{j}-\bar{\omega}_{k}+\bar{\omega}_{l}=\bar{\omega}_{i} \end{array}}}.\end{aligned}$$ This expression for the nonlinear polarization accounts for the fact that the nonlinear susceptibility is invariant to frequency permutations. The first term in Eq.  represents SPM effects of the pulse envelopes, the second and third terms describe the XPM between modes with the same frequency and XPM between pulses propagating at different frequencies, respectively, whereas the last term describes FWM processes. If one inserts in Eq.  the linear and nonlinear polarizations given by Eq.  and Eq. , respectively, then discards the fast time-varying terms, one obtains the following system of coupled equations that governs the dynamics of the mode envelopes: $$\begin{aligned} \label{cmetimefin} &\rho_{i}\frac{\displaystyle \partial A_{n_{i}\rho_{i}}^{(i)}}{\displaystyle \partial z} = i\sum_{q\geq1}\frac{\beta^{(q)i}_{n_{i}\rho_{i}}}{q!}\left(i\frac{\partial}{\partial t}\right)^{q}A_{n_{i}\rho_{i}}^{(i)} \notag \\ &+i\sum_{q\geq1}\sum\limits_{\scriptsize{(m_{i}\sigma_{i})\neq(n_{i}\rho_{i})}}\frac{\beta^{(q)ii}_{n_{i}\rho_{i},m_{i}\sigma_{i}}}{q!}\left(i\frac{\partial}{\partial t}\right)^{q}A_{m_{i}\sigma_{i}}^{(i)} \notag \\ &+i\frac{\vartheta_{n_{i}\rho_{i}}^{i}(z)}{v_{g,n_{i}}^{i}}A_{n_{i}\rho_{i}}^{(i)}+i\sum\limits_{\scriptsize{(m_{i}\sigma_{i})\neq(n_{i}\rho_{i})}}\frac{\vartheta_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}(z)}{\sqrt{v_{g,n_{i}}^{i}v_{g,m_{i}}^{i}}}A_{m_{i}\sigma_{i}}^{(i)} \notag \\ &+\frac{3i\bar{\omega}_{i}}{16\epsilon_{0}a^{2}}\Bigg\{\sum_{m_{i}\sigma_{i}}\Bigg[\frac{\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}(z)}{v_{g,m_{i}}^{i}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{m_{i}\sigma_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)} \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} (p_{i}\varrho_{i})\neq(m_{i}\sigma_{i}) \\ p_{i}>m_{i} \end{array}}}\frac{2\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{i}\varrho_{i}}^{i}(z)}{v_{g,p_{i}}^{i}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{p_{i}\varrho_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)} \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum\limits_{\scriptsize{p_{j}\varrho_{j}}}\frac{2\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{j}\varrho_{j}}^{ij}(z)}{v_{g,p_{j}}^{j}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{p_{j}\varrho_{j}}^{(j)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}\Bigg] \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} p_{j}q_{k}m_{l} \\ \varrho_{j}\tau_{k}\sigma_{l} \end{array}}}e^{i\Delta\bar{\beta}_{n_{i}p_{j}q_{k}m_{l}}z} \frac{2\Gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z)}{\sqrt{v_{g,p_{j}}^{j}v_{g,q_{k}}^{k}v_{g,m_{l}}^{l}v_{g,n_{i}}^{i}}}\notag \\ &~~~~~\left.\times A_{p_{j}\varrho_{j}}^{(j)}A_{q_{k}\tau_{k}}^{(k)*}A_{m_{l}\sigma_{l}}^{(l)}\right\vert_{j\neq k\neq l\neq i}\Bigg\},~i=1,\ldots,4,\end{aligned}$$ where $\Delta\bar{\beta}_{n_{i}p_{j}q_{k}m_{l}}=\varrho_{j}\bar{\beta}_{p_{j}}-\tau_{k}\bar{\beta}_{q_{k}}+\sigma_{l}\bar{\beta}_{m_{l}}-\rho_{i}\bar{\beta}_{n_{i}}$ is the wavevector mismatch. The coefficients $\vartheta_{n_{i}\rho_{i}}^{i}$ and $\vartheta_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}$ represent the wavevector shift of the optical mode $(n_{i},\rho_{i})$ and the linear coupling constant between modes $(n_{i},\rho_{i})$ and $(m_{i},\sigma_{i})$, induced by the linear perturbations, respectively, $\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}$ and $\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{i}\varrho_{i}}^{i}$ describe SPM and XPM-induced coupling between modes with the same frequency, $\bar{\omega}_{i}$, respectively, $\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{j}\varrho_{j}}^{ij}$ represents the XPM-induced coupling between modes with frequencies $\bar{\omega}_{i}$ and $\bar{\omega}_{j}$, and $\Gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}$ is related to the FWM interaction among the pulses. All these nonlinear coefficients have the meaning of $z$-dependent effective cubic susceptibilities. The linear and nonlinear coefficients in Eqs.  are given by the following relations: \[couplcoefflinnl\] $$\begin{aligned} \label{couplcoeffLinS} &\vartheta_{n_{i}\rho_{i}}^{i}(z) = \frac{\bar{\omega}_{i}a}{4\bar{W}_{n_{i}}^{i}}\int_{S}[\delta \epsilon_{\mathrm{fc}}(\mathbf{r})+\delta \epsilon_{\mathrm{loss}}(\mathbf{r})]\vert\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})\vert^{2}dS, \\ \label{couplcoeffLinX} &\vartheta_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}(z) = \frac{\bar{\omega}_{i}e^{i(\sigma_{i}\bar{\beta}_{m_{i}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{4\sqrt{\bar{W}_{n_{i}}^{i}\bar{W}_{m_{i}}^{i}}} \notag \\ &~~\times\int_{S}[\delta \epsilon_{\mathrm{fc}}(\mathbf{r})+\delta \epsilon_{\mathrm{loss}}(\mathbf{r})]\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i})\cdot\mathbf{e}_{m_{i}\sigma_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaSPM} &\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}(z) = \frac{\epsilon_{0}^{2}a^{4}e^{i(\sigma_{i}\bar{\beta}_{m_{i}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{\bar{W}_{m_{i}}^{i}\sqrt{\bar{W}_{m_{i}}^{i}\bar{W}_{n_{i}}^{i}}}\int_{S}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots\mathbf{e}_{m_{i}\sigma_{i}}(\bar{\omega}_{i})\mathbf{e}_{m_{i}\sigma_{i}}^{*}(\bar{\omega}_{i})\mathbf{e}_{m_{i}\sigma_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaXPM} &\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{i}\varrho_{i}}^{i}(z) = \frac{\epsilon_{0}^{2}a^{4}e^{i(\sigma_{i}\bar{\beta}_{m_{i}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{\bar{W}_{p_{i}}^{i}\sqrt{\bar{W}_{m_{i}}^{i}\bar{W}_{n_{i}}^{i}}}\int_{S}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots\mathbf{e}_{p_{i}\varrho_{i}}(\bar{\omega}_{i})\mathbf{e}_{p_{i}\varrho_{i}}^{*}(\bar{\omega}_{i})\mathbf{e}_{m_{i}\sigma_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaXPMdiffFreq} &\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{j}\varrho_{j}}^{ij}(z) = \frac{\epsilon_{0}^{2}a^{4}e^{i(\sigma_{i}\bar{\beta}_{m_{i}}-\rho_{i}\bar{\beta}_{n_{i}})z}}{\bar{W}_{p_{j}}^{j}\sqrt{\bar{W}_{m_{i}}^{i}\bar{W}_{n_{i}}^{i}}}\int_{S}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{j},\bar{\omega}_{i})\vdots\mathbf{e}_{p_{j}\varrho_{j}}(\bar{\omega}_{j})\mathbf{e}_{p_{j}\varrho_{j}}^{*}(\bar{\omega}_{j})\mathbf{e}_{m_{i}\sigma_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaFWM} &\Gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z) = \frac{\epsilon_{0}^{2}a^{4}}{\sqrt{\bar{W}_{p_{j}}^{j}\bar{W}_{q_{k}}^{k}\bar{W}_{m_{l}}^{l}\bar{W}_{n_{i}}^{i}}}\int_{S}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{k},\bar{\omega}_{l})\vdots\mathbf{e}_{p_{j}\varrho_{j}}(\bar{\omega}_{j})\mathbf{e}_{q_{k}\tau_{k}}^{*}(\bar{\omega}_{k})\mathbf{e}_{m_{l}\sigma_{l}}(\bar{\omega}_{l})dS.\end{aligned}$$ While Eqs.  seem complicated, in cases of practical interest they can be considerably simplified. To be more specific, these equations describe a multitude of optical effects pertaining to both linear and nonlinear gratings, including linear coupling between modes with the same frequency, nonlinear coupling between modes with the same frequency, due to SPM and XPM effects, XPM-induced coupling between modes with different frequency, and FWM interactions. In most experimental set-ups, however, not all these linear and nonlinear effects occur simultaneously as in a generic case not all of them lead to efficient pulse interactions. These ideas becomes clear if one inspects the exponential factors in Eqs. -. Thus, they vary over a characteristic length comparable to the lattice constant of the PhC, namely much more rapidly as compared to the spatial variation rate of the pulse envelopes. As a result, except for the mode $(n_{i},\rho_{i})$, these linear and nonlinear coefficients cancel. There are, however, particular cases when some of these interactions are phase-matched and consequently are resonantly enhanced. To be more specific, the integrals in Eqs. - are periodic functions of $z$, with period $a$, so that it is possible that a Fourier component of these integrals phase-matches a specific linear or nonlinear interaction between modes (e.g., the linear coupling between two modes with the same frequency and SPM- or XPM-induced nonlinear coupling between modes). In this study, we do not consider such accidental phase-matching of mode interactions. With this in mind, we discard all terms in Eqs.  that average to zero to obtain the final form of the coupled-mode equations for the pulse envelopes: $$\begin{aligned} \label{cmetimefinav} i&\bigg[\rho_{i}\frac{\displaystyle \partial A_{n_{i}\rho_{i}}^{(i)}}{\displaystyle \partial z} +\frac{\delta^{i}_{n_{i}}(z)}{v_{g,n_{i}}^{i}}\frac{\partial A_{n_{i}\rho_{i}}^{(i)}}{\partial t}\bigg]-\frac{\delta^{i}_{n_{i}}(z)\bar{\beta}_{2,n_{i}}}{2}\frac{\partial^{2} A_{n_{i}\rho_{i}}^{(i)}}{\partial t^{2}}\notag \\ &+\frac{\bar{\omega}_{i}\delta n_{\mathrm{fc}}\bar{\kappa}_{n_{i}}^{i}(z)}{n v_{g,n_{i}}^{i}}A_{n_{i}\rho_{i}}^{(i)}+\frac{ic\bar{\kappa}_{n_{i}}^{i}(z)}{2n v_{g,n_{i}}^{i}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{n_{i}\rho_{i}}^{(i)} \notag \\ &+\gamma_{n_{i}\rho_{i}}^{i}(z)\vert A_{n_{i}\rho_{i}}^{(i)}{\vert}^{2}A_{n_{i}\rho_{i}}^{(i)} +\sum\limits_{\scriptsize{\begin{array}{c} (p_{i}\varrho_{i})\neq(n_{i}\rho_{i}) \\ p_{i}>n_{i} \end{array}}}2\gamma_{n_{i}\rho_{i},p_{i}\varrho_{i}}^{i}(z) \notag \\ &\times\vert A_{p_{i}\varrho_{i}}^{(i)}{\vert}^{2}A_{n_{i}\rho_{i}}^{(i)}+\sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum\limits_{\scriptsize{p_{j}\varrho_{j}}}2\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}}^{ij}(z) \vert A_{p_{j}\varrho_{j}}^{(j)}{\vert}^{2}A_{n_{i}\rho_{i}}^{(i)} \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} p_{j}q_{k}m_{l} \\ \varrho_{j}\tau_{k}\sigma_{l} \end{array}}}2e^{i\Delta\bar{\beta}_{n_{i}p_{j}q_{k}m_{l}}z}\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z) \notag \\ &~~~~~\left.\times A_{p_{j}\varrho_{j}}^{(j)}A_{q_{k}\tau_{k}}^{(k)*}A_{m_{l}\sigma_{l}}^{(l)}\right\vert_{j\neq k\neq l\neq i}=0,~i=1,\ldots,4,\end{aligned}$$ where the new parameters introduced in this equation are defined as: \[couplcoefflinnlav\] $$\begin{aligned} \label{couplcoeffLinS} &\bar{\kappa}_{n_{i}}^{i}(z) = \frac{\epsilon_{0}an^{2}}{2\bar{W}_{n_{i}}^{i}}\int_{S_{\mathrm{nl}}}\vert\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})\vert^{2}dS, \\ \label{GammaSPMav} &\gamma_{n_{i}\rho_{i}}^{i}(z) = \frac{3\bar{\omega}_{i}\epsilon_{0}a^{2}}{16v_{g,n_{i}}^{i^{\scriptstyle 2}}}\frac{1}{\bar{W}_{n_{i}}^{i^{\scriptstyle 2}}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i})\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaXPMav} &\gamma_{n_{i}\rho_{i},p_{i}\varrho_{i}}^{i}(z) = \frac{3\bar{\omega}_{i}\epsilon_{0}a^{2}}{16v_{g,n_{i}}^{i}v_{g,p_{i}}^{i}}\frac{1}{\bar{W}_{n_{i}}^{i}\bar{W}_{p_{i}}^{i}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{i},-\bar{\omega}_{i},\bar{\omega}_{i})\vdots\mathbf{e}_{p_{i}\varrho_{i}}(\bar{\omega}_{i})\mathbf{e}_{p_{i}\varrho_{i}}^{*}(\bar{\omega}_{i})\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaXPMdiffFreqav} &\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}}^{ij}(z) = \frac{3\bar{\omega}_{i}\epsilon_{0}a^{2}}{16v_{g,n_{i}}^{i}v_{g,p_{j}}^{j}}\frac{1}{\bar{W}_{n_{i}}^{i}\bar{W}_{p_{j}}^{j}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{j},\bar{\omega}_{i})\vdots\mathbf{e}_{p_{j}\varrho_{j}}(\bar{\omega}_{j})\mathbf{e}_{p_{j}\varrho_{j}}^{*}(\bar{\omega}_{j})\mathbf{e}_{n_{i}\rho_{i}}(\bar{\omega}_{i})dS, \\ \label{GammaFWMav} &\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z) = \frac{3\bar{\omega}_{i}\epsilon_{0}a^{2}}{16{(v_{g,p_{j}}^{j}v_{g,q_{k}}^{k}v_{g,m_{l}}^{l}v_{g,n_{i}}^{i})}^{\frac{1}{2}}} \notag \\ &~~\frac{1}{{(\bar{W}_{p_{j}}^{j}\bar{W}_{q_{k}}^{k}\bar{W}_{m_{l}}^{l}\bar{W}_{n_{i}}^{i})}^{\frac{1}{2}}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{n_{i}\rho_{i}}^{*}(\bar{\omega}_{i})\cdot \hat{\chi}^{(3)}(\bar{\omega}_{j},-\bar{\omega}_{k},\bar{\omega}_{l}) \notag \\ &~~\vdots\mathbf{e}_{p_{j}\varrho_{j}}(\bar{\omega}_{j})\mathbf{e}_{q_{k}\tau_{k}}^{*}(\bar{\omega}_{k})\mathbf{e}_{m_{l}\sigma_{l}}(\bar{\omega}_{l})dS.\end{aligned}$$ In these equations, $S_{\mathrm{nl}}(z)$ is the transverse surface of the region filled with nonlinear material. Note that the exponential factor in the term describing the FWM does not average to zero because the FWM interaction is assumed to be nearly phase-matched and therefore the exponential factor varies over a characteristic length that is much larger than the lattice constant, $a$. Importantly, the linear and nonlinear effects in Eq.  appear as being inverse proportional to the $v_{g}$ and $v_{g}^{2}$, respectively. In other words, one does not need to rely on any phenomenological considerations to describe slow-light effects, as they are naturally captured by our model. Carriers dynamics ----------------- The last step in our derivation of the theoretical model describing FWM in Si-PhCWGs is to determine the influence of photogenerated FCs on pulse dynamics. To this end, we first find the rate at which electron-hole pairs are generated optically, *via* degenerate and nondegenerate TPA, and as a result of FWM. More specifically, we first multiply Eqs. , after all linear terms have been discarded, by $A_{n_{i}\rho_{i}}^{(i)*}$, then multiply the complex conjugate of Eqs.  by $A_{n_{i}\rho_{i}}^{(i)}$, and sum the results over all carrier frequencies and modes. The outcome of these simple manipulations can be cast as: $$\begin{aligned} \label{PowTPAXAMFWMtime} &\frac{\displaystyle \partial}{\displaystyle \partial z}\sum_{i=1}^{4}\sum_{n_{i}\rho_{i}}\rho_{i}\vert A_{n_{i}\rho_{i}}^{(i)}\vert^{2}=- \frac{3}{8\epsilon_{0}a^{2}}\sum_{i=1}^{4}\sum_{n_{i}\rho_{i}}\bar{\omega}_{i}\notag \\ &\times\mathfrak{Im}\Bigg\{\sum_{m_{i}\sigma_{i}}\Bigg[\frac{\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}}^{i}(z)}{v_{g,m_{i}}^{i}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{m_{i}\sigma_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}A_{n_{i}\rho_{i}}^{(i)*} \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} (p_{i}\varrho_{i})\neq(m_{i}\sigma_{i}) \\ p_{i}>m_{i} \end{array}}}\frac{2\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{i}\varrho_{i}}^{i}(z)}{v_{g,p_{i}}^{i}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{p_{i}\varrho_{i}}^{(i)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}A_{n_{i}\rho_{i}}^{(i)*} \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum\limits_{\scriptsize{p_{j}\varrho_{j}}}\frac{2\Gamma_{n_{i}\rho_{i},m_{i}\sigma_{i}p_{j}\varrho_{j}}^{ij}(z)}{v_{g,p_{j}}^{j}\sqrt{v_{g,m_{i}}^{i}v_{g,n_{i}}^{i}}} \vert A_{p_{j}\varrho_{j}}^{(j)}{\vert}^{2}A_{m_{i}\sigma_{i}}^{(i)}A_{n_{i}\rho_{i}}^{(i)*}\Bigg] \notag \\ &+\sum\limits_{\scriptsize{\begin{array}{c} p_{j}q_{k}m_{l} \\ \varrho_{j}\tau_{k}\sigma_{l} \end{array}}}e^{i\Delta\bar{\beta}_{n_{i}p_{j}q_{k}m_{l}}z} \frac{2\Gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z)}{\sqrt{v_{g,p_{j}}^{j}v_{g,q_{k}}^{k}v_{g,m_{l}}^{l}v_{g,n_{i}}^{i}}}\notag \\ &~~~~~\left.\times A_{p_{j}\varrho_{j}}^{(j)}A_{q_{k}\tau_{k}}^{(k)*}A_{m_{l}\sigma_{l}}^{(l)}A_{n_{i}\rho_{i}}^{(i)*}\right\vert_{j\neq k\neq l\neq i}\Bigg\}.\end{aligned}$$ The sum in the l.h.s. of this equation represents the rate at which optical power is transferred to FCs. This power is absorbed by carriers generated in the silicon slab, in the infinitesimal volume $dV(z)=A_{\mathrm{nl}}(z)dz$, where $A_{\mathrm{nl}}(z)$ is an effective area. This area is defined in terms of the Poynting vector of the field propagating inside the silicon slab, $$\begin{aligned} \label{effArea} A_{\mathrm{nl}}(z) = \frac{\big[\int_{S_{\mathrm{nl}}}\vert\langle \mathbf{E}(\mathbf{r},t)\times\mathbf{H}(\mathbf{r},t)\rangle_{t}\vert dS\big]^{2}}{\int_{S_{\mathrm{nl}}}\vert\langle \mathbf{E}(\mathbf{r},t)\times\mathbf{H}(\mathbf{r},t)\rangle_{t}\vert^{2} dS}.\end{aligned}$$ In this equation, $\langle f\rangle_{t}$ means the time average of $f$. Using Eq. , and taking into account the fact that $A_{n_{i}\rho_{i}}^{(i)}$ varies in time much slower than $e^{-i\bar{\omega}_{i} t}$, one can express Eq.  in the following form: $$\begin{aligned} \label{effAreaSimpl} A_{\mathrm{nl}}(z) = \frac{\displaystyle \bigg[\int_{S_{\mathrm{nl}}}\bigg\lvert\sum_{i=1}^{4}\sum_{n_{i}\rho_{i}}\frac{\vert A_{n_{i}\rho_{i}}^{(i)}\vert^{2}}{\bar{P}_{n_{i}}} \mathfrak{Re}(\mathbf{e}_{n_{i}\rho_{i}}\times\mathbf{h}_{n_{i}\rho_{i}}^{*})\bigg\rvert dS\bigg]^{2}}{\displaystyle \int_{S_{\mathrm{nl}}}\bigg\lvert\sum_{i=1}^{4}\sum_{n_{i}\rho_{i}}\frac{\vert A_{n_{i}\rho_{i}}^{(i)}\vert^{2}}{\bar{P}_{n_{i}}} \mathfrak{Re}(\mathbf{e}_{n_{i}\rho_{i}}\times\mathbf{h}_{n_{i}\rho_{i}}^{*})\bigg\rvert^{2} dS}.\end{aligned}$$ In spite of the fact that it might seem difficult to use this formula to calculate the effective area, we will show in the next section that in cases of practical interest it can be simplified considerably. We also stress that Eq.  gives the effective transverse area of the region in which FCs are generated, so that it should not be confused with the modal effective area. In fact, since in the FWM process there are several co-propagating beams, a single effective modal area is not well defined. The energy transferred to FCs when an electron-hole pair is generated *via* absorption of two photons with frequencies $\bar{\omega}_{i}$ and $\bar{\omega}_{j}$ is equal to $\hbar(\bar{\omega}_{i}+\bar{\omega}_{j})$. Using this result and neglecting again all terms in Eq.  that average to zero, it can be easily shown that the carriers dynamics are governed by the following rate equation: $$\begin{aligned} \label{FCdyn} &\frac{\partial N}{\partial t} = -\frac{N}{\tau_{c}}+\frac{1}{\hbar A_{\mathrm{nl}}(z)}\sum_{i=1}^{4}\sum_{n_{i}\rho_{i}}\bigg\{\frac{\gamma_{n_{i}\rho_{i}}^{\prime\prime i}(z)}{\bar{\omega}_{i}}\vert A_{n_{i}\rho_{i}}^{(i)}{\vert}^{4} \notag \\ &~~~+\sum\limits_{\scriptsize{\begin{array}{c} (p_{i}\varrho_{i})\neq(n_{i}\rho_{i}) \\ p_{i}>n_{i} \end{array}}}\frac{2\gamma_{n_{i}\rho_{i},p_{i}\varrho_{i}}^{\prime\prime i}(z)}{\bar{\omega}_{i}}\vert A_{p_{i}\varrho_{i}}^{(i)}{\vert}^{2}\vert A_{n_{i}\rho_{i}}^{(i)}{\vert}^{2} \notag \\ &~~~+\sum\limits_{\scriptsize{\begin{array}{c} j=1 \\ j\neq i \end{array}}}^{4}\sum\limits_{\scriptsize{p_{j}\varrho_{j}}}\frac{4\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}}^{\prime\prime ij}(z)}{\bar{\omega}_{i}+\bar{\omega}_{j}} \vert A_{p_{j}\varrho_{j}}^{(j)}{\vert}^{2}\vert A_{n_{i}\rho_{i}}^{(i)}{\vert}^{2} \notag \\ &~~~+\sum\limits_{\scriptsize{\begin{array}{c} p_{j}q_{k}m_{l} \\ \varrho_{j}\tau_{k}\sigma_{l} \end{array}}}\mathfrak{Im}\bigg[e^{i\Delta\bar{\beta}_{n_{i}p_{j}q_{k}m_{l}}z}\frac{4\gamma_{n_{i}\rho_{i},p_{j}\varrho_{j}q_{k}\tau_{k}m_{l}\sigma_{l}}^{jkl}(z)}{\bar{\omega}_{i}+\bar{\omega}_{k}} \notag \\ &~~~~~~~\left.\times A_{p_{j}\varrho_{j}}^{(j)}A_{q_{k}\tau_{k}}^{(k)*}A_{m_{l}\sigma_{l}}^{(l)}A_{n_{i}\rho_{i}}^{(i)*}\right\vert_{j\neq k\neq l\neq i}\bigg]\bigg\},\end{aligned}$$ where $\tau_{c}\approx\SI{500}{\pico\second}$ [@mch09oe] is the FC recombination time in Si-PhCWGs and $\zeta^{\prime}$ ($\zeta^{\prime\prime}$) means the real (imaginary) part of the complex number, $\zeta$. Degenerate four-wave mixing {#sDFWM} =========================== The system of coupled nonlinear partial differential equations, Eqs.  and Eq. , fully describes the FWM of optical pulses and FCs dynamics and represents the main result derived in this study. In practical experimental set-ups, however, the most used pulse configuration is that of degenerate FWM. In this particular case, the optical frequencies of the two pump pulses are the same, $\bar{\omega}_{1}=\bar{\omega}_{2}\equiv\omega_{p}$, whereas the two generated pulses, the signal and the idler, have frequencies $\bar{\omega}_{3}\equiv\omega_{s}$ and $\bar{\omega}_{4}\equiv\omega_{i}$, respectively. Moreover, we assume that all modes are forward-propagating modes and that at each carrier frequency there is only one guided mode in which the optical pulses that enter in the FWM process can propagate – others, should they exist, would not be phase-matched – so that we set $N_{i}=1$, $i=1,\ldots,4$. Under these circumstances, Eqs.  and Eq.  can be simplified to: \[FWMpsi\] $$\begin{aligned} \label{FWMp} i&\bigg[\frac{\displaystyle \partial A_{p}}{\displaystyle \partial z} +\frac{\delta_{p}(z)}{v_{g,p}}\frac{\partial A_{p}}{\partial t}\bigg]-\frac{\delta_{p}(z)\bar{\beta}_{2,p}}{2}\frac{\partial^{2} A_{p}}{\partial t^{2}}\notag \\ &+\frac{\omega_{p}\delta n_{\mathrm{fc}}\bar{\kappa}_{p}(z)}{n v_{g,p}}A_{p}+\frac{ic\bar{\kappa}_{p}(z)}{2n v_{g,p}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{p} \notag \\ &+\left[\gamma_{p}(z)\vert A_{p}{\vert}^{2}+2\gamma_{ps}(z)\vert A_{s}{\vert}^{2}+2\gamma_{pi}(z)\vert A_{i}{\vert}^{2}\right]A_{p} \notag \\ &+2e^{i\Delta\bar{\beta}z}\gamma_{psi}(z)A_{s}A_{i}A_{p}^{*}=0, \\\label{FWMs} i&\bigg[\frac{\displaystyle \partial A_{s}}{\displaystyle \partial z} +\frac{\delta_{s}(z)}{v_{g,s}}\frac{\partial A_{s}}{\partial t}\bigg]-\frac{\delta_{s}(z)\bar{\beta}_{2,s}}{2}\frac{\partial^{2} A_{s}}{\partial t^{2}}\notag \\ &+\frac{\omega_{s}\delta n_{\mathrm{fc}}\bar{\kappa}_{s}(z)}{n v_{g,s}}A_{s}+\frac{ic\bar{\kappa}_{s}(z)}{2n v_{g,s}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{s} \notag \\ &+\left[\gamma_{s}(z)\vert A_{s}{\vert}^{2}+2\gamma_{sp}(z)\vert A_{p}{\vert}^{2}+2\gamma_{si}(z)\vert A_{i}{\vert}^{2}\right]A_{s} \notag \\ &+e^{-i\Delta\bar{\beta}z}\gamma_{spi}(z)A_{p}^{2}A_{i}^{*}=0, \\\label{FWMi} i&\bigg[\frac{\displaystyle \partial A_{i}}{\displaystyle \partial z} +\frac{\delta_{i}(z)}{v_{g,i}}\frac{\partial A_{i}}{\partial t}\bigg]-\frac{\delta_{i}(z)\bar{\beta}_{2,i}}{2}\frac{\partial^{2} A_{i}}{\partial t^{2}}\notag \\ &+\frac{\omega_{i}\delta n_{\mathrm{fc}}\bar{\kappa}_{i}(z)}{n v_{g,i}}A_{i}+\frac{ic\bar{\kappa}_{i}(z)}{2n v_{g,i}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{i} \notag \\ &+\left[\gamma_{i}(z)\vert A_{i}{\vert}^{2}+2\gamma_{ip}(z)\vert A_{p}{\vert}^{2}+2\gamma_{is}(z)\vert A_{s}{\vert}^{2}\right]A_{i} \notag \\ &+e^{-i\Delta\bar{\beta}z}\gamma_{ips}(z)A_{p}^{2}A_{s}^{*}=0,\end{aligned}$$ $$\begin{aligned} \label{FWMFCdyn} &\frac{\partial N}{\partial t} = -\frac{N}{\tau_{c}}+\frac{1}{\hbar A_{\mathrm{nl}}(z)}\bigg\{\sum_{\mu=p,s,i}\bigg[\frac{\gamma_{\mu}^{\prime\prime}(z)}{\omega_{\mu}}\vert A_{\mu}{\vert}^{4} \notag \\ &~+\sum\limits_{\scriptsize{\begin{array}{c} \nu=p,s,i \\ \nu\neq \mu \end{array}}}\frac{4\gamma_{\mu\nu}^{\prime\prime}(z)}{\omega_{\mu}+\omega_{\nu}} \vert A_{\mu}{\vert}^{2}\vert A_{\nu}{\vert}^{2}\bigg]+\frac{1}{\omega_{p}}\mathfrak{Im}\big[2\gamma_{psi}(z) \notag \\ &~\times {A_{p}^{*}}^{2}A_{s}A_{i}e^{i\Delta\bar{\beta}z}+[\gamma_{spi}(z)+\gamma_{ips}(z)] A_{p}^{2}A_{s}^{*}A_{i}^{*}e^{-i\Delta\bar{\beta}z}\big]\big\},\end{aligned}$$ where $\Delta\bar{\beta}=\beta_s+\beta_i-2\beta_p$. The coefficients of the linear and nonlinear terms in Eqs.  and Eq.  are: ![Energy diagrams representing the nonlinear optical processes included in Eqs. . (a) SPM and degenerate TPA corresponding to $\gamma_{\mu}^{\prime}$ and $\gamma_{\mu}^{\prime\prime}$, respectively. (b) XPM and XAM corresponding to $\gamma_{\mu\nu}^{\prime}$ and $\gamma_{\mu\nu}^{\prime\prime}$, respectively. Two possible ways of energy transfer that can occur during a degenerate FWM process: (c) two pump photons generate a signal and an idler photon, a process described by $\gamma_{psi}$; (d) the reverse process, described by $\gamma_{ips}$ and $\gamma_{spi}$, in which a signal and an idler photon generate two pump photons.[]{data-label="fig:endiagr"}](Figure4.eps){width="8cm"} \[FWMcouplcoefflinnlav\] $$\begin{aligned} \label{FWMcouplcoeffLinS} &\bar{\kappa}_{\mu}(z) = \frac{\epsilon_{0}an^{2}}{2\bar{W}_{\mu}}\int_{S_{\mathrm{nl}}}\vert\mathbf{e}_{\mu}(\omega_{\mu})\vert^{2}dS, \\ \label{FWMGammaSPMav} &\gamma_{\mu}(z) = \frac{3\omega_{\mu}\epsilon_{0}a^{2}}{16v_{g,\mu}^{2}}\frac{1}{\bar{W}_{\mu}^{2}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{\mu}^{*}(\omega_{\mu}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\omega_{\mu},-\omega_{\mu},\omega_{\mu})\vdots\mathbf{e}_{\mu}(\omega_{\mu})\mathbf{e}_{\mu}^{*}(\omega_{\mu})\mathbf{e}_{\mu}(\omega_{\mu})dS, \\ \label{FWMGammaXPMdiffFreqav} &\gamma_{\mu\nu}(z) = \frac{3\omega_{\mu}\epsilon_{0}a^{2}}{16v_{g,\mu}v_{g,\nu}}\frac{1}{\bar{W}_{\mu}\bar{W}_{\nu}}\int_{S_{\mathrm{nl}}}\mathbf{e}_{\mu}^{*}(\omega_{\mu}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\omega_{\nu},-\omega_{\nu},\omega_{\mu})\vdots\mathbf{e}_{\nu}(\omega_{\nu})\mathbf{e}_{\nu}^{*}(\omega_{\nu})\mathbf{e}_{\mu}(\omega_{\mu})dS, \\ \label{FWMpsiGammaFWMav} &\gamma_{psi}(z) = \frac{3\omega_{p}\epsilon_{0}a^{2}}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{S_{\mathrm{nl}}}\mathbf{e}_{p}^{*}(\omega_{p})\cdot \hat{\chi}^{(3)}(\omega_{s},-\omega_{p},\omega_{i})\vdots\mathbf{e}_{s}(\omega_{s})\mathbf{e}_{p}^{*}(\omega_{p})\mathbf{e}_{i}(\omega_{i})dS, \\ \label{FWMspiGammaFWMav} &\gamma_{spi}(z) = \frac{3\omega_{s}\epsilon_{0}a^{2}}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{S_{\mathrm{nl}}}\mathbf{e}_{s}^{*}(\omega_{s})\cdot \hat{\chi}^{(3)}(\omega_{p},-\omega_{i},\omega_{p})\vdots\mathbf{e}_{p}(\omega_{p})\mathbf{e}_{i}^{*}(\omega_{i})\mathbf{e}_{p}(\omega_{p})dS, \\ \label{FWMipsGammaFWMav} &\gamma_{ips}(z) = \frac{3\omega_{i}\epsilon_{0}a^{2}}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{S_{\mathrm{nl}}}\mathbf{e}_{i}^{*}(\omega_{i})\cdot \hat{\chi}^{(3)}(\omega_{p},-\omega_{s},\omega_{p})\vdots\mathbf{e}_{p}(\omega_{p})\mathbf{e}_{s}^{*}(\omega_{s})\mathbf{e}_{p}(\omega_{p})dS,\end{aligned}$$ where $\mu$ and $\nu\neq\mu$ take one of the values $p$, $s$, and $i$ and the frequency degeneracy at the pump frequency has been taken into account. Note that, as expected, when the nonlinear coefficients $\gamma$’s are real quantities, namely when nonlinear optical absorption effects can be neglected, the optical pumping term in Eq.  vanishes. Moreover, since in experiments usually $P_{p}\gg P_{s},P_{i}$, the effective area given by Eq.  can be reduced to the following simplified form: $$\label{effAreaSimplSW} A_{\mathrm{nl}}(z) = \frac{\displaystyle \left(\int_{S_{\mathrm{nl}}}\big\vert \mathfrak{Re}\left[\mathbf{e}_{p}(\omega_{p})\times\mathbf{h}_{p}^{*}(\omega_{p})\right]\big\vert dS\right)^{2}}{\displaystyle \int_{S_{\mathrm{nl}}}\big\vert \mathfrak{Re}\left[\mathbf{e}_{p}(\omega_{p})\times\mathbf{h}_{p}^{*}(\omega_{p})\right]\big\vert^{2} dS}.$$ The types of nonlinear interactions incorporated in our theoretical model described by Eqs.  are summarized in Fig. \[fig:endiagr\] *via* the energy diagrams defined by the frequencies of the specific pairs of interacting photons. Thus, as per Fig. \[fig:endiagr\](a), the terms proportional to the $\gamma_{\mu}^{\prime}$ and $\gamma_{\mu}^{\prime\prime}$ coefficients describe SPM and degenerate TPA effects, respectively, whereas Fig. \[fig:endiagr\](b) illustrates XPM and XAM (also called nondegenerate TPA) interactions whose strength is proportional to $\gamma_{\mu\nu}^{\prime}$ and $\gamma_{\mu\nu}^{\prime\prime}$, respectively. Finally, there are two distinct types of FWM processes, represented in Fig. \[fig:endiagr\](c) and Fig. \[fig:endiagr\](d). In the first case two pump photons combine and generate a pair of photons, one at the signal frequency and the other one at the idler, a process described by the term proportional to $\gamma_{psi}$. The reverse process, represented by the $\gamma_{ips}$ and $\gamma_{spi}$ terms, corresponds to the case in which a signal and an idler photon combine to generate a pair of photons at the pump frequency. As Eqs.  show, the linear and nonlinear optical coefficients of the waveguide depend on the index of refraction of silicon, both explicitly and implicitly *via* the optical modes of the waveguide. In our calculations the implicit modal frequency dispersion is not taken into account because it cannot be incorporated in the PWE method used to compute the modes. On the other hand, the explicit material dispersion is accounted for *via* the following Sellmeier equation describing the frequency dependence of the index of refraction of silicon [@p98book]: $$n^2(\lambda)=\epsilon+\frac{A}{\lambda^2}+\frac{B\lambda_1^2}{\lambda_1^2-\lambda^2}, \label{index}$$ where $\lambda_1=\SI{1.1071}{\micro\meter}$, $\epsilon=11.6858$, $A=\SI{0.939816}{\square\micro\meter}$, and $B=\num{8.10461e-3}$. ![(a) Dependence of $A_{\mathrm{nl}}$ on $z$, determined for the odd (solid line) and even (dashed line) modes for several values of the group-index, $n_{g}$. (b) Frequency dispersion of $\tilde{A}_{\mathrm{nl}}$ calculated for the two modes, in the spectral domain where they are guiding modes.[]{data-label="fig:area"}](Figure5.eps){width="8cm"} The system of coupled equations, Eqs.  and Eq. , form the basis for our analysis of degenerate FWM in silicon PhC waveguides. In our simulations, based on numerical integration of this system of equations using a standard split-step Fourier method combined with a fifth-order Runge-Kuta method for the integration of the linear, carriers dependent terms, the $z$-dependence of the coefficients in these equations is rigorously taken into account. However, one can significantly decrease the simulation time by averaging these fast-varying coefficients over a lattice constant, as this way the integration step for the resulting, averaged system can be increased considerably. The derivation of this averaged model is presented in the Appendix. ![(a) Dependence of $\kappa$ on $z$, determined for the odd (solid line) and even (dashed line) modes for several values of the group-index, $n_{g}$. (b) Frequency dispersion of $\tilde{\kappa}$ calculated for the two modes, in the spectral domain where they are guiding modes.[]{data-label="fig:kappa"}](Figure6.eps){width="8cm"} One of the key differences between our theoretical description of FWM processes in Si-PhCWGs and the widely used models for FWM in waveguides with uniform cross-section, such as optical fibers or silicon photonic wires, is that the linear and nonlinear waveguide coefficients are periodic functions of the distance along the waveguide. In what follows, we discuss this feature of the FWM in more detail, starting with the effective area, $A_{\mathrm{nl}}$, defined by Eq. . The dependence of this area on the longitudinal distance, $z$, is presented in Fig. \[fig:area\](a), where $z$ spans the length of a unit cell. As we have discussed, a physical characteristic of slow-light modes is their increased spatial extent. This property is clearly illustrated in Fig. \[fig:area\](a), which shows that in the case of the even and odd modes the effective area increases by almost a factor of two when the group index varies from and from , respectively. This property is also illustrated by the frequency dispersion of the effective area, averaged over a unit cell, as per Fig. \[fig:area\](b). Thus, it can be seen in this figure that the effective area has a maximum at $k_z\approx 0.3(2\pi/a)$ for the even mode and at the edge of the Brillouin zone for both modes, namely in the regions of slow light indeed. ![Dependence of $\delta$ on $z$, determined for several values of $n_{g}$. Solid and dashed lines correspond to the odd and even mode, respectively.[]{data-label="fig:delta"}](Figure7.eps){width="8cm"} The $z$-dependence of the spatial mode overlap, $\kappa$, and the frequency dispersion of its spatial average over a unit cell, $\tilde{\kappa}$, are plotted in Figs. \[fig:kappa\](a) and \[fig:kappa\](b), respectively. These figures show that the mode overlap varies more strongly with $z$ in the case of the even mode, whereas in both cases the mode overlap variation increases as the group-index, $n_{g}$, increases. Interestingly enough, the averaged overlap coefficient of the even mode has a maximum at $k_z\approx 0.3(2\pi/a)$, i.e. $\lambda\approx\SI{1.52}{\micro\meter}$, which coincides with a minimum of its $v_{g}$. Note also that whereas $\kappa(z)$ can be larger than unity within the unit cell, its average, $\tilde{\kappa}<1$. This result is expected because $\tilde{\kappa}$ quantifies the mode overlap with the slab waveguide. ![(a), (b) Dependence of $\gamma^{\prime}$ and $\gamma^{\prime\prime}$ on $z$, respectively, determined for $k_z=0.35(2\pi/a)$. (c), (d) Frequency dispersion of spatially averaged values of $\gamma^{\prime}(z)$ and $\gamma^{\prime\prime}(z)$, respectively, determined both for the even and odd modes.[]{data-label="fig:gamma"}](Figure8.eps){width="8cm"} In Fig. \[fig:delta\] we present the dependence on $z$ of another physical quantity that characterizes the linear optical properties of the PhC waveguide, namely its dispersive properties. This parameter, $\delta$, quantifies the extent to which the $z$-dependent dispersion coefficients differ from their averaged values. Similarly to the mode overlap coefficient $\kappa$, $\delta(z)$ shows a more substantial changes with $z$ in the case of the even mode as compared to the odd one and an increase of the amplitude of these oscillations with the increase of $n_{g}$. Moreover, as it has been demonstrated in the preceding section, the average of $\delta(z)$ over a unit cell is equal to unity. The $z$-dependence of the nonlinear waveguide coefficient that characterizes the strength of SPM and TPA effects and the wavelength dependence of its average over a unit cell are plotted in the top and bottom panels of Fig. \[fig:gamma\], respectively. One relevant result illustrated by these plots is that the nonlinear waveguide coefficient increases considerably as the GV of the optical mode is tuned to the slow-light regime. Indeed, for the case presented in Figs. \[fig:gamma\](a) and \[fig:gamma\](b) the group-index of the odd and even modes are $n_{g}=6.4$ and $n_{g}=11.5$, respectively. This phenomenon is better illustrated by the wavelength dependence of the spatially averaged values of $\gamma^{\prime}(z)$ and $\gamma^{\prime\prime}(z)$, which are shown in Figs. \[fig:gamma\](c) and \[fig:gamma\](d), respectively. Thus, these plots indicate that the nonlinear waveguide coefficient increases by more than an order of magnitude as the wavelength is tuned from the fast-light to the slow-light regime, the nonlinear interactions being enhanced correspondingly. Phase-matching condition {#sPMC} ======================== Before we use the theoretical model we developed to investigate the properties of FWM in Si-PhCWGs, we derive and discuss the conditions in which optimum nonlinear pulse interaction can be achieved. In particular, efficient FWM is achieved when the interacting pulses are phase-matched, namely when the total (linear plus nonlinear) wavevector mismatch is equal to zero. In the most general case this phase-matching condition depends in an intricate way on the peak power of the pump, $P_{p}$, signal, $P_{s}$, and idler, $P_{i}$, as well as on the linear and nonlinear coefficients of the waveguide [@a13book]. This complicated relation takes a very simple form when one considers an experimental set-up most used in practice, namely when the pump is much stronger than the signal and idler, $P_{p}\gg P_{s},P_{i}$. Under these circumstances, the phase-matching condition can be expressed as: $$\label{phase1} 2\gamma_{p}^{\prime} P_p-2\beta_p+\beta_s+\beta_i=0.$$ In order to determine the corresponding wavelengths of the optical pulses, this relation must be used in conjunction with the energy conservation relation, that is $2\omega_p=\omega_s+\omega_i$. An alternative phase-matching condition, less accurate but easier to use in practice, can be derived by expanding the propagation constants, $\beta_{s,i}(\omega)$, in Taylor series around the pump frequency, $\omega_p$: ![(a), (b) Wavelength diagrams defined by Eq.  and Eq. , respectively. In both panels dashed lines correspond to $\omega_p=\omega_{s}=\omega_{i}$.[]{data-label="fig:wavediagr"}](Figure9.eps){width="8cm"} $$\begin{aligned} \label{taylor} \beta_{s,i}(\omega)&=\sum_{n\geq0}\frac{(\omega-\omega_p)^{n}}{n!}\left.\frac{d^{n}\beta_{s,i}}{d\omega^{n}}\right|_{\omega=\omega_{p}}.\end{aligned}$$ Inserting these expressions in Eq.  and neglecting all terms beyond the fourth-order, one arrives to the following relation: $$\label{phase2} 2\gamma_{p}^{\prime}P_p+\beta_{2p}(\Delta\omega)^2+\frac{1}{12}\beta_{4p}(\Delta\omega)^4=0,$$ where $\Delta \omega\equiv\vert\omega_p-\omega_s\vert=\vert\omega_p-\omega_i\vert$. The wavelength diagrams presented in Figs. \[fig:wavediagr\](a) and \[fig:wavediagr\](b) display the triplets of wavelengths for which the phase-matching conditions expressed by Eq.  and Eq. , respectively, are satisfied. These wavelength diagrams were calculated only for the even mode because only this mode possesses spectral regions with anomalous dispersion \[cf. Fig. \[fig:disp\](b)\], which is a prerequisite condition for phase-matching the FWM. More specifically, efficient FWM can be achieved if the pump wavelength ranges from $\lambda_p=$ . Moreover, the diagrams in Fig. \[fig:wavediagr\] show that the predictions based on Eq.  and Eq.  are in good agreement, especially when $\Delta\omega$ is small. They start to agree less as $\Delta\omega$ increases because the contribution of the terms discarded when the series expansion of $\beta_{s,i}(\omega)$ is truncated increases as $\Delta\omega$ increases. Figure \[fig:wavediagr\] also suggests that the spectral domain in which efficient FWM is achieved depends on the pump power, $P_{p}$. To be more specific, it can be seen that for $P_{p}\lesssim\SI{0.7}{\watt}$, a spectral gap opens where the phase-matching condition cannot be satisfied. The spectral width of this gap increases when $P_{p}$ decreases as the the waveguide was not designed to possess phase-matched modes in the linear regime. Moreover, the diagrams presented in Fig. \[fig:wavediagr\] show that in the fast-light regime the wavelengths defined by the phase-matching condition depend only slightly on $P_{p}$, whereas a much stronger dependence is observed when the wavelengths of the signal and idler lie in slow-light spectral domains. Results and discussion {#sResults} ====================== In this section we illustrate how our theoretical model can be used to investigate various phenomena related to FWM in Si-PhCWGs. In particular, we will compare the pulse interaction in slow- and fast-light regimes, calculate the FWM gain, and investigate the influence of various waveguide parameters on the FWM process. The choice of the values of physical parameters of the co-propagating pulses and that of the input pump power has been guided by the exact phase-matching condition given by Eq. . In all our calculations we assumed that the pulses propagate in the even mode and, unless otherwise specified, the following values for the pulse and waveguide parameters have been used in all our simulations: the input peak pump power, $P_p=10^2P_s=\SI{5}{\watt}$, the input pulse width, $T_p=T_{s}=\SI{7}{\pico\second}$, and the intrinsic waveguide loss coefficient, $\alpha_{\mathrm{in}}=\SI{50}{\decibel\per\centi\meter}$. ![Pulse evolution in the time domain. Left (right) panels correspond to fast-light (slow-light) regimes, the group-index of the pulses being: $n_{g,i}=9.48$ ($n_{g,i}=20.3$), $n_{g,p}=8.64$ ($n_{g,p}=8.69$), $n_{g,s}=10.37$ ($n_{g,s}=23.3$).[]{data-label="fig:pulsetime"}](Figure10.eps){width="8cm"} Let us consider first the evolution of the envelopes of the pulses in the time domain, both in the slow- and fast-light regimes, as illustrated in Fig. \[fig:pulsetime\]. The triplet of wavelengths for which the phase-matching condition is satisfied is $\lambda_p=\SI{1554}{\nano\meter}$, $\lambda_s=\SI{1536}{\nano\meter}$, and $\lambda_i=\SI{1571}{\nano\meter}$ in the fast-light regime, whereas in the slow-light regime the wavelengths are $\lambda_p=\SI{1559}{\nano\meter}$, $\lambda_s=\SI{1524}{\nano\meter}$, and $\lambda_i=\SI{1597}{\nano\meter}$. We stress that in both cases the pump pulse propagates in the fast-light regime, whereas the signal and idler are both generated either in the fast- or slow-light regime. Under these circumstances, one expects that the pump evolution in the time domain is similar in the two cases, a conclusion validated by the plots shown in Figs. \[fig:pulsetime\](c) and \[fig:pulsetime\](d). However, the dynamics of the signal and idler are strikingly different when they propagate in the slow-light or fast-light regimes. There are several reasons that account for these differences. First, whereas the FCA coefficient, $\alpha_{\mathrm{fc}}$, has similar values in the two cases, the FCA and intrinsic losses are much larger in the slow-light regime because the strength of both these effects is inverse proportional to $v_{g}$. This is reflected in Fig. \[fig:pulsetime\] as a much more rapid decay in the slow-light regime of the signal and idler pulses. Second, it can be seen that in the slow-light regime the idler pulse grows at a faster rate. This is again a manifestation of slow-light effects. In particular, the nonlinear coefficient $\gamma_{ips}$, which determines the FWM gain, is inverse proportional to $(v_{g,i}v_{g,s})^{1/2}$ \[see Eq. \]. As a consequence, the FWM gain is strongly enhanced when both the signal and idler propagate in the slow-light regime. ![From top to bottom, the left (right) panels show the evolution of the spectra of the idler, pump, and signal in the case of fast-light (slow-light) regimes. The waveguide and pulse parameters are the same as in Fig. \[fig:pulsetime\].[]{data-label="fig:pulsespctr"}](Figure11.eps){width="8cm"} The slow-light effects are reflected not only in the characteristics of the time-domain propagation of the pulses but they also affect the evolution of the pulse spectra. In order to illustrate this idea, we plot in Fig. \[fig:pulsespctr\] the $z$-dependence of the spectra of the pulses. Similarly to the time-domain dynamics, the spectra of the pump are almost the same in the slow- and fast-light regimes as its GV does not differ much between the two cases. The most noteworthy differences between the slow- and fast-light scenarios can again be observed in the case of the idler and signal. Thus, in the slow-light regime the idler decays faster due to increased losses and grows and broadens more significantly because of enhanced FWM gain and FCD effects, respectively. The influence of FCD on the spectral features of the pulses can also be seen in the case of the signal and, to a smaller extent, the pump. More specifically, Eq.  shows that the index of refraction of the waveguide decreases due to the generation of FCs. This in turn leads to a phase-shift and, consequently, a blue-shift of the pulse [@bkr04oe]. Interestingly enough, one can also see in Fig. \[fig:pulsespctr\] that as the frequency of the pulses shifts during their propagation new spectral peaks are forming at the initial wavelengths for which the phase-matching condition was satisfied.  \[Fig. \[fig:pulsetime\](b)\], whereas the remaining triplet of phase-matched wavelengths is $\lambda_p=\SI{1556}{\nano\meter}$, $\lambda_s=\SI{1530}{\nano\meter}$, and $\lambda_i=\SI{1582}{\nano\meter}$ (red).[]{data-label="fig:gain"}](Figure12.eps){width="8cm"} In order to gain a deeper insight into the influence of slow-light effects on the FWM process, we computed the $z$-dependence of the pulse energies when the frequencies of the signal and idler were tuned in the slow-light regions of the even mode of the waveguide. We considered two scenarios, namely these energies were calculated by including FWM terms in Eqs.  and Eq.  and, in the other case, by setting them to zero, that is $\gamma_{psi}=\gamma_{spi}=\gamma_{ips}=0$. In the former case, the FWM terms are responsible for transferring energy from the pump pulse to the signal and idler. Therefore, a suitable quantity to characterize the efficiency of this energy transfer is what we call the FWM enhancement factor, $\eta$, which in the case of the signal is defined as $\eta_{s}=10\log[(E_{SXF}-E_{SX})/E_{s,in}]$. Here, $E_{SXF}$ and $E_{SX}$ are the signal energies calculated by taking into account, in one case, SPM, XPM, and FWM effects, and only SPM and XPM terms in the other case (i.e. FWM terms are neglected in the latter case), and $E_{s,in}$ is the input energy of the signal. The results of these calculations are summarized in Fig. \[fig:gain\]. In particular, it can be clearly seen in Fig. \[fig:gain\](a) that the FWM enhancement factor is strongly dependent on pulse propagation regime. To be more specific, as the signal and idler are shifting in the slow-light regime a smaller amount of energy is transferred from the pump pulse to the signal. There are two effects whose combined influence leads to this behavior. First, as we discussed, the pulses experience larger optical losses in the slow-light regime and therefore the signal losses energy at higher rate. Equally important, as the pulses are tuned in the slow-light regime the walk-off parameter, $\Delta$, defined as $\Delta=1/v_{g,s}-1/v_{g,p}$, increases, meaning that the pulses interact for a shorter time and consequently less energy is transferred to the signal. These conclusions are clearly validated by the results summarized in Fig. \[fig:gain\](b), where we plot the energy of the signal vs. the propagation distance, determined for several values of the walk-off parameter. In addition, this figure somewhat surprisingly suggests that the FWM process is more efficient in the fast-light regime, which is again due to the fact that the pump and signal overlap over longer time. ![(a) Dependence of loss factor, $\Lambda$, on the propagation distance, $z$, determined for different values of the group-index, $n_{g}$. (b) Dependence of $\Lambda$ on $n_{g}$, determined for different values of $z$. The slopes of the curves corresponding to $z=\SI{10}{\micro\meter}$ and $z=\SI{70}{\micro\meter}$ are shown in the inset.[]{data-label="fig:loss"}](Figure13.eps){width="8cm"} It is well known that in the slow-light regime linear optical effects are enhanced by a factor of $c/v_{g}$, whereas cubic nonlinear interactions increase by a factor of $(c/v_{g})^{2}$. For example, FCA and TPA are proportional to $v_{g}^{-1}$ and $v_{g}^{-2}$, respectively. Our theoretical model predicts, however, that when the mutual interaction between FCs and the optical field is taken into account these scaling laws can significantly change. This can be understood as follows: the amount of FCs generated via TPA, $N$, is proportional to $v_{g}^{-2}$ and since FCA is proportional to the product $v_{g}^{-1}N$, it scales with the GV as $v_{g}^{-3}$. In order to validate this argument we have determined the optical total loss experienced by a pulse when it propagates in the presence of TPA and FCA or, in a different scenario, when only TPA is present \[the latter case is realized by simply setting $\alpha_{\mathrm{fc}}=0$ in Eqs. \]. Moreover, to simplify our analysis, we consider the propagation of only one pulse by setting all parameters describing XPM and FWM interactions to zero. Finally, we also reduced the input power to in order to avoid strong SPM-induced pulse reshaping. Under these conditions, the effect of the FCA on the pulse dynamics can be conveniently characterized by introducing a loss factor, $\Lambda$, defined as $\Lambda=10\log[(E_{T}-E_{TF})/E_{in}]$, where $E_{TF}$ and $E_{T}$ are the pulse energies in the case when both TPA and FCA terms are included in the model and when only TPA is present, respectively, and $E_{in}$ is the input energy of the pulse. The results of these calculations are presented in Fig. \[fig:loss\]. The variation of the loss factor, $\Lambda$, with the propagation distance, determined for several values of the GV is presented in Fig. \[fig:loss\](a). As one would have expected, the loss factor increases with the group-index, $n_{g}$, which is a reflection of the fact that the FC-induced losses increase with the decrease of the GV. One can observe, however, that when the propagation distance is larger than about $139a$ the loss factor begins to decrease when the GV decreases. This behavior is a direct manifestation of slow-light effects, namely as the frequency is tuned to the slow-light regime the optical losses increase significantly irrespective of the fact that only TPA is considered or both TPA and FCA effects are incorporated in the numerical simulations. This subtle dependence of FC-induced losses on $v_{g}$ is perhaps better reflected by the plots presented in Fig. \[fig:loss\](b). Thus, for several values of the propagation distance, we have determined the variation of $\Lambda$ with the group-index, $n_{g}$. Then, by calculating the slope of the function $\Lambda(n_{g})$ represented on a logarithmic scale one can determine how FC-induced losses scale with $v_{g}$ \[cf. the inset in Fig. \[fig:loss\](b)\]. The results of this analysis clearly demonstrate that FC losses are proportional to $v_{g}^{-3}$, which agrees with the predictions of our qualitative evaluation of this dependence. We stress that for large $n_{g}$ (i.e., small $v_{g}$) the $v_{g}^{-3}$ dependence no longer holds at large propagation distance, chiefly because the pulse is strongly reshaped in the slow-light regime due to enhanced nonlinear optical effects, and thus its peak power is no longer exclusively determined by optical losses. Conclusion {#sConcl} ========== In conclusion, we have derived a rigorous theoretical model, which describes pulsed four-wave-mixing in one-dimensional photonic crystal slab waveguides made of silicon. Our theoretical model rigorously incorporate all key linear and nonlinear optical effects affecting the optical pulse dynamics, including modal dispersion, free-carrier dispersion, free-carrier absorption, self- and cross-phase modulation, two-photon absorption, cross-absorption modulation, and four-wave mixing. In addition, the mutual interaction between photogenerated free-carriers and optical field is incorporated in our theoretical analysis in a natural way by imposing the conservation the total energy of the optical field and free-carriers. Importantly, our theoretical formalism allows one to derive rigorous formulae for the optical coefficients characterizing the linear and nonlinear optical properties of the photonic crystal waveguides, avoiding thus any of the approximations that are commonly used in the investigation of nonlinear pulse dynamics in semiconductor waveguides based on photonic crystals. As a practical application of the theoretical results developed in this study, we have used our theoretical model to investigate the properties of degenerate four-wave-mixing of optical pulses propagating in photonic crystal waveguides made of silicon, with a special focus being on highlighting the differences between the pulse dynamics in the slow- and fast light regimes. This analysis has revealed not only that linear and nonlinear effects are enhanced in the slow-light regime by a factor of $n_{g}$ and $n_{g}^{2}$, respectively, but also that these scaling laws are markedly affected in the presence of free-carriers. Moreover, since our study has been performed in a very general framework, i.e. generic optical properties of the waveguides (multi-mode waveguides) and pulse configuration (multi-frequency optical field), our findings can also be used to describe many phenomena not considered in this work. For example, important nonlinear effects, including stimulated and spontaneous Raman scattering, coherent anti-Stokes Raman scattering, and third-harmonic generation, can be included in our model by simply adding the proper nonlinear polarizations. This work was supported by the Engineering and Physical Sciences Research Council, grant No EP/J018473/1. The work of S. L. was supported through a UCL Impact Award graduate studentship. N. C. P. acknowledges support from European Research Council / ERC Grant Agreement no. ERC-2014-CoG-648328. Appendix: Averaged model describing degenerate four-wave-mixing {#sAppendix .unnumbered} =============================================================== The spatial scale over which the envelope of picosecond pulses varies is much larger than the lattice constant of the PhC and therefore for such optical pulses one can simplify the system of equations governing the pulse interaction, i.e. Eqs.  and Eq. , by taking the average over one lattice constant. Under these conditions, the corresponding system of coupled equations can be cast in the following form: \[FWMpsiAv\] $$\begin{aligned} \label{FWMpAv} i&\bigg(\frac{\displaystyle \partial A_{p}}{\displaystyle \partial z} +\frac{1}{v_{g,p}}\frac{\partial A_{p}}{\partial t}\bigg)-\frac{\bar{\beta}_{2,p}}{2}\frac{\partial^{2} A_{p}}{\partial t^{2}}+\frac{\omega_{p}\delta n_{\mathrm{fc}}\tilde{\kappa}_{p}}{n v_{g,p}}A_{p}\notag \\ &+\frac{ic\tilde{\kappa}_{p}}{2n v_{g,p}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{p}+2e^{i\Delta\bar{\beta}z}\tilde{\gamma}_{psi}A_{s}A_{i}A_{p}^{*} \notag \\ &+\left(\tilde{\gamma}_{p}\vert A_{p}{\vert}^{2}+2\tilde{\gamma}_{ps}\vert A_{s}{\vert}^{2}+2\tilde{\gamma}_{pi}\vert A_{i}{\vert}^{2}\right)A_{p}=0, \\ \label{FWMsAv}i&\bigg(\frac{\displaystyle \partial A_{s}}{\displaystyle \partial z} +\frac{1}{v_{g,s}}\frac{\partial A_{s}}{\partial t}\bigg)-\frac{\bar{\beta}_{2,s}}{2}\frac{\partial^{2} A_{s}}{\partial t^{2}}+\frac{\omega_{s}\delta n_{\mathrm{fc}}\tilde{\kappa}_{s}}{n v_{g,s}}A_{s}\notag \\ &+\frac{ic\tilde{\kappa}_{s}}{2n v_{g,s}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{s}+e^{-i\Delta\bar{\beta}z}\tilde{\gamma}_{spi}A_{p}^{2}A_{i}^{*} \notag \\ &+\left(\tilde{\gamma}_{s}\vert A_{s}{\vert}^{2}+2\tilde{\gamma}_{sp}\vert A_{p}{\vert}^{2}+2\tilde{\gamma}_{si}\vert A_{i}{\vert}^{2}\right)A_{s}=0, \\ \label{FWMiAv}i&\bigg(\frac{\displaystyle \partial A_{i}}{\displaystyle \partial z} +\frac{1}{v_{g,i}}\frac{\partial A_{i}}{\partial t}\bigg)-\frac{\bar{\beta}_{2,i}}{2}\frac{\partial^{2} A_{i}}{\partial t^{2}}+\frac{\omega_{i}\delta n_{\mathrm{fc}}\tilde{\kappa}_{i}}{n v_{g,i}}A_{i}\notag \\ &+\frac{ic\tilde{\kappa}_{i}}{2n v_{g,i}}(\alpha_{\mathrm{fc}}+\alpha_{\mathrm{in}})A_{i}+e^{-i\Delta\bar{\beta}z}\tilde{\gamma}_{ips}A_{p}^{2}A_{s}^{*} \notag \\ &+\left(\tilde{\gamma}_{i}\vert A_{i}{\vert}^{2}+2\tilde{\gamma}_{ip}\vert A_{p}{\vert}^{2}+2\tilde{\gamma}_{is}\vert A_{s}{\vert}^{2}\right)A_{i}=0,\end{aligned}$$ $$\begin{aligned} \label{FWMFCdynAv} &\frac{\partial N}{\partial t} = -\frac{N}{\tau_{c}}+\frac{1}{\hbar} \bigg\{\sum_{\mu=p,s,i}\bigg[\frac{\Upsilon_{\mu}^{\prime\prime}}{\omega_{\mu}}\vert A_{\mu}{\vert}^{4} \notag \\ &~+\sum\limits_{\scriptsize{\begin{array}{c} \nu=p,s,i \\ \nu\neq \mu \end{array}}}\frac{4\Upsilon_{\mu\nu}^{\prime\prime}}{\omega_{\mu}+\omega_{\nu}} \vert A_{\mu}{\vert}^{2}\vert A_{\nu}{\vert}^{2}\bigg]+\frac{1}{\omega_{p}}\mathfrak{Im}\big[2\Upsilon_{psi} \notag \\ &~\times {A_{p}^{*}}^{2}A_{s}A_{i}e^{i\Delta\bar{\beta}z}+(\Upsilon_{spi}+\Upsilon_{ips}) A_{p}^{2}A_{s}^{*}A_{i}^{*}e^{-i\Delta\bar{\beta}z}\big]\big\}.\end{aligned}$$ The coefficients of the linear and nonlinear terms in these equations are given by the following formulae: \[FWMcouplcoefflinnlAv\] $$\begin{aligned} \label{FWMcouplcoeffLinSAv} &\tilde{\kappa}_{\mu} = \frac{\epsilon_{0}n^{2}}{2\bar{W}_{\mu}}\int_{V_{\mathrm{nl}}}\vert\mathbf{e}_{\mu}(\omega_{\mu})\vert^{2}dV, \\ \label{FWMGammaSPMAv} &\tilde{\gamma}_{\mu} = \frac{3\omega_{\mu}\epsilon_{0}a}{16v_{g,\mu}^{2}}\frac{1}{\bar{W}_{\mu}^{2}}\int_{V_{\mathrm{nl}}}\mathbf{e}_{\mu}^{*}(\omega_{\mu}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\omega_{\mu},-\omega_{\mu},\omega_{\mu})\vdots\mathbf{e}_{\mu}(\omega_{\mu})\mathbf{e}_{\mu}^{*}(\omega_{\mu})\mathbf{e}_{\mu}(\omega_{\mu})dV, \\ \label{FWMGammaXPMdiffFreqAv} &\tilde{\gamma}_{\mu\nu} = \frac{3\omega_{\mu}\epsilon_{0}a}{16v_{g,\mu}v_{g,\nu}}\frac{1}{\bar{W}_{\mu}\bar{W}_{\nu}}\int_{V_{\mathrm{nl}}}\mathbf{e}_{\mu}^{*}(\omega_{\mu}) \notag \\ &~~\cdot\hat{\chi}^{(3)}(\omega_{\nu},-\omega_{\nu},\omega_{\mu})\vdots\mathbf{e}_{\nu}(\omega_{\nu})\mathbf{e}_{\nu}^{*}(\omega_{\nu})\mathbf{e}_{\mu}(\omega_{\mu})dV, \\ \label{FWMpsiGammaFWMAv} &\tilde{\gamma}_{psi} = \frac{3\omega_{p}\epsilon_{0}a}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{V_{\mathrm{nl}}}\mathbf{e}_{p}^{*}(\omega_{p})\cdot \hat{\chi}^{(3)}(\omega_{s},-\omega_{p},\omega_{i})\vdots\mathbf{e}_{s}(\omega_{s})\mathbf{e}_{p}^{*}(\omega_{p})\mathbf{e}_{i}(\omega_{i})dV, \\ \label{FWMspiGammaFWMAv} &\tilde{\gamma}_{spi} = \frac{3\omega_{s}\epsilon_{0}a}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{V_{\mathrm{nl}}}\mathbf{e}_{s}^{*}(\omega_{s})\cdot \hat{\chi}^{(3)}(\omega_{p},-\omega_{i},\omega_{p})\vdots\mathbf{e}_{p}(\omega_{p})\mathbf{e}_{i}^{*}(\omega_{i})\mathbf{e}_{p}(\omega_{p})dV, \\ \label{FWMipsGammaFWMAv} &\tilde{\gamma}_{ips} = \frac{3\omega_{i}\epsilon_{0}a}{16v_{g,p}{(v_{g,s}v_{g,i})}^{\frac{1}{2}}}\frac{1}{\bar{W}_{p}{(\bar{W}_{s}\bar{W}_{i})}^{\frac{1}{2}}} \notag \\ &~~\times\int_{V_{\mathrm{nl}}}\mathbf{e}_{i}^{*}(\omega_{i})\cdot \hat{\chi}^{(3)}(\omega_{p},-\omega_{s},\omega_{p})\vdots\mathbf{e}_{p}(\omega_{p})\mathbf{e}_{s}^{*}(\omega_{s})\mathbf{e}_{p}(\omega_{p})dV, \\ \label{FWMUpsilonFWMAv}&\Upsilon_{\varpi}=\frac{1}{a}\int_{z_{0}}^{z_{0}+a}\frac{\gamma_{\varpi}(z)}{A_{\mathrm{nl}}(z)}dz.\end{aligned}$$ where $V_{\mathrm{nl}}$ is the volume occupied by silicon in a unit cell of the PhC waveguide and $z_{0}$ an arbitrary distance. Finally, $A_{\mathrm{nl}}(z)$ in Eq.  is given by Eq. , whereas the index $\varpi$ takes any of the following values: $p$, $s$, $i$, $ps$, $si$, $ip$, $psi$, $spi$, or $ips$.whereas the index $\varpi$ takes any of the following values: $p$, $s$, $i$, $ps$, $si$, $ip$, $psi$, $spi$, and $ips$. Note that in deriving the averaged equations governing the pulse and FC dynamics, Eqs.  and Eq. , we have assumed that the FWM process is nearly phase-matched, namely $\Delta\bar{\beta}\ll 1/a$. In other words, the phase of the exponential factors in these equations vary much slower with the distance, $z$, as compared to the variation of the dielectric constant of the PhC waveguide. ![Comparison between pulse evolution as described by the full and averaged model is presented in the left and right panels, respectively. The group-index of the pulses are $n_{g,i}=20.3$, $n_{g,p}=8.69$, and $n_{g,s}=23.3$ and correspond to the slow-light propagation scenario presented in Fig. \[fig:pulsetime\]. The bottom panel shows the $z$-dependence of the normalized pulse amplitude, $\Psi_{\mu}(z)=A_{\mu}(z_{0}+z)/A_{\mu}(z_{0})$, $\mu=p,s,i$, calculated for the unit cell starting at $z_{0}=200a$.[]{data-label="fig:pulsetimecomp"}](Figure14.eps){width="8cm"} A comparison between the predictions of the full and averaged models is illustrated in Fig. \[fig:pulsetimecomp\]. Thus, we have considered the slow-light pulse dynamics presented in Fig. \[fig:pulsetime\] and determined the pulse evolution using both the full and averaged models. As it can be seen, both models predict a similar pulse dynamics for the entire propagation length, $z=1000a$. This result is expected as the envelope of picosecond pulses, as are those chosen in our simulations, spans a large number of unit cells and therefore the pulse amplitude is only slightly affected by the local inhomogeneity of the index of refraction. The fast variation with $z$ of the pulse envelope is shown in Fig. \[fig:pulsetimecomp\](g), where we plot the $z$-dependence of the normalized pulse amplitude, $\Psi_{\mu}(z)=A_{\mu}(z_{0}+z)/A_{\mu}(z_{0})$, $\mu=p,s,i$, calculated for the unit cell starting at $z_{0}=200a$. It can be seen in this figure that the pulse envelope varies at a spatial scale commensurable with the lattice constant yet the amplitude of these variations is much smaller than the pulse peak amplitude. The magnitude of these variations, however, would comparatively become more significant should the pulse duration would be brought to the femtosecond range. [99i]{} R. Ho, K. W. Mai, and M. A. Horowitz, Proc. IEEE **89**, 490 (2001). K. K. Lee, D. R. Lim, H. C. Luan, A. Agarwal, J. Foresi, and L. C. Kimerling,  **77**, 1617 (2000). R. U. Ahmad, F. Pizzuto, G. S. Camarda, R. L. Espinola, H. Rao, and R. M. Osgood, **14**, 65 (2002). R. Claps, D. Dimitropoulos, V. Raghunathan, Y. Han, and B. Jalali, **11**, 1731 (2003). R. Espinola, J. I. Dadap, R. M. Osgood, S. J. McNab, and Y. A. Vlasov, **12**, 3713 (2004). H. S. Rong, R. Jones, A. S. Liu, O. Cohen, D. Hak, A. Fang, and M. Paniccia,  **433**, 725 (2005). M. A. Foster, A. C. Turner, J. E. Sharping, B. S. Schmidt, M. Lipson, and A. L. Gaeta,  **441**, 960 (2006). X. Liu, R. M. Osgood, Y. A. Vlasov, and W. M. J. Green, **4**, 557 (2010). G. Cocorullo, M. Iodice, I. Rendina, and P. M. Sarro, **7**, 363 (1995). A. Liu, R. Jones, L. Liao, D. Samara-Rubio, D. Rubin, O. Cohen, R. Nicolaescu, and M. Paniccia,  **427**, 615 (2004). Q. Xu, B. Shmidt, S. Pradhan, and M. Lipson,  **435**, 325 (2005). X. Chen, N. C. Panoiu, I. W. Hsieh, J. I. Dadap, and R. M. Osgood, **18**, 2617 (2006). M. Mohebbi, **20**, 921 (2008). O. Boyraz, P. Koonath, V. Raghunathan, and B. Jalali, **12**, 4094 (2004). I. W. Hsieh, X. Chen, X. P. Liu, J. I. Dadap, N. C. Panoiu, C. Y. Chou, F. Xia, W. M. Green, Y. A. Vlasov, and R. M. Osgood, **15**, 15242 (2007). N. C. Panoiu, X. Liu, and R. M. Osgood,  **34**, 947 (2009). N. C. Panoiu, X. Chen, and R. M. Osgood,  **31**, 3609 (2006). H. Fukuda, K. Yamada, T. Shoji, M. Takahashi, T. Tsuchizawa, T. Watanabe, J. Takahashi, and S. Itabashi, **13**, 4629 (2005). R. Espinola, J. Dadap, R. M. Osgood, S. McNab, and Y. Vlasov, **13**, 4341 (2005). M. A. Foster, A. C. Turner, R. Salem, M. Lipson, and A. L. Gaeta, **15**, 12949 (2007). S. Zlatanovic, J. S. Park, S. Moro, J. M. C. Boggio, I. B. Divliansky, N. Alic, S. Mookherjea, and S. Radic, **4**, 561 (2010). J. B. Driscoll, N. Ophir, R. R. Grote, J. I. Dadap, N. C. Panoiu, K. Bergman, and R. M. Osgood, **20**, 9227 (2012). R. M. Osgood, N. C. Panoiu, J. I. Dadap, X. Liu, X. Chen, I-W. Hsieh, E. Dulkeith, W. M. J. Green, and Y. A. Vlassov, **1**, 162 (2009). E. Yablonovitch,  **58**, 2059 (1987). S. John,  **58**, 2486 (1987). A. Mekis, J. C. Chen, I. Kurland, S. Fan, P. R. Villeneuve, and J. D. Joannopoulos,  **77**, 3787 (1996). S. Y. Lin, E. Chow, V. Hietala, P. R. Villeneuve, J. D. Joannopoulos, **282**, 274 (1998). T. Baba, N. Fukaya, and J. Yonekura, **35**, 654 (1999). S. G. Johnson, S. Fan, P. R. Villeneuve, J. D. Joannopoulos, and L. A. Kolodziejski,  **60**, 5751 (1999). A. Chutinan and S. Noda,  **62**, 4488 (2000). J. S. Foresi, P. R. Villeneuve, J. Ferrera, E. R. Thoen, G. Steinmeyer, S. Fan, J. D. Joannopoulos, L. C. Kimerling, H. I. Smith, and E. P. Ippen,  **390**, 143 (1997). O. Painter, R. K. Lee, A. Scherer, A. Yariv, J. D. O’Brien, P. D. Dapkus, and I. Kim, **284**, 1819 (1999). Y. Akahane, T. Asano, B. S. Song, and S. Noda,  **425**, 944 (2003). J. P. Reithmaier, G. Sek, A. Loffler, C. Hofmann, S. Kuhn, S. Reitzenstein, L. V. Keldysh, V. D. Kulakovskii, T. L. Reinecke, and A. Forchel,  **432**, 197 (2004). T. Yoshie, A. Scherer, J. Hendrickson, G. Khitrova, H. M. Gibbs, G. Rupper, C. Ell, O. B. Shchekin, and D. G. Deppe,  **432**, 200 (2004). S. Fan, P. R. Villeneuve, J. D. Joannopoulos, M. J. Khan, C. Manolatou, and H. A. Haus,  **59**, 15882 (1999). A. Chutinan, M. Mochizuki, M. Imada, and S. Noda,  **79**, 2690 (2001). M. Soljacic and J. D. Joannopoulos, **3**, 211 (2004). T. F. Krauss, **2**, 448 (2008). T. Baba, **2**, 465 (2008). M. Notomi, K. Yamada, A. Shinya, J. Takahashi, C. Takahashi, and I. Yokohama,  **87**, 253902 (2001). M. Soljacic, S. G. Johnson, S. Fan, M. Ibanescu, E. Ippen, and J. D. Joanopoulos,  **19**, 2052 (2002). N. C. Panoiu, M. Bahl, and R. M. Osgood,  **28**, 2503 (2003). N. C. Panoiu, M. Bahl, and R. M. Osgood, **12**, 1605 (2004). Y. A. Vlasov, M. O. Boyle, H. F. Hamann, and S. J. McNab,  **438**, 65 (2005). J. F. McMillan, X. Yang, N. C. Panoiu, R. M. Osgood, and C. W. Wong,  **31**, 1235 (2006). B. Corcoran, C. Monat, C. Grillet, D. J. Moss, B. J. Eggleton, T. P. White, L. O’Faolain, and T. F. Krauss, **3**, 2503 (2009). I. H. Rey, Y. Lefevre, S. A. Schulz, N. Vermeulen, and T. F. Krauss,  **84**, 035306 (2011). J. F. McMillan, M. Yu, D. L. Kwong, and C. W. Wong, **18**, 15484 (2010). M. Santagiustina, C. G. Someda, G. Vadala, S. Combrie, and A. De Rossi, **18**, 21024 (2010). C. Monat, M. Ebnali-Heidari, C. Grillet, B. Corcoran, B. J Eggleton, T. P. White, L. O’Faolain, J. Li, and T. F. Krauss, **18**, 22915 (2010). J. Li, L. O’Faolain, I. H. Rey, and T. F. Krauss, **19**, 4458 (2011). T. Chen, J. Sun, and L. Li, **20**, 20043 (2012). S. Lavdas, S. Zhao, J. B. Driscoll, R. R. Grote, R. M. Osgood, and N. C. Panoiu,  **39**, 401 (2014). J. D. Joannopoulos, S. G. Johnson, J. N. Winn, and R. D. Meade, *Photonic Crystals: Molding the Flow of Light*, 2nd ed. (Princeton University Press, Princeton, NJ, 2008). S. G. Johnson and J. D. Joannopoulos, **8**, 173 (2001). X. Chen, N. C. Panoiu, and R. M. Osgood,  **42**, 160 (2006). N. C. Panoiu, J. F. McMillan, and C. W. Wong, **16**, 257 (2010). A. W. Snyder, **18**, 383 (1970). J. E. Sipe, C. M. de Sterke, and B. J. Eggleton,  **49**, 1437 (2002). R. A. Soref and B. R. Bennett,  **23**, 123 (1987). R. W. Boyd, *Nonlinear Optics*, 3rd ed. (Academic Press, San Diego, 2008). J. Zhang, Q. Lin, G. Piredda, R. W. Boyd, G. P. Agrawal, and P. M. Fauchet,  **91**, 071113(1-3) (2007). D. Michaelis, U. Peschel, C. Wachter, and A. Brauer,  **68**, 065601(R)(1-4) (2003). T. Kamalakis and T. Sphicopoulos,  **43**, 923 (2007). A. W. Snyder and J. D. Love, *Optical Waveguide Theory* (Chapman and Hall, London, 1983). P. N. Butcher and D. Cotter, *The Elements of Nonlinear Optics* (Cambridge University Press, 1991). C. Monat, B. Corcoran, M. Ebnali-Heidari, C. Grillet, B. J. Eggleton, T. P. White, L. O’Faolain, and T. F. Krauss, **17**, 2944 (2009). E. D. Palik, *Handbook of Optical Constants of Solids* (Academic Press, San Diego, 1998). G. P. Agrawal, *Nonlinear Fiber Optics*, 5th ed. (Academic Press, San Diego, 2013).

ArXiv

--- abstract: 'We discuss the statistical mechanics of a system of self-gravitating fermions in a space of dimension $D$. We plot the caloric curves of the self-gravitating Fermi gas giving the temperature as a function of energy and investigate the nature of phase transitions as a function of the dimension of space. We consider stable states (global entropy maxima) as well as metastable states (local entropy maxima). We show that for $D\ge 4$, there exists a critical temperature (for sufficiently large systems) and a critical energy below which the system cannot be found in statistical equilibrium. Therefore, for $D\ge 4$, quantum mechanics cannot stabilize matter against gravitational collapse. This is similar to a result found by Ehrenfest (1917) at the atomic level for Coulombian forces. This makes the dimension $D=3$ of our universe very particular with possible implications regarding the anthropic principle. Our study enters in a long tradition of scientific and philosophical papers who studied how the dimension of space affects the laws of physics.' author: - 'Pierre-Henri Chavanis' title: 'Statistical mechanics and thermodynamic limit of self-gravitating fermions in $D$ dimensions' --- Laboratoire de Physique Théorique, Université Paul Sabatier,\ 118 route de Narbonne 31062 Toulouse, France.\ Introduction {#sec_introduction} ============ The statistical mechanics of systems with long-range interactions is currently a topic of active research [@dauxois]. Among long-range interactions, the gravitational force plays a fundamental role. Therefore, the developement of a statistical mechanics for self-gravitating systems is of considerable interest [@paddy]. In this context, a system of self-gravitating fermions enclosed within a box provides an interesting model which can be studied in great detail [@ht; @pt]. This model incorporates an effective small-scale cut-off played by the Pauli exclusion principle and a large scale cut-off played by the confining box (other forms of confinement could also be considered). The statistical mechanics of this system is rigorously justified and presents a lot of interesting features which are of interest in statistical mechanics [@houches] and astrophysics [@chavcape]. Its detailed study is therefore important at a conceptual and practical level. In a preceding paper [@pt], we have discussed the nature of phase transitions in the self-gravitating Fermi gas in a space of dimension $D=3$. Our study was performed in both microcanonical and canonical ensembles and considered an arbitrary degree of degeneracy relative to the system size. This study completes previous investigations by Hertel & Thirring [@ht] who worked in the canonical ensemble and considered small system sizes. At high temperatures and high energies, the system is in a gaseous phase and quantum effects are completely negligible. At some transition temperature $T_{t}$ or transition energy $E_{t}$ (for sufficiently large system sizes), a first order phase transition is expected to occur and drive the system towards a condensed phase. However, gaseous states are still metastable below this transition point and gravitational collapse will rather occur at a smaller critical temperature $T_{c}$ (Jeans temperature) [@aa] or critical energy $E_{c}$ (Antonov energy) [@antonov; @lbw; @paddy] at which the metastable branch disappears (spinodal point). The end-state of the collapse is a compact object with a “core-halo” structure. Typically, it consists of a degenerate nucleus surrounded by a “vapour”. The nucleus (condensate) resembles a white dwarf star [@chandra2]. At non-zero temperature, this compact object is surrounded by a dilute atmosphere. Therefore, when quantum mechanics is properly accounted for, there exists an equilibrium state (global maximum of entropy or free energy) for each value of accessible energy and temperature. The condensate results from the balance between gravitational contraction and quantum pressure. As first noticed by Fowler [@fowler] in his classical theory of white dwarf stars, quantum mechanics is able to stabilize matter against gravitational collapse. One object of this paper is to show that this conclusion is no more valid in a space of dimension $D\ge 4$. For a system of mass $M$ enclosed within a box of radius $R$, there exists a critical temperature (for sufficiently large $R$) and a critical energy below which the system cannot be found at statistical equilibrium. This is like the Antonov instability for self-gravitating classical particles in $D=3$ [@antonov; @lbw; @paddy] but it now occurs for fermions. Therefore, quantum mechanics cannot arrest gravitational collapse in $D\ge 4$. This result is connected to our previous observation [@langevin] that a classical white dwarf star (a polytrope of index $n_{3/2}=D/2$) becomes unstable for $D\ge 4$. Interestingly, this result is similar to that of Ehrenfest [@ehrenfest] who considered the stability of atomic structures (in Bohr’s model) for different dimensions of space and concludes that $D<4$ is required for stability. In this paper, we determine the caloric curve of the self-gravitating Fermi gas for an arbitrary dimension of space and an arbitrary degree of degeneracy (or system size). We exhibit particular dimensions that play a special role in the problem. The dimension $D=2$ is critical because the results established for $D\neq 2$ cannot be directly extended to $D=2$ [@sc1]. Furthermore, in $D=2$ the radius of a white dwarf star is independent on its mass and given in terms of fundamental constants by $R=0.27 \ h m^{-3/2} G^{-1/2}$. The dimension $D=4$ is also critical because it is the dimension at which classical white dwarf stars become unstable. At this particular dimension, their mass is independent on radius and can be expressed in terms of fundamental constants as $M=1.44\ 10^{-2}h^{4}m^{-5}G^{-2}$. Mathematically, this is similar to Chandrasekhar’s limiting mass [@chandra1] for relativistic white dwarf stars in $D=3$. The dimension $D=2(1+\sqrt{2})$ is also particular because at this dimension, the white dwarf stars cease to be self-confined and have infinite mass. Finally, $D=10$ is the dimension at which the caloric curve of classical isothermal spheres loses its characteristic spiral nature [@sc1]. Although we systematically explore all dimensions of space in order to have a complete picture of the problem, only dimensions $D=1$, $D=2$ and $D=3$ are [*a priori*]{} of physical interest. The dimension $D=1$ is considered in cosmology and in connexion with shell models, and the dimension $D=2$ can be useful to describe filaments or ring structures with high aspect ratio. Two-dimensional gravity is also of interest for its properties of conformal invariance and for its relation with two-dimensional turbulence [@houches]. Non-integer dimensions can arise if the system has a fractal nature. The paper is organized as follows. In Sec. \[sec\_fermions\], we determine the thermodynamical parameters of the self-gravitating Fermi gas in dimension $D$. The Fermi-Dirac entropy is introduced from a combinatorial analysis. In Sec. \[sec\_lim\], we consider asymptotic limits corresponding to the classical self-gravitating gas and to completely degenerate structures (white dwarfs). We emphasize the importance of metastable states in astrophysics and explain how they can be taken into account in the theory (see also [@meta]). We also discuss the thermodynamic limit of the self-gravitating quantum gas and compare it with the thermodynamic limit of the self-gravitating classical gas in the dilute limit [@vs]. In Sec. \[sec\_caloric\], we provide a gallery of caloric curves of the self-gravitating Fermi gas in different dimensions of space. Rigorous mathematical results on the existence of solutions of the Fermi-Poisson equation have been obtained by Stańczy [@stanczy]. Finally, in the conclusion, we place our study in a more general perspective. We give a short historical account of scientific and philosophical papers who studied the role played by the dimension of space in determining the form of the laws of physics. These works tend to indicate that the dimension $D=3$ of our universe is very particular. This is also the result that we reach in our study. These remarks can have implications regarding the anthropic principle. Thermodynamics of self-gravitating $D$-fermions {#sec_fermions} =============================================== The Fermi-Dirac distribution {#sec_fd} ---------------------------- We consider a system of $N$ fermions interacting via Newtonian gravity in a space of dimension $D$. We assume that the mass of the configuration is sufficiently small so as to ignore relativistic effects. Let $f({\bf r},{\bf v},t)$ denote the distribution function of the system, i.e. $f({\bf r},{\bf v},t)d^{D}{\bf r} d^{D}{\bf v}$ gives the mass of particles whose position and velocity are in the cell $({\bf r},{\bf v};{\bf r}+d^{D}{\bf r},{\bf v}+d^{D}{\bf v})$ at time $t$. The integral of $f$ over the velocity determines the spatial density $$\rho=\int f d^{D}{\bf v}, \label{fd1}$$ and the total mass of the configuration is given by $$M=\int \rho d^{D}{\bf r}, \label{fd2}$$ where the integral extends over the entire domain. On the other hand, in the meanfield approximation, the total energy of the system can be expressed as $$E={1\over 2}\int fv^{2}d^{D}{\bf r}d^{D}{\bf v}+{1\over 2}\int\rho\Phi d^{D}{\bf r}=K+W, \label{fd3}$$ where $K$ is the kinetic energy and $W$ the potential energy. The gravitational potential $\Phi$ is related to the density by the Newton-Poisson equation $$\Delta\Phi=S_{D} G\rho, \label{fd4}$$ where $S_{D}=2\pi^{D/2}/\Gamma(D/2)$ is the surface of a unit sphere in a space of dimension $D$ and $G$ is the constant of gravity (which depends on the dimension of space). We now wish to determine the most probable distribution of self-gravitating fermions at statistical equilibrium. To that purpose, we divide the individual phase space $\lbrace {\bf r},{\bf v}\rbrace$ into a very large number of microcells with size $(h/m)^D$ where $h$ is the Planck constant (the mass $m$ of the particles arises because we use ${\bf v}$ instead of ${\bf p}$ as a phase space coordinate). A microcell is occupied either by $0$ or $1$ fermion (or $g=2s+1$ fermions if we account for the spin). We shall now group these microcells into macrocells each of which contains many microcells but remains nevertheless small compared to the phase-space extension of the whole system. We call $\nu$ the number of microcells in a macrocell. Consider the configuration $\lbrace n_i \rbrace$ where there are $n_1$ fermions in the $1^{\rm st}$ macrocell, $n_2$ in the $2^{\rm nd}$ macrocell etc..., each occupying one of the $\nu$ microcells with no cohabitation. The number of ways of assigning a microcell to the first element of a macrocell is $\nu$, to the second $\nu -1$ etc. Since the particles are indistinguishable, the number of ways of assigning microcells to all $n_i$ particles in a macrocell is thus $${1\over n_i!}{\times} {\nu!\over (\nu-n_i)!}. \label{fd5}$$ To obtain the number of microstates corresponding to the macrostate $\lbrace n_i \rbrace$ defined by the number of fermions $n_i$ in each macrocell (irrespective of their precise position in the cell), we need to take the product of terms such as (\[fd5\]) over all macrocells. Thus, the number of microstates corresponding to the macrostate $\lbrace n_i \rbrace$, i.e. the probability of the state $\lbrace n_i \rbrace$, is $$W(\lbrace n_i \rbrace)=\prod_i {\nu!\over n_i!(\nu-n_i)!}. \label{fd6}$$ This is the Fermi-Dirac statistics. As is customary, we define the entropy of the state $\lbrace n_i \rbrace$ by $$S(\lbrace n_i \rbrace)=\ln W(\lbrace n_i \rbrace). \label{fd7}$$ It is convenient here to return to a representation in terms of the distribution function giving the phase-space density in the $i$-th macrocell $$f_i=f({\bf r}_i,{\bf v}_i)={n_i \ m\over \nu \ ({h\over m})^D}={n_i\eta_0\over \nu}, \label{fd8}$$ where we have defined $\eta_0=m^{D+1}/h^D$, which represents the maximum value of $f$ due to Pauli’s exclusion principle. Now, using the Stirling formula, we have $$\ln W(\lbrace n_i \rbrace)\simeq \sum_i \nu (\ln\nu-1)-\nu\biggl\lbrace {f_i\over \eta_0}\biggl\lbrack \ln\biggl ({\nu f_i\over \eta_0}\biggr )-1\biggr\rbrack +\biggl (1-{f_i\over \eta_0}\biggr )\biggl\lbrack\ln\biggl\lbrace \nu\biggl (1-{f_i\over \eta_0}\biggr )\biggr\rbrace-1\biggr\rbrack\biggr\rbrace. \label{fd9}$$ Passing to the continuum limit $\nu\rightarrow 0$, we obtain the usual expression of the Fermi-Dirac entropy $$S=-k_B\int \biggl\lbrace {f\over\eta_{0}}\ln {f\over\eta_{0}}+\biggl (1- {f\over\eta_{0}}\biggr)\ln \biggl (1- {f\over\eta_{0}}\biggr)\biggr\rbrace\ {d^{D}{\bf r}d^{D}{\bf v}\over ({h\over m})^D}. \label{fd10}$$ If we take into account the spin of the particles, the above expression remains valid but the maximum value of the distribution function is now $\eta_{0}=g m^{D+1}/h^{D}$, where $g=2s+1$ is the spin multiplicity of the quantum states (the phase space element has also to be multiplied by $g$). An expression of entropy similar to (\[fd10\]), but arising for a completely different reason, has been introduced by Lynden-Bell in the context of the violent relaxation of collisionless stellar systems [@lb; @cs; @dubrovnik]. In that context, $\eta_0$ represents the maximum value of the initial distribution function and the actual distribution function (coarse-grained) must always satisfy $\overline{f}\le \eta_0$ by virtue of the Liouville theorem. This is the origin of the “effective” exclusion principle in Lynden-Bell’s theory, which has nothing to do with quantum mechanics. Since the particles (stars) are distinguishable classical objects (but subject to an exclusion principle in the collisionless regime), Lynden-Bell’s statistics corresponds to a $4$-th form of statistics (in addition to the Maxwell-Boltzmann, Fermi-Dirac and Bose-Einstein statistics). However, for a single type of phase element $\eta_0$, it leads to the same results as the Fermi-Dirac statistics. We also recall that in the non-degenerate (or classical) limit $f\ll\eta_0$, the Fermi-Dirac entropy (\[fd10\]) reduces to the Boltzmann entropy $$S=-k_B\int {f\over m}\biggl\lbrack \ln\biggl ({f h^D\over g m^{D+1}}\biggr )-1\biggr \rbrack d^{D}{\bf r}d^{D}{\bf v}. \label{fd11}$$ Now that the entropy has been precisely justified, the statistical equilibrium state (most probable state) of self-gravitating fermions is obtained by maximizing the Fermi-Dirac entropy (\[fd10\]) at fixed mass (\[fd2\]) and energy (\[fd3\]): $${\rm Max}\quad S[f]\quad | \quad E[f]=E, M[f]=M. \label{fd12}$$ Introducing Lagrange multipliers $1/T$ (inverse temperature) and $\mu$ (chemical potential) to satisfy these constraints, and writing the variational principle in the form $$\delta S-{1\over T}\ \delta E+{\mu\over T} \delta N=0, \label{fd13}$$ we find that the [*critical points*]{} of entropy correspond to the Fermi-Dirac distribution $$f={\eta_{0}\over 1+\lambda e^{\beta m ({v^{2}\over 2}+\Phi)}}, \label{fd14}$$ where $\lambda=e^{-\beta \mu}$ is a strictly positive constant (inverse fugacity) and $\beta={1\over k_B T}$ is the inverse temperature. Clearly, the distribution function satisfies $f\le \eta_{0}$, which is a consequence of Pauli’s exclusion principle. So far, we have assumed that the system is isolated so that the energy is conserved. If now the system is in contact with a thermal bath (e.g., a radiation background) fixing the temperature, the statistical equilibrium state minimizes the free energy $F=E-TS$, or maximizes the Massieu function $J=S-\beta E$, at fixed mass and temperature: $${\rm Max}\quad J[f]\quad |\quad M[f]=M. \label{fd15}$$ Introducing Lagrange multipliers and writing the variational principle in the form $$\delta J+{\mu\over T} \delta N=0, \label{fd16}$$ we find that the [*critical points*]{} of free energy are again given by the Fermi-Dirac distribution (\[fd14\]). Therefore, the critical points (first variations) of the variational problems (\[fd12\]) and (\[fd15\]) are the same. However, the stability of the system (regarding the second variations) can be different in microcanonical and canonical ensembles. When this happens, we speak of a situation of [*ensemble inequivalence*]{} [@pt]. The stability of the system can be determined by a graphical construction, by simply plotting the caloric curve/series of equilibria $\beta(E)$ and using the turning point method of Katz [@katz; @katz2]. Thermodynamical parameters {#sec_para} -------------------------- Integrating the distribution function (\[fd14\]) over velocity, we find that the density of particles is related to the gravitational potential by $$\rho={\eta_{0}S_{D}2^{{D/2}-1}\over (\beta m)^{D/2}}I_{D/2-1}(\lambda e^{\beta m\Phi}), \label{p1}$$ where $I_{n}$ denotes the Fermi integral $$I_{n}(t)=\int_{0}^{+\infty}{x^{n}\over 1+t e^{x}}dx. \label{p2}$$ We recall the identity $$I'_{n}(t)=-{n\over t}I_{n-1}(t), \qquad (n>0), \label{p3}$$ which can be established from (\[p2\]) by an integration by parts. The gravitational potential is now obtained by substituting Eq. (\[p1\]) in the Poisson equation (\[fd4\]). We introduce the rescaled distance $\xi=\lbrack {S_{D}^{2} 2^{D/2-1}G\eta_{0}/(\beta m)^{D/2-1}}\rbrack^{1/2}r$ and the variables $\psi=\beta m (\Phi-\Phi_{0})$ and $k=\lambda e^{\beta m \Phi_{0}}$, where $\Phi_{0}$ is the central potential. Thus, we get the $D$-dimensional Fermi-Poisson equation $${1\over\xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\psi\over d\xi}\biggr )=I_{D/2-1}(ke^{\psi(\xi)}), \label{p4}$$ $$\psi(0)=\psi'(0)=0. \label{p5}$$ As is well-known, self-gravitating systems at non-zero temperature have the tendency to evaporate. Therefore, there is no equilibrium state in a strict sense and the statistical mechanics of self-gravitating systems is essentially an out-of-equilibrium problem. However, the evaporation rate is small in general and the system can be found in a quasi-equilibrium state for a relatively long time. In order to describe the thermodynamics of the self-gravitating Fermi gas rigorously, we shall use an artifice and enclose the system within a spherical box of radius $R$ (the box typically represents the size of the cluster under consideration). In that case, the solution of Eq. (\[p4\]) is terminated by the box at the normalized radius $$\alpha=\biggl \lbrack {S_{D}^{2} 2^{D/2-1} G\eta_{0}\over (\beta m)^{D/2-1}}\biggr \rbrack^{1/2} R. \label{p6}$$ For a spherically symmetric configuration, the Gauss theorem can be written $${d\Phi\over dr}={GM(r)\over r^{D-1}}, \label{p7}$$ where $M(r)=\int_{0}^{r}\rho S_{D}r^{D-1}dr$ is the mass within the sphere of radius $r$. Applying this result at $r=R$ and using the variables introduced previously we get $$\eta\equiv {\beta GMm\over R^{D-2}}=\alpha\psi'_{k}(\alpha). \label{p8}$$ This equation relates the dimensionless box radius $\alpha$ and the uniformizing variable $k$ to the dimensionless inverse temperature $\eta$. According to Eqs. (\[p6\]) and (\[p8\]), $\alpha$ and $k$ are related to each other by the relation $\alpha^{2}\eta^{D/2-1}=\mu$ or, explicitly, $$\alpha^{D+2\over D-2}\psi'_{k}(\alpha)=\mu^{2\over D-2}, \label{p9}$$ where $$\mu=\eta_{0}\sqrt{S_{D}^{4}2^{D-2}G^{D}M^{D-2}R^{D(4-D)}}, \label{p10}$$ is the degeneracy parameter [@cs]. It should not be confused with the chemical potential. We shall give a physical interpretation of this parameter in Sec. \[sec\_d3\]. The calculation of the energy is a little more involved. First, we introduce the local pressure $$p={1\over D}\int f v^{2}d^{D}{\bf v}. \label{p11}$$ Using the Fermi-Dirac distribution function (\[fd14\]), we find that $$p={\eta_{0}S_{D}2^{{D/2}}\over D(\beta m)^{D/2+1}}I_{D/2}(k e^{\psi}). \label{p12}$$ The kinetic energy $K=(D/2)\int p d^{D}{\bf r}$ can thus be written $${K R^{D-2}\over GM^{2}}={\alpha^{4+4D-D^{2}\over D-2}\over \mu^{4\over D-2}}\int_{0}^{\alpha}I_{D/2}(ke^{\psi_{k}(\xi)})\xi^{D-1}d\xi. \label{p13}$$ In order to determine the potential energy, we use the $D$-dimensional version of the Virial theorem [@langevin]. For $D\neq 2$, it reads $$2K+(D-2)W=DV_{D}R^{D}p(R), \label{p14}$$ where $V_{D}=S_{D}/D$ is the volume of a hypersphere with unit radius (the case $D=2$ will be considered specifically in Sec. \[sec\_d2\]). Using the expression of the pressure (\[p12\]) at the box radius $R$, we get $${W R^{D-2}\over GM^{2}}={2\over D(D-2)}{\alpha^{2(D+2)\over D-2}\over \mu^{4\over D-2}}I_{D/2}(k e^{\psi(\alpha)})-{2KR^{D-2}\over (D-2)GM^{2}}. \label{p15}$$ Combining Eqs. (\[p13\]) and (\[p15\]), we finally obtain $$\Lambda\equiv -{E R^{D-2}\over GM^{2}}={4-D\over D-2}{\alpha^{4+4D-D^{2}\over D-2}\over \mu^{4\over D-2}}\int_{0}^{\alpha}I_{D/2}(ke^{\psi_{k}(\xi)})\xi^{D-1}d\xi-{2\over D(D-2)}{\alpha^{2(D+2)\over D-2}\over \mu^{4\over D-2}}I_{D/2}(k e^{\psi(\alpha)}). \label{p16}$$ For $D=3$, Eqs. (\[p8\]) and (\[p16\]) return the expressions derived in [@cs; @pt]. For a given value of $\mu$ and $k$, we can solve the ordinary differential equation (\[p4\]) until the value of $\alpha$ at which the condition (\[p9\]) is satisfied. Then, Eqs. (\[p8\]) and (\[p16\]) determine the temperature and the energy of the configuration. By varying the parameter $k$ (for a fixed value of the degeneracy parameter $\mu$), we can determine the full caloric curve/series of equilibria $\beta(E)$. Extending the results of [@pt] in $D$ dimensions, the entropy of each configuration, parameterized by $\alpha$, is given by $${S\over Nk_{B}}=-{4+4D-D^{2}\over D(4-D)}\Lambda\eta+\psi_{k}(\alpha)+{\eta\over D-2}+\ln k-{2(D-2)\over D^{2}(4-D)}{\alpha^{2D\over D-2}\over \mu^{2\over D-2}}I_{D/2}(ke^{\psi_{k}(\alpha)}), \label{p17}$$ and the free energy by $$F=E-TS. \label{p18}$$ In the microcanonical ensemble, a solution is stable if it corresponds to a maximum of entropy $S[f]$ at fixed mass and energy. In the canonical ensemble, the condition of stability requires that the solution be a minimum of free energy $F[f]$ at fixed mass and temperature. This meanfield approach is [*exact*]{} in a thermodynamical limit such that $N\rightarrow +\infty$ with $\mu$, $\eta$, $\Lambda$ fixed. If we fix $\eta_{0}$ (i.e. $\hbar$) and $G$, this implies that $RN^{(D-2)/(D(4-D))}$, $TN^{-4/(D(4-D))}$, $EN^{-(4D-D^{2}+4)/(D(4-D))}$, $SN^{-1}$ and $JN^{-1}$ approach a constant value for $N\rightarrow +\infty$ (the free energy $F$ scales as $N^{(4D-D^{2}+4)/(D(4-D))}$). This is the quantum thermodynamic limit (QTL) for the self-gravitating gas [@pt; @rieutord]. The usual thermodynamic limit $N,R\rightarrow +\infty$ with $N/R^{D}$ constant is clearly not relevant for inhomogeneous systems whose energy is non-additive. Asymptotic limits {#sec_lim} ================= The non degenerate limit $(\mu=\infty)$ {#sec_class} --------------------------------------- Before considering the case of an arbitrary degree of degeneracy, it may be useful to discuss first the non degenerate limit corresponding to a classical isothermal gas ($\hbar\rightarrow 0$). For $f\ll\eta_{0}$, the distribution function (\[fd14\]) reduces to the Maxwell-Boltzmann formula $$f={\eta_{0}\over\lambda}e^{-\beta m({v^{2}\over 2}+\Phi)}, \label{cc1}$$ which can be written more conveniently as $$f=\biggl ({\beta m\over 2\pi}\biggr )^{D/2}\rho({\bf r})\ e^{-\beta m {v^{2}\over 2}}. \label{cc1bis}$$ The density profile can be written $$\rho=\rho_{0}e^{-\psi(\xi)}, \label{cc2}$$ where $\rho_{0}$ is the central density, $\xi$ is the normalized distance $$\xi=(S_{D} G\beta m\rho_{0})^{1/2}r, \label{cc3}$$ and $\psi$ is the solution of the $D$-dimensional Emden equation $${1\over \xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\psi\over d\xi}\biggr )=e^{-\psi}, \label{cc4}$$ with boundary conditions $$\psi(0)=\psi'(0)=0. \label{cc5}$$ This equation can be obtained from Eq. (\[p4\]) by taking the limit $k\rightarrow +\infty$ and using the limiting form of the Fermi integral $$I_{n}(t)\sim {1\over t}\Gamma(n+1), \qquad (t\rightarrow +\infty). \label{cc6}$$ From Eq. (\[cc1bis\]), we check that the local equation of state of a classical self-gravitating isothermal gas is $p({\bf r})={\rho({\bf r})\over m}k_{B}T$ whatever the dimension of space. The thermodynamical parameters are given by $$\eta=\alpha\psi'(\alpha), \label{cc7}$$ $$\Lambda={D(4-D)\over 2(D-2)} {1\over\alpha\psi'(\alpha)}-{1\over D-2}{e^{-\psi(\alpha)}\over\psi'(\alpha)^{2}}, \label{cc8}$$ $${S-S_{0}\over Nk_{B}}=-{D-2\over 2}\ln\eta-2\ln\alpha+\psi(\alpha)+{\eta\over D-2}-2\Lambda\eta, \label{cc9}$$ $${S_{0}\over Nk_{B}}=\ln\mu+\ln\biggl ({2\pi^{D/2}\over S_{D}}\biggr )+1-{D\over 2}, \label{cc9bis}$$ where $\alpha=(S_{D} G\beta m\rho_{0})^{1/2}R$ is the normalized box radius. For $D=2$, the thermodynamical parameters can be calculated analytically [@sc1]. Introducing the pressure at the box $P=p(R)$, the global equation of state of the self-gravitating gas can be written $${PV\over Nk_{B}T}={1\over D}{\alpha^{2}\over \eta}e^{-\psi(\alpha)}. \label{cc11}$$ We recall that the foregoing expressions can be expressed in terms of the value of the Milne variables $u_{0}=u(\alpha)$ and $v_{0}=v(\alpha)$ at the normalized box radius [@aa; @grand]. The structure and the stability of classical isothermal spheres in $D$ dimensions have been studied in detail in [@sc1]. The classical thermodynamic limit (CTL) of self-gravitating systems, or dilute limit [@vs], is such that $N\rightarrow +\infty$ with $\eta$, $\Lambda$ fixed. If we take $\beta\sim 1$, this implies that $R\sim N^{1/(D-2)}$ and $E,S,J,F\sim N$. The physical distinction between the quantum thermodynamic limit (QTL) and the classical thermodynamic limit (CTL) is related to the existence of long-lived gaseous metastable states as discussed in [@rieutord; @meta]. The completely degenerate limit {#sec_deg} ------------------------------- For $\beta\rightarrow +\infty$ (i.e., $T=0$), the distribution function (\[fd14\]) reduces to a step function: $f=\eta_{0}$ if $v\le v_{F}$ and $f=0$ if $v\ge v_{F}$, where $v_{F}({\bf r})=\sqrt{2(\mu/m-\Phi)}$ is the local Fermi velocity. In that case, the density and the pressure can be explicitly evaluated: $$\rho=\int_{0}^{v_{F}}\eta_{0}S_{D}v^{D-1}dv=\eta_{0}S_{D}{v_{F}^{D}\over D}, \label{d1}$$ $$p={1\over D}\int_{0}^{v_{F}}\eta_{0}S_{D}v^{D+1}dv=\eta_{0}{S_{D}\over D}{v_{F}^{D+2}\over D+2}. \label{d2}$$ Eliminating the Fermi velocity between these two expressions, we find that the equation of state of a cold Fermi gas in $D$ dimensions is $$p=K\rho^{1+2/D}, \qquad K={1\over D+2}\biggl ({D\over\eta_{0}S_{D}}\biggr )^{2/D}. \label{d3}$$ This equation of state describes a $D$-dimensional classical white dwarf star (throughout this paper, we shall call “white dwarf star”, or “fermion ball”, a completely degenerate self-gravitating system. This terminology will be extended to any dimension of space). In $D=3$, classical white dwarf stars are equivalent to polytropes with index $n=3/2$ [@fowler]. In $D$ dimensions, classical “white dwarf stars” are equivalent to polytropes with index [@langevin]: $$n_{3/2}={D\over 2}. \label{d4}$$ The structure and the stability of polytropic spheres in $D$ dimensions have been studied in detail in [@langevin]. It is shown that a polytrope of index $n$ is self-confined for $n<n_{5}=(D+2)/(D-2)$ and stable for $n<n_{3}=D/(D-2)$. Therefore, white dwarf stars ($n=n_{3/2}=D/2$) are self-confined only for $D<2(1+\sqrt{2})$ and they are stable only for $D\le 4$. For $D>4$, quantum mechanics is not able to stabilize matter against gravitational collapse. Thus, $D=4$ is a critical dimension regarding gravitational collapse. $D=2$ is also critical [@sc1]. Therefore, the dimension of space of our universe $2<D=3<4$ lies between two critical dimensions. We now introduce dimensionless parameters associated with $n_{3/2}$ polytropes which will be useful in the sequel. Their density profile can be written $$\rho(r)=\rho_{0}\ \theta^{D/2}(\xi), \label{d5}$$ where $\rho_{0}$ is the central density, $\xi$ is the normalized distance $$\xi=\biggl \lbrack {2S_{D}G\rho_{0}^{(D-2)\over D}\over K(D+2)}\biggr \rbrack^{1/2} r, \label{d6}$$ and $\theta$ is solution of the $D$-dimensional Lane-Emden equation $${1\over \xi^{D-1}}{d\over d\xi}\biggl (\xi^{D-1}{d\theta\over d\xi}\biggr )=-\theta^{D/2}, \label{d7}$$ with boundary conditions $$\theta(0)=1, \qquad \theta'(0)=0. \label{d8}$$ This equation can be obtained from Eq. (\[p4\]) by taking the limit $k\rightarrow 0$ and using the limiting form of the Fermi integral $$I_{n}(t)\sim {(-\ln t)^{n+1}\over n+1}, \qquad (t\rightarrow 0). \label{d9}$$ For $D<2(1+\sqrt{2})$, the solution of the Lane-Emden equation (\[d7\]) vanishes at a finite distance $\xi_{1}$ defining the radius $R_{*}$ of the white dwarf star ([*complete polytrope*]{}). Using the results of [@langevin], the mass-radius relation of $D$-dimensional white dwarf stars is given by $$M^{D-2\over D}R_{*}^{4-D}={K(D+2)\over 2 G S_{D}^{2/D}}\omega_{D/2}^{D-2\over D}, \label{d10}$$ where we have defined $$\omega_{D/2}=-\xi_{1}^{D+2\over D-2}\theta'(\xi_{1}). \label{d11}$$ For $2<D<4$, the mass $M$ decreases with the radius $R_{*}$ while for $D<2$ and for $4<D<2(1+\sqrt{2})$ it increases with the radius (see Fig. \[MRD\]). The mass-radius relation (\[d10\]) exhibits the two critical dimensions of space $D=2$ and $D=4$ discussed previously. For $D=2$, the radius is independent on mass and for $D=4$, the mass is independent on radius (see Sec. \[sec\_caloric\]). The energy of a self-confined white dwarf star is $$E=-\lambda_{D/2}{ GM^{2}\over R_{*}^{D-2}}, \label{d14}$$ where $$\lambda_{D/2}={D(4-D)\over (D-2)(4+4D-D^{2})}. \label{d15}$$ We note that the energy of a white dwarf star vanishes for $D=4$. According to Poincaré’s theorem [@chandra2], this determines the onset of instability. We thus recover the fact that complete white dwarf stars are unstable for $D>4$ [@langevin]. For $D>2(1+\sqrt{2})$, the density of a $n_{3/2}$ polytrope never vanishes (as $n_{3/2}>n_{5}$) and we need to confine the system within a box of radius $R$ ([*incomplete polytrope*]{}) to avoid the infinite mass problem. In that case, the white dwarf star exerts a pressure against the box. White dwarf stars with $R_{*}>R$ when $D<2(1+\sqrt{2})$ are also incomplete. They are arrested by the box at the normalized radius $\xi=\alpha$ with $\alpha=\lbrace 2S_{D}G\rho_{0}^{(D-2)/D}/\lbrack K(D+2)\rbrack\rbrace^{1/2}R$. As shown in [@langevin], the normalized mass and the normalized energy of the configuration parameterized by $\alpha$ are given by $$\eta_{P}\equiv {M\over S_{D}}\biggl\lbrack {2S_{D}G\over K(D+2)}\biggr \rbrack^{D\over D-2}{1\over R^{D(D-4)\over D-2}}=-\alpha^{D+2\over D-2}\theta'(\alpha), \label{d17}$$ $$\Lambda\equiv -{ER^{D-2}\over GM^{2}}={-2\over D^{2}-4D-4}\biggl\lbrace {D(4-D)\over 2(D-2)}\biggl \lbrack 1+(D-2){\theta(\alpha)\over\alpha\theta'(\alpha)}\biggr \rbrack+{2-D\over 2+D}{\theta(\alpha)^{D+2\over 2}\over \theta'(\alpha)^{2}}\biggr\rbrace. \label{d18}$$ In the present context, the normalized mass $\eta_P$ is related to the degeneracy parameter $\mu$ by the relation $$\eta_{P}=\biggl ({2\mu\over D}\biggr )^{2\over D-2}. \label{d19}$$ On the other hand, using Eqs. (\[d10\]) and (\[d14\]), the normalized mass and the normalized energy of a self-confined white dwarf star with $R_{*}<R$ (complete polytrope) are given by $$\eta_{P}=\omega_{D/2}\biggl ({R_{*}\over R}\biggr )^{(D-4)D\over D-2} \label{d20}$$ $$\Lambda=\lambda_{D/2}\biggl ({R\over R_{*}}\biggr )^{D-2}. \label{d21}$$ Eliminating $R_{*}$ between these two relations, we obtain the “mass-energy” relation $$\Lambda \eta_{P}^{(D-2)^{2}\over D(D-4)}=\lambda_{D/2} (\omega_{D/2})^{(D-2)^{2}\over D(D-4)}, \label{d22}$$ which will be useful in our subsequent analysis. Caloric curves in various dimensions {#sec_caloric} ==================================== Series of equilibria and metastable states {#sec_seq} ------------------------------------------ We shall now determine the caloric curve $\beta(E)$ of the self-gravitating Fermi gas as a function of the degeneracy parameter $\mu$ for any dimension of space $D$. This study has already been performed for $D=3$ in [@pt]. The critical points of the Fermi-Dirac entropy $S[f]$ at fixed $E$ and $M$ (i.e., the distribution functions $f({\bf r},{\bf v})$ which cancel the first order variations of $S$ at fixed $E$, $M$) form a series of equilibria parameterized by the uniformizing variable $k$. At each point in the series of equilibria corresponds a temperature $\beta$ and an energy $E$ determined by Eqs. (\[p8\]) and (\[p16\]). In this approach, $\beta$ is the Lagrange multiplier associated with the conservation of energy in the variational problem (\[fd13\]). It has also the interpretation of a kinetic temperature in the Fermi-Dirac distribution (\[fd14\]). We can thus plot $\beta(E)$ along the series of equilibria. There can be several values of temperature $\beta$ for the same energy $E$ because the variational problem (\[fd12\]) can have several solutions: a local entropy maximum (metastable state), a global entropy maximum, and one or several saddle points. We shall represent all these solutions on the caloric curve because local entropy maxima (metastable states) are in general more physical than global entropy maxima for the timescales achieved in astrophysics. Indeed, the system can remain frozen in a metastable gaseous phase for a very long time. This is the case, in particular, for globular clusters and for the gaseous phase of fermionic matter (at high energy and high temperature). The time required for a metastable gaseous system to collapse is in general tremendously long and increases exponentially with the number $N$ of particles (thus, $t_{life}\rightarrow +\infty$ in the thermodynamic limit $N\rightarrow +\infty$) [@meta]. This is due to the long-range nature of the gravitational potential. Therefore, metastable states are in reality stable states. At high temperatures and high energies, the global entropy maximum is not physically relevant [@ko; @ispolatov; @grand; @rieutord]. Condensed objects (e.g., planets, stars, white dwarfs, fermion balls,...) only form below a critical energy $E_{c}$ (Antonov energy) [@antonov; @lbw; @paddy] or below a critical temperature $T_{c}$ (Jeans temperature) [@aa], when the gaseous metastable phase ceases to exist (spinodal point). The case $2<D<4$ {#sec_d3} ---------------- We start to describe the structure of the caloric curve of the self-gravitating Fermi gas for $2<D<4$ (specifically $D=3$). Let us first consider the Fermi gas at $T=0$ (white dwarf stars). The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD3\]. In the present context, it gives the energy of the star as a function of its mass. Since the curve does not present turning points, all the white dwarf star configurations are stable. According to Eq. (\[d10\]), for $2<D<4$, the mass $M$ of a complete white dwarf star is a decreasing function of its radius $R_*$. Therefore, if the system is enclosed within a box, there exists a characteristic mass $$M_*(R)={\chi_{D}\over \eta_{0}^{2\over D-2}G^{D\over D-2}}R^{-{D(4-D)\over D-2}} \label{gtq1}$$ such that for $M>M_*(R)$ the star is self-confined ($R_{*}<R$) and for $M<M_*(R)$, it is restricted by the box. In terms of the dimensionless mass $\eta_P$, complete $n_{3/2}$ polytropes correspond to $\eta_P\ge \omega_{D/2}$ and incomplete $n_{3/2}$ polytropes to $\eta_P\le \omega_{D/2}$. For $2<D<4$, there exists a stable equilibrium at $T=0$ for all mass $M$. We now briefly describe the caloric curve for arbitrary temperature and energy. A more complete description is given in [@pt] for $D=3$. First, we note that, according to Eqs. (\[p10\]), (\[d3\]) and (\[d10\]), $$\mu=\mu_*(D) \biggl ({R\over R_*}\biggr )^{D(4-D)\over 2}, \label{gtq2}$$ where $$\mu_{*}(D)\equiv {D\over 2}(\omega_{D/2})^{D-2\over 2}. \label{gtq3}$$ Therefore, the degeneracy parameter $\mu$ can be seen as the ratio (with some power) between the size of the system $R$ and the size $R_{*}$ of a white dwarf star with mass $M$. Accordingly, a small value of $\mu$ corresponds to a large “effective” cut-off (played by Pauli’s exclusion principle) or, equivalently, to a small system size. Alternatively, a large value of $\mu$ corresponds to a small “effective” cut-off or a large system size. This gives a physical interpretation to the degeneracy parameter. For $\mu\rightarrow +\infty$ (i.e. $\hbar\rightarrow 0$), we recover classical isothermal spheres. In that case, the caloric curve $\beta(E)$ forms a spiral. For finite values of $\mu$, the spiral unwinds due to the influence of degeneracy and gives rise to a rich variety of caloric curves (Fig. \[LHD3\]). For large systems, the caloric curve has a $Z$-shape (“dinosaur’s neck”) and for small systems it has a $N$-shape. The phase transitions in the self-gravitating Fermi gas for $D=3$ and the notion of metastable states, spinodal points, critical points, collapse, explosion, and hysteresis are discussed in [@pt; @ispolatov; @rieutord; @meta]. Similar notions are discussed in [@stahl] for a hard spheres gas. The ground state of the self-gravitating Fermi gas ($T=0$) corresponds to a white dwarf star configuration. For given $\mu$, its structure (radius, energy) is determined by the intersection between the $\Lambda-\eta_P$ curve in Fig. \[LHpD3\] and the line defined by Eq. (\[d19\]). The “white dwarf” is complete ($R_{*}<R$) for $\mu>\mu_{*}(D)$ and incomplete ($R_{*}>R$) otherwise. For $\mu>\mu_{*}(D)$, the normalized energy of the white dwarf is given by $$\Lambda_{max}(D,\mu)=\lambda_{D/2}\biggl ({\mu\over\mu_{*}}\biggr )^{2(D-2)\over D(4-D)}. \label{gtq4}$$ This is the ground state of the self-gravitating Fermi gas corresponding to the asymptote in Fig. \[LHD3bis\] (this asymptote exists for all curves in Fig. \[LHD3\] but is outside the frame). For classical particles ($\hbar=0$), there is no equilibrium state if energy and temperature are below a critical threshold [@antonov; @lbw]. In that case, the system undergoes gravitational collapse and forms binaries (in microcanonical ensemble) or a Dirac peak (in canonical ensemble); see Appendices A and B of [@sc1] and [@grand; @rieutord; @meta]. For self-gravitating fermions, an equilibrium state exists for all values of temperature and for all accessible energies ($E\ge E_{ground}$). Gravitational collapse is arrested by quantum pressure as first realized by Fowler [@fowler]. We shall now show that this claim ceases to be true in dimension $D\ge 4$. The case $4<D<2(1+\sqrt{2})$ {#sec_d41} ---------------------------- We now consider the case $4<D<2(1+\sqrt{2})$ (specifically $D=4.1$). Let us first describe the Fermi gas at $T=0$. The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD4.1\]. For $D>4$, the curves $\eta_{P}(\alpha)$ and $\Lambda(\alpha)$ associated to $n_{3/2}$ polytropes have their extrema at the same point (see Appendix C of [@langevin]). Therefore, the $\Lambda$-$\eta_{P}$ curve presents a cusp at $(\Lambda_{0},\eta_{P,c})$. Past this point in the series of equilibria, $n_{3/2}$ polytropes are unstable. According to Eq. (\[d10\]), for $D>4$, the radius $R_*$ of a self-confined white dwarf star increases with its mass. For $M<M_*(R)$ there exists self-confined white dwarf star configurations. In terms of the dimensionless mass $\eta_P$, this corresponds to $\eta_P\le \omega_{D/2}$ (see Fig. \[LHpD4.1\]). However, such configurations are unstable since they lie after the turning point [@langevin]. Therefore, only incomplete (box confined) white dwarf stars can be stable in $D>4$. Inspecting Fig. \[LHpD4.1\] again, we observe that these configurations exist only below a critical mass $$M_c(R)=\eta_{P,c}(D) S_{D}R^{D(D-4)\over D-2}\biggl\lbrack {K(D+2)\over 2S_{D}G}\biggr \rbrack^{D\over D-2}. \label{gtq5}$$ For $M>M_c(R)$, there is no equilibrium state at $T=0$ for $D>4$. In terms of the dimensionless mass $\eta_P$, equilibrium states exist only for $\eta_{P}<\eta_{P,c}(D)$. The caloric curve for arbitrary value of temperature and energy is represented in Fig. \[LHD4.1\]. For $\mu\rightarrow +\infty$, we recover the classical spiral [@sc1]. For finite values of $\mu$, there exists equilibrium solutions at all temperatures only if $\eta_P<\eta_{P,c}(D)$. Using Eq. (\[d19\]), this corresponds to $$\mu<{D\over 2}\eta_{P,c}(D)^{D-2\over 2}\equiv \mu_{c}(D). \label{gtq6}$$ If $\mu>\mu_{c}(D)$, or equivalently if $M>M_*(R)$, there exists a minimum energy $E_c=-\Lambda_c GM^2/R^{D-2}$ (which appears to be positive) and a minimum temperature $T_c=GM/(\eta_c R^{D-2})$ below which there is no equilibrium state (the values of $\eta_{c}$ and $\Lambda_{c}$ depend on $D$ and $\mu$). In that case, the system is expected to collapse. This is similar to the Antonov instability (gravothermal catastrophe) for classical particles [@antonov; @lbw]. Since we deal here with self-gravitating fermions, we could expect that quantum pressure would arrest the collapse. Our study shows that this is not the case for $D>4$. Quantum mechanics cannot stabilize matter against gravitational collapse anymore. The case $D=4$ {#sec_d4} -------------- The dimension $D=4$ is special because it is the dimension of space above which quantum pressure cannot balance gravity anymore. Therefore, $D=4$ is critical and it deserves a particular attention. First, consider the Fermi gas at $T=0$. It corresponds to a polytrope of index $n_{3/2}=n_3$ [@langevin]. The $\Lambda-\eta_{P}$ curve defined by Eqs. (\[d17\]), (\[d18\]) and (\[d22\]) is represented in Fig. \[LHpD4\]. Since the curve is monotonic the box-confined $n_{3/2}$ polytropes are stable and the complete $n_{3/2}$ polytropes are marginally stable. For $D=4$, the mass of a self-confined white dwarf star is independent on its radius, see Eq. (\[d10\]). It can be expressed in terms of fundamental constants as $$M_{limit}={\omega_{2}\over g S_{4}^{2}}{h^{4}\over m^{5}G^{2}}\simeq 1.44\ 10^{-2}\ {h^{4}\over m^{5}G^{2}}, \label{gtq7}$$ where $\omega_{2}\simeq 11.2$ (we have taken $g=2$ in the numerical application). Mathematically, this is similar to Chandrasekhar’s limiting mass for relativistic white dwarf stars equivalent to $n=3$ polytropes in $D=3$ [@chandra1]. However, it is here a purely classical (i.e. nonrelativistic) result. Relativistic effects will be considered in a forthcoming paper [@relativity]. The energy of the self-confined white dwarf stars is $E=0$. Considering Fig. \[LHpD4\] again, we see that incomplete white dwarf stars exist only for $M<M_{limit}$. In terms of the dimensionless mass $\eta_P$, this corresponds to $\eta<\eta_{P,c}=\omega_{2}\simeq 11.2$. For $M>M_{limit}$, there is no equilibrium state at $T=0$. The caloric curve for arbitrary value of temperature and energy is represented in Fig. \[LHD4\] (see an enlargement in Fig. \[LHD4ZOOM\]). Its description is similar to that of Sec \[sec\_d41\]. For $M>M_{limit}$, or equivalently $\mu\ge \mu_{c}=2\omega_{2}\simeq 22.4.$ there exists a minimum energy $E_c=-\Lambda_c GM^2/R^{2}$ and a minimum temperature $T_c=GM/(\eta_c R^{2})$ below which there is no equilibrium state. The case $D\ge 2(1+\sqrt{2})$ {#sec_d48} ----------------------------- The caloric curves for $D\ge 2(1+\sqrt{2})$ are similar to those of Secs. \[sec\_d41\] and \[sec\_d4\]. There are, however, two main differences. For $D\ge 10$, the classical spiral ceases to exist [@sc1]. Thus, the caloric curve does not wind up as $\mu\rightarrow +\infty$ contrary to Fig. \[LHD4ZOOM\]. On the other hand, for $D\ge 2(1+\sqrt{2})$, it is not possible to construct self-confined white dwarf stars [@langevin]. This is just a mathematical curiosity since complete white dwarfs stars are unstable for $D>4$ anyway. This property changes the unstable branch of the $\Lambda-\eta_{P}$ diagram without consequence on the caloric curves. The $\Lambda-\eta_{P}$ diagram is represented Figs. \[LHpDn5\] and \[LHpD5.1\]. For $D>2(1+\sqrt{2})$, it displays an infinity of cusps towards the singular solution ($\Lambda_{s}$,$\eta_{P,s}$), see Fig. \[LHpD5.1\]. For $D=2(1+\sqrt{2})$, there is just one cusp (see Fig. \[LHpDn5\]) and the Lane-Emden equation (\[d7\]) can be solved analytically. This corresponds to the $D$-dimensional Schuster solution obtained for $n=n_{5}$ [@langevin]. In that case, we find explicitly $$\theta_{5}={1\over \bigl \lbrack 1+{\xi^{2}\over 4(2+\sqrt{2})}\bigr \rbrack^{\sqrt{2}}}. \label{fw1}$$ The normalized mass and the normalized energy can be expressed as $$\eta_{P}={\alpha^{2+\sqrt{2}}\over 2(1+\sqrt{2})\bigl \lbrack 1+{\alpha^{2}\over 4(2+\sqrt{2})}\bigr \rbrack^{1+\sqrt{2}}}, \label{fw2}$$ $$\Lambda_{5}=-2(1+\sqrt{2})\biggl\lbrack 1+{\alpha^{2}\over 4(2+\sqrt{2})}\biggr\rbrack^{2(1+\sqrt{2})}{1\over \alpha^{2(2+\sqrt{2})}}\int_{0}^{\alpha}{\xi^{1+2\sqrt{2}}\over \bigl\lbrack 1+{\xi^{2}\over 4(2+\sqrt{2})}\bigr\rbrack^{2(1+\sqrt{2})}}d\xi. \label{fw3}$$ The case $D=2$ {#sec_d2} -------------- Let us now consider smaller dimensions of space. The dimension $D=2$ is critical concerning gravitational collapse as discussed in [@sc1]. For $D=2$, the relevant Fermi integrals are $I_{0}$ and $I_{1}$. By definition, $$I_{0}(t)=\int_{0}^{+\infty}{dx\over 1+t e^{x}}. \label{t1}$$ Changing variables to $y=e^{x}$, we easily find that $$I_{0}(t)=\ln\biggl (1+{1\over t}\biggr ). \label{t2}$$ Therefore, the Fermi-Poisson equation (\[p4\]) becomes $${1\over\xi}{d\over d\xi}\biggl (\xi {d\psi\over d\xi}\biggr )=\ln\bigl (1+k^{-1}e^{-\psi}\bigr ). \label{t3}$$ $$\psi(0)=\psi'(0)=0. \label{t4}$$ On the other hand, using the identity (\[p3\]), giving $$I'_{1}(t)=-{1\over t}\ln\biggl (1+{1\over t}\biggr ), \label{t5}$$ one finds that $$I_{1}(t)=-\int_{-{1/t}}^{0}{\ln(1-x)\over x}dx=-{\rm Li}_{2}\bigl (-{1\over t}\bigr ), \label{t6}$$ where ${\rm Li}_{2}$ is the dilogarithm. Consider first the Fermi gas at $T=0$. In $D=2$, a white dwarf star is equivalent to a polytrope with index $n_{3/2}=1$. The Lane-Emden equation can then be solved analytically and we obtain $\theta=J_{0}(\xi)$, where $J_{0}$ is the Bessel function of zeroth order. The density drops to zero at $\xi_{1}=\alpha_{0,1}\simeq 2.40$, the first zero of $J_{0}$. Considering the mass-radius relation (\[d10\]) in $D=2$, we see that the radius is independant on mass. Therefore, complete white dwarf stars in two dimensions all have the same radius. It can be written in terms of fundamental constants as $$R_{*}={\xi_1\over 2\pi}\biggl ({h^{2}\over g m^{3}G}\biggr )^{1/2}=0.27\ {h\over m^{3/2}G^{1/2}}. \label{t7}$$ The relation between the mass and the central density of the white dwarf star is $$M={\rho_{0}\over 4\pi^{2}}{h^{2}\over g m^{3}G}\xi_{1}|\theta'_{1}|, \label{t8}$$ where $\theta'_{1}=J_{0}'(\alpha_{0,1})\simeq -0.52$. Thus, the density profile of a two-dimensional white dwarf star can be written $$\rho(r)=\rho_{0}J_{0}\bigl ({\xi_{1}r\over R_{*}}\bigr ). \label{t9}$$ This is similar to the vorticity profile of a minimum enstrophy vortex in 2D hydrodynamics [@jfm; @gt]. The energy of a complete polytrope of index $n$ in $D=2$ is $E=-(n-1)GM^{2}/8+(1/2)GM^{2}\ln (R_{*}/R)$ with the convention $\Phi(R)=0$ [@langevin]. Therefore the energy of a 2D white dwarf star is $$E={1\over 2}GM^{2}\ln \bigl ({R_{*}\over R}\bigr ). \label{t10}$$ Two-dimensional white dwarf stars exist for any mass $M$ and they are stable. Noting that $R_{*}/R=(\mu_{*}/\mu)^{1/2}=\xi_1/\sqrt{\mu}$ where $\mu=4\pi^{2}\eta_{0}GR^{2}$, we can write the normalized energy of the self-confined white dwarf star as $$\Lambda={1\over 2}\ln\biggl ({\sqrt{\mu}\over\xi_1}\biggr ). \label{t11}$$ Let us now consider the case of incomplete white dwarf stars that are confined by the box ($R_*>R$). This corresponds to $\mu<\xi_1^2$. Using Eq. (\[d6\]), we find that $\alpha=\sqrt{\mu}$. Then, using the results of [@langevin], we find that the normalized energy of a box-confined white dwarf star in two dimensions is $$\Lambda=-{1\over 2}{J_{0}(\sqrt{\mu})\over\sqrt{\mu}J_{1}(\sqrt{\mu})}. \label{t12}$$ We now consider the self-gravitating Fermi gas at finite temperature $T\neq 0$. According to Eq. (\[p6\]) we have $\alpha=\sqrt{\mu}$. Using Eq. (\[p8\]), we obtain $$\eta\equiv \beta GMm=\sqrt{\mu}\ \psi'(\sqrt{\mu}). \label{t12bis}$$ We need to calculate the energy specifically because the expression (\[p16\]) breaks down in $D=2$. The kinetic energy $K=\int p d^{2}{\bf r}$ can be written $${K\over GM^{2}}={1\over \eta^{2}}\int_{0}^{\sqrt{\mu}}I_{1}(k e^{\psi})\xi d\xi. \label{t13}$$ On the other hand, using an integration by parts, the potential energy is given by $$W=-{1\over 4\pi G}\int (\nabla\Phi)^{2}d^{2}{\bf r}, \label{t14}$$ where we have taken $\Phi(R)=0$. Introducing the dimensionless quantities defined in Sec. \[sec\_para\], we get $${W\over GM^{2}}=-{1\over 2\eta^{2}}\int_{0}^{\sqrt{\mu}}\psi'(\xi)^{2}\xi d\xi. \label{t15}$$ Summing Eqs. (\[t13\]) and (\[t15\]), the total normalized energy of the Fermi gas in two dimensions is $$\Lambda\equiv -{E\over GM^{2}}=-{1\over \eta^{2}}\int_{0}^{\sqrt{\mu}}I_{1}(k e^{\psi})\xi d\xi+{1\over 2\eta^{2}}\int_{0}^{\sqrt{\mu}}\psi'(\xi)^{2}\xi d\xi. \label{t16}$$ The corresponding caloric curve is plotted in Fig. \[LHD2\]. For $\mu\rightarrow +\infty$, we recover the classical caloric curve displaying a critical temperature $k_{B}T_{c}=GMm/4$ [@sc1]. Below $T_{c}$, a classical gas experiences a gravitational collapse and develops a Dirac peak [@sc1]. When quantum mechanics is taken into account, the collapse stops when the system becomes degenerate. The Dirac peak is replaced by a fermion ball surrounded by a dilute halo. At $T=0$, we have a pure Fermi condensate without halo. This is the ground state of the self-gravitating Fermi gas corresponding to the vertical asymptotes in Fig. \[LHD2\]. For $\mu<\xi_1^2$ (incomplete white dwarf stars), the minimum energy is given by Eq. (\[t12\]) and for $\mu<\xi_1^2$ (complete white dwarf stars) by Eq. (\[t11\]). This discussion concerning the difference between Dirac peaks (for classical particles) and fermion balls (for quantum particles) in the canonical ensemble remains valid for $2\le D<4$. Note also that there is no collapse (gravothermal catastrophe) in the microcanonical ensemble in $D=2$ [@klb; @sc1]. The case $D<2$ {#sec_d1} -------------- We finally conclude by the case $D<2$ (specifically $D=1$). First, we consider the Fermi gas at $T=0$. The $\Lambda-\eta_{P}$ curve which gives the energy of the star as a function of its mass is represented in Fig. \[LHpD1\]. Since the curve does not present turning points, all the white dwarf star configurations are stable. According to Eq. (\[d10\]), for $D<2$, the mass $M$ of a complete white dwarf star increases with its radius $R_*$. Therefore, for $M<M_*(R)$ the star is self-confined and for $M>M_*(R)$ it is restricted by the box. There exists a stable equilibrium state at $T=0$ for all mass. In terms of the dimensionless mass $\eta_P$, complete $n_{3/2}$ polytropes correspond to $\eta_P\le \omega_{D/2}$ and incomplete $n_{3/2}$ polytropes to $\eta_P\ge \omega_{D/2}$. This situation is reversed with respect to that of Fig. \[LHpD3\]. The caloric curve for arbitrary temperature and energy is represented in Fig. \[LHD1\]. For $\mu\rightarrow +\infty$ (i.e. $\hbar\rightarrow 0$), we recover the curve obtained in [@sc1] for classical isothermal systems. The caloric curve $\beta(E)$ is monotonic. Therefore, there is no phase transition for $D<2$. Thus, the change in the caloric curve due to quantum mechanics is not very important since an equilibrium state (global maximum of entropy or free energy) already exists for any accessible energy $E$ and any temperature $T$ in classical mechanics. Quantum mechanics, however, changes the ground state of the system. The ground state of the self-gravitating Fermi gas ($T=0$) corresponds to a white dwarf star configuration. Its structure (radius, energy) is determined by the intersection between the $\Lambda-\eta_P$ curve in Fig. \[LHpD1\] and the line defined by Eq. (\[d19\]). The “white dwarf” is complete ($R_{*}<R$) for $\mu>\mu_{*}(D)$ and incomplete ($R_{*}>R$) otherwise. For $\mu>\mu_{*}(D)$, the normalized energy of the white dwarf is given by Eq. (\[gtq4\]). This is the ground state of the self-gravitating Fermi gas corresponding to the asymptote in Fig. \[LHD1\]. In $D=1$, it is possible to obtain more explicit results. Using the results of [@langevin], for $n_{3/2}=1/2$ polytropes, we have $\xi_{1}=(3\pi/4)^{1/2}\Gamma(5/3)/\Gamma(7/6)\simeq 1.49$ and $|\theta'_{1}|=2/\sqrt{3}\simeq 1.15$. Therefore, $\omega_{1/2}=0.349$ and $\mu_{*}=0.846$. For $\mu>\mu_{*}=0.846$, the normalized energy of a complete white dwarf star (ground state) is $$\Lambda_{min}=-{3\over 7}\biggl ({\mu_{*}\over \mu}\biggr )^{2/3}. \label{bff}$$ Conclusion {#sec_conclusion} ========== In this paper, we have studied how the dimension of space affects the nature of phase transitions in the self-gravitating Fermi gas. Since this model has a fundamental interest in astrophysics [@chavcape] and statistical mechanics [@houches], it is important to explore its properties thoroughly even if we sacrifice for practical applications. It is well-known in statistical mechanics that the dimension of space plays a crucial role in the problem of phase transitions. For example, concerning the Ising model, the behaviour in $D=1$ and $D\ge 2$ is radically different [@huang]. We have reached a similar conclusion for the self-gravitating Fermi gas. The solution of the problem in $D<2$ does not yield any phase transition. In $D=2$, phase transitions appear in the canonical ensemble but not in the microcanonical ensemble. In $D>2$, phase transitions appear both in microcanonical and canonical ensembles in association with gravitational collapse. The beauty of self-gravitating systems, and other systems with long-range interactions, is their simplicity since the mean-field approximation is exact in any dimension. Therefore, the mean-field theory does [*not*]{} predict any phase transition for the self-gravitating Fermi gas in $D=1$, contrary to the Ising model. At a more philosophical level, several scientists have examined the role played by the dimension of space in determining the form of the laws of physics. This question goes back to Ptolemy who argues in his treatise [*On dimensionality*]{} that no more than three spatial dimensions are possible in Nature. In the $18^{\rm th}$ century, Kant realizes the deep connection between the inverse square law of gravitation and the existence of three spatial dimensions. In the twentieth century, Ehrenfest [@ehrenfest] argues that planetary orbits, atoms and molecules would be unstable in a space of dimension $D\ge 4$. Other investigations on dimensionality are reviewed in the paper of Barrow [@barrow]. Although we ignored this literature at the begining, our study clearly enters in this type of investigations. We have found that the self-gravitating Fermi gas possesses a rich structure and displays several characteristic dimensions $D=2$, $D=4$, $D=2(1+\sqrt{2})$ and $D=10$. Moreover, as already noted in [@langevin], the dimension $D=4$ is critical because at that dimension quantum mechanics cannot stabilize matter against gravitational collapse, contrary to the situation in $D=3$. Interestingly, this result is similar to that of Ehrenfest although it applies to white dwarf stars instead of atoms. The dimension $D=2$ is also critical as found in [@sc1] and in different domains of physics. Therefore, the dimension of our (macroscopic) universe $D=3$ plays a very special role regarding the laws of physics (this is illustrated in Fig. \[MRD\]). Following the far reaching intuition of Kant, we can wonder whether the three space dimensions are a consequence of Newton’s inverse square law, rather than the opposite. We note also that extra-dimensions can appear at the microscale, an idea originating from Kaluza-Klein theory. This idea took a renaissance in modern theories of grand unification. Our approach shows that already at a simple level, the coupling between Newton’s equations (gravitation) and Fermi-Dirac statistics (quantum mechanics) reveals a rich structure as a function of $D$. Relativistic effects will be considered in a forthcoming paper [@relativity]. Finally, our study can shed light on the mathematical properties of the Vlasov-Poisson system. Indeed, there is a close connexion between collisionless stellar systems and self-gravitating fermions [@lb; @csr; @cs; @dubrovnik]. For example, the fact that the Vlasov equation does not blow up (i.e., experiences gravitational collapse) in $D=3$ for non singular initial conditions can be related to a sort of exclusion principle, as in quantum mechanics. Due to the Liouville theorem in $\mu$-space, the distribution function must remain smaller that its maximum initial value $f\le \eta_{0}$ and this prevents complete collapse [@cs; @robert], unlike for collisional stellar systems [@lbw] described by the Landau-Poisson system. Since quantum mechanics cannot arrest gravitational collapse in $D\ge 4$ (for sufficiently low energies), this suggests that the Vlasov-Poisson system can probably blow up for $D\ge 4$. This remark could be of interest for mathematicians. [**Acknowledgements**]{} I acknowledge interesting discussions with P. Biler, T. Nadzieja and R. Stańczy. I also thank B. Douçot for encouragements. [10]{}

ArXiv

--- author: - 'Amit Vurgaft, Shai B. Elbaz and Amir D. Gat' bibliography: - 'Bib\_File.bib' date: 2019 title: 'Forced motion of a cylinder within a liquid-filled elastic tube' --- Introduction ============ This work studies the dynamic response of a liquid-filled tube due to the forced axial motion of an internal rigid cylinder. This configuration is relevant to various minimally invasive medical procedures in which slender devices are inserted into fluid-filled biological vessels. For example, recent technologies for percutaneous revascularization involve insertion of cylindrical devices into blocked blood vessels [@rogers2007overview; @davis2015no]. Similar methods are used in the field of interventional radiology, such as laser angioplasty [@serruys1993quantitative], microvascular plug deployment [@pellerin2014microvascular] and removing blood clots by thrombolysis and thrombectomy [@dunn2015thrombectomy]. Additional relevant procedures are endoscopies of body organs which contain liquid, e.g. cystoscopy [@chew1996urethroscopy], as well as the frequently used procedure of urinary catheterization [@nacey1993evolution]. The examined configuration is actuated by an external force which induces a viscous flow-field, applying fluidic stress on the fluid-solid interface and creating deformation of the tube, thus modifying the flow-field. This interaction between viscous and elastic effects is relevant to various research fields [@duprat2015fluid], including locomotion at low-Reynolds-numbers [@wiggins1998flexive; @camalet2000generic], flow in flexible and collapsible tubes [@heil1996stability; @heil1998stokes; @marzo2005three] and the dynamics of membrane-bound particles [@vlahovska2011dynamics; @abreu2014fluid] among many others [@lister2013viscous; @hewitt2015elastic; @elbaz2016axial]. Fluid-solid interactions, in geometries similar to the one investigated here, have been extensively studied in the context of medical operations and biological flows. For example, previous works analyzing the fluid mechanics of catheterized arteries include [@karahalios1990some], who investigated flow in an axisymmetric cross-section of a catheterized artery, and estimated the shear stress at the artery wall due to catheterization. [@sarkar2001nonlinear] modeled the pulsating blood flow in the annular cross-section between a catheter and an elastic tube and calculated the induced pressure gradient along the elastic tube. [@vajravelu2011mathematical] modelled a non-Newtonian Herschel-Bulkley fluid flow in an elastic tube, representing a catheterized artery. [@kumar] showed that the effective viscosity, flow rate and arterial wall shear stress are significantly altered in the catheterized site. Other relevant works studied the motion of closely-fitting solids in elastic tubes filled with viscous fluid, as a model of blood cells in narrow capillaries. Such problems involve the effects of the hydrodynamic stress generated in the lubricating film between the two bodies and the elastic stress which develops as a result of the contact of the particle and the tube. This type of analysis was first done by [@lighthill1968pressure] who provided analytical solutions for the pressure-field in such configurations, as well as predicting a necking phenomena next to the contact point. Later, [@tozeren1982flow] found the force required to maintain the motion of the particle, in addition to calculation of the fluid pressure-field and the elastic deformation. [@tani2017motion] examined the friction force between the inner solid and the elastic tube in the case of a dry sphere-tube contact and in the case of lubrication by a thin fluid layer. Another recent relevant work is [@park2018viscous], who presented both analysis and experimental data of viscous flow in a bio-inspired soft valve configuration. In this context, the authors studied a cylinder and a concentric sphere with a narrow gap between them, and obtained the effect of the sphere on the nonlinear relation between the externally applied pressure difference and the flow rate. The aim of this work is to analyze the motion of a solid cylinder within a liquid-filled tube, due to a prescribed external force, in relevance to various medical procedures. In §2 we begin by defining the geometrical and physical properties of the examined configuration, as well as stating the governing lubrication equation and integral constraints of the system. Analysis of the simpler case of small linearized elastic deformations is presented in §3. We then present the large deformation nonlinear dynamics for insertion (§4) and extraction (§5) of the cylinder from the elastic tube. In §6 we provide concluding remarks. Problem formulation & governing equations ========================================= We study a Newtonian, incompressible, creeping flow due to the motion of a rigid cylinder within a liquid-filled elastic tube, as shown in figure \[figure\_extension\]. The inner cylinder represents a simplified model of a minimally invasive medical device and the linearly elastic tube is a simplified model of a bounding biological vessel (artery, urethra, etc.). An external force is applied on the cylinder, which consequently moves in relation to the tube. ![**Illustration of the examined configuration (at rest) and definition of the coordinate system.** A rigid tube is located within a semi-closed liquid-filled elastic tube. The cylinder can be extracted from, or inserted into, the tube by an external axial force. $h_0$ is the initial gap between the cylinder and the tube, $l_p$ is the length of the inserted tube, $r_p$ is the radius of the tube, $l$ is the length from the inlet to the constriction closing the tube, $r_i$ and $r_o$ are the inner and outer radii of the tube, respectively.[]{data-label="figure_extension"}](Figure_1.eps){width="100.00000%"} Assuming axisymmetry, we denote a cylindrical coordinate system $(x,r)$ with origins at the center of the opening of the tube. We denote time $t$, liquid velocity $(u,v)$, liquid gauge pressure $p$, liquid viscosity $\mu $, liquid density $\rho $ and volume flux ${v_q(t)}$ entering or existing the tube through the inlet. We denote the length of the cylinder which penetrated into the tube ${l_p}$, Young’s modulus of the tube $E$, tube radial deformation ${d_{r}}$, tube inner radius ${r_i}$ and outer radius ${r_o}$, tube length $l$, tube wall thickness $w={r_o}-{r_i}$, inner cylinder radius at rest ${r_p}$. The gap between the cylinder and the tube at rest is ${h_0}={r_i}-{r_p}$. ${f_e}(t)$ denotes external force applied on the cylinder. $\tilde z$ is an auxiliary moving coordinate located at the tip of the penetrating cylinder, and related to the $x$ coordinate by $\tilde z = {l_p}(t)-x$. Hereafter, normalized variables are denoted by uppercase letters and characteristic parameters are denoted by lowercase letters with asterisks (e.g. if $f$ is a dimensional variable, ${f^{\rm{*}}}$ is the characteristic value of $f$ and $F = f/{f^{\rm{*}}}$ is the corresponding normalized variable). We define the normalized coordinates $X,R$ (starting at the tube inlet), the normalized moving coordinate $\tilde Z$ (starting at the penetrated side of the cylinder), and time $T$ $$X = \frac{x}{l},\quad R=\frac{r}{r_i},\quad \tilde Z = \frac{{\tilde z}}{{{l_p}}}=\frac{{{l_p}-x}}{{{l_p}}},\quad T=\frac{t}{t^*}. \label{scaling_1}$$ The normalized radial deformation $D_r$, pressure $P$, external force $F_e$, and penetrated length $L_p$ are $${D_r} = \frac{{{d_r}}}{{d_r^*}}, \quad P = \frac{p}{{p^ * }},\quad {F_e} = \frac{{{f_e}}}{{f_e^ * }},\quad {L_p} = \frac{{{l_p}}}{{l_p^ * }} \label{scaling_2}$$ where the relations between $d_{r}^*$, $p^*$, $l_p^*$, $f_e^*$ and $t^*$ are derived for various limits in the following sections. The ratio of initial gap $h_0$ to characteristic radial deformations is denoted by $${\lambda _h} = \frac{{{h_0}}}{{d_r^ * }}.$$ Small ratios required in applying the lubrication approximation are $$\frac{h_0}{r_i} \ll 1,\quad\frac{{{d_{r}^*+h_0}}}{{{l_p^*}}} \ll 1,\quad \frac{{{d_{r}^*+h_0}}}{{{l_p^*}}}\frac{\rho u^* (h_0+d_r^*)}{\mu}\ll1 \label{UP200}$$ corresponding to slender configuration in both axial and radial coordinates, along with negligible inertial effects. The elastic shell model requires the small ratios, $$\frac{w}{{{r_i}}} \ll 1,\quad \frac{{{d_{r}^ *} }}{{{r_i}}} \ll 1$$ corresponding to thin wall thickness and small elastic deformations. The ratio between the penetrated length of the cylinder and the length of the tube is defined by $${\varepsilon_l} = \frac{{{l_p^ *} }}{l}<1.$$ For the majority of this work, we will use the leading-order relation between the pressure inside the tube and its radial deformation [@timoshenko1959theory], $$P(X,T)= \frac{{Ewd^*_r}}{{{p^* r_i^2}}}{D_r}(x,t), \label{pdr}$$ and thus $p^*=d_r^*Ew/r_i^2$, and $P=D_r$ for all cases hereafter. In order to obtain the governing equations, we apply the lubrication approximation for flow in the gap region between the cylinder and the tube. Normalizing the lubrication equations according to (\[scaling\_1\])-(\[scaling\_2\]), neglecting $O(h_0/r_i)$ terms, and using standard procedures [@leal2007advanced], yields the relevant Reynolds equation $$\begin{gathered} {\frac{{\partial {D_r}}}{{\partial T}}} -\frac{t^*Ew h_0^3}{r_i^2 12\mu (l_p^*)^2}\frac{\partial }{\partial \tilde Z}\left[\frac{1}{L_p^2 }\frac{\partial P}{\partial \tilde Z}\left( 1+\frac{D_r}{\lambda_h}\right)^3 \right] +\lambda_h\frac{\partial }{\partial \tilde Z} \left[ \frac{\partial {L_p}}{\partial T} \frac{1}{2L_p}\left( 1 +\frac{D_r}{\lambda_h}\right) \right] =0. \label{Def_equation}\end{gathered}$$ The Reynolds equation is supplemented by two integral constraints. The first is integral mass conservation, given in scaled form by $$L_p -L_p(0) - \frac{2 d_r^* l}{ r_i l_p^*} \int_0^1 {D_r dX} + \frac{{h_0^3{p^*}{t^*}}}{{6{r_p}\mu {l_p^*}^2}} \left(\int_0^T \frac{1}{L_p}\frac{\partial P}{\partial \tilde Z}\bigg|_{\tilde Z = 1}d\tilde T \right) =0, \label{Mass_eq}$$ where gauge pressure at the inlet is zero, $P(X=0)={D_r}(X = 0) = 0$. The first term in (\[Mass\_eq\]) is the volume of liquid displaced by the advancing cylinder, the second term is the volume change due to solid deformation and the last term is the total volume of the liquid which exited the system through the annular inlet since $T=0$. The second constraint is scaled integral momentum equation on the inner cylinder, given by $$F_e-\frac{\pi r_p^2 p^*}{f_e^*}P(\tilde Z=0)-\frac{2\pi r_p \mu {l_p^*}^2}{ d_r^* f_e^* t^*}\left[ {L_p}\frac{{\partial {L_p}}}{{\partial T}}\int_0^1\frac{d\tilde{Z}}{\lambda_h+D_r} \right]=0, \label{Force_eq}$$ where the first term is the external force, the second term is the pressure at the base of the penetrating cylinder ($\tilde Z=0$) and the third term is shear stress of the cylinder wall ($R=R_p$). Since a general analytic solution of the above equations is not available, approximated solutions for various limits will be pursued in the following sections in order to provide insight regarding the examined configuration. We begin by examining the simplest linearized limit. Linearized insertion & extraction dynamics, ${|f_e|}\ll E h_0 w $ {#SDFLH} ================================================================= In this section we begin studying this problem by examining the simplified linearized case of small deformations where ${\lambda _h} \gg 1$. For $\lambda_h=h_0/d_r^*\gg1$ and $h_0 l/r_i l_p^*\lesssim 1$, order of magnitude analysis of (\[Mass\_eq\]) yields $$t^*=\frac{{6{r_p}\mu {l_p^*}^2}}{h_0^3{p^*}}, \label{tstar}$$ and substituting (\[tstar\]) into (\[Force\_eq\]) yields $p^*=f_e^*/\pi r_p^2$ and $P(\tilde Z=0)=F_e+O(h_0^2/r_p^2)$. In order to obtain approximated solution of (\[Def\_equation\]), under the integral constraints (\[Mass\_eq\]) and (\[Force\_eq\]), we introduce the regular asymptotic expansions $$P\left( {\tilde Z,T} \right)=D_r\left( {\tilde Z,T} \right) =D_{r,0}+\lambda_h^{-1}D_{r,1}$$ and $${L_p}\left( T \right)=L_{p,0}+\lambda_h^{-1}L_{p,1}.$$ \[asymptotic\_expan\] Substituting the expansions (\[asymptotic\_expan\]) and collecting terms for each order, the leading order of (\[Def\_equation\])-(\[Force\_eq\]) is given by $$\quad \frac{\partial^2 P_0}{\partial \tilde Z^2}=0, \quad { {L_{p,0}}} - {L_p(0)} +\int_0^T \frac{1}{L_{p,0}}\frac{\partial P_0}{\partial \tilde Z}\bigg|_{\tilde Z = 1}d\tilde T = 0, \label{leadingO}$$ along with the boundary conditions $P_0(\tilde Z=0,T)=F_e$ and $P_0(\tilde Z=1,T)= 0$. The leading order equations are readily solved, yielding $${{L_{p,0}}(T) = \sqrt {2 {\int\limits_0^T {{F_e}(\tilde{T})d\tilde{T}} } + {L_p(0)}^2} },\quad {P_0}\left( {\tilde Z,T} \right) = \left( {1 - \tilde Z} \right){F_e}. \label{(A)}$$ Substituting the expansions (\[asymptotic\_expan\]) and collecting terms of order $O\left({\lambda_h}^{-1}\right)$ of (\[Def\_equation\])-(\[Force\_eq\]), yields the first-order equations $$\frac{\partial ^2 P_1}{\partial \tilde{Z}^2} = 0, \quad { {L_{p,1}}} - {\int_0^1 {{D_{r,0}}dX} }- {\int \limits_0^T \frac{{\partial {P_0}}}{{\partial \tilde Z}}\bigg|_{\tilde Z = 1}\frac{{{{ {{L_{p,1}}} }}}}{{{{ {{L_{p,0}^2}} }}}}d\tilde T} = 0, \label{M1}$$ along with the homogeneous boundary conditions $P_1(\tilde{Z}=0,T)=0$ and $P_1(\tilde{Z}=1,T) = 0$. In order to solve (\[M1\]), the integral of the leading order deformation is split into two regions: $$\int_0^1 {{D_r}dX} = {\varepsilon_l}{L_p}\int_0^1 {{D_r}d\tilde Z} + \int_{{\varepsilon_l}{L_p}}^1 {{D_r}dX}, \label{spldef}$$ where the first region represents the penetrated region (see region $\tilde z >0$ in Figure 1) and the second region represents the unpenetrated region of the tube in which the pressure is approximately constant and equals $P_0(\tilde Z=0,T)=F_e$. Substituting (\[pdr\]) into (\[spldef\]), yields $$\begin{split} \int_0^1 {{D_{r,0}}dX} = {\varepsilon_l}{ {L_{p,0}}}\int_0^1 {\left( {1 - \tilde Z} \right){F_e }d\tilde Z} + {F_e}\left( {1 - {\varepsilon_l}{{ {{L_{p,0}}} }}} \right)= F_e\left( {1 - \frac{{{\varepsilon_l}}}{2}{{ {{L_{p,0}}} }}} \right). \label{Dr to P} \end{split}$$ Differentiating (\[M1\]) with respect to $T$ after substituting the expression obtained in (\[Dr to P\]), as well as the solution of $P_{0}(\tilde{Z},T)$ from (\[(A)\]), yields the ODE governing ${{L_{p,1}}}$ $$\begin{split} \frac{{\partial {{ {{L_{p,1}}} }}}}{{\partial T}} + \frac{{ {{ {{F_{e}}} }}}}{{L_{p,0}^2}}{ {L_{p,1}}} = \frac{\partial}{\partial T}\left[F_e\left(1-\varepsilon_p\frac{L_{p,0}}{2}\right)\right] \label{B} \end{split}$$ and thus the first-order correction, which accounts for effects of elasticity, is $$L_{p,1}=e^{-\int_{0}^{T}{\frac{F_e}{L_{p,0}^2}d\tau}}\int_{0}^{T}{\frac{\partial}{\partial T}\left[F_e\left(1-\varepsilon_p \frac{L_{p,0}}{2}\right)\right]e^{\int_{0}^{\tilde{T}}{\frac{F_e}{L_{p,0}^2}d\tau}}d\tilde{T}},\quad P_1\left({\tilde{Z},T} \right)= 0. \label{UP300}$$ ![**Motion due to various external force profiles.** Smooth lines denote (rigid) leading order solution (\[(A)\]) and dashed lines denote first order solutions which include effects of elasticity (\[UP300\]). In all panels $L_{p}(T=0)=0.5$, $\varepsilon_p=0.1$, $h_0/r_p=0.1$ and $\lambda_h=10$. The examined force profiles (see inserts) are ${F_e(T)} =5\cdot H(T-0.2)$ (a), $-0.5\cdot \tanh(2T)$ (b), $ 1.5\cdot{\exp [{{{{{\left( {-T - 0.5 } \right)}^2}}}/{{2\cdot{0.05 ^2}}}}]}$ (c), $ -0.4\cdot {\exp[{{{{{\left( {-T - 0.5 } \right)}^2}}}/{{2\cdot{0.05 ^2}}}}]}$ (d). The leading order motion is always is the direction of the external force. However, elastic effect yield extrema points for non-monotonous forces (panels c,d), along with a motion of the cylinder in a direction opposite to the external force.[]{data-label="PUSH_SMALL_ALL"}](Figure_2.eps){width="\textwidth"} We illustrate these results in Figure \[PUSH\_SMALL\_ALL\], showing the motion of the cylinder for various external force profiles. The leading order solutions $L_{p,0}$ given by (\[(A)\]) represent reference rigid configurations and are marked by solid lines. The effect of elasticity is presented by the difference between the leading and first-order solutions, $L_{p,0}+\lambda_h^{-1}L_{p,1}$ given by (\[UP300\]) and marked by dashed lines. The actuating external force vs. time profiles are presented as inserts within the panels. For all panels, normalized initial insertion of the cylinder is $L_{p}(T=0)=0.5$, ratio of cylinder to tube length is $\varepsilon_p=l_p^*/l=0.1$, initial gap to tube radius ratio is $h_0/r_p=0.1$ and initial gap to radial displacement ratio is $\lambda_h=h_0/d_r^*=10$. Panels (a) and (b) present the motion of the inner cylinder due to a sudden positive forcing (${F_e} =5 H(T-0.2)$) and gradual negative forcing ($F_e=-0.5 \tanh(2T)$). For positive forcing, as expected, elasticity increases the penetration of the cylinder into the tube. While the positive pressure reduces viscous resistance via increased gap thickness, the dominant effect is related to deformation created ahead of the inner cylinder, which allows penetration of the cylinder due to displacement of liquid via the increased cross-sectional areas. This is evident also in panel (b), where elasticity decrease $L_p$, even even though negative pressure increases viscous resistance in this case. Panels (c) and (d) present forcing patterns of respectively positive and negative Gaussian profiles with mean of 0.5 and variance of $0.05^2$. In this case, tube elasticity creates a maxima point, and thus change the direction of motion of the inner cylinder as the elastic energy stored in the tube is gradually released. The linearized limit used in this section allowed short preliminary examination of the system dynamics for a simplified case, showing the effect of elasticity on viscous resistance, as well as the effect of potential elastic energy. We proceed to examine non-linear configurations in the following sections. Nonlinear insertion dynamics, ${f_e}\gg E h_0 w $ {#insertionPart} ================================================= ![**Illustration of the configuration for an external insertion force, in the limit of large deformation.** We distinguish between three domains: the unpeeled region, the peeled region and the uniformly pressurized region. These domains change in time due to the advance of the cylinder, $l_{p}(t)$, and the propagation of the peeling front, ${\tilde z_{F}}(t)$.[]{data-label="figure_compress"}](Figure_3.eps){width="90.00000%"} This section examines a sufficiently large positive external force which inserts the cylinder into the tube and creates large deformations compared with the initial gap between the inner cylinder and the external tube (i.e. $\lambda_h=d_r^*/h_0\gg1$). In this limit, as shown below, the Reynolds’ evolution equation (\[Def\_equation\]) is reduced to a Porous-Medium-Equation, a nonlinear diffusion equation characterized by solutions with distinct non-smooth fronts. The non-smooth front separates between a pressurized deformed region and a region with trivial solution of zero gauge pressure and deformation in which the liquid pressure did not propagate yet. We refer the reader to [@vazquez2007porous] for a detailed discussion of such equations. This section will focus on dynamics involving such a front, and thus an additional geometric division is required, separating the region before the front and after the front. This division is presented in figure \[figure\_compress\]. The location of the front is denoted hereafter as ${\tilde z_F}$, and in normalized form by $${\tilde Z_F} = \frac{{{{\tilde z}_F}}} {{{l_p}}}.$$ The three different domains are presented in Figure \[figure\_compress\] and include: (I.) The unpeeled region ${{\tilde z}_F}(t) < \tilde z \le {l_p}(t)$ to which the deformation front has not yet reached.(II.) The peeled lubrication region $0 < \tilde z \le {{\tilde z}_F}(t)$, and (III.) ${l_p}(t) < x \le l$, the uniformly pressurized bulk of the tube. Scaling & derivation of governing equations ------------------------------------------- The above physical limit allows to simplify the governing Reynolds’ equation, as well as the integral mass and momentum conservation constraints (\[Mass\_eq\])-(\[Force\_eq\]). The compact support of the fluidic front (see discussion below (\[Def\_equation\_nonlinear\_OUTER\_PME\])) allows to simplify integral mass conservation to $$\begin{split} {L_p} - {L_p(0)} = \frac{2 r_i d_r^* l}{r_p^2 l_p^*} \left[(1-\varepsilon_l L_p) P(\tilde Z =0)+\varepsilon_l {{L_p}\int_0^{{{\tilde Z}_F}} {{D_{r}}d\tilde Z} } \right]. \label{3.57} \end{split}$$ For the integral force conservation equation, the effect of shear at the peeled region II can be neglected compared with the unpeeled region I, thus simplifying (\[Force\_eq\]) to $${{F_e} - \frac{{\pi {r_p}^2{p^ *} }}{{{f_e^ *} }}{P(\tilde{Z}=0)} - \frac{{2\pi {r_p}\mu {l_p^*}^2}}{{{f_e^ *} {h_0}{t^*}}}\left[ {{L_p}\frac{{\partial {L_p}}}{{\partial T}}(1 - {{\tilde Z}_F})} \right] = 0}. \label{F_e_gov}$$ Together with (\[pdr\]) and $p^*=f_e^*/\pi r_p^2$ we can simplify (\[3.57\]) to $${{P(\tilde{Z}=0)}= \frac{{E{w}{{l_p^*}}}}{{2l{p^ * r_p} }}\left[ {\frac{{{L_p} - {L_p(0)}}}{{1 - {\varepsilon_l}{L_p}}}} \right] - {\varepsilon_l}\left[ {\frac{{{L_p}\int_0^{{{\tilde Z}_F}} {{D_{r}}d\tilde Z} }}{{1 - {\varepsilon_l}{L_p}}}} \right]}, \label{P_t_gov}$$ and scaling of (\[P\_t\_gov\]) therefore yields $l_p^*=p^ * 2l{ r_p}/Ew$. Substituting (\[P\_t\_gov\]) into (\[F\_e\_gov\]), the set of equations governing the large deformation limit are $$\begin{split} {F_e} = \left[ {\frac{{{L_p}\left( {1 - {\varepsilon_l}\int_0^{{{\tilde Z}_F(T)}} {{D_{r}}d\tilde Z} } \right) - {L_p(0)}}}{{1 - {\varepsilon_l}{L_p}}}} \right] + \Pi_1\left[ {{L_p}\frac{{\partial {L_p}}}{{\partial T}}(1 - {{\tilde Z}_F(T)})} \right], \label{ODE} \end{split}$$ and $${\frac{{\partial {D_r}}}{{\partial T}}} =\Pi_2\frac{\partial }{\partial \tilde Z}\left(\frac{1}{L_p^2 }\frac{\partial P}{\partial \tilde Z} D_r^3 \right) -\frac{\partial }{\partial \tilde Z} \left( \frac{\partial {L_p}}{\partial T} \frac{D_r}{2L_p} \right), \label{Def_equation_nonlinear}$$ where, by using the obtained relations between the characteristic values ($f_e=p^* \pi r_p^2 $, $l_p^* =p^*2lr_p/Ew$ and $d_r^*=p^*r_i^2/Ew$) $\Pi_1$ and $\Pi_2$ can be presented in terms of $f_e^*$, $t^*$ and known geometric and physical parameters of the problem, $$\Pi_1=\frac{8l^2 \mu f_e^*}{\pi r_p h_0 (Ew)^2 t^* },\quad \Pi_2=\frac{t^* f_e^*r_i^4}{\pi 48\mu l^2r_p^4}.$$ \[GOV\_NON\_INSERT\] For a prescribed $f_e^*$, two different time-scales are evident in (\[GOV\_NON\_INSERT\]). The first is $${t_{\textrm{shear}}^ *} =\frac{ 8l^2 \mu f_e^*}{\pi r_p h_0 (Ew)^2 },$$ which is the time-scale in which shear is a leading-order term in (\[ODE\]), (shear is described by the second RHS term). For $t^*\ll {t_{\textrm{shear}}^ *}$, pressure effects can be neglected in (\[ODE\]) and the dominant balance would be between the external force and viscous shear. The second time-scale is $${t_{\textrm{peeling}}^ *} = \frac{\pi 48\ \mu l^2 r_p^4}{f_e^* r_i^4 }, \label{t_peeling}$$ and represents the time-scale in which pressure-driven flow is a dominant term in region II (see first RHS term in (\[Def\_equation\_nonlinear\])). Matched asymptotics ------------------- The value of $\Pi_1$ decreases with $t^*$ while $\Pi_2$ increases with $t^*$, thus indicating that early time-dynamics are governed by shear, while late-time dynamics are governed by pressurization of the liquid within the tube. Hereafter, in order to proceed, we limit our focus to configurations with the small ratios of $$\label{ass2} \varepsilon_t=\frac{t_{\textrm{shear}}^ *}{{t_{\textrm{peeling}}^ *}}\ll1,\quad \varepsilon_l=\frac{l_p^*}{l}\ll1.$$ Additionally, this section examines the response of the configuration to a specific case of suddenly applied external force of the form $$F_e=H(T),$$ where $H(T)$ is the heaviside function. ### Outer region For time-scales of $t^*=t^*_{\textrm{peeling}}$, we obtain $\Pi_2=1$ and $\Pi_1=\varepsilon_t$ and thus the leading order (\[GOV\_NON\_INSERT\]) is simplified to $$\begin{split} {F_e} = \left[ {\frac{{{L_p}\left( {1 - {\varepsilon_l}\int_0^{{{\tilde Z}_F(T)}} {{D_{r}}d\tilde Z} } \right) - {L_p(0)}}}{{1 - {\varepsilon_l}{L_p}}}} \right] + O(\varepsilon_t), \label{ODE_OUTER} \end{split}$$ and $${\frac{{\partial {D_r}}}{{\partial T}}} =\frac{\partial }{\partial \tilde Z}\left(\frac{1}{L_p^2 }\frac{\partial P}{\partial \tilde Z} D_r^3 \right) -\frac{\partial }{\partial \tilde Z} \left( \frac{\partial {L_p}}{\partial T} \frac{D_r}{2L_p} \right). \label{Def_equation_nonlinear_OUTER}$$ \[GOV\_NON\_INSERT\_OUTER\] Within the outer-region, we apply regular asymptotics with regard to the second small parameter $\varepsilon_l$ (which is $\gg \varepsilon_t$). The asymptotic expansions for both $L_p$ and $P=D_r$ are defined by $$L_p=L_{p,0}+\varepsilon_l L_{p,1},\quad P=D_r=D_{r,0}+\varepsilon_l D_{r,1}.$$ Applying standard asymptotic procedure on (\[ODE\_OUTER\]) yields $$L_p=F_e+L_p(0)+\varepsilon_l (F_e+L_p(0))\left( \int_0^{\tilde Z_{F(T)}}{D_{r,0}d\tilde Z} -F_e\right) \label{asym_LP}$$ which depends only on the leading-order deformation $D_{r,0}$. Substituting $L_p$ from (\[asym\_LP\]) into (\[Def\_equation\_nonlinear\_OUTER\]), yields the PDE governing $D_{r,0}$, $$\frac{\partial D_{r,0}}{\partial T} =\frac{1}{4(F_e+L_p(0))^2 }\frac{\partial ^2}{\partial \tilde Z^2}\left( D_{r,0}^4 \right), \label{Def_equation_nonlinear_OUTER_PME}$$which is a Porous-Medium-Equation of order 4, supplemented by the initial condition of $D_{r,0}(\tilde Z,0)=0$ and boundary condition of $P_0=D_{r,0}(0,T)=1$ (we set $F_e=H(T>0)=1$). Self-similar treatment of the above equation (\[Def\_equation\_nonlinear\_OUTER\_PME\]) is possible, following the approach presented by [@zel1950towards] and [@barenblatt1952some]. Defining the self-similar variable $$\eta=2(F_e+L_p(0))\frac{Z}{T^{1/2}} \label{eta_def}$$ yields an ODE for $f(\eta)=D_{r,0}$ $$(f^4(\eta))''+\frac{1}{2}\eta f'(\eta)=0,$$ along with initial and boundary conditions $$\quad f(0)=1,\quad f(\eta_f)=0.$$ \[selfsim\] Equation (\[selfsim\]) can be solved numerically [see @vazquez2007porous; @elbaz2016axial], yielding $$\eta _f=1.704,\quad \int_0^{\eta_f} {f(\eta)d\eta}=1.305. \label{numvals}$$ Substituting (\[eta\_def\]) into (\[numvals\]), we obtain the front location $$Z_f=0.852\frac{T^{1/2}}{F_e+L_p(0)} \label{front}$$ as well as the mass-flux entering region II $$\int_0^{\tilde Z_f} {D_{r,0}d\tilde Z}=\frac{\partial \tilde Z}{\partial \eta} \int_0^{\eta_f} {f(\eta)d\eta} =0.652\frac{T^{1/2}}{F_e+L_p(0)}. \label{mass_ZT}$$ Substituting (\[mass\_ZT\]) into (\[asym\_LP\]), yields the solution for $L_p$ for the outer-region $$L_p=F_e+L_p(0)+\varepsilon_l \left[ 0.652(T^{1/2}) -F_e(F_e+L_p(0))\right].$$ The initial condition $L_p(0)$ of the outer region needs to be related, via matching, to the inner-region solution derived in the next subsection. ### Inner region We introduce the rescaled inner region coordinate $$\bar T=\frac{T}{\varepsilon_t}.$$ For inner region time-scale $t^*\sim \varepsilon_t t^*_{\textrm{peeling}}$, we can estimate the location of the peeling front is $\tilde Z_F(\varepsilon_t \bar T)\sim O(\varepsilon_t)$. Thus, the leading-order $O(1)$ balance in (\[ODE\]) yields the inner-region solution $\bar L_p$ $$\bar L_p\frac{\partial {\bar L_p}}{\partial \bar T} = - \bar L_p +\bar L_p(0)+F_e ,$$ which is the Abel equation of the second kind [@zaitsev2002handbook], for which a closed-form solution for Heaviside-function force, ${F_e}(T) =H(T)$ is available. The inner-region dynamics are thus given by $${L_p}(\bar T) = (L_p(0)+F_e)\left( {1 + {\rm{W}}\left[ -{ \frac{1}{L_p(0)+F_e}{\exp{\left(-1-\frac{\bar T+C}{L_p(0)+F_e} \right)}}} \right]} \right),$$ where $W(T)$ is the Lambert-$W$ function [@weisstein2002lambert]. The initial condition ${L_p}\left( {T = 0} \right) = {L_p(0)}$ gives the constant $C$ $$C = -(F_e+L_p(0))\rm{Ln}(F_e)-L_p(0).$$ ### Uniform solution Matching between the inner and outer regions is required in order to obtain a uniform asymptotic solution. This yields the requirement $$\lim_{\bar T\rightarrow\infty} \bar L_p = \lim_{ T\rightarrow0} L_p$$ and thus $\bar L_p(0)+F_e=L_p(0)+F_e$, and $\bar L_p(0)=L_p(0)$. The composite expansion is therefore given by $\bar L_p (\bar T)+L_p (T)-(L_p(0)+F_e)$, yielding $$\begin{gathered} L_{p,\rm{uniform}}=(L_p(0)+F_e)\Bigg\{ 1+\varepsilon_l \left[ \frac{0.652(T^{1/2})}{F_e+L_p(0)} -F_e\right]\\ + {\rm{W}}\left[ -{ \frac{1}{L_p(0)+F_e}{\exp{\left(-1-\frac{ T/\varepsilon_t+C}{L_p(0)+F_e} \right)}}} \right] \Bigg\}, \label{unifromLP}\end{gathered}$$ which is the uniform solution representing the response of the configuration to a sudden external load. The above solution incorporates the effect of a propagating front. However, the location of the front will reach the end of the tube for $T>[(F_e+L_p(0))/0.852]^2$ (see equation (\[front\])). Before proceeding to discuss and present the uniform solution (\[unifromLP\]), we will approximate the dynamics of the post-peeling regime and connected it to the uniform solution presented above. ### Post-peeling dynamics After the peeling front reached the outlet, the entire tube is peeled and the viscous resistance can be approximated by the quasi-steady deformation solution of the Reynolds’ equation (\[Def\_equation\_nonlinear\_OUTER\_PME\]) $$D_r\approx(1-\tilde Z)^{1/4} P(0) \label{postP}$$ for $t^*\sim t^*_{\rm{peeling}}$, integral force balance yields $P(0)=F_e$ and integral conservation of mass yields $$\frac{\partial L_p}{\partial T}=\varepsilon_l\frac{F_e^4}{L_p(t) } \label{postLP}$$ which is obtained by substituting (\[postP\]) into the second RHS term of (\[Mass\_eq\]) and deriving with regards to time. Integrating (\[postLP\]), we can obtain the post-peeling solution $$L_p=\sqrt{L_p^2(T=T_B)+2\varepsilon_lT F_e^4} \label{postpeeling}$$ valid for $T>T_B=[(F_e+L_p(0))/0.852]^2$. Summary of nonlinear insertion dynamics --------------------------------------- We illustrate these results in Figure \[UNI\], which presents the location of the cylinder within the tube ($L_p$) vs. time, for the normalized parameters $F_e=H(T)$, $\varepsilon_l=0.1$, $\varepsilon_t=0.05$, and initial condition $L_p(0)=0.1$. The figure presents the uniform solution (\[unifromLP\]) for $0<T<T_b$ and the post-peeling solution (\[postpeeling\]) for $T>T_B$, where $T_b=[(F_e+L_p(0))/0.852]^2$. The response of the examined configuration to a suddenly applied insertion force, at the large deformation limit ($\lambda_h=d_r^*/h_0\gg1$), is shown to involve three distinct regimes. In the first regime (inner-regime, §4.2.2) the motion of the cylinder is governed by balance between shear stress, fluidic pressure and the external force, while the effect of viscous peeling is negligible. As $t$ increases, the effect of shear decreases while the fluidic pressure increases. In the second regime (outer-regime, §4.2.1) the cylinder decelerates and the external force is balanced by fluidic pressure ahead of the cylinder, while effects of shear are negligible. The motion of the cylinder in this regime is governed by balance between the fluidic volume displaced by insertion of the cylinder and the fluidic volume required to peel the inner cylinder from the external tube. Finally, as the peeling front reaches the inlet of the tube, a third post-peeling regime is obtained (§4.2.4). In this regime the cylinder motion is governed by balance between balance between the volume displaced by the cylinder and the fluidic flow outside of the configuration. The next section we proceed to examine the opposite case of extraction of the cylinder from the tube. ![**The location of the cylinder within the tube $L_p$, and liquid pressure ahead of the cylinder $P(Z=0)$, vs. time $T$.** The presented results are for configuration defined by $L_p(T=0)=0.1$, $\varepsilon_l=0.1$, $\varepsilon_t=0.05$ and $F_e=H(T)$. The white, light grey and dark grey parts denote the inner, outer and post-peeling regions, respectively.[]{data-label="UNI"}](Figure_4.eps){width="100.00000%"} Non-linear extraction dynamics, $f_e\lessapprox - E w h_0$ ========================================================== This section will examine the forced extraction of an inner cylinder from a liquid-filled tube. In this case a negative external force ${f_e}(t)<0$ creates a negative gauge pressure within the tube, and thus negative deformations. We will focus on the non-linear limit involving negative deformations of the order of $h_0$ or greater (where $h_0$ is the initial gap between the cylinder and the tube). The dynamics during extraction can be described by the previously derived results for insertion for the small deformation limit ($d^*_r / h_0 \ll 1$, see §3), as well as for the initial shear-dominated inner-region large deformation limit ($d^*_r / h_0 \gg 1$, $t^*\sim 8l^2\mu f_e^*/\pi r_p h_0 E^2 w^2$ see §4.2.2). However, outer-region solutions at the large deformation limit ($d^*_r/h_0\gg1$, $t^*\gg 8l^2\mu f_e^*/\pi r_p h_0 E^2 w^2$) exhibit essentially different dynamics. The negative gauge pressure created during extraction closes the gap between the cylinder and the tube, and may create contact between the two solids. This section will examine the case of nearly contacting cylinder and tube (see Fig \[PullPD\]a, §5.1) and the case of contact (see Fig \[PullPD\]b, §5.2). ![**Illustration of the configuration at extraction for (a) near contact and (b) contact cases.** Panel (a) presents the near contact case, where the gap at $h_0+d_r(\tilde z=0)\rightarrow 0^+$, thus nearly separating the liquid into two regions. The slope at $\tilde z=0$ is given by $\alpha$. Panel (b) presents contact, where the gap $h_0+d_r(\tilde z=0)=0$, and additional friction forces $(f_r,f_z)$ act on the elastic tube. The forces ($q_t,q_0$) are projections of the friction into direction normal and tangent to the surface of the elastic tube. \[PullPD\]](Figure_5.eps){width="70.00000%"} Near contact, $f_e\rightarrow (-2\pi Ew h_0r_p^2/r_i^2)^+$ ---------------------------------------------------------- For the limit of near contact between the tube and the cylinder (see Fig. \[PullPD\]a) we define the normalized gap at $X=\varepsilon_l L_p$ by $$D_r|_{X=\varepsilon L_p}+1=\varepsilon_{NC}H_{NC},\quad \varepsilon_{NC}=\ll1$$ where $H_{NC}$ is the scaled minimal gap and the small parameter $\varepsilon_{NC}$ will be later related to the extraction force (alternatively, $H_{NC}=(d_r|_{x=l_p} +h_0)/(\varepsilon_{NC}h_0)\sim O(1)$). This limit can be leverage to the simplification of the governing equations, allowing analytical treatment. While the governing integral conservation equations of mass and momentum are largely unchanged, the equations governing the flow field and deformation are significantly modified. ### Governing equations The rapid change in fluid pressure near $X=\varepsilon_l L_p$ for $\varepsilon_{NC}\rightarrow 0$ requires to include elastic bending effects in the $X$ direction. Following [@timoshenko1959theory], the deformation of a circular cylindrical tube loaded axisymmetrically is described by $$K\frac{{{\partial ^4}{D_r}\left( {X,T} \right)}}{{\partial {X^4}}} + {D_r}\left( {X,T} \right)\sim P\left( {X,T} \right), \label{Nor_tim}$$ where $$K \equiv \frac{{w^2 r_i^2}}{{12l^4\left( {1 - {\nu ^2}} \right)}},$$ the deformation $D$ is scaled by $d_r^*=h_0$ and the pressure $P$ is scaled by $p^*=Ewh_0/r_i^2$. The external force $F_e$ is thus scaled by $f^*=\pi r_p^2 p^*$. Equation (\[Nor\_tim\]) replaces the previous relation (\[pdr\]) between fluid pressure and elastic deformation. The limit of near contact is singular with regard to fluid resistance, and thus viscous resistance is defined by the small finite parameter $\varepsilon_{NC}$. Applying a Taylor series around $X=\varepsilon_lL_p$ allows to simplify the pressure-flux relation to $$\begin{gathered} \dot Q= -\frac{\pi r_i p^* h_0^3}{3\mu\dot q^* l}\frac{\partial P}{\partial X} \left(D_r+1 \right)^3 \\ \approx -\frac{\pi r_i p^* h_0^3}{3\mu\dot q^* l}\frac{\partial P}{\partial X} \left(D_r|_{X=\varepsilon_l L_p}+1 +\frac{\partial D_r}{\partial X}\Big|_{X=\varepsilon L_p}(X-\varepsilon_l L_p)\right)^3 \end{gathered}$$ where $\dot Q$ is scaled by $\dot q^*=\pi r_i p^* h_0^3/3\mu\dot q^* l$. Integrating from $X=\varepsilon L_p$ to $X$, and extracting the pressure difference in term of flux $\dot Q$, yields $$\begin{gathered} P|_{X=\varepsilon_l L_p}-P|_{X}= \left(D_r|_{X=\varepsilon_l L_p}+1\right)^{-2}\left(-2\frac{\partial D_r}{\partial X}\right)^{-1} \dot Q -\\ \left(D_r|_{X=\varepsilon_l L_p}+1 +\frac{\partial D_r}{\partial X}\Big|_{X=\varepsilon_l L_p}(X-\varepsilon_l L_p)\right)^{-2}\left(-2\frac{\partial D_r}{\partial X}\right)^{-1} \dot Q .\end{gathered}$$ For $\varepsilon_{NC}\rightarrow0$ the pressure difference $P|_{X=\varepsilon_l L_p}-P|_{X}$ asymptotes to a constant, defined only by the conditions near $X=\varepsilon_l L_p$. Applying $P(X=0)=0$, and $D_r|_{X=\varepsilon_l L_p}+1=H_{NC}\varepsilon_{NC}\ll1$, the following relation is obtained $$\dot Q \approx-2\varepsilon_{NC}^2\left(\frac{\partial D_r}{\partial X} H_{NC}P\right)\Bigg|_{X=\varepsilon_l L_p} ,$$ representing the flux only by the conditions at $X=\varepsilon_l L_p$ (where $P(X=0)=0$). Thus, the governing integral mass and momentum conservation equations for extraction are $${L_p} - {L_p(0)} =\frac{2 r_i h_0 l}{r_p^2 l_p^*}\int_0^1D_rdX-\frac{h_0^4 E w t^*\varepsilon_{NC}^2}{3\mu l_p^* r_p^2 l r_i} \int\limits_0^T \left(\frac{\partial D_r}{\partial X} H_{NC}P\right)\Bigg|_{X=\varepsilon_l L_p} {d\tilde T },\label{ext_vol2}$$ and $${{F_e} - {P|_{X=\varepsilon_l L_p}} - \frac{2\mu (l_p^*)^2 r_i^2}{r_p Ewh_0^2t^*}\left[ {{L_p}\frac{{\partial {L_p}}}{{\partial T}}\int_0^1\frac{d\tilde Z}{1+D_r}} \right] = 0}. \label{F_e_gov_22}$$ These equations are similar to §4.1, with the exception of the last RHS in (\[ext\_vol2\]), representing mass flux $\dot Q$. Scaling of (\[ext\_vol2\]) and (\[F\_e\_gov\_22\]) yields $$l_p^*=\frac{2 r_i h_0 l}{r_p^2 },$$ and two dimensionless ratios $$\tilde \Pi_1=\frac{h_0^3 E w \varepsilon_{NC}^2 t^*}{6\mu l^2 r_i^2},\quad\tilde \Pi_2=\frac{8\mu l^2 r_i^4}{r_p^5 Ewt^*}.$$ Similarly to the case of nonlinear insertion dynamics, the ratio $\tilde \Pi_1$ decreases with $t^*$ while $\tilde \Pi_1$ increases with $t^*$, suggesting different early-time and late-time dynamics. Thus two corresponding time-scales are evident. The first is the time-scale of the initial shear-dominated regime $$t^*_{\textrm{shear}}=\frac{8\mu l^2 r_i^4}{r_p^5 Ew}.$$ The second time-scale is of the late-time motion due to flow through near-contact gap between the cylinder and the tube, $$t^*_{\textrm{NC}}=\frac{6\mu l^2 r_i^2 }{Ew h_0^3 \varepsilon_{NC}^2}.\label{tshear}$$ For the examined configuration, the ratio between the two time-scales is a geometrically small parameter, given by $$\varepsilon_t=\frac{t^*_{\textrm{shear}}}{t^*_{\textrm{NC}}}=\frac{8 h_0^3 r_i^2 \varepsilon_{NC}^2}{6r_p^5}\ll1$$ and so $\varepsilon_t \ll \varepsilon_{NC}^2$ . Substituting $t^*=t^*_{NC}$, yields the governing equations $$L_p-L_p(0) =(1-\varepsilon_l L_p) P|_{X=\varepsilon_l L_p} - \int\limits_0^T \left(\frac{\partial D_r}{\partial X} H_{NC}P\right)\Bigg|_{X=\varepsilon_l L_p} {d\tilde T } ,\label{ext_vol}$$ and $$F_e - P|_{X=\varepsilon_l L_p} -\varepsilon_t\left( L_p\frac{\partial L_p}{\partial T}\int_0^1\frac{d\tilde Z}{1+D_r}\right)= 0. \label{F_e_gov_2}$$ We proceed by asymptotic expansions, based on perturbations from the exact contact state $\varepsilon_{NC}=0$. We thus define the external extraction force by $${F_e} =F_{e,0}(1- {\varepsilon _{NC}}),$$ where $F_{e,0}=F_{e,contact}$ is the external extraction force creating exact contact $D_r|_{X=\varepsilon_l L_p}=-1$, and thus zero flux $\dot Q=0$. In addition, we define the expansions for the fluidic pressure $${P} = {P_{0}} + {\varepsilon _{NC}}{P_{1}}.$$ \[LLL\] We apply a matched asymptotic scheme, and begin by solving for the outer-region $t\sim t^*_{NC}$ . ### Outer-region The leading-order solution is the exact contact case, where $F_e=F_{e,contact}$ , and after the initial transition regime (§4.2.2), the contact point between the solids separate the fluid into two domains, and thus $\dot Q=0$. The outer-region force balance (\[Force\_eq\]) is simplified to $$P_0|_{X=\varepsilon_l L_p}=F_{e,0},\quad P_1|_{X=\varepsilon_l L_p}=-F_{e,0} \label{F.B.0}$$ and the pressure-field is $$P_0(X,T)+\varepsilon_{NC} P_1(X,T)= H(X-\varepsilon L_p) \left( F_{e,0}+\varepsilon_{NC} F_{e,1}\right), \label{HeaviP0}$$ where $H$ is the Heaviside function. The solid deformation is governed by equation (\[Nor\_tim\]), along with the boundary conditions, ${{{\partial ^2}{D_r}}}{{/\partial {X^2}}} | _{X=0} = {{{\partial ^3}{D_r}}}{{/\partial {X^3}}}| _{X=0} = {D_r}| _{X=1} = {{\partial {D_r}}}{{/\partial X}}| _{X=1} = 0.$ Deformation patterns may be obtained numerically and analytically. However, since the singularity for $\varepsilon_{NC}\rightarrow 0$ dictates that flux is governed by the conditions near $X=\varepsilon _l L_p$, only expressions for $D_{r}|_{X=\varepsilon_l L_p}$ and $\partial D_{r}/\partial X|_{X=\varepsilon_l L_p}$ are required. Due to linearity and symmetry considerations, the gap slope at $X=\varepsilon_l L_p$ can be simplified to $$D_{r}|_{X=\varepsilon_l L_p}=\frac{1}{2}\left(P|_{X<\varepsilon_l L_p}+P|_{X>\varepsilon_l L_p})\right),\label{gap}$$ and $$\frac{\partial D_r}{\partial X}\Bigg|_{X=\varepsilon_l L_p}=\frac{\left(P|_{X>\varepsilon_l L_p}-P|_{X<\varepsilon_l L_p})\right)}{(64K)^{1/4}}.\label{drdx}$$ Thus, the requirement of $D_{r,0}|_{X=\varepsilon_l L_p}=-1$ yields $$F_{e,0}=P_{0}(\tilde{Z}=0)=-2.\label{pres}$$ and $H_{NC}$ is $$H_{NC}=1$$ and $$\frac{\partial D_{r,0}}{\partial X}\Bigg|_{X=\varepsilon_l L_p}+\varepsilon_{NC} \frac{\partial D_{r,1}}{\partial X}\Bigg|_{X=\varepsilon_l L_p}=\frac{ {F_{e,0}}(1-\varepsilon _{NC})}{64K^{1/4}}.$$ Substituting (\[gap\])-(\[pres\]) into (\[ext\_vol\]), $L_p$ to order $\varepsilon_{NC}^2$ is thus $$L_p =L_p(0)+ F_e -\frac{ F_e^2}{(64K)^{1/4}}T .\label{ext_vol}$$ ### Matched solution The inner solution is identical to the insertion case, and thus is given in §4.2.2. Matching the inner and outer solutions for near-contact extraction yields the requirement $$\lim_{\bar T\rightarrow\infty} \bar L_p = \lim_{ T\rightarrow0} L_p$$ and thus $\bar L_p(0)+F_e=L_p(0)+F_e$, and $\bar L_p(0)=L_p(0)$. The composite expansion is therefore given by $\bar L_p (\bar T)+L_p (T)-(L_p(0)+F_e)$, and the uniform solution is $$\begin{gathered} L_{p,\rm{uniform}}=(L_p(0)+F_e)\Bigg\{1- \frac{1}{L_p(0)+F_e}\frac{ F_e^2}{(64K)^{1/4}}T\\ + {\rm{W}}\left[ -{ \frac{1}{L_p(0)+F_e}{\exp{\left(-1-\frac{ T/\varepsilon_t-(F_e+L_p(0))\rm{Ln}(F_e)-L_p(0)}{L_p(0)+F_e} \right)}}} \right] \Bigg\}, \label{unifromLP}\end{gathered}$$ which is presents the response of the configuration to a sudden external load. The uniform solution is plotted in figure 6 for several configurations. In all cases, the parameters of $\varepsilon_t=0.1$, $K=0.3$ and $F_e=-1.9$ (corresponding to $\varepsilon_{NC}=0.05$) are used. Panel 6(a) presents the motion of the cylinder for the initial conditions $L_p(T=0)=5$ (smooth line), along with the inner solution (dashed line) and outer solution (dashed-dotted line). After the initial inner dynamics, the singular effect of the near-contact dominants the motion and creates a steady extraction speed independent of $L_p$. Panel (b) presents various initial values of $L_p(T=0)$, showing that the inner-regions changes with $L_p(T=0)$, but not the extraction speed at the outer region. For the case of $L_p(T=0)=2\approx -F_e$, we see that total extraction from the tube occurs before reaching the outer-region. Thus, the rapid extraction eliminates deceasing the pressure sufficiently to create near-contact dynamics. ![**The location of the cylinder within the tube $L_p$ vs. time $T$ for the near contact limit.** All configurations are defined by $\varepsilon_t=0.1$, $K=0.3$ and $F_e=-1.9$ (corresponding to $\varepsilon_{NC}=0.05$). In panel (a) $L_p(T=0)=5$ and the different elements of the composite expansion are presented. Panel (b) presents various initial values of $L_p(T=0)$.[]{data-label="UNI_PULL"}](Figure_6.eps){width="90.00000%"} Contact $f_e <- 2\pi Ew h_0r_p^2/r_i^2$ --------------------------------------- Equation (\[gap\]) indicates that for $F_e<-2$ (or in dimensionless form $f_e < -2\pi Ew h_0r_p^2/r_i^2$), there will be contact between the inner cylinder and the outer elastic tube. In such a case, additional forces will be applied at $\tilde Z=0$, as illustrated in Fig. \[PullPD\](b), which will modify the system dynamics. The contact prevents fluid from exiting the tube, and applies normal and tangential forces on the elastic tube. Thus, the outer-region governing equations are simply $$L_p-L_p (0)=\int_{1-\varepsilon_l L_p}^{1}{D_r dX}$$ and $$F_e-P|_{X=\varepsilon_l L_p}+F_z=0$$ where $F_z=f_z/f^*$ is the additional force due to drag in the $\tilde Z$ direction, and using the scaling used in §5.1 ($d_r^*=h_0$, $l_p^*={2 r_i h_0 l}/{r_p^2 }$, $p^*=Ewh_0/r_i^2$ and $f_e^*=\pi r_p^2 p^*$). In dimensional terms, the friction force $f_z=q_t \cos ⁡\alpha-q_0 \sin\alpha$, where $\tan\alpha=-(\partial d_r/\partial x)|_{x=l_p}$. Since $\alpha\ll1$, we can approximate $\alpha\approx -(\partial d_r/\partial x)|_{x=l_p}$ and obtain $$f_z=2\pi r_i q_0 \left(\mu_f+\frac{\partial d_r}{\partial x}\bigg|_{x=l_p} \right).\label{fz}$$ Calculation of the normal force acting on the elastic tube, $q_0$, is obtained from the requirement of $d_r|_{x=l_p}=-h_0$, where the deformation $d_r|_{x=l_p}$ is given by $$d_r|_{x=l_p}=-q_0\frac{ r_i^2}{Ew }\left(\frac{{3\left( {1 - {\nu ^2}} \right)}}{{16 w^2 r_i^2}}\right)^{1/4}+p|_{x=l_p} \frac{r_i^2}{2Ew}\label{q0}$$ representing the linear summation of the deformation due to a a localized normal force $q_0$ [first RHS term, @timoshenko1959theory] and the deformation due to liquid pressure (second RHS term, see equation (\[gap\])). We now calculate the normal force $q_0$ by requiring $d_r|_{x=l_p}=-h_0$ in equation (\[q0\]). The obtained $q_0$ is substituted into (\[fz\]), along with ${\partial d_r}{\partial x}|_{x=l_p}$ calculated from (\[drdx\]) (the localized force creates a symmetrical deformation, and does not affect the slope). This procedure yields the dimensional friction force $f_z$ $$f_z≈-\frac{2\pi Ew}{r_i}\left(h_0+p|_{x=l_p}\frac{r_i^2}{2Ew}\right)\left(\mu_f \left(\frac{{64w^2 r_i^2}}{{12\left( {1 - {\nu ^2}} \right)}}\right)^{1/4}+ p|_{x=l_p}\frac{r_i^2} { Ew} \right),$$ or in normalized form, $F_z$ $$F_z=-2\left( 1+\frac{P|_{X=\varepsilon_lL_p}}{2}\right)\left(\hat \Pi_1 +\hat\Pi_2P|_{X=\varepsilon_lL_p}\right)$$ where $$\hat \Pi_1=\frac{\mu_f }{r_p^2}\left(\frac{{16 w^2 r_i^6}}{{3\left( {1 - {\nu ^2}} \right)}}\right)^{1/4} ,\quad \hat \Pi_2=\frac{r_i h_0}{r_p^2}.$$ The obtained dimensionless ratios $\hat \Pi_1$ and $\hat \Pi_2$ are geometrically small parameters, and thus friction does not have a leading order effect on the system dynamics. For consistency, since terms of similar orders were previously neglected, $\hat \Pi_2$ is neglected hereafter. However, for some configurations $\hat \Pi_1\gg \hat \Pi_2$. Keeping $O(\hat\Pi_1)$ terms yields expressions for the fluid pressure ahead of the cylinder $$P|_{X>\varepsilon_l L_p }=\frac{F_e-2\hat \Pi_1}{1+\hat \Pi_1}$$ and the outer-region penetration length, $$L_p=L_p(0)+\frac{F_e-2\hat\Pi_1}{1+\hat\Pi_1}.$$ Thus, friction effects are of order of the small parameter $\hat\Pi_1$, and only slightly reduce the liquid pressure and the penetration length $L_p$. Following similar asymptotic matching procedure to that presented in §5.1, the uniform solution for contact is given by $$\begin{gathered} L_{p,\rm{uniform}}=\left(L_p(0)+\hat F_e\right)\Bigg\{1\\ + {\rm{W}}\left[ -{ \frac{1}{L_p(0)+\hat F_e}{\exp{\left(-1-\frac{ \hat T-(\hat F_e+L_p(0))\rm{Ln}\left(\hat F_e\right)-L_p(0)}{L_p(0)+\hat F_e} \right)}}} \right] \Bigg\}, \label{unifromLP}\end{gathered}$$ where $\hat F_e = ({F_e-2\hat\Pi_1})/({1+\hat\Pi_1})$ and $\hat T = t/t^*_{\textrm{shear}}$ (where $t^*_{\textrm{shear}}$ is defined in (\[tshear\])). The uniform solution (\[unifromLP\]) reaches a steady-state of $L_p\rightarrow L_p(0)+\hat F_e$. If $L_p(0)+\hat F_e<0$, the steady-state solution is not physical, indicating that the inner cylinder completely exits the tube. Thus, steady-state contact will lock the inner cylinder within the elastic tube only for the range of forces $$2\hat \Pi_1-L_p(0)(1+\hat \Pi_1)<F_e<-2$$ Outside of this range the inner cylinder will exit the tube either due to fluid entering the tube $F_e>-2$ or sliding while in contact $F_e\leq 2\hat\Pi_1-L_p(0)(1+\hat\Pi_1)$. (In dimensional force, and omitting the small $O(\hat \Pi_1)$ terms, this range is given by $-\pi r_p^4Ew l_p(0)/2r_i^2 l<f_e<-2\pi Ewh_0 r_p^2/r_i^2$.) The results presented in this subsection will be discussed further in the following section, summarizing the non-linear extraction section §5. Summary of extraction dynamics ------------------------------ Figure \[contact\_SUM\] summarizes the different dynamics obtained for the case of non-linear extraction. For near contact (smooth line, $F_e>-2$) the viscous resistance in the region near $\tilde Z=0$ is singular and determines the liquid mass flux, and thus the motion of the inner cylinder. After a shear-dominant early dynamics (which occurs for all cases), the cylinder moves in constant speed due to the steady conditions near $\tilde Z=0$. Increasing the extraction force to $2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)<F_e<-2$ creates steady-state contact (dashed line) in which the extraction force is balanced with the force due to liquid pressure at a constant penetration length $L+p=L_p(0)+({F_e-2\hat\Pi_1})/({1+\hat\Pi_1})$. Friction creates only weak $O(\hat \Pi_1)$ effects. Extracting the tube with a greater force $F_e\leq 2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)$ completely removes the inner cylinder from the tube before a steady-state can be reached (dashed-dotted line), and thus loackage of the inner cylinder is limited to a specific range of extraction forces. ![**Extraction dynamics for near-contact (smooth line, $F_e>-2$), locked contact (dashed line, $2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)<F_e<-2$) and extracted contact (dashed-dotted line, $F_e\leq 2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)$) configurations**. For $F_e<-2$, the negative gauge pressure reduces, but not eliminates, the gap between the cylinder and the tube. Pressure-driven flow through this gap allows the gradual extraction of the inner cylinder. For $2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)<F_e<-2$, after an initial transient motion, the cylinder is locked within the tube and negative gauge pressure creates contact between cylinder and the tube. Thus, there is no flow, and the inner cylinder is locked at a constant location in which there is a balance between the external force, friction, and the negative gauge pressure. Finally, for $F_e\leq 2\tilde\Pi_1-L_p(0)(1+\tilde\Pi_1)$, the inner cylinder is extracted in the initial transient motion. For all presented cases $K=0.3$, $L_p(0)=3$, $\varepsilon_t=0.01$ and $\tilde \Pi_1=0.1$.[]{data-label="contact_SUM"}](Figure_7.eps){width="90.00000%"} Concluding remarks ================== This work studied the low-Reynolds number fluid mechanics of a basic configuration, consisting of a slender cylinder inserted into a fluid-filled elastic tube, which is relevant to various minimally invasive medical procedures. Governing integro-differential equations were derived by applying the lubrication approximation and thin shell elastic model. Solutions for various limits were obtained by scaling analysis and regular and singular asymptotic schemes. Table 1 summarizes the different regions with regard to the value of $f_e$, the external force acting to extract or insert the cylinder to the tube. In addition, Table 1 presents comments and descriptions on the dominant mechanisms in regions with no approximate solutions. ---------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ $f_e\gg E h_0w$ Examined in §4. Nonlinear dynamics involving three distinct regimes (with additional simplifying assumptions detailed in (\[ass2\])). An early-time regime (§4.2.2) governed by balance between shear stress, fluidic pressure and the external force. Intermediate regime (§4.2.1) governed by the external force, fluid pressure and viscous-peeling of the inner cylinder from the external tube. Late-time regime (§4.2.4) in which pressure-driven viscous flow exiting the tube determines the motion of the inner cylinder. $f_e\sim E h_0w$ Not examined in this work. Nonlinear insertion dynamics create positive deformations which reduce viscous resistance, but do not involve a distinct propagation of a peeling front. $|f_e|\ll E h_0w$ Examined in §3. Linearized dynamics representing a rigid configuration in leading order (see (\[(A)\])), with small corrections due to elastic effects (see (\[UP300\])). The small deformations create elastic potential energy, which may lead to motion in a direction opposite to the external transient force. $ -\frac{2\pi Ew h_0r_p^2}{r_i^2}< f_e \sim -E h_0 w $ Not examined in this work. Nonlinear extraction dynamics create negative deformations, thus increasing the viscous resistance. The front of the cylinder does not have a singular dominant effect on viscous resistance. $f_e\rightarrow \left(-\frac{2\pi Ew h_0r_p^2}{r_i^2}\right)^+$ Examined in §5.1. Elastic deformation creates near contact between the tube and the elastic cylinder at $x=l_p$. The viscous resistance in the region near $x=l_p$ is singular and dominates the mass-flow outside of the cylinder. After an early time region similar to §4.2.2, the cylinder decelerates and exits the tube at a constant speed determined by the conditions near $x=l_p$ (see (\[unifromLP\])). In this case the dynamics of the configuration are highly sensitive to the geometry at the tip of the penetrating cylinder. $-\frac{\pi r_p^4Ew l_p(0)}{2r_i^2 l}<f_e<-\frac{2\pi Ewh_0 r_p^2}{r_i^2}$ Examined in §5.2. In this case the extraction, and negative deformation, create contact between the inner cylinder and the elastic tube. After an early time region similar to §4.2.2, the cylinder decelerates and reaches a steady-state of balance between the external force, the fluid pressure and friction. In this range of force the inner cylinder remains at a constant position within the tube. $f_e <-\frac{\pi r_p^4Ew l_p(0)}{2r_i^2 l}$ Examined in §5.2. Similar to the previous case, however, for this range of extracting forces the cylinder is completely extracted from the tube before a steady-state balance is reached. ---------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ : Summary of results for different values of $f_e$, the external insertion or extraction force.

ArXiv

--- abstract: 'We study various SDP formulations for [Vertex Cover]{} by adding different constraints to the standard formulation. We show that [Vertex Cover]{} cannot be approximated better than $2-o(1)$ even when we add the so called pentagonal inequality constraints to the standard SDP formulation, en route answering an open question of Karakostas [@Karakostas]. We further show the surprising fact that by strengthening the SDP with the (intractable) requirement that the metric interpretation of the solution is an $\ell_1$ metric, we get an exact relaxation (integrality gap is 1), and on the other hand if the solution is arbitrarily close to being $\ell_1$ embeddable, the integrality gap may be as big as $2-o(1)$. Finally, inspired by the above findings, we use ideas from the integrality gap construction of Charikar [@Char02] to provide a family of simple examples for negative type metrics that cannot be embedded into $\ell_1$ with distortion better than $8/7-\eps$. To this end we prove a new isoperimetric inequality for the hypercube.' author: - Hamed Hatami - Avner Magen - Evangelos Markakis - | \ \ bibliography: - 'tensor.bib' title: 'Integrality gaps of semidefinite programs for Vertex Cover and relations to $\ell_1$ embeddability of negative type metrics' --- Introduction ============ A [vertex cover]{} in a graph $G=(V,E)$ is a set $S \subseteq V$ such that every edge $e \in E$ intersects $S$ in at least one endpoint. Denote by $\vc(G)$ the size of the minimum vertex cover of $G$. It is well-known that the minimum vertex cover problem has a $2$-approximation algorithm, and it is widely believed that for every constant $\eps>0$, there is no ($2-\eps$)-approximation algorithm for this problem. Currently the best known hardness result for this problem shows that $1.36$-approximation is NP-hard [@DS02]. If we were to assume the Unique Games Conjecture [@Khot], the problem would be essentially settled as $2-\Omega(1)$ would then be NP-hard [@KhotRegev]. In a seminal paper, Goemans and Williamson [@GW95] introduced semidefinite programming as a tool for obtaining approximation algorithms. Since then semidefinite programming has been applied to various approximation problems and has become an important technique, and indeed the best known approximation algorithms for many problems are obtained by solving an SDP relaxation of them. The best known algorithms for [Vertex Cover]{} compete in “how big is the little o” in the $2-o(1)$ factor. The best two are in fact based on SDP relaxations: Halperin [@H02] gives a $(2-\log \log \Delta / \log \Delta)$-approximation where $\Delta$ is the maximal degree of the graph while Karakostas obtains a $(2-\Omega(1/\sqrt{\log n}))$-approximation [@Karakostas]. The standard way to formulate the [Vertex Cover]{} problem as a quadratic integer program is the following: $$\begin{array}{clcl} \Min & \sum_{i \in V} (1+x_0 x_i)/2 &\qquad &\\ {\rm s.t.} & (x_i-x_0)(x_j-x_0) = 0 & &\forall \ ij \in E \\ & x_i \in \{-1,1\} & & \forall \ i \in \{0\} \cup V, \end{array}$$ where the set of the vertices $i$ for which $x_i=x_0$ correspond to the vertex cover. By relaxing this integer program to a semidefinite program, the scalar variable $x_i$ now becomes a vector $\bv_i$ and we get: $$\label{SDP-VC1} \begin{array}{clcl} \Min & \sum_{i \in V} (1+\bv_0\bv_i)/2 &\qquad &\\ {\rm s.t.} & (\bv_i-\bv_0)\cdot(\bv_j-\bv_0) = 0 & &\forall \ ij \in E \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V. \end{array}$$ Kleinberg and Goemans [@KG95] proved that SDP (\[SDP-VC1\]) has integrality gap of $2-o(1)$. Specifically, given $\epsilon > 0$, they construct a graph $G_\eps$ for which $\vc(G_\eps)$ is at least $(2-\eps)$ times larger than the optimal solution to the SDP. They also suggested the following strengthening of SDP (\[SDP-VC1\]) and left its integrality gap as an open question: $$\label{SDP-VC2} \begin{array}{clcl} \Min & \sum_{i \in V} (1+\bv_0\bv_i)/2 &\qquad &\\ {\rm s.t.} & (\bv_i-\bv_0)\cdot (\bv_j-\bv_0) = 0 & &\forall \ ij \in E \\ & (\bv_i-\bv_k)\cdot (\bv_j-\bv_k) \ge 0 & &\forall \ i,j,k \in \{0\} \cup V \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V. \end{array}$$ Charikar [@Char02] answered this question by showing that the same graph $G_\eps$ but a different vector solution satisfies SDP (\[SDP-VC2\])[^1] and gives rise to an integrality gap of $2-o(1)$ as before. The following is an equivalent formulation to SDP (\[SDP-VC2\]): $$\label{SDP-VC3} \begin{array}{clcl} \Min & \sum_{i \in V} 1-\|\bv_0-\bv_i\|^2/4 &\qquad &\\ {\rm s.t.} & \|\bv_i-\bv_0\|^2+\|\bv_j-\bv_0\|^2 = \|\bv_i-\bv_j\|^2 & &\forall \ ij \in E \\ & \|\bv_i-\bv_k\|^2+\|\bv_j-\bv_k\|^2 \ge \|\bv_i-\bv_j\|^2 & &\forall \ i,j,k \in \{0\} \cup V \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V \\ \end{array}$$ #### Viewing SDPs as relaxations over $\ell_1$ The above reformulation reveals a connection to metric spaces. The second constraint in SDP (\[SDP-VC3\]) says that $\|\cdot\|^2$ induces a metric on $\{\bv_i:i \in \{0\} \cup V\}$, while the first says that $\bv_0$ is on the shortest path between the images of every two neighbours. This suggests a more careful study of the problem from the metric viewpoint which is the purpose of this article. Such connections are also important in the context of the [Sparsest Cut]{} problem, where the natural SDP relaxation was analyzed in the breakthrough work of Arora, Rao and Vazirani [@ARV] and it was shown that its integrality gap is at most $O(\sqrt {\log n})$. This later gave rise to some significant progress in the theory of metric spaces [@CGR; @ALN]. For a metric space $(X,d)$, let $c_1(X,d)$ denote the minimum distortion required to embed $(X,d)$ into $\ell_1$ (see [@Mat02] for the related definitions). So $c_1(X,d)=1$ if and only if $(X,d)$ can be embedded isometrically into $\ell_1$. Consider a vertex cover $S$ and its corresponding solution to SDP (\[SDP-VC2\]), i.e., $\bv_i=1$ for every $i \in S \cup \{0\}$ and $\bv_i=-1$ for every $i \not\in S$. The metric defined by $\|\cdot\|^2$ on this solution (i.e., $d(i,j)=\|\bv_i-\bv_j\|^2$) is isometrically embeddable into $\ell_1$. Thus we can strengthen SDP (\[SDP-VC2\]) by allowing any arbitrary list of valid inequalities in $\ell_1$ to be added. For example the triangle inequality, is one type of such constraints. The next natural inequality of this sort is the [pentagonal inequality]{}: A metric space $(X,d)$ is said to satisfy the pentagonal inequality if for $S,T \subset X$ of sizes 2 and 3 respectively it holds that $\sum_{i\in S,j\in T}d(i,j) \ge \sum_{i,j\in S}d(i,j) + \sum_{i,j\in T}d(i,j)$. Note that this inequality does no longer apply to every metric, but it does to ones that are $\ell_1$ embeddable. This leads to the following natural strengthening of SDP (\[SDP-VC3\]): $$\label{SDP-VC4} \begin{array}{clcl} \Min & \sum_{i \in V} 1-\|\bv_0-\bv_i\|^2/4 &\qquad &\\ {\rm s.t.} & \|\bv_i-\bv_0\|^2+\|\bv_j-\bv_0\|^2 = \|\bv_i-\bv_j\|^2 & &\forall \ ij \in E \\ & \begin{array}{cl} \sum_{i \in S,j \in T} \|\bv_i-\bv_j\|^2 \ge & \sum_{i,j \in S}\|\bv_i-\bv_j\|^2+\\ &\sum_{i,j \in T}\|\bv_i-\bv_j\|^2 \end{array} & & \begin{array}{l} \forall \ S,T \subseteq \{0\} \cup V, \\ |S|=2, |T|=3 \end{array} \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V \\ \end{array}$$ In Theorem \[Thm:Pentagonal\], we prove that SDP (\[SDP-VC4\]) has an integrality gap of $2-\eps$, for every $\eps>0$. It is interesting to note that for the classical problem of [Sparsest Cut]{}, it is not known how to show a nonconstant integrality gap against pentagonal (or any other $k$-gonal) inequalities, although recently a nonconstant integrality gap was shown by Khot and Vishnoi and later by Devanur [*et al.*]{} [@KhotVishnoi; @DKSV06] in the presence of the triangle inequalities[^2]. One can actually impose any $\ell_1$-constraint not only for the metric defined by $\{\bv_i : i \in V \cup \{0\}\}$, but also for the one that comes from $\{\bv_i : i \in V \cup \{0\}\} \cup \{-\bv_i : i \in V\cup \{0\}\} $. This fact is used in [@Karakostas] where the triangle inequality constraints on this extended set are added, achieving an integraility gap of at most $2 -\Omega(\frac{1}{\sqrt{\log n}})$. It is also asked whether the integrality gap of this strengthening breaks the “$2-o(1)$ barrier”. In Section \[Sec:strongerSDP\] we answer this question in the negative. #### Integrality gap with respect to $\ell_1$ embeddability At the extreme, strengthening the SDP with $\ell_1$-valid constraints, would imply the condition that the metric defined by $\|\cdot\|$ on $\{\bv_i:i\in \{0\} \cup V\}$, namely $d(i,j)=\|\bv_i-\bv_j\|^2$ is $\ell_1$ embeddable. Doing so leads to the following intractable program (which we refer to as SDP for convenience): $$\label{SDP-l1} \begin{array}{clcl} \Min & \sum_{i \in V} 1-\|\bv_0-\bv_i\|^2/4 &\qquad &\\ {\rm s.t.} & \|\bv_i-\bv_0\|^2+\|\bv_j-\bv_0\|^2 = \|\bv_i-\bv_j\|^2 & &\forall \ ij \in E \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V \\ &c_1(\{\bv_i:i \in \{0\} \cup V\},\|\cdot\|^2) = 1 && \end{array}$$ In [@ACMM05], it is shown that an SDP formulation of [Minimum Multicut]{}, even with the constraint that the $\|\cdot\|^2$ distance over the variables is isometrically embeddable into $\ell_1$, still has a large integrality gap. For the [Max Cut]{} problem, which is more intimately related to our problem, it is easy to see that the $\ell_1$ embeddability condition does not prevent the integrality gap of $8/9$; it is therefore tempting to believe that there is a large integrality gap for SDP (\[SDP-l1\]) as well. Surprisingly, SDP (\[SDP-l1\]) has no gap at all; in other words, as we show in Theorem \[Thm:SDP-l1\], the answer to SDP (\[SDP-l1\]) is exactly the size of the minimum vertex cover. A consequence of this fact is that any feasible solution to SDP (\[SDP-VC2\]) that surpasses the minimum vertex cover induces an $\ell_2^2$ distance which is not isometrically embeddable into $\ell_1$. This includes the integrality gap constructions of Kleinberg and Goemans’, and that of Charikar’s for SDPs (\[SDP-VC2\]) and (\[SDP-VC3\]) respectively. The construction of Charikar is more interesting in the sense that the obtained $\ell_2^2$ distance is a metric (from now on we refer to it as a [negative type metric]{}; see [@dela97] for background and nomenclature). In contrast to Theorem \[Thm:SDP-l1\], we show in Theorem \[Thm:SDP-weakl1\] that if we relax the last constraint in SDP (\[SDP-l1\]) to $c_1(\{\bv_i:i \in \{0\} \cup V\},\|\cdot\|^2) \le 1+\delta$ for any constant $\delta > 0$, then the integrality gap may “jump” to $2-o(1)$. Compare this with a problem such as [Sparsest Cut]{} in which an addition of such a constraint immediately implies integrality gap at most $1+\delta$. #### Negative type metrics that are not $\ell_1$ embeddable Inspired by the above results, we construct in Theorem \[Thm:distortion\] a simple negative type metric space $(X,\|\cdot\|^2)$ that does not embed well into $\ell_1$. Specifically, we get $c_1(X) \ge \frac{8}{7}-\eps$ for every $\eps>0$. In order to show this we prove a new isoperimetric inequality for the hypercube $Q_n = \{-1,1\}^n$, which we believe is of independent interest. This theorem generalizes the standard one, and under certain conditions provides better guarantee for edge expansion: \[Thm:isoper\](Generalized Isoperimetric inequality) For every set $S \subseteq Q_n$, $$|E(S,S^c)| \ge |S|(n-\log_2 |S|)+p(S).$$ where $p(S)$ denotes the number of vertices $\bu \in S$ such that $-\bu \in S$. Khot and Vishnoi [@KhotVishnoi] constructed an example of an $n$-point negative type metric that for every $\delta>0$ requires distortion at least $(\log \log n)^{1/6-\delta}$ to embed into $\ell_1$. Krauthgamer and Rabani [@KrauthgamerRabani] showed that in fact Khot and Vishnoi’s example requires a distortion of at least $\Omega(\log \log n)$. Later Devanur [*et al.*]{} [@DKSV06] showed an example which suffers an $\Omega (\log \log n)$ distortion even on average when embedded into $\ell_1$ (we note that our example is also “bad” on average). Although the above examples require nonconstant distortion to embed into $\ell_1$, we believe that Theorem \[Thm:distortion\] is interesting for the following reasons: (i) Khot and Vishnoi’s example is quite complicated, and there is no good explanation to the fact that triangle inequality holds (citing the authors “this is where the magic happens”). Simple constructions such as the one we obtain may give a better understanding of the problem and lead to simpler constructions of negative type metrics that behave poorly in the above sense (ii) there are not many known examples of negative type metrics that require a constant $c>1$ distortion to embed into $\ell_1$, and finding such examples is challenging and desirable. In fact before Khot and Vishnoi’s result, the best known lower bounds (see [@KhotVishnoi]) were due to Vempala, $10/9$ for a metric obtained by a computer search, and Goemans, $1.024$ for a metric based on the Leech Lattice (compare these to the $8/7-\epsilon$ bound of Theorem \[Thm:distortion\]). We mention that by [@ALN] every negative type metric embeds into $\ell_1$ with distortion $O(\sqrt{\log{n}}\log{\log{n}})$. Preliminaries and notation \[sec:notation\] =========================================== A vertex cover of a graph $G$ is a set of vertices that touch all edges. An independent set in $G$ is a set $I \subseteq V$ such that no edge $e\in E$ joins two vertices in $I$. We denote by $\alpha(G)$ the size of the maximum independent set of $G$. Vectors are always denoted in bold font (such as $\bv$, $\bw$, etc.); $\|\bv\|$ stands for the Euclidean norm of $\bv$, $\bu\cdot \bv$ for the inner product of $\bu$ and $\bv$, and $\bu \otimes \bv$ for their tensor product. Specifically, if $\bv, \bu \in \R^n$, $\bu \otimes \bv$ is the vector with coordinates indexed by ordered pairs $(i,j)\in [n]^2$ that assumes value $\bu_i \bv_j$ on coordinate $(i,j)$. Similarly, the tensor product of more than two vectors is defined. It is easy to see that $(\bu \otimes \bv) . (\bu' \otimes \bv') = (\bu\cdot \bu') (\bv\cdot \bv')$. For two vectors $\bu \in \R^n$ and $\bv \in \R^m$, denote by $(\bu,\bv) \in \R^{n+m}$ the vector whose projection to the first $n$ coordinates is $\bu$ and to the last $m$ coordinates is $\bv$. Next, we give a few basic definitions and facts about finite metric spaces. A metric space $(X,d_X)$ embeds with distortion at most $D$ into $(Y,d_Y)$ if there exists a mapping $\phi : X \mapsto Y$ so that for all $a,b \in X$ $\gamma\cdot d_X(a,b) \le d_Y(\phi(a),\phi(b)) \le \gamma D \cdot d_X(a,b)$, for some $\gamma > 0$. We say that $(X,d)$ is $\ell_1$ embeddable if it can be embedded with distortion 1 into $\R^m$ equipped with the $\ell_1$ norm. An $\ell_2^2$ distance on $X$ is a distance function for which there there are vectors $\bv_x \in \R^m$ for every $x\in X$ so that $d(x,y) = \|\bv_x - \bv_y\|^2$. If, in addition, $d$ satisfies triangle inequality, we say that $d$ is an $\ell_2^2$ metric or [negative type metric]{}. It is well known [@dela97] that every $\ell_1$ embeddable metric is also a negative type metric. $\ell_1$ and Integrality Gap of SDPs for Vertex Cover – an “all or nothing” phenomenon ====================================================================================== It is well known that for the [Sparsest Cut]{} problem, there is a tight connection between $\ell_1$ embeddability and integrality gap. In fact the integrality gap is bounded above by the least $\ell_1$ distortion of the SDP solution. At the other extreme stand problems like [Max Cut]{} and [Multi Cut]{}, where $\ell_1$ embeddability does not provide any strong evidence for small integrality gap. In this section we show that [Vertex Cover]{} falls somewhere between these two classes of $\ell_1$-integrality gap relationship, and it witnesses a sharp transition in integrality gap in the following sense: while $\ell_1$ embeddability prevents [*any*]{} integrality gap, allowing a small distortion, say $1.001$ does not prevent integrality gap of $2-o(1)$! \[Thm:SDP-l1\] For a graph $G=(V,E)$, the answer to the SDP formulated in SDP (\[SDP-l1\]) is the size of the minimum vertex cover of $G$. Let $d$ be the metric solution of SDP (\[SDP-l1\]). We know that $d$ is the result of an $\ell_2^2$ unit representation (i.e., it comes from square norms between unit vectors), and furthermore it is $\ell_1$ embeddable. By a well known fact about $\ell_1$ embeddable metrics (see, eg, [@dela97]) we can assume that there exist $\lambda_t>0$ and $f_t:\{0\} \cup V \rightarrow \{-1,1\}$, $t=1,\ldots,m$, such that $$\label{distances} \|\bv_i-\bv_j\|^2=\sum_{t=1}^{m} \lambda_t |f_t(i)-f_t(j)|,$$ for every $i,j \in \{0\} \cup V$. Without loss of generality, we can assume that $f_t(0) = 1$ for every $t$. For convenience, we switch to talk about [Independent Set]{} and its relaxation, which is the same as SDP (\[SDP-l1\]) except for the objective function that becomes $\rm Max \sum_{i \in V} \|\bv_0-\bv_i\|^2/4$. Obviously, the theorem follows from showing that this is an exact relaxation. We argue that (i) $I_t=\{i \in V: f_t(i)=-1\}$ is a (nonempty) independent set for every $t$, and (ii) $\sum \lambda_t = 2$. Assuming these two statements we get $$\begin{aligned} \sum_{i \in V} \frac{\|\bv_i - \bv_0\|^2}{4} &=&\sum_{i \in V} \frac{\sum_{t=1}^{m} \lambda_t |1-f_t(i)|}{4}= \sum_{t=1}^m \frac{\lambda_t |I_t|}{2} \le \max_{t \in [m]} |I_t| \le \alpha(G),\end{aligned}$$ and so the relaxation is exact and we are done. We now prove the two statements. The first is rather straightforward: For $i,j \in I_t$, (\[distances\]) implies that $d(i,0)+d(0,j)>d(i,j)$. It follows that $ij$ cannot be an edge else it would violate the first condition of the SDP. (We may assume that $I_t$ is nonempty since otherwise the $f_t(\cdot)$ terms have no contribution in (\[distances\]).) The second statement is more surprising and uses the fact that the solution is optimal. The falsity of such a statement for the problem of [Max Cut]{} (say) explains the different behaviour of the latter problem with respect to integrality gaps of $\ell_1$ embeddable solutions. We now describe the proof. Let $\bv_i'=(\sqrt{\lambda_1/2}f_1(i),\ldots,\sqrt{\lambda_m/2}f_m(i),0)$. From (\[distances\]) we conclude that $\|\bv_i'-\bv_j'\|^2 = \|\bv_i-\bv_j\|^2$, hence there exists a vector $\bw = (w_1,w_2,...,w_{m+1}) \in \R^{m+1}$ and an orthogonal transformation $T$, such that $$\bv_i=T\left(\bv_i'+\bw\right).$$ Since the constraints and the objective function of the SDP are invariant under orthogonal transformations, without loss of generality we may assume that $$\bv_i=\bv_i'+\bw,$$ for $i \in V \cup \{0\}$. We know that $$\label{normv_i} 1=\|\bv_i\|^2=\|\bv_i'+\bw\|^2= w_{m+1}^2+\sum_{t=1}^m (\sqrt{\lambda_t/2}f_t(i)+w_t)^2.$$ Since $\|\bv'_i\|^2=\|\bv'_0\|^2=\sum_{t+1}^m \lambda_t/2$, for every $i \in V \cup \{0\}$, from (\[normv\_i\]) we get $\bv'_0 \cdot \bw=\bv'_i\cdot \bw$. Summing this over all $i \in V$, we have $$|V|(\bv'_0\cdot \bw)=\sum_{i\in V}\bv'_i\cdot \bw =\sum_{t=1}^m (|V|-2|I_t|)\sqrt{\lambda_t/2} w_t,$$ or $$\sum_{t=1}^m |V|\sqrt{\lambda_t/2}w_t=\sum_{t=1}^m (|V|-2|I_t|)\sqrt{\lambda_t/2} w_t ,$$ and therefore $$\label{sumzero} \sum_{t=1}^m |I_t|\sqrt{\lambda_t/2} w_t=0.$$ Now (\[normv\_i\]) and (\[sumzero\]) imply that $$\label{coreineq} \max_{t \in [m]} |I_t| \ge \sum_{t=1}^m (\sqrt{\lambda_t/2}f_t(0)+ w_t)^2 |I_t| = \sum_{t=1}^m \left(\frac{\lambda_t |I_t|}{2} + w_t^2|I_t|\right) \ge \sum_{t=1}^m \frac{\lambda_t |I_t|}{2}.$$ As we have observed before $$\sum_{t=1}^m\frac{\lambda_t |I_t|}{2} = \sum_{i \in V} \frac{\|\bv_i - \bv_0\|^2}{4}$$ which means (as clearly $\sum_{i \in V} \frac{\|\bv_i - \bv_0\|^2}{4} \ge \alpha(G)$) that the inequalities in (\[coreineq\]) must be tight. Now, since $|I_t|>0$ we get that $\bw = \bo$ and from (\[normv\_i\]) we get the second statement, i.e., $\sum \lambda_t = 2$. This concludes the proof. Now let us replace the last constraint in SDP (\[SDP-l1\]), $c_1(\{\bv_i:i \in \{0\} \cup V\},\|\cdot\|^2) = 1$, with a weaker condition $c_1(\{\bv_i:i \in \{0\} \cup V\},\|\cdot\|^2) \le 1+\delta$, for arbitrary $\delta>0$. $$\begin{array}{clcl} \Min & \sum_{i \in V} 1-\|\bv_0-\bv_i\|^2/4 &\qquad &\\ {\rm s.t.} & \|\bv_i-\bv_0\|^2+\|\bv_j-\bv_0\|^2 = \|\bv_i-\bv_j\|^2 & &\forall \ ij \in E \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V \\ &c_1(\{\bv_i:i \in \{0\} \cup V\},\|\cdot\|^2) \le 1+\delta && \end{array}$$ \[Thm:SDP-weakl1\] For every $\eps>0$, there is a graph $G$ for which $\frac{\vc(G)}{\sd(G)} \ge 2-\eps$, where $\sd(G)$ is the solution to the above SDP For the proof we show that the negative type metric implied by Charikar’s solution (after adjusting the parameters appropriately) requires distortion of at most $1+\delta$. We postpone the proof to the appendix. Integrality Gap against the stronger Semi Definite formulations =============================================================== In this section we discuss the integrality gap for stronger semi-definite formulations of vertex cover. In particular we show that Charikar’s construction satisfies both SDPs (\[SDP-Karakostas\]) and (\[SDP-VC4\]). We start by describing this construction. Charikar’s construction \[Charikar\] ------------------------------------ The graphs used in the construction are the so called Hamming graphs. These are graphs with vertices $\{-1,1\}^n$ and two vertices are adjacent if their Hamming distance is exactly an even integer $d=\gamma n$. A result of Frankl and Rödl [@FR87] shows that $\vc(G) \ge 2^n-(2-\delta)^n$, where $\delta>0$ is a constant depending only on $\gamma$. Kleinberg and Goemans [@KG95] showed that by choosing proper $n$ and $\gamma$, this graph gives an integrality gap of $2-\eps$ for SDP (\[SDP-VC1\]). Charikar [@Char02] showed that in fact $G$ implies the same result for the SDP formulation in (\[SDP-VC2\]) too. To this end he introduced the following solution to SDP (\[SDP-VC2\]): For every $\bu_i \in \{-1,1\}^n$, define $\bu_i'=\bu_i/\sqrt{n}$, so that $\bu_i'\cdot \bu_i'=1$. Let $\lambda=1-2\gamma$, $q(x) = x^{2t} + 2t\lambda^{2t-1}x$ and define $\by_0=(0,\ldots,0,1)$, and $$\by_i={\sqrt{1-\beta^2 \over q(1)}} \left(\underbrace{\bu_i' \otimes \ldots \otimes \bu_i'}_{\mbox{$2t$ times}}, \sqrt{2t\lambda^{2t-1}}\bu_i',0\right) + \beta \by_0,$$ where $\beta$ will be determined later. Note that $\by_i$ is normalized to satisfy $\|\by_i\|=1$. Moreover $\by_i$ is defined so that $\by_i\cdot \by_j$ takes its minimum value when $ij \in E$, i.e., when $\bu_i'\cdot \bu_j'=-\lambda$. As is shown in [@Char02], for every $\eps>0$ we may set $t = \Omega({1\over \eps}), \beta=\Theta(1/t), \gamma = {1\over 4t}$ to get that $(\by_0-\by_i)\cdot (\by_0-\by_j)=0$ for $ij \in E$, while $(\by_0-\by_i)\cdot (\by_0-\by_j)\ge 0$ always. Now we verify that all the triangle inequalities, i.e., the second constraint of SDP (\[SDP-VC2\]) are satisfied: First note that since every coordinate takes only two different values for the vectors in $\{\by_i: i \in V\}$, it is easy to see that $c_1(\{\by_i: i \in V\},\|\cdot\|^2)=1$. So the triangle inequality holds when $i,j,k \in V$. When $i=0$ or $j=0$, the inequality is trivial, and it only remains to verify the case that $k=0$, i.e., $(\by_0-\by_i)\cdot (\by_0-\by_j) \ge 0$, which was already mentioned above. Now $ \sum_{i \in V} (1+\by_0\cdot \by_i)/2 =\frac{1+\beta}{2} \cdot |V| =\left(\frac{1}{2}+O(\eps) \right)|V|$, where by the result of Frankl and Rödl $\vc(G)=(1-o(1))|V|$. Karakostas’ and Pentagonal SDP formulations {#Sec:strongerSDP} ------------------------------------------- Karakostas suggests the following SDP relaxation, that is the result of adding to SDP (\[SDP-VC3\]) the triangle inequalities applied to the set $\{\bv_i : i \in V \cup \{0\}\} \cup \{-\bv_i : i \in V\cup \{0\}\} $. $$\label{SDP-Karakostas} \begin{array}{clcl} \Min & \sum_{i \in V} (1+\bv_0\bv_i)/2 &\qquad &\\ {\rm s.t.} & (\bv_i-\bv_0)\cdot (\bv_j-\bv_0) = 0 & &\forall \ ij \in E \\ & (\bv_i - \bv_k) \cdot (\bv_j - \bv_k) \ge 0 & &\forall \ i,j,k \in V \\ & (\bv_i + \bv_k) \cdot (\bv_j - \bv_k) \ge 0 & &\forall \ i,j,k \in V \\ & (\bv_i + \bv_k) \cdot (\bv_j + \bv_k) \ge 0 & &\forall \ i,j,k \in V \\ &\| \bv_i \| = 1 & & \forall \ i \in \{0\} \cup V. \end{array}$$ We prove that this variant has integrality gap $2-o(1)$ by showing that Charikar’s construction satisfies SDP (\[SDP-Karakostas\]). We postpone the proof to the appendix. \[Thm:Karakostas\] The integrality gap of SDP (\[SDP-Karakostas\]) is bigger than $2-\eps$, for any $\eps >0$. By now we know that taking all the $\ell_1$ constraints leads to an exact relaxation, but clearly one that is not tractable. Our goal here is to explore the possibility that stepping towards $\ell_1$ embeddability while still maintaining computational feasibility would considerably reduce the integrality gap. A canonical set of valid inequalities for $\ell_1$ metrics is the so called [*Hypermetric inequalities*]{}. Metrics that satisfy all these inequalities are called [*hypermetrics*]{}. Again, taking all these constraints is not feasible, and yet we do not know whether this may lead to a better integrality gap (notice that we do not know that Theorem \[Thm:SDP-l1\] remains true if we replace the $\ell_1$ embeddability constraints with a hypermetricity constraint). See [@dela97] for a related discussion about hypermetrics. We instead consider the effect of adding a small number of such constraints. The simplest hypermetric inequalities beside triangle inequalities are the [*pentagonal*]{} inequalities. These inequalities consider two sets of points in the space of size 2 and 3, and require that the sum of the distances between points in different sets is at least the sum of the distances within sets. Formally, let $S,T \subset X$, $|S|=2,|T|=3$, then we have the inequality $\sum_{i\in S,j\in T}d(i,j) \ge \sum_{i,j\in S}d(i,j) + \sum_{i,j\in T}d(i,j)$. To appreciate this inequality it is useful to describe where it fails. Consider the graph metric of $K_{2,3}$. Here, the LHS of the inequality is 6 and the RHS is 8, hence $K_{2,3}$ violates the pentagonal inequality. In the following theorem we show that this “next level” strengthening past the triangle inequalities fails to reduce the integrality gap significantly. \[Thm:Pentagonal\] The integrality gap of SDP (\[SDP-VC4\]) is at least $2-\eps$ for any $\eps >0$. We give here an outline of the proof (the complete proof appears in the appendix). We resort again to Charikar’s construction. Recall that by ignoring $\by_0$ the metric space defined by $d(i,j) = \|\by_i-\by_j\|^2$ is $\ell_1$ embeddable. Therefore, the only $\ell_1$-valid inequalities that may be violated are ones containing $\by_0$. Hence, we wish to consider a pentagonal inequality containing $\by_0$ and four other vectors, denoted by $\by_1,\by_2,\by_3,\by_4$. Assume first that the partition of the five points in the inequality puts ${\by_0}$ together with two other points; then, using the fact that $d(0,1) = d(0,2) = d(0,3) = d(0,4)$ and triangle inequality we get that such an inequality must hold. It therefore remains to consider a partition of the form $(\{\by_1,\by_2,\by_3\},\{\by_4,\by_0\})$, in other words we need to show that: $$d(1,2)+ d(1,3) +d(2,3) + d(0,4) \le d(1,4) +d(2,4)+d(3,4)+d(0,1)+d(0,2)+d(0,3)$$ Let $q(x) = x^{2t}+2t\lambda^{2t-1}x$. Recall that every $\by_i$ is associated with a $\{-1,1\}^n$ unit vector $\bu_i$ and its scaled counterpart $\bu'_i$. After substituting each $\by_i$ as a function of $\bu'_i$, the inequality gets the form $$\label{ineq2} E = q(\bu'_1\cdot \bu'_2)+q(\bu'_1\cdot \bu'_3)+q(\bu'_2\cdot \bu'_3)-q(\bu'_1\cdot \bu'_4)-q(\bu'_2\cdot \bu'_4)-q(\bu'_3\cdot \bu'_4) \ge-2q(1)/(1+\beta)$$ The rest of the proof analyzes the minima of the function $E$ and ensures that (\[ineq2\]) is satisfied at those minima. We proceed by first partitioning the coordinates of the original hypercube into four sets according to the sign of $\bu_1,\bu_2$ and $\bu_3$ on these coordinates. We let $P_0$ be the set of coordinates in which all three vectors assume negative value, and $P_1 (P_2,P_3)$ be the coordinates on which $\bu_1 (\bu_2,\bu_3)$ is positive and the other two vectors negative. Without loss of generality the union of these four sets is the set of all coordinates. Next, $\bu_4$ is considered. Using the convexity of the polynomial $q$ we show that we may assume that $\bu_4$ is either all $1$ or all $-1$ on each set $P_i$. Stronger properties of $q$ ensure that $\bu_4$ is $-1$ on the $P_0$ coordinates. The cases left to check now are characterized by whether $\bu_4$ is $1$ or $-1$ on each of $P_1,P_2,P_3$. By symmetry, all we need to know is the number of blocks $P_i$ on which $\bu_4$ takes the value $1$. Hence we are left with four cases and we use calculus arguments to analyze each case separately. Our analysis shows that in all cases the function $E$ is minimized when $\bu_4$ identifies with one of $\bu_1,\bu_2,\bu_3$; but then it can be easily seen that the pentagonal inequality reduces to a triangle inequality which we know is valid. Lower bound for embedding negative type metrics into $\ell_1$ {#sec:l2metl1} ============================================================= While, in view of Theorem \[Thm:SDP-weakl1\], Charikar’s metric does not supply an example that is far from $\ell_1$, we may still (partly motivated by Theorem \[Thm:SDP-l1\]) utilize the idea of “tensoring the cube” and then adding some more points in order to achieve negative type metrics that are not $\ell_1$ embeddable. Our starting point is an isoperimetric inequality on the cube that generalizes the standard one, and under certain conditions provides better edge expansion guarantee. Such a setting is also relevant in [@KhotVishnoi; @KrauthgamerRabani] where harmonic analysis tools are used to bound expansion; there tools are unlikely to be applicable to our case where the interest and improvements lie in the constants. (Generalized Isoperimetric inequality) For every set $S \subseteq Q_n$, $$|E(S,S^c)| \ge |S|(n-\log_2 |S|)+p(S).$$ where $p(S)$ denotes the number of vertices $\bu \in S$ such that $-\bu \in S$. We use induction on $n$. Divide $Q_n$ into two sets $V_{1}=\{\bu:\bu_1=1\}$ and $V_{-1}=\{\bu:\bu_1=-1\}$. Let $S_{1}=S \cap V_{1}$ and $S_{-1}=S \cap V_{-1}$. Now, $E(S,S^c)$ is the disjoint union of $E(S_1,V_1 \setminus S_1)$, $E(S_{-1},V_{-1} \setminus S_{-1})$, and $E(S_1,V_{-1} \setminus S_{-1}) \cup E(S_{-1},V_{1} \setminus S_1)$. Define the operator $\widehat{\cdot}$ on $Q_n$ to be the projection onto the last $n-1$ coordinates, so for example $\widehat{S_1}=\{\bu \in Q_{n-1}: (1,\bu) \in S_1\}$. It is easy to observe that $$|E(S_1,V_{-1}\setminus S_{-1}) \cup E(S_{-1},V_{1} \setminus S_1)|=|\widehat{S_1} \Delta \widehat{S_{-1}}|.$$ We now argue that $$\label{divide} p(S) + |S_1| - |S_{-1}| \le p(\widehat{S_1})+p(\widehat{S_{-1}})+|\widehat{S_1} \Delta \widehat{S_{-1}}|.$$ To prove (\[divide\]), for every $\bu \in \{-1,1\}^{n-1}$, we show that the contribution of $(1,\bu)$, $(1,-\bu)$, $(-1,\bu)$, and $(-1,-\bu)$ to the right hand side of (\[divide\]) is at least as large as their contribution to the left hand side: This is trivial if the contribution of these four vectors to $p(S)$ is not more than their contribution to $p(\widehat{S_{1}})$, and $p(\widehat{S_{-1}})$. We therefore assume that the contribution of the four vectors to $p(S)$, $p(\widehat{S_{1}})$, and $p(\widehat{S_{-1}})$ are $2$, $0$, and $0$, respectively. Then without loss of generality we may assume that $(1,\bu),(-1,-\bu) \in S$ and $(1,-\bu),(-1,\bu) \not\in S$, and in this case the contribution to both sides is $2$. By induction hypothesis and (\[divide\]) we get $$\begin{aligned} |E(S,S^c)| &=&|E(\widehat{S_1},Q_{n-1} \setminus \widehat{S_1}| + |E(\widehat{S_{-1}},Q_{n-1} \setminus \widehat{S_{-1}}| + |\widehat{S_1} \Delta \widehat{S_{-1}}| \\ &\ge & |S_1|(n-1-\log_2 |S_1|)+p(\widehat{S_1})+|S_{-1}|(n-1-\log_2 |S_{-1}|)+ p(\widehat{S_{-1}}) +|\widehat{S_1} \Delta \widehat{S_{-1}}|\\ & \ge & |S|n-|S|-(|S_1|\log_2 |S_1|+|S_{-1}|\log_2 |S_{-1}|) +p(\widehat{S_1})+p(\widehat{S_{-1}})+|\widehat{S_1} \Delta \widehat{S_{-1}}| \\ & \ge & |S|n-(2|S_{-1}|+|S_1|\log_2 |S_1|+|S_{-1}|\log_2 |S_{-1}|) + p(S).\end{aligned}$$ Now the lemma follows from the fact that $2|S_{-1}|+|S_1|\log_2 |S_1|+|S_{-1}|\log_2 |S_{-1}| \le |S| \log_2 |S|$, which can be obtained using easy calculus. We call a set $S \subseteq Q_n$ [*symmetric*]{} if $-\bu \in S$ whenever $\bu \in S$. Note that $p(S)=|S|$ for symmetric sets $S$. For every symmetric set $S \subseteq Q_n$ $$|E(S,S^c)| \ge |S|(n-\log_2 |S|+1).$$ The corollary above implies the following Poincaré inequality. \[Poincare\] (Poincaré inequality for the cube and an additional point) Let $f:Q_n \cup \{{\bf 0}\} \rightarrow \R^m$ satisfy that $f(\bu)=f(-\bu)$ for every $\bu \in Q_n$. Then the following Poincaré inequality holds. $$\label{poincareEq} {1\over 2^n}\cdot {8\over 7}(4 \alpha+1/2) \sum_{\bu,\bv \in Q_n}\|f(\bu)-f(\bv)\|_1 \le \alpha\sum_{\bu\bv \in E} \|f(\bu)-f(\bv)\|_1 + \frac{1}{2}\sum_{\bu\in Q_n}\|f(\bu)-f({\bf 0})\|_1$$ where $\alpha=\frac{\ln 2}{14 - 8\ln 2}.$ It is well known that instead of considering $f:V\rightarrow \ell_1$, it is enough to prove the above inequality for $f:V\rightarrow\{0,1\}$. Further, we may assume without loss of generality that $f({\bf 0}) = 0$. Associating $S$ with $\{\bu:f(\bu)=1\}$, Inequality (\[poincareEq\]) reduces to $$\label{boolpoinc} {1\over 2^n}{8\over 7}(4 \alpha+1/2) |S||S^c| \le \alpha |E(S,S^c)| + |S|/2,$$ where $S$ is a symmetric set, owing to the condition $f(\bu) = f(-\bu)$. From the isoperimetric inequality of Theorem \[Thm:isoper\] we have that $|E(S,S^c)| \ge |S|(x+1)$ for $x= n - \log_2|S|$ and so $$\left(\frac{\alpha (x+1)+1/2}{1-2^{-x}}\right) \frac{1}{2^n} |S||S^c)| \le \alpha |E(S,S^c)| + |S|/2.$$ It can be verified (See Lemma \[calculus\]) that $\frac{\alpha (x+1)+1/2}{1-2^{-x}}$ attains its minimum in $[1,\infty)$ at $x=3$ whence $\frac{\alpha (x+1)+1/2}{1-2^{-x}} \ge {4\alpha+1/2 \over 7/8}$, and Inequality (\[boolpoinc\]) is proven. \[Thm:distortion\] Let $V=\{\tilde{\bu}: \bu \in Q_n\} \cup \{{\bf 0}\}$, where $\tilde{\bu}=\bu \otimes \bu$. Then for the semi-metric space $X=(V, \|\cdot\|^2)$ we have $c_1(X) \ge \frac{8}{7}-\eps$, for every $\eps>0$ and sufficiently large $n$. We start with an informal description of the proof. The heart of the argument is showing that the cuts that participate in a supposedly good $\ell_1$ embedding of $X$ cannot be balanced on one hand, and cannot be imbalanced on the other. First notice that the average distance in $X$ is almost double that of the distance between $\bo$ and any other point (achieving this in a cube structure without violating the triangle inequality was where the tensor operation came in handy). For a cut metric on the points of $X$, such a relation only occurs for very imbalanced cuts; hence the representation of balanced cuts in a low distortion embedding cannot be large. On the other hand, comparing the (overall) average distance to the average distance between neighbouring points in the cube shows that any good embedding must use cuts with very small edge expansion, and such cuts in the cube must be balanced (the same argument says that one must use the dimension cuts when embedding the hamming cube into $\ell_1$ with low distortion). The fact that only [*symmetric cuts*]{} participate in the $\ell_1$ embedding (or else the distortion becomes infinite due to the tensor operation) enables us to use the stronger isoperimetric inequality which leads to the current lower bound. We proceed to the proof itself. We may view $X$ as a distance function with points in $ \bu \in Q_n \cup \{{\bf 0}\}$, and $d(\bu,\bv) = \|\tilde{\bu} - \tilde{\bv}\|^2$. We first notice that $X$ is indeed a metric space, i.e., that triangle inequalities are satisfied: notice that $X \setminus \{\bo\}$ is a subset of $\{-1,1\}^{n^2}$. Therefore, the square Euclidean distances is the same (upto a constant) as their $\ell_1$ distance. Hence, the only triangle inequality we need to check is $\|\tilde{\bu}-\tilde{\bv}\|^2 \le \|\tilde{\bu}-{\bf 0}\|^2 + \|\tilde{\bv}-{\bf 0}\|^2$, which is implied by the fact that $\tilde{\bu}\cdot \tilde{\bv} =(\bu\cdot \bv)^2$ is always nonnegative. For every $\bu,\bv \in Q_n$, we have $d(\bu,{\bf 0}) = \|\tilde{\bu}\|^2 = \tilde{\bu} \cdot \tilde{\bu} = (\bu\cdot \bu)^2=n^2$, and $d(\bu,\bv) = \|\tilde{\bu}-\tilde{\bv}\|^2=\|\tilde{\bu}\|^2+\|\tilde{\bv}\|^2-2(\tilde{\bu} \cdot \tilde{\bv}) = 2n^2-2(\bu\cdot \bv)^2$. In particular, if $\bu \bv \in E$ we have $d(\bu,\bv) = 2n^2-2(n-2)^2 = 8(n-1)$. We next notice that $$\sum_{\bu,\bv \in Q_n}d(\bu,\bv)= 2^{2n}\times 2n^2-2\sum_{\bu,\bv}(\bu \cdot \bv)^2= 2^{2n}\times 2n^2-2\sum_{\bu,\bv}(\sum_i \bu_i\bv_i)^2 =2^{2n}(2n^2-2n),$$ as $\sum_{\bu,\bv} \bu_i\bv_i\bu_j\bv_j$ is $2^{2n}$ when $i=j$, and $0$ otherwise. Let $f$ be a nonexpanding embedding of $X$ into $\ell_1$. Notice that $$d(\bu,-\bu) = 2n^2-2(\bu\cdot \bv)^2 = 0,$$ and so any embedding with finite distortion must satisfy $f(\bu)=f(-\bu)$. Therefore Inequality (\[poincareEq\]) can be used and we get that $$\label{poincfrac} \frac{\alpha \sum_{\bu\bv \in E}\|f(\tilde{\bu})-f(\tilde{\bv})\|_1+ \frac{1}{2}\sum_{\bu \in Q_n} \|f(\tilde{\bu})-f({\bf 0})\|_1} {\frac{1}{2^n}\sum_{\bu,\bv \in Q_n}\|f(\tilde{\bu})-f(\tilde{\bv})\|_1} \ge {8\over 7}(4 \alpha+1/2).$$ On the other hand, $$\label{fractionOriginal} \frac {\alpha \sum_{\bu\bv \in E}d(\bu,\bv) + {1 \over 2}\sum_{\bu\in Q_n} d(\bu,{\bf 0})}{{1 \over 2^n}\sum_{\bu,\bv \in Q_n}d(\bu,\bv)}= \frac{8\alpha(n^2-n)+n^2}{2n^2-2n}= 4 \alpha+1/2 + o(1).$$ The discrepancy between (\[poincfrac\]) and (\[fractionOriginal\]) shows that for every $\eps>0$ and for sufficiently large $n$, the required distortion of $V$ into $\ell_1$ is at least $8/7-\eps$. Conclusion ========== We have considered the metric characterization of SDP relaxations of [Vertex Cover]{} and specifically related the amount of “$\ell_1$ information” that is enforced with the resulting integrality gap. We showed that a $2-o(1)$ integrality gap survives in the feasible extreme of this range, while no integrality gap exists in the most powerful (and not feasible) extreme, i.e., when $\ell_1$ embeddability of the solution is enforced. We further demonstrated that integrality gap is not a continuous function of the possible distortion that is allowed, as it jumps from 1 to $2-o(1)$ when the allowed distortion changes from 1 to $1+\delta$. These results motivated us to find a negative type metric that does not embed well to $\ell_1$, which is a fairly elusive object. The natural extensions of these results are to (i) check whether the addition of more $k$-gonal inequalities (something that can be done efficiently for any finite number of such inequalities) can reduce the integrality gap or prove otherwise. We in fact conjecture that the integrality gap is still $2-o(1)$ when we impose the condition that the solution is a Hypermetric. It is interesting to note that related questions are discussed in the context of LP relaxations of [Vertex Cover]{} in [@ABLT05] (ii) use the nonembeddability construction and technique in Section \[sec:l2metl1\] to find negative type metrics that incur more significant distortion when embedded into $\ell_1$. It is interesting to investigate whether (and how) our findings are connected to the question of the power of Lift and Project methods; specifically the one that is defined with the Positive Semi Definiteness constraints, also known as $\mbox{LS}^+$ (see [@AAT05] for relevant discussion). Notice that $k$ rounds of LS$^+$ will imply all $k$-gonal inequalities, but may be much stronger. In fact, we do not even know whether applying two rounds of LS$^+$ does not lead to an integrality gap of $2-\Omega(1)$. Last, we suggest looking at connections of $\ell_1$-embeddability and integrality gaps for other NP-hard problems. Under certain circumstances, such connections may be used to convert hardness results of combinatorial problems into hardness results of approximating $\ell_1$ distortion. Acknowledgment {#acknowledgment .unnumbered} ============== Special thanks to George Karakostas for very valuable discussions. Appendix ======== Proof of Theorem \[Thm:SDP-weakl1\] \[Sec:proof-SDP-weakl1\] ------------------------------------------------------------ Let $\by_i$ and $\bu_i'$ be defined as in Section \[Charikar\]. To prove Theorem \[Thm:SDP-weakl1\], it is sufficient to prove that $c_1(\{\by_i:i \in \{0\} \cup V\},\|\cdot\|^2)=1+o(1)$. Note that every coordinate of $\by_i$ for all $i \in V$ takes at most two different values. It is easy to see that this implies $c_1(\{\by_i:i \in V\},\|\cdot\|^2)= 1$. In fact $$f:\by_i \mapsto {1-\beta^2 \over q(1)} \left(\frac{2}{n^t}\underbrace{\bu_i' \otimes \ldots \otimes \bu_i'}_{\mbox{$2t$ times}},\frac{2}{\sqrt{n}}2t\lambda^{2t-1}\bu_i'\right),$$ is an isometry from $(\{\by_i:i \in V\},\|\cdot\|^2)$ to $\ell_1$. For $i \in V$, we have $$\|f(\by_i)\|_1 = {1-\beta^2 \over q(1)} \left(\frac{2}{n^t} \times \frac{n^{2t}}{n^t}+\frac{2}{\sqrt{n}}2t\lambda^{2t-1}\frac{1}{\sqrt{n}}+0\right) ={1-\beta^2 \over q(1)}\times(2+4t\lambda^{2t-1})$$ Since $\beta=\Theta(\frac{1}{t})$, recalling that $\lambda=1-\frac{1}{2t}$, it is easy to see that for every $i\in V$, $\lim_{t \rightarrow \infty} \|f(\by_i)\|_1=2$. On the other hand for every $i \in V$ $$\lim_{t \rightarrow \infty}\|\by_i-\by_0\|^2=\lim_{t \rightarrow \infty}2-2(\by_i\cdot \by_0)= \lim_{t\rightarrow \infty}2-2\beta=2.$$ So if we extend $f$ to $\{\by_i: i \in V \cup \{0\}\}$ by defining $f(\by_0)={\bf 0}$, we obtain a mapping from $(\{\by_i: i \in V \cup \{0\}\}, \|\cdot\|^2)$ to $\ell_1$ whose distortion tends to $1$ as $t$ goes to infinity Proof of Theorem \[Thm:Karakostas\] ----------------------------------- We show that the Charikar’s construction satisfies formulation (\[SDP-Karakostas\]). By [@Char02] and from the discussion in Section \[Charikar\], it follows that all edge constraints and triangle inequalities of the original points hold. Hence we need only consider triangle inequalities with at least one nonoriginal point. By homogeneity, we may assume that there is exactly one such point. Since all coordinates of $\by_i$ for $i>0$ assume only two values with the same absolute value, it is clear that not only does the metric they induce is $\ell_1$ but also taking $\pm \by_i$ for $i>0$ gives an $\ell_1$ metric; in particular all triangle inequalities that involve these vectors are satisfied. In fact, we may fix our attention to triangles in which $\pm \by_0$ is the middle point. This is since $$(\pm \by_i - \pm \by_j) \cdot (\by_0 - \pm \by_j) = (\pm \by_j - \by_0) \cdot (\mp \by_i - \by_0).$$ Consequently, and using symmetry, we are left with checking the nonnegativity of $(\by_i + \by_0) \cdot (\by_j + \by_0)$ and $(-\by_i - \by_0)\cdot (\by_j - \by_0)$. $$(\by_i + \by_0) \cdot (\by_j + \by_0) = 1 + \by_0 \cdot (\by_i+\by_j)+\by_i \cdot \by_j \ge 1 + 2\beta + \beta^2 - (1-\beta^2) = 2\beta(1+\beta) \ge 0.$$ Finally, $$(-\by_i - \by_0)\cdot (\by_j - \by_0) = 1 +\by_0 \cdot (\by_i-\by_j) -\by_i \cdot \by_j = 1-\by_i \cdot \by_j \ge 0$$ as $\by_i,\by_j$ are of norm 1. Proof of Theorem \[Thm:Pentagonal\] ----------------------------------- Again we show that the metric space used in Charikar’s construction satisfies the pentagonal inequalities. As explained in the outline of the proof in Section \[Sec:strongerSDP\], we need to consider only pentagonal inequalities in which the partition of the vectors is of the form $(\{\by_1,\by_2,\by_3\}, \{\by_4,\by_0 \})$. Therefore we need to show that: $$d(1,2)+ d(1,3) +d(2,3) + d(0,4) \le d(1,4) +d(2,4)+d(3,4)+d(0,1)+d(0,2)+d(0,3)$$ As the vectors are of unit norm, it is clear that $d(0,i) = 2-2\beta$ for all $i>0$ and that $d(i,j) = 2-2\by_i\by_j$. Recall that every $\by_i$ is associated with a $\{-1,1\}$ vector $\bu_i$ and with its normalized multiple $\bu'_i$. Also, it is simple to check that $\by_i\cdot \by_j = \beta^2+(1-\beta^2)q(\bu'_i\cdot \bu'_j)/q(1)$ where $q(x) = x^{2t}+2\lambda^{2t-1}x$. After substituting this in the previous expression, it is easy to see that our goal is then to show: $$\label{ineq.} E = q(\bu'_1\cdot \bu'_2)+q(\bu'_1\cdot \bu'_3)+q(\bu'_2\cdot \bu'_3)-q(\bu'_1\cdot \bu'_4)-q(\bu'_2\cdot \bu'_4)-q(\bu'_3\cdot \bu'_4) \ge-2q(1)/(1+\beta)$$ We partition the coordinates of the original hypercube into four sets according to the values assumed by $\bu_1,\bu_2$ and $\bu_3$. Assume without loss of generality that in any coordinate at most one of these get the value -1 (otherwise multiply the values of the coordinate by $-1$). We get four sets, $P_0$ for the coordinates in which all three vectors assume value -1, and $P_1,P_2,P_3$ for the coordinates in which exactly $\bu_1,\bu_2,\bu_3$ respectively assumes value 1. We now consider $\bu_4$. We argue that without loss of generality we may assume that $\bu_4$ is “pure” on each of the $P_0,P_1,P_2,P_3$; in other words it is either all 1 or all $-1$ on each one of the them. Assume for sake of contradiction that there are $w$ coordinates in $P_0$ on which $\bu_4$ assumes value $-1$, and that $0<w<|P_0|$. Let $\bu_4^+$ (similarly $\bu_4^-$) be identical to $\bu_4$ except we replace one 1 in $P_0$ by $-1$ (replace one $-1$ in $P_0$ by 1). We show that replacing $\bu_4$ by $\bu_4^+$ or by $\bu_4^-$ we decrease the expression $E$. This means that the original $\bu_4$ could not have been a choice that minimized $E$ and the claim follows. Let $p_i = \bu_i \cdot \bu_4$, $p_i^+ = \bu_i' \cdot (\bu^+_4)'$ and $p_i^- = \bu'_i \cdot (\bu^-_4)'$ for $i = 1,2,3$. Notice that the above replacement only changes the negative terms in (\[ineq.\]) so our goal now is to show that $\sum_{i=1}^3 q(p_i) < \max\{\sum_{i=1}^3 q(p_i^+) , \sum_{i=1}^3 q(p_i^-)\}$. $$\max\{\sum_{i=1}^3 q(p_i^+) , \sum_{i=1}^3 q(p_i^-)\} \ge {\sum_{i=1}^3 q(p_i^+) + \sum_{i=1}^3 q(p_i^-) \over 2} =$$ $$\sum_{i=1}^3 {q(p_i^+) + q(p_i^-) \over 2} > \sum_{i=1}^3 q\left({p_i^+ + p_i^- \over 2}\right) = \sum_{i=1}^3 q(p_i),$$ where the second last inequality is using the (strict) convexity of $q$. This of course applies to $P_1,P_2$ and $P_3$ in precisely the same manner. The above characterization significantly limits the type of configurations we need to check but regretfully, there are still quite a lot of cases to check. For $P_0$, we can in fact say something stronger than we do for $P_1,P_2,P_3$: \[Prop:w\] If there is a violating configuration, there is one with $\bu_4$ that has all the $P_0$ coordinates set to $-1$. This is not a surprising fact; in fact if $q$ was a monotone increasing function this would be obvious, but of course the whole point behind $q$ is that it brings to minimum some intermediate value ($-\lambda$) and hence can not be increasing. The convexity of $q$ is also not enough, and one should really utilize the exact properties of $q$. We postpone the proof till the end and continue our analysis assuming the proposition. The cases left to check now are characterized by whether $\bu_4$ is $1$ or $-1$ on each of $P_1,P_2,P_3$. By symmetry all we really need to know is $$\xi(\bu_4) = |\{i ~:~ \bu_4 \mbox { is 1 on }P_i\}|$$ If $\xi(\bu_4)=1$ it means that $\bu_4$ is the same as one of $\bu_1,\bu_2$ or $\bu_4$ hence the pentagonal inequality reduces to the triangle inequality, which we have already shown is valid. If $\xi(\bu_4)=3$, it is easy to see that in this case $\bu'_1 \bu'_4 = \bu'_2 \bu'_3$, and likewise $\bu'_2 \bu'_4 = \bu'_1 \bu'_3$ and $\bu'_3 \bu'_4 = \bu'_1 \bu'_2$ hence $E$ is $0$ for these cases, which means that the inequality \[ineq.\] is satisfied. We are left with the cases $\xi(\bu_4) \in \{0,2\}$. [**Case 1:**]{} $\xi(\bu_4) = 0$ Let $x = {2 \over n}|P_1|,y = {2 \over n}|P_2|, z = {2 \over n}|P_3|$. Notice that $x+y+z = {2 \over n}(|P_1|+|P_2|+|P_3|) \le 2$, as these sets disjoint. Now, think of $$E = E(x,y,z) =q(1-(x+y)) +q(1-(x+z)) + q(1-(y+z)) -q(1-x) -q(1-y) -q(1-z)$$ as a function from $\R^3$ to $\R$, and we will show the (stronger than necessary) claim that $E$ achieves its minimum in $\{(x,y,z)\in \R^3 : x+y+z \le 2\}$ at points where either $x,y$ or $z$ are zero. Assume without loss of generality that $0\leq x\leq y \leq z$. We consider the function $g(\delta) = E(x-\delta,y+\delta,z)$. It is easy to see that $g'(0) = q'(1-(x+z))-q'(1-(y+z))-q'(1-x)+q'(1-y)$. Our goal is to show that $g'(0)$ is nonpositive, and in fact that $g'(\delta) \le 0$ for every $\delta \in [0,x]$. This, by the Mean Value Theorem implies that $$E(0,x+y,z) \le E(x,y,z),$$ and in particular that in this case we may assume that $x=0$. This means that $\by_1 = \by_4$ which reduces to the triangle inequality on $\by_0,\by_2,\by_3$. Note that in $q'(1-(x+z))-q'(1-(y+z))-q'(1-x)+q'(1-y)$, the two arguments in the terms with positive sign have the same average as the arguments in the terms with negative sign, namely $\mu = 1-(x+y+z)/2$. We now have $g'(0) = q'(\mu+b)-q'(\mu+s)-q'(\mu-s)+q'(\mu-b)$, where $b = (x-y+z)/2, s = (-x+y+z)/2$. $$\begin{aligned} g'(0) & = &[q'(\mu+b)+q'(\mu-b) - q'(\mu+s)-q'(\mu-s)]\\ & = & 2t[(\mu+b)^{2t-1} + (\mu-b)^{2t-1} - (\mu+s)^{2t-1} - (\mu-s)^{2t-1}]\\ & = & 4t\sum_{i \mbox{ even}} {2t-1 \choose i} \mu^{2t-1-i}(b^i - s^i)\end{aligned}$$ Notice that $\mu = 1-(x+y+z)/2 \ge 0$. Further, since $x\le y$, we get that $s \ge b\ge 0$. This means that $g'(0) \le 0$. It can be easily checked that the same argument holds if we replace $x,y$ by $x-\delta$ and $y+\delta$. Hence $g'(\delta) \le 0$ for every $\delta \in [0,x]$, and we are done. [**Case 2:**]{} $\xi(\bu_4) = 2$ The expression for $E$ is now: $$E(x,y,z) = q(1-(x+y))+q(1-(x+z))+q(1-(y+z))-q(1-x)-q(1-y)-q(1-(x+y+z))$$ Although $E(x,y,z)$ is different than in Case 1, the important observation is that if we consider again the function $g(\delta) = E(x-\delta, y+\delta,z)$ then the derivative $g'(\delta)$ is the same as in Case 1 and hence the same analysis shows that $E(0,x+y,z)\le E(x,y,z)$. Therefore we may assume that $x=0$. This means that $\by_2$ identifies with $\by_4$ and the inequality reduces to the triangle inequality on $\by_0,\by_1,\by_3$. It now remains to prove Proposition \[Prop:w\]: [Proof of Proposition \[Prop:w\]]{} Fix a configuration for $\bu_1,\bu_2,\bu_3$ and as before let $x = \frac{2}{n}|P_1|$, $y = \frac{2}{n}|P_2|$, $z = \frac{2}{n}|P_3|$, and $w = \frac{2}{n}|P_0|$, where $w>0$. Consider a vector $\bu_4$ that has all $-1$’s in $P_0$. Let $H_i = \frac{2}{n}H(\bu_i,\bu_4)$, where $H(\bu_i,\bu_4)$ is the Hamming distance from $\bu_4$ to $\bu_i$, $i=1,2,3$. It suffices to show that replacing the $P_0$-part of $\bu_4$ with $1$’s (which means adding $w$ to each $H_i$) does not decrease the LHS of \[ineq.\], i.e., that: $$\label{Eq:w} q(1-H_1) + q(1-H_2)+ q(1-H_3) \geq q(1-(H_1+w)) +q(1-(H_2+w))+q(1-(H_3+w))$$ Because of the convexity of $q$ as explained before, the cases that we need to consider are characterized by whether $\bu_4$ is $1$ or $-1$ on each of $P_1,P_2,P_3$. By symmetry there are $4$ cases to check, corresponding to the different values of $\xi(\bu_4)$. In some of these cases, we use the following argument: consider the function $g(\delta) = q(1-(H_1+\delta)) +q(1-(H_2+\delta))+q(1-(H_3+\delta))$, where $\delta\in [0,w]$. Let $a_i = 1- (H_i+\delta)$. The derivative $g'(\delta)$ is: $$g'(\delta) = -(q'(a_1)+q'(a_2)+q'(a_3)) = -2t(a_1^{2t-1}+a_2^{2t-1}+a_3^{2t-1}+3\lambda^{2t-1})$$ If we show that the derivative is negative for any $\delta\in [0,w]$, that would imply that $g(0)\geq g(w)$ and hence we are done since we have a more violating configuration if we do not add $w$ to the Hamming distances. [**Case 1:**]{} $\xi(\bu_4) = 0$ In this case $H_1 = x$, $H_2 = y$, $H_3 = z$. Note that $x+y+z+w = 2$. Hence, if $H_i\geq 1$ for some $i$, say for $H_1$, then $H_2+\delta\leq 1$ and $H_3+\delta\leq 1$. This implies that $a_2\geq 0$ and $a_3\geq 0$. Thus $$g'(\delta)\leq -( -1 +3\lambda^{2t-1}) \leq 1-3/e<0$$ since $\lambda^{2t-1} = (1-\frac{1}{2t})^{2t-1}\geq 1/e$. Hence we are done. Therefore, we can assume that $H_i< 1$ for all $i$, i.e., $1-H_i\geq 0$. We now compare the LHS and RHS of (\[Eq:w\]). In particular we claim that each term $q(1-H_i)$ is at least as big as the corresponding term $q(1-(H_i+w))$. This is because of the form of the function $q$. Note that $q$ is increasing in $[0,1]$ and also that the value of $q$ at any point $x\in [0,1]$ is greater than the value of $q$ at any point $y\in [-1,0)$. Therefore since $1-H_i>0$ and since we only subtract $w$ from each point, it follows that (\[Eq:w\]) holds. [**Case 2:**]{} $\xi(\bu_4) = 1$ Assume without loss of generality that $\bu_4$ is $1$ on $P_1$ only. In this case, $H_1 = 0$, $H_2 = x+y$ and $H_3 = x+z$. The LHS of inequality (\[Eq:w\]) is now: $LHS = q(1) + q(1-(x+y)) + q(1-(x+z))$, whereas the RHS is: $$RHS = q(1-w) + q(1-(x+y+w)) + q(1-(x+z+w)) = q(1-w) + q(-1+z)+ q(-1+y)$$ by using the fact that $x+y+w = 2-z$. Let $\alpha_1 = 1$, $\alpha_2 = 1-(x+y)$, $\alpha_3 = 1-(x+z)$. The LHS is the sum of the values of $q$ at these points whereas the RHS is the sum of the values of $q$ after shifting each point $\alpha_i$ to the left by $w$. Let $\alpha'_i = \alpha_i-w$. The difference $\Delta = q(1) - q(1-w)$ will always be positive since $q(1)$ is the highest value that $q$ achieves in $[-1,1]$. Therefore to show that (\[Eq:w\]) holds it is enough to show that the potential gain in $q$ from shifting $\alpha_2$ and $\alpha_3$ is at most $\Delta$. Suppose not and consider such a configuration. This means that either $q(\alpha'_2)>q(\alpha_2)$ or $q(\alpha'_3)>q(\alpha_3)$ or both. We will consider the case that both points achieve a higher value after being shifted. The same arguments apply if we have only one point that improves its value after subtracting $w$. Hence we assume that $q(\alpha'_2)>q(\alpha_2)$ and $q(\alpha'_3)>q(\alpha_3)$. Before we proceed, we state some properties of the function $q$, which can be verified by simple calculations: \[Cl:q\] The function $q$ is decreasing in $[-1,-\lambda]$ and increasing in $[-\lambda, 1]$. Furthermore, for any $2$ points $x,y$ such that $x\in [-1,2-3\lambda]$ and $y\geq 2-3\lambda$, $q(y)\geq q(x)$. Using the above claim, we can argue about the location of $\alpha_2$ and $\alpha_3$. If $\alpha_2\geq 2-3\lambda\geq -\lambda$, then $q(\alpha_2)\geq q(\alpha'_2)$. Thus both $\alpha_2$ and $\alpha_3$ must belong to $[-1,2-3\lambda] = [-1, -1+\frac{3}{2t}]$. We will restrict further the location of $\alpha_2$ and $\alpha_3$ by making some more observations about $q$. The interval $[-1,2-3\lambda]$ is the union of $A_1 = [-1,-\lambda]$ and $A_2 = [-\lambda, 2-3\lambda]$ and we know $q$ is decreasing in $A_1$ and increasing in $A_2$. We claim that $\alpha_2, \alpha_3$ should belong to $A_1$ in the worst possible violation of (\[Eq:w\]). To see this, suppose $\alpha_2\in A_2$ and $\alpha_3\in A_2$ (the case with $\alpha_2\in A_2$, $\alpha_3\in A_1$ can be handled similarly). We know that $q$ is the sum of a linear function and the function $x^{2t}$. Hence when we shift the $3$ points to the left, the difference $q(1) - q(1-w)$ is at least as big as a positive term that is linear in $w$. This difference has to be counterbalanced by the differences $q(\alpha'_2) - q(\alpha_2)$ and $q(\alpha'_3)-q(\alpha_3)$. However the form of $q$ ensures that there is a point $\zeta_2\in A_1$ such that $q(\alpha_2) = q(\zeta_2)$ and ditto for $\alpha_3$. Hence by considering the configuration where $\alpha_2\equiv \zeta_2$ and $\alpha_3\equiv\zeta_3$ we will have the same contribution from the terms $q(\alpha'_2) - q(\alpha_2)$ and $q(\alpha'_3)-q(\alpha_3)$ and at the same time a smaller $w$. Therefore we may assume that $w\leq |A_1| = \frac{1}{2t}$, which is a very small number. By substituting the value of $q$, (\[Eq:w\]) is equivalent to showing that: $$1-(1-w)^{2t} + 6t\lambda^{2t-1}w \geq (\alpha_2-w)^{2t} - \alpha_2^{2t} + (\alpha_3-w)^{2t} - \alpha_3^{2t}$$ It is easy to see that the difference $1-(1-w)^{2t}$ is greater than or equal to the difference $(\alpha_2-w)^{2t} - \alpha_2^{2t}$. Hence it suffices to show: $$6t\lambda^{2t-1}w \geq (\alpha_3-w)^{2t} - \alpha_3^{2t}$$ Since $w$ is small, we estimate the difference $(\alpha_3-w)^{2t} - \alpha_3^{2t}$ using the first derivative of $x^{2t}$ (the lower order terms are negligible). Thus the RHS of the above inequality is at most $2t|\alpha_3|^{2t-1}w$, which is at most $2tw$. But the LHS is: $$6t\lambda^{2t-1}w \geq 6t/e w > 2tw$$ Therefore no configuration in this case can violate (\[Eq:w\]) and we are done. [**Case 3:**]{} $\xi(\bu_4) = 2$ Assume that $\bu_4$ is $1$ on $P_1$ and $P_2$. Now $H_1 = y$, $H_2 = x$, $H_3 = x+y+z$. The LHS and RHS of (\[Eq:w\]) are now: $$\begin{aligned} LHS &=& q(1-y) + q(1-x) + q(1-(x+y+z))\\ RHS & = & q(1-(y+w)) + q(1-(x+w)) + q(-1) \end{aligned}$$ As in case 2, let $\alpha_1 = 1-y$, $\alpha_2 = 1-x$ and $\alpha_3 = 1-(x+y+z)$ be the $3$ points before shifting by $w$. First note that either $\alpha_1> 0$ or $\alpha_2> 0$. This comes from the constraint that $x+y+z+w = 2$. Assume that $\alpha_1> 0$. Hence $q(\alpha_1) - q(\alpha_1-w)>0$. If $\alpha_2\not\in[-1,2-3\lambda]$ then we would be done because by Claim \[Cl:q\], $q(\alpha_2) - q(\alpha_2-w)>0$. Therefore the only way that (\[Eq:w\]) can be violated is if the nonlinear term $(\alpha_3-w)^{2t} - \alpha_3^{2t}$ can compensate for the loss for the other terms. It can be easily checked that this cannot happen. Hence we may assume that both $\alpha_2,\alpha_3\in[-1,2-3\lambda]$ and that $q(\alpha_2-w)>q(\alpha_2)$, $q(\alpha_3-w)> q(\alpha_3)$. The rest of the analysis is based on arguments similar to case 2 and we omit it from this version. [**Case 4:**]{} $\xi(\bu_4) = 3$ This case can also be done using similar arguments with case 2 and 3. A technical lemma ----------------- \[calculus\] The function $f(x)=\frac{\alpha(x+1)+1/2}{1-2^{-x}}$ for $\alpha=\frac{\ln 2}{14 - 8\ln 2}$ attains its minimum in $[1,\infty]$ at $x=3$. The derivative of $f$ is $$\frac{1-2^{-x} - (\alpha(x+1)+1/2) \ln(2) 2^{-x}}{(1-2^{-x})^2}.$$ It is easy to see that $f'(3)=0$, $f(1)=4\alpha+1>8/7$, and $\lim_{x \rightarrow \infty} f(x)=\infty$. So it is sufficient to show that $$g(x)=1-2^{-x} - (\alpha(x+1)+1/2) \ln(2) 2^{-x},$$ is an increasing function in the interval $[1,\infty)$. To show this note that $$g'(x)=2^{-x}\ln(2)\left(1-\alpha+\alpha x\ln(2)+\alpha\ln(2)\right)>0,$$ for $x \ge 1$. [^1]: To be more precise, Charikar’s proof was for a slightly weaker formulation than (\[SDP-VC2\]) but it is not hard to see that the same construction works for SDP (\[SDP-VC2\]) as well. [^2]: As Khot and Vishnoi note, and leave as an open problem, it is possible that their example satisfies some or all $k$-gonal inequalities.

ArXiv

=-2.5cm =-1.5cm =24.7cm =16.5cm [**A Phase in a Coherent State Wave Function – Is It Always Irrelevant?**]{}\ [**W. Berej and P. Rozmej**]{}\ [ *Theoretical Physics Department, University Maria Curie-Skłodowska, Pl. M.Curie-Skłodowskiej 1, Pl-20031 Lublin, Poland\ E-mail: [email protected], [email protected]*]{}\ The coherent states of a harmonic oscillator were first constructed by Schrödinger$^{1,2}$ as “the most classical” states of the oscillator. The properties of these states have been studied in a systematic way by Glauber$^{3}$, who showed their importance for the quantum mechanical treatment of optical coherence and who introduced the name “coherent state”. The description of the coherent states is now a standard subject in certain textbooks on quantum mechanics$^{4,5}$, monographs on quantum optics$^{6,7,8}$ and more specialized texts devoted to coherent states of different kinds$^{9,10,11}$. They were also discussed in many articles in this journal$^{12}$. As the applications of the coherent states usually do not require the explicit introduction of coordinate variable, the wave function for the coherent state is written down only in some texts$^{3,5,8,9,10}$. However, a phase factor, properly defined at the beginnig of the discussion of the coherent states, is eventually omitted in some sources$^{3,8,10}$. In this note we would like to point out the circumstance in which the knowledge of this unimodular factor in the coordinate representation of the coherent state is essential. Let’s recall first the basic facts concerning the coherent states$^{13}$. Following Glauber$^{3}$, the coherent state $|\alpha\rangle$ is conveniently defined as an eigenstate of the harmonic oscillator anihilation operator $\hat{a}$ with complex eigenvalue $\alpha$: $$\hat{a}\,|\alpha\rangle = \alpha\,|\alpha\rangle.$$ Using this definition and its adjoint form, one can calculate the expectation values of the position and momentum $$\begin{aligned} \langle\alpha|x|\alpha\rangle &\!\! =\!\! & (2\hbar/m\omega)^{\frac{1}{2}}\,\mbox{Re}\,\alpha\nonumber\\ \langle\alpha|p|\alpha\rangle & = & (2\hbar m\omega)^{\frac{1}{2}}\,\mbox{Im}\,\alpha.\end{aligned}$$ It is also not difficult to find the Fock representation of the normalized state $|\alpha\rangle$ in the basis of states $|n\rangle$ labelled by the occupation number $n$: $$|\alpha\rangle = e^{-\frac{1}{2}|\alpha|^{2}}\,\sum_{n=0}^{\infty} \frac{\alpha^{n}}{\sqrt{n!}}|n\rangle,$$ where the phase is defined by choosing $$\langle 0|\alpha\rangle= e^{-\frac{1}{2}|\alpha|^{2}}.$$ In the further discussion of the coherent states one can check their well known properties: they minimize the uncertainty relation and evolve in time, following the motion of a classical oscillator. The effect of the time evolution operator acting on the coherent state $|\alpha\rangle$ is easily calculated by making use of expansion (3): $$\exp(-iHt/\hbar)|\alpha\rangle = e^{-\frac{1}{2}i\omega t}|\alpha e^{-i\omega t}\rangle,$$ i.e. the coherent state remains coherent at all times. One can immediately see that expectation values (2) of the position and momentum in the evolving coherent state carry out a simple harmonic motion. Property (5) allows one to write down the wave function obeying the Schrödinger equation and the initial condition $|\psi(0)\rangle=|\alpha\rangle$, simply adding the factor with the zero-point energy in the exponent and substituting $\alpha e^{-i\omega t}$ for $\alpha$ in the coordinate representation of the coherent state $|\alpha\rangle$. However, to exploit this property one must know the full coherent state wave function $\psi_{\alpha}(x)= \langle x|\alpha\rangle$ with a proper phase factor defined by condition (4). As we shall see below, the phase depends on the complex number $\alpha$ labelling the coherent state and the substitution $\alpha\longrightarrow\alpha e^{-i\omega t}$ makes it time dependent. Therefore, the omission of the phase factor leads to the wave function which is not a solution of the Schrödinger equation. During our study of the dynamics of wave packets$^{14}$, among the cited references$^{3-12}$ we found only one text$^{5}$ which provides this phase with a detailed derivation$^{15}$. To the end of this note we derive the needed phase factor starting from definition (1). It is a simple exercise involving calculation of the standard integrals with the Gaussian functions in their integrands, though certain care must be taken because $\alpha$ is a complex number. At the same time we shall correct some errors existing in the literature. The eigenvalue problem of Eq. (1) in the coordinate representation is a first-order differential equation for the wave function $\psi_{\alpha}(x)$. Its general solution has the following form: $$\psi_{\alpha}(x) = A\exp\!\left[-\left(\sqrt{\frac{m\omega}{2\hbar}}x- \alpha\right)^{\!2\,} \right],$$ where $A$ is a normalization constant. Thus, the corresponding probability density has a familiar, Gaussian shape. The normalization condition yields$^{16}$ $$A = \left(\frac{m\omega}{\pi\hbar}\right)^{\!\frac{1}{4}} \exp\left\{\frac{1}{4}(\alpha-\alpha^{\ast})^{2}+i\phi\right\},$$ where $\phi$ is a real phase to be established from condition (4). Carrying out one more integration, in which the explicit expression for the ground state wave function of the harmonic oscillator is used, one obtains: $$i\phi = \frac{1}{4}(\alpha^{2}-\alpha^{\ast\,2}).$$ Therefore, the normalized coherent state wave function with a phase factor chosen to ensure (4), may be written as: $$\psi_{\alpha}(x) = \left(\frac{m\omega}{\pi\hbar}\right)^{\!\frac{1}{4}} \exp\!\left(\frac{1}{2}\alpha^{2}-\frac{1}{2}|\alpha|^{2}\right)\, \exp\!\left[-\left(\sqrt{\frac{m\omega}{2\hbar}}x- \alpha\right)^{\!2\,}\right].$$ After the introduction of $\alpha = \alpha_{1} + i\alpha_{2}$, this may be rewritten in a more readable form: $$\psi_{\alpha}(x) = \left(\frac{m\omega}{\pi\hbar}\right)^{\!\frac{1}{4}} \exp\left(-i\alpha_{1}\alpha_{2}\right)\, \exp\!\left[-\left(\sqrt{\frac{m\omega}{2\hbar}}x- \alpha_{1}\right)^{\!2\,}\right]\, \exp\!\left(2i\alpha_{2}\sqrt{\frac{m\omega}{2\hbar}}x\right).$$ Finally, one can introduce the expectation values of Eq. (2), for brevity denoted by $\langle x\rangle, \langle p\rangle$: $$\psi_{\alpha}(x) = \left(\frac{m\omega}{\pi\hbar}\right)^{\!\frac{1}{4}} \exp\!\left(-i\frac{\langle x\rangle\langle p\rangle}{2\hbar}\right)\, \exp\!\left[-\frac{m\omega}{2\hbar}(x-\langle x\rangle)^{2}\right]\, \exp\!\left(\frac{i}{\hbar}\langle p\rangle x\right).$$ The above epressions may be found in some research papers$^{17}$. We see that the coherent state wave function differs from a Gaussian wave packet with the mean momentum $\langle p\rangle$ by the phase factor $\exp\!\left[-i\langle x\rangle\langle p\rangle/(2\hbar)\right]$ which must be included if the quantum evolution of $\psi_{\alpha}(x)$ is considered and Eq. (5) is used. When for special initial conditions the substitution $\langle x\rangle\longrightarrow x_{0}\cos\omega t, \langle p\rangle\longrightarrow -m\omega x_{0}\sin\omega t$ is made in Eq. (11), one immediately obtains the well known wave packet oscillating without change of its shape$^{18}$: $$\psi(x,t) = \left(\frac{m\omega}{\pi\hbar}\right)^{\!\frac{1}{4}}\ \!\exp\!\left[-\frac{m\omega}{2\hbar}(x\!-\!x_{0}\cos\omega t)^{2}\right]\, \!\exp\!\left[-i\!\left(\frac{1}{2}\omega t\! +\! \frac{m\omega}{\hbar}x_{0}x\sin\omega t \!-\!\frac{m\omega}{4\hbar}x_{0}^{2}\sin2\omega t\right)\right].$$ [**Literature**]{} 1. E. Schrödinger, “Der Stetige Übergang von der Mikro- zur Makromechanik”, Naturwissenschaften [**14**]{}, 664–666 (1926). 2. See the interesting historical paper: F. Steiner, “Schrödinger’s Discovery of Coherent States”, Physica B [**151**]{}, 323–326 (1988). 3. R.J. Glauber, “Coherent and Incoherent States of the Radiation Field”, Phys. Rev. [**131**]{}, 2766–2788 (1963). 4. E. Merzbacher, [*Quantum Mechanics*]{} (Wiley, New York, 1970), second edition, pp. 362–370. 5. C. Cohen-Tannoudji, B. Diu, and F. Laloe, [*Quantum Mechanics*]{}, Vol. 1, (Wiley, New York, 1978), Complement G$_{V}$. 6. J.R. Klauder and E. Sudarshan, [*Fundamentals of Quantum Optics*]{}, (Benjamin, New York, 1968), Chap. 7. 7. W.H. Louisell, [*Quantum Statistical Properties of Radiation*]{} (Wiley, New York, 1973), pp. 104–110. 8. L. Mandel and E. Wolf, [*Optical Coherence and Quantum Optics*]{} (Cambridge University Press, Cambridge, 1995), Chap. 11. 9. J.R. Klauder and B.-S. Skagerstam, [*Coherent States. Applications in Physics and Mathematical Physics*]{} (World Scientific, Singapore, 1985). 10. A.M. Perelomov, [*Generalized Coherent States and Their Applications*]{} (Springer, Berlin, 1986). 11. W.-M. Zhang, D.H. Feng, and R. Gilmore, “Coherent States: Theory and Some Applications”, Rev. Mod. Phys. [**62**]{}, 867–927 (1990). 12. We mention only two of them: P. Carruthers and M.M. Nieto, “Coherent States and the Forced Harmonic Oscillator”, Am. J. Phys. [**33**]{}, 537–544 (1965); S. Howard and S.K. Roy, “Coherent States of a Harmonic Oscillator”, [*ibid.*]{} [**55**]{},1109–1117 (1987). 13. Sections III and IV of Glauber’s paper (Ref. 3) contain an excellent discussion of the properties of the coherent states and should be an obligatory reading for anyone interested in this subject. A complete and thorough discussion may also be found in Ref. 5. The authors of this textbook choose as the defining condition of the coherent state the requirement that the mean values $\langle x\rangle, \langle p\rangle, \langle H\rangle$ are as close as possible to the classical values. 14. R. Arvieu, P. Rozmej, and W. Berej, “Time-dependent partial waves and vortex rings in the dynamics of wavepackets”, J. Phys. A [**30**]{}, 5381–5392 (1997) . 15. The correct phase factor is also given in Ref. 7. However, the order of the reasoning is reversed there: the phase is chosen somewhat arbitrarily and then $\langle x|\alpha\rangle$ serves as the generating function for oscillator eigenfunctions. 16. The reader of Ref. 3 should be warned that the coherent state wave funcions given by the formula (3.29) of this paper are actually not normalized: the exponential present in our expression (7) is missing there. In the normalization constant written in Ref. 8 there is a sign error in the exponent. 17. For example: M.M. Nieto and L.M. Simmons Jr., “Coherent States for General Potentials I”, Phys. Rev. D [**20**]{}, 1321–1331 (1979); S.-J. Chang and K.-J. Shi, “Evolution and Exact Eigenstates of a Resonant Quantum System”, Phys. Rev. A [**34**]{}, 7–22 (1986). 18. L.I. Schiff, [*Quantum Mechanics*]{} (Mc-Graw Hill, New York, 1968), third edition, pp. 74–76.

ArXiv

--- address: 'LPTHE, Universities of Paris VI & VII and CNRS, 75252 Paris 75005, France' author: - 'Gavin P. Salam' title: 'CAESAR: Computer Automated Resummations' --- Global properties of energy-momentum flow in final-states, such as event shapes and jet-resolution thresholds, offer a good compromise between simplicity and sensitivity to the dynamics of QCD. Thanks to the former, it has been possible to make a wide range of theoretical predictions for them, including both fixed-order and resummed perturbative calculations as well as non-perturbative model calculations. This set of theoretical tools has been essential in helping us to learn about QCD from the extensive experimental data, having led in particular to numerous of measurements of the running coupling, ${\alpha_s}$ [@Bethke:2002rv], tests of the colour factors of QCD [@SU3] and highly instructive studies concerning hadronisation (for reviews, see [@Beneke; @DasSalReview]). Most of the investigations so far have been carried out for observables that are sensitive to the deviation from two-jet ($1+1$-jet) structure in ${e^+e^-}$ (DIS). This fairly straightforward environment has been of considerable value, especially in developing and refining our ideas about hadronisation. However the most stringent tests of the understanding thus gained, as well as possible novel extensions, are likely to be found in the context of multi-jet events, including those at hadron-hadron colliders. A difficulty in such studies is that the next-to-leading logarithmic (NLL) resummed calculations, needed in the limit of only soft and collinear emissions, have up to now usually been carried out by hand, a tedious and error-prone procedure, which has to be repeated separately for each new observable in each process. The complexity of such calculations can increase substantially as one goes to multi-jet processes [@eeKout]. This is to be contrasted with fixed-order calculations, usually embodied in the form of a ‘fixed-order Monte Carlo’ program (FOMC, [e.g. ]{}[@NLOJET]), which given a subroutine for the observable, provides, numerically, the leading and next-to-leading order (LO, NLO) coefficients of the perturbative series. Though the internal complexity of FOMCs increases for processes with extra legs, the principles of their use remain identical regardless of the process. This makes it possible, even for those not expert in higher order QCD calculations, to obtain fixed-order predictions for arbitrary observables, in a wide range of processes. It is natural therefore to ask whether analogous programs could be constructed for resummed calculations. To help explain a solution to this question, proposed in [@caesar], it is useful to define the problem a little more precisely. We will consider a general event shape (or jet-resolution threshold), defined by some function $V(q_1,\ldots,q_N)$ of the momenta $q_1,\ldots,q_N$ of the $N$ particles in an event. If the observable is such that it vanishes smoothly in the limit of $n$ narrow jets, then we call it an $(n+1)$-jet observable, and resummation will be needed in the $n$-jet limit. The resummed calculations should provide results for the probability $f(v)$ that the event shape has a value smaller than $v$. For small $v$, this probability can often (exceptions are discussed below) be written in the ‘exponentiated’ form $$\label{eq:vProb-general} f(v) \simeq \exp\left[ L g_1({\alpha_s}L) + g_2({\alpha_s}L) + {\alpha_s}g_3({\alpha_s}L) + \cdots\right]\,, \qquad\quad L = \ln \frac1v\,,$$ where $Lg_1({\alpha_s}L)$ contains LL terms, ${\alpha_s}^n L^{n+1}$, $g_2({\alpha_s}L)$ the NLL ones and so forth. Usually NLL accuracy represents the state of the art (though a related kind of resummation has recently been performed to NNLL accuracy (${\alpha_s}g_3$) for the ${e^+e^-}$ energy-energy correlation [@deFlorian:2004mp]). In analogy with FOMCs, which calculate the LO and NLO contributions to $f(v)$, the aim of an automated resummation program should be the calculation (separately) of the resummed LL and NLL contributions, [i.e. ]{}the $g_1({\alpha_s}L)$ and $g_2({\alpha_s}L)$ functions, free of any contamination from potentially spurious higher orders, and in a form such that they can be expanded, allowing matching to fixed-order calculations.[^1] The construction of an automated resummer, as least in our approach [@caesar], proceeds however in quite a different fashion from an FOMC (usually just a Monte Carlo integrator with appropriate weights). This is because, in order to separate out the structure of (\[eq:vProb-general\]) it is useful to have some analytical understanding of the observable — for example regarding its behaviour in the presence of a soft and collinear emission, nearly always of the form $$\label{eq:simple-second-time} V(\{{\tilde p}\}, k)= d_{\ell}\left(\frac{k_t}{Q}\right)^{a_\ell} e^{-b_\ell\eta}\, g_\ell(\phi)\>,$$ where $\{{\tilde p}\}$ are the hard momenta in the event and $k$ is a soft emission that is collinear to the hard momentum (leg) with index $\ell$, with a transverse momentum, rapidity and azimuthal angle $k_t$, $\eta$ and $\phi$ relative to leg $\ell$. For a large class of observables, given the coefficients $a_\ell$ and $b_\ell$, the function $g_1({\alpha_s}L)$ has a known simple analytical form. So an automated resummer can determine the LL terms by verifying the parametrisation (\[eq:simple-second-time\]) and determining the coefficients. This sounds remarkably simple, and indeed there is a ‘catch’: one has also to establish that a given observable belongs to the class for which the LL terms truly are just determined by the values of the $a_\ell, b_\ell$. Understanding precisely what that class is has been one of the significant developments of [@caesar], with the introduction of a concept that we have called recursive infrared and collinear (rIRC) safety. Normal infrared and collinear safety (IRC) states that when an emission is very soft and/or collinear it should have no effect on the observable. The recursive variant essentially extends this condition to two-scale problems (as is relevant when resumming) and states that given an ensemble of arbitrarily soft and collinear emissions, the addition of further emissions of similar softness or collinearity should not change the value of the observable by more than a factor of order one. Furthermore the addition of relatively much softer or more collinear emissions (whether with respect to the hard leg or one of the other emissions) should not change the value of the observable by more than some power of the relative extra softness or collinearity. While at first sight this seems similar to normal IRC safety, a number of observables have been found to be IRC safe but rIRC unsafe. Recursive IRC safety is a sufficient condition for the form (\[eq:vProb-general\]) to hold. While in principle one could also imagine resumming rIRC unsafe observables, we are not aware of any cases where this has been done. At NLL accuracy, it is not sufficient to know just the observable’s behaviour when there is a single soft and collinear emission — additionally one needs to know how the observable behaves when there are multiple soft and collinear emissions of roughly similar hardness (its ‘multiple-emission dependence’). Analytically this usually requires an in-depth analysis of the observable. However, as was shown in [@numsum], there exists a general solution to the problem, namely that at NLL all multiple-emission dependence can be taken into account via a contribution $-\ln {{\cal F}}(R')$ to $g_2$, $$\begin{gathered} \label{eq:cF-evshps-anyn} {{\cal F}}(R') = \lim_{\epsilon\to0} \frac{\epsilon^{R'}}{R'} \sum_{m=0}^{\infty} \frac{1}{m!} \left( \prod_{i=1}^{m+1} \sum_{\ell_i=1}^n R_{\ell_i}' \int_{\epsilon}^{1} \frac{d\zeta_i}{\zeta_i} \int_0^{2\pi} \frac{d\phi_i}{2\pi} \right) \delta\!\left(\ln \zeta_1\right) \times \\ \times \exp\left(-R'\ln \lim_{{\bar v}\to0} \frac{V(\{\tilde p\},\kappa_1(\zeta_1 {\bar v}) , \ldots, \kappa_{m+1}(\zeta_{m+1} {\bar v}))}{\bar v}\right),\end{gathered}$$ where $R'({\alpha_s}L)=-\partial_L L g_1({\alpha_s}L)$ ($R'+1$ determines the average number of ‘relevant’ emissions), $R'_\ell$ is the part of $R'$ associated with leg $\ell$, and $\kappa_i(\bar v)$ is any soft momentum collinear to leg $\ell_i$ that satisfies $V(\{\tilde p\},\kappa_i(\bar v)) = \bar v$. Since ${{\cal F}}(R')$ can be calculated by Monte Carlo integration, one eliminates the need for any analytical study of the observable (the various limits can be taken numerically), thus providing the main NLL element of an automated resummation. One unsatisfactory issue that remains with the above approach is connected with the problem of globalness. It was shown in [@NG1] that for observables that are sensitive to emissions only in a restricted angular region (for example a single jet), additional classes of NLL terms arise, termed non-global logarithms (NGLs), so far calculated only in the large-${N_c}$ limit (though see also [@weigert]). Such NGLs depend non-trivially on the details of the boundaries of the restricted region in which the observable is sensitive to emissions, and these boundaries can be quite complex. Accordingly, for the time being, we have chosen to consider only the simpler case of (continuously) global observables, which have the characteristic that the $d_\ell$ are all non-zero and the $a_\ell$ are all equal. These observables’ resummations are free of NGLs. The resulting resummation program is called [<span style="font-variant:small-caps;">caesar</span>]{}, the Computer Automated Expert Semi-Analytical Resummer, and many results obtained with it can be found at [`http://qcd-caesar.org`](http://qcd-caesar.org), for ${e^+e^-}$, DIS and hadron-hadron collisions. Particularly topical are hadron-collider dijet event-shapes, the subject also of recent experimental [@D0Thrust] and theoretical fixed-order [@Nagy03] work. These are of interest for a variety of reasons, ranging from the possibility of ‘standard’ measurements such as ${\alpha_s}$, to more novel studies examining for example the underlying event with techniques such as those discussed in [@KoutZ0], or the soft (colour evolution) anomalous dimension matrices of [@Sterman4Legs]. The event shape studied in [@D0Thrust; @Nagy03] (a transverse thrust) has the drawback that its resummation involves NGLs, and so is beyond the current scope of [<span style="font-variant:small-caps;">caesar</span>]{}. To allow more straightforward compatibility with resummed calculations, it is instead preferable to specifically design a set of global dijet event shapes (and jet rates). This has been the subject of work in [@BSZhh]. One of the main practical issues discussed there is the question of how to reconcile the requirement of globalness with the limited experimental detector coverage at large rapidities. Fortunately this problem can be circumvented in a variety of ways. This opens the road at hadron-colliders to full resummed studies matched with fixed order calculations [@NLOJET], with the prospect of significantly extending the very fruitful studies carried out in ${e^+e^-}$ and DIS over the past years. **Acknowledgements.** The work discussed here has been carried out in collaboration with Andrea Banfi and Giulia Zanderighi. I with to thank them also for helpful comments on this write-up. [99]{} S. Bethke, Nucl. Phys. Proc. Suppl.  [**121**]{} (2003) 74 \[hep-ex/0211012\]. S. Kluth et al., Eur. Phys. J. C [**21**]{} (2001) 199 and references therein. M. Beneke, Phys. Rept.  [**317**]{} (1999) 1. M. Dasgupta and G. P. Salam, J. Phys. G [**30**]{} (2004) R143 \[hep-ph/0312283\]. A. Banfi, G. Marchesini, Yu. L. Dokshitzer and G. Zanderighi, JHEP [**0007**]{} (2000) 002 \[hep-ph/0004027\]; JHEP [**0105**]{} (2001) 040 \[hep-ph/0104162\]. Z. Nagy, Phys. Rev. Lett.  [**88**]{} (2002) 122003 \[hep-ph/0110315\]. A. Banfi, G. P. Salam and G. Zanderighi, Phys. Lett. B [**584**]{} (2004) 298 \[hep-ph/0304148\]; hep-ph/0407286. D. de Florian and M. Grazzini, hep-ph/0407241. G. Abbiendi [*et al.*]{}, Comput. Phys. Commun.  [**67**]{} (1992) 465. G. Corcella [*et al.*]{}, JHEP [**0101**]{} (2001) 010 \[hep-ph/0011363\]. T. Sjöstrand, Comput. Phys. Commun.  [**82**]{} (1994) 74; T. Sjöstrand [*et al.*]{}, Comput. Phys. Commun.  [**135**]{} (2001) 238; \[hep-ph/0010017\]. A. Banfi, G. P. Salam and G. Zanderighi, JHEP [**0201**]{} (2002) 018 \[hep-ph/0112156\]. M. Dasgupta and G. P. Salam, Phys. Lett. B [**512**]{} (2001) 323 \[hep-ph/0104277\]; JHEP [**0203**]{} (2002) 017 \[hep-ph/0203009\]. H. Weigert, Nucl. Phys. B [**685**]{} (2004) 321 \[hep-ph/0312050\]. I. A. Bertram \[D0 Collaboration\], Acta Phys. Polon. B [**33**]{} (2002) 3141. Z. Nagy, Phys. Rev. D [**68**]{} (2003) 094002 \[hep-ph/0307268\]. A. Banfi, G. Marchesini, G. Smye and G. Zanderighi, JHEP [**0108**]{} (2001) 047 \[hep-ph/0106278\]. J. Botts and G. Sterman, Nucl. Phys. B [**325**]{} (1989) 62; N. Kidonakis and G. Sterman, Nucl. Phys. B [**505**]{} (1997) 321 \[hep-ph/9705234\]; N. Kidonakis, G. Oderda and G. Sterman, Nucl. Phys. B [**531**]{} (1998) 365 \[hep-ph/9803241\]; G. Oderda, Phys. Rev. D [**61**]{} (2000) 014004 \[hep-ph/9903240\]. A. Banfi, G. P. Salam and G. Zanderighi, hep-ph/0407287. [^1]: This can be contrasted with the output from exclusive Monte Carlo event generators like Herwig [@Herwig] or Pythia [@Pythia], which embody resummation for almost any observable via parton branching algorithms, and which provide an overall result for $f(v)$, generally correct to LL accuracy, sometimes to NLL, but for which it is difficult to guarantee the accuracy, extract separately the LL and NLL terms, perform matching to fixed order, and unambiguously separate the perturbative and non-perturbative effects.

ArXiv

--- abstract: | The act of measurement on a quantum state is supposed to “collapse” the state into one of several eigenstates of the operator corresponding to the observable being measured. This measurement process is sometimes described as outside standard quantum-mechanical evolution and not calculable from Schrödinger’s equation [@Borg]. The Weisskopf-Wigner approach is an attempt to derive the probability of an irreversible process, [*[viz]{}*]{}, the emission of a photon by an excited atom as it decays into a lower-energy state, by an innovative approximation. This approximation method turns what should be a completely reversible process into a spontaneous collapse, in agreement with experiment. The general reason why this method works is this: the decay of an atom with the emission of a photon which disappears into a continuum of states means the emitted energy gets “lost” in the continuum. It cannot easily make its way back to the emitter, hence the state of the atom collapses from the “excited” to the “ground” state. This paper takes a similar tack, but makes no such approximations. We tackle the evolution of the state of an electron as it goes through double slits. We model the actual process of checking whether the electron has gone through one particular slit, including the fact that the process of amplification of the detection requires external energy. The process of measurement, described exactly by Schrödinger’s equation with a sensibly chosen interaction (hermitian) Hamiltonian mathematically “collapses the wave-function”, if the measurement succeeds. However, depending on the structure of the continuum of states of the measuring apparatus, the “collapse” will eventually reverse in a small system. The time-scale of such reversal in typical macroscopic systems will be longer than the age of the Universe, hence for all effective purposes, the wave-function has indeed “collapsed” upon measurement. This calculation supports the viewpoint that the Observer as well as the Quantum System can be thought of as part of a larger quantum state, which hasn’t really collapsed at all, simply moved around in Hilbert space in response to an interaction Hamiltonian. The consequences of the act of measurement can indeed be modeled by standard quantum mechanics. No additional assumptions are needed to produce decoherence and wave-function collapse. The approach to treating the observer as part of the measurement process is quite well-known, researched intensively and described lucidly in [@Coleman]. A very similar problem has been studied intensively by Allahverdyan et. al[@Balian1; @Balian2; @Balian3], using the Liouiville-von Neumann density matrix formalism. While this paper describes a study of the Schrödinger equation with a time-dependent Hamiltonian directly, it is different in that we do not require a Curie-Weiss type of paramagnetic-ferromagnetic transition to achieve similar results. In addition, a central point in our thesis is that one needs to add energy into the system to make a measurement. However, it is similar in that the macroscopic number of variables that are affected are the root of the rapid collapse. author: - Satish Ramakrishna title: 'A microscopic model of wave-function collapse' --- The study of quantum measurement has had a long history, beginning with the earliest discoveries in quantum mechanics. The studies of decoherence in quantum computers represent practical realizations of the measurement problem, as wave-functions representing quantum states of computers slowly transition from superposed to mixed states due to interactions with the environment. In [@Balian1; @Balian2], the authors study a spin system coupled to a large magnetic system in a metastable state, near the Curie-Weiss transition and further coupled to a heat bath. They demonstrate that the spin system’s density matrix, when traced over its environment moves from possessing off-diagonal elements to one with purely diagonal elements, consistent with a mixed (collapsed) state. The collapse happens very quickly, at a rate proportional to the macroscopic number of degrees of freedom in the magnet that undergoes the transition. However, the “registration” of this collapse, at the instrument pointer, is found to be much slower, driven by the relaxation of the diagonal elements in the density matrix. In this paper, we study the classic double-slit experiment, with a photomultiplier tube and add an important ingredient - measurement is a process where external energy is used to detect and amplify weak signals, which are then registered on a macroscopic device. This energy could come through the energy of a photon that scatters off an electron, or the subsequent photo-amplification in our process. Such amplification is necessary to prevent noise from being mistaken for a measurement. The double-slit experiment ========================== Consider the description of the standard double-slit experiment. A source of electrons is on the left in Fig. 1. It shoots electrons, at an extremely slow rate, at an absorbing screen (on the right). Interposed between the screen and the electron gun is another absorbing screen, which has two slits (labeled $1$ and $2$). Between the slits and assumed to be aimed at slit $1$, there is an intense source of high-energy “locator” photons, that can scatter (if they interact) off the electrons. Most of the ones that scatter off electrons are picked up by the input maw of a photo-multiplier tube. The energy of the photons that scatter off the electrons is $\omega_p$, while the energy of the electrons is $\omega_e$ (we use natural units throughout the calculation, so $\hbar = 1$).  These “locator” photons are captured by the photomultiplier tube after scattering. We don’t need, for the purposes of the calculation, to understand the details of how the exact process of amplification occurs. We just need to realize that the end result is a bright spot, that emits several (a macroscopic number of) photons of several frequencies, from a point on a display tube. This point on the display tube is, therefore, directly connected to the reception of a single locator photon at the input end of the amplifier. The physical description of the Hamiltonian applicable to the problem is as follows. There is a “free” term that represents the free electron, with energy $\omega_e$. This electron could be at slit $1$ or at slit $2$, these are represented by subscripts on the number operator for the electrons. There is also a “free” term that represents the free photons, of frequency $\omega_p$ (for the locator photon that directly scatters off the electron at slit $1$), while $\omega_i$ represents the frequencies of the photons that are subsequently produced by the initially scattered photon (and thence) at the photomultiplier tube. We ignore photons from the intense source that do not scatter off the electron, since the photomultiplier input ignores the direct beam and only looks for scattered photons. We are modeling the process where the electron interacts with one photon at the slit, which, then at the photomultiplier tube, cascades, producing $N$ other photons in each state $i$ with each subsequent interaction. In a real photo-multiplier tube, the locator photons excite a few electrons, each of which then produces a few more electrons. The process cascades and a large current (millions of times larger than the initially excited current) is produced. In our case, we assume the locator photons excite a few photons, which each excite a few more photons and so on, to produce a macroscopic number of photons in a continuum of photon energy states. This constitutes “measurement” by the device. We consider the initial wave-function of the electron to have collapsed into one that is definitely localized at slit $1$ when the amplitude for the superposed state falls very close to 0. There is a subtle point to consider. We purposely select, out of all the possible ways to select independent vectors in the Hilbert space of the problem, vectors that correspond to our particular situation. In particular, one might ask why we assume that the electron has indeed gone through slit $1$ when we see a macroscopic signal from the photomultiplier tube at slit $1$. The answer is that we “define” the situation as such - we declare the result of the measurement is that [if]{} we see a macroscopic signal at slit $1$, it implies that the electron indeed went through slit $1$. Our observation that the interference pattern on the screen gets replaced by a “clump” pattern confirms that this assignment is right. To put it pithily, we have learnt that when we see a car headed towards us, it is useful to step out of the way. In the density-matrix formalism, this just means that we have zero off-diagonal elements post-measurement[@Balian2]. In our approach below, we choose independent vectors in Hilbert space corresponding to what we define as the inference of a measurement. In addition, our choice of basis vectors in Hilbert space is determined by how we interpret the results of our measurements - our theory decides what we have observed. We write the hermitian Hamiltonian as $$\begin{aligned} {\cal H} = {\cal H}_{electron} + {\cal H}_{photon} + {\cal H}_{int}^{(1)} + {\cal H}_{int}^{(2)}+ {\cal H}_{int}^{(3)}\end{aligned}$$ where each piece is explained below. First, we start with the Hamiltonian for the “free” electrons and photons. $$\begin{aligned} {\cal H}_{electron} = \omega_e \left( c_1^{\dagger} c_1 + c_2^{\dagger} c_2 \right) \nonumber \\ {\cal H}_{photon} = \omega_p \gamma_0^{\dagger} \gamma_0 + \sum_{i=1}^{N} \omega_i \gamma_i^{\dagger} \gamma_i \end{aligned}$$ Then we add interactions between the electron and the first locator photon (we assume the interaction parameter $\Gamma_1$ is real in this example). This interaction is modeled as between the electron at slit \# $1$ and the photon in mode $0$. In order to ensure energy conservation, energy needs to be supplied through the interaction. One may consider the interaction parameter as an auxiliary field, with a time dependence, which allows the photon to be produced in the presence of the electron at slit $1$ and emerge with energy $\omega_p$. $$\begin{aligned} {\cal H}_{int}^{(1)} = \Gamma_1 e^{-i \omega_p t} \gamma_0^{\dagger} c_1^{\dagger} c_1 + \Gamma_1 e^{i \omega_p t} \gamma_0 c_1^{\dagger} c_1\end{aligned}$$ Next, we add the interactions that lead to the creation of child photons (in any of $R$ photon states) from this first locator photon at the photomultiplier tube ($N$ is a multiplication fraction, usually $2-3$). Again, energy is inserted into the system through the interaction vertex, the first “enhancement” being modeled by another auxiliary field with time-dependence. $$\begin{aligned} {\cal H}_{int}^{(2)} =\sum_{i=1}^R \bigg( \Gamma_2^i e^{-i N \omega_i t + i \omega_p t} (\gamma_i^{\dagger})^N \gamma_0 + \Gamma_2^i e^{i N \omega_i t - i \omega_p t}\gamma_0^{\dagger} \gamma_i^N \bigg)\end{aligned}$$ Next, these photons create other photons in the available states, with similar multiplication fractions, again using the external measurement system to add energy. $$\begin{aligned} {\cal H}_{int}^{(3)} = \sum_{i=1, j=1; j>j; i \ne j}^R \bigg( s \: e^{i \omega_i t-i M_1 \omega_j t} (\gamma_j^{\dagger})^{M_1} \gamma_i + s \: e^{-i \omega_i t+i M_1 \omega_j t} \gamma_i^{\dagger} (\gamma_j)^{M_1} \bigg) \end{aligned}$$ This Hamiltonian is explicitly time-dependent and energy is inserted from the outside through the interaction. The scale of the couplings and photo-multiplication is, $\Gamma_2 \approx \Gamma_1 \approx s$, while $N, M_1$ are of ${\mathcal O}(1-10)$. The number of modes $R$ of photons that are subsequently produced, is macroscopic and large. As one can see, once a photon interacts with the electron, it subsequently produces more photons and the process cascades exponentially within the states available to the photons - subsequent processes distribute this energy into other states in the continuum. We will now write down some key states in the Fock space of the system. First, we represent the state of the unperturbed system as $$\begin{aligned} \ket{\psi}_0 = \frac{\ket{1}_e + \ket{2}_e}{\sqrt{2}} \: \ket{n_0=0, n_1=0, n_2=0, ... , n_Q = 0}\end{aligned}$$ The first $0$ in the ending ket represents that there are no photons scattered at slit $1$, hence the electron is still in a superposition of two states (either at slit $1$ or slit $2$). The second list of $0$’s implies that there are no photons in the photomultiplier tube either. The various perturbed states are as follows. In all these states, if a scattering event has occurred relative to slit $1$, the electron is now firmly in state $\ket{1}_e$. In fact, the state $\ket{2}_e \ket{n_0=1; n_1, n_2, ..., n_R}$ is not part of the possible Fock space of the system - physically, we cannot have a scattering event at slit $1$ while the electron is at slit $2$! We could, of course, expect noise, but that can be measured prior to the experiment and appropriately adjusted for. The first of the perturbed states is where the initial photon is scattered, while the others are where subsequent photomultiplier events have occured with the further production of other photons. These states are $$\begin{aligned} \ket{\psi}_1 \equiv \ket{\psi}_{1,1;0...0} = \ket{1}_e \ket{n_0=1; n_1=0, n_2 = 0, ..., n_R=0} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ \ket{\psi}_2^i \equiv \ket{\psi}_{1,0;0...N...0} = \ket{1}_e \ket{n_0=0; n_1=0, n_2 = 0, .,n_i=N,.., n_R=0} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ . \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ . \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ . \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ \ket{\psi}_3^{ij} \equiv \ket{\psi}_{1,0;0,..., n_i=N-1,...n_j=M_1,...,0} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\nonumber \\ = \ket{1}_e \ket{n_0=0; n_1=0, ..., n_i=N-1,...,n_j=M_1,...,n_R=0} \nonumber \\ \ket{\psi}_4^{ijk} \equiv \ket{\psi}_{1,0;0..., n_i=N-2,...n_j=M_1,...,n_k=M_1,...,0} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ = \ket{1}_e \ket{n_0=0; n_1=0, ..., n_i=N-2,...,n_j=M_1,...,n_k=M_1,...,n_R=0} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\end{aligned}$$ there are of course many more possible end states and we can construct the series for the system’s state $\ket{\psi}$ with more terms, as in Equation (8) below. We assume, in what follows that the states are labelled by $i$ and there are a macroscopic number $R$ of them. Chosen this way, these states are themselves orthogonal to each other and since they are in the occupation number representation, can be enumerated and are complete. In general, we can write the state of the system at any time as $$\begin{aligned} \ket{\psi} = a e^{-i \omega_e t} \ket{\psi}_0 + b e^{-i (\omega_e + \omega_p) t} \ket{\psi}_1 + \sum_{i = 1}^R c_i e^{-i (\omega_e + N \omega_i )t} \ket{\psi}_2^i \nonumber \\ + \sum_{ij; i<>j} d_{ij} \ket{\psi}_3^{ij} e^{-i(\omega_e + (N-1) \omega_i+M_1 \omega_j) t} \nonumber \\ + \sum_{ijk; i<>j<>k} d_{ijk} \ket{\psi}_4^{ijk} e^{-i(\omega_e + (N-2) \omega_i+M_1 \omega_j+ M_1 \omega_j t ) t} + ...\end{aligned}$$ where we have explicitly included the “free” time-dependence of the states in the exponentials multiplying every term. We can now write down Schrödinger’s equation for this general state, remembering that the initial state is described by $a=1, b=0, c_i=0 \: \forall i, d_{\vec n}=0 \: \forall {\vec n}$. $$\begin{aligned} i \frac{\partial \ket{\psi}}{\partial t} = {\cal H} \ket{\psi} \end{aligned}$$ We successively pre-multiply the Schrödinger’s equation as follows with the bra-vectors corresponding to the below kets, 1. $\ket{\psi}_0 e^{- i \omega_e t} = \frac{\ket{1}_e + \ket{2}_e}{\sqrt{2}} \: \ket{n_0=0; n_1=0, n_2=0, ... , n_R = 0} e^{- i \omega_e t}$ 2. $\ket{\psi}_1 e^{- i \omega_e t - i \omega_p t} = \ket{1}_e \ket{n_0=1; n_1=0, n_2 = 0, ..., n_R=0} e^{- i \omega_e t - i \omega_p t}$ 3. $\ket{\psi}_2^i e^{- i \omega_e t - i N \omega_i t} =\ket{1}_e \ket{n_0=0; n_1=0, n_2 = 0, .,n_i=N,.., n_R=0} e^{- i \omega_e t - i N \omega_i t}$ 4. $\ket{\psi}_3^{ij} e^{- i \omega_e t - i (N-1) \omega_i t-M_1 \omega_j t} \\ =\ket{1}_e \ket{0; n_1= 0, ..., n_i=N-1,..., n_j=M_1, ...,n_R=0} e^{- i \omega_e t - i (N-1) \omega_i t-i M_1 \omega_j t}$ We can continue this process for all the successive parameters. After some algebra, we are left with $$\begin{aligned} i \frac{\partial a}{\partial t} = \frac{\Gamma_1}{\sqrt{2}} \: b \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ i \frac{\partial b}{\partial t} = \frac{\Gamma_1}{\sqrt{2}} \: a + \sum_{i=1}^R \Gamma_2^{(i)} \sqrt{N!} c_i \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \: \:\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ i \frac{\partial c_i}{\partial t} = \Gamma_2^{(i)} \sqrt{N!} \: b + \sum_{j=1; j \ne i}^{R} s \sqrt{N} \sqrt{M!} \: d_{ij} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ i \frac{\partial d_{ij}}{\partial t} = s \sqrt{N} \sqrt{M_1!} \: c_i + \sum_{k <>i,j} s \sqrt{N-1} \sqrt{M_1!} \: d_{ijk}\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \end{aligned}$$ In particular, generalizing, we could write the last three equations above in a convenient matrix form as $$\begin{aligned} i \frac{\partial b}{\partial t} = \frac{\Gamma_1}{\sqrt{2}} \: a \: + \sqrt{N!} \: { \bar \Gamma_2} .{\bar c} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ i \frac{\partial \bar c}{\partial t} = {\bar \Gamma_2} \sqrt{N!} \: b \: + s \sqrt{N}\: \sqrt{M_1!} \: {\bar { \bar {\cal B}}} .{\bar d_2} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ i \frac{\partial \bar d_2}{\partial t} = s \sqrt{N} \: \sqrt{M_1!} \:{\bar {\bar {\cal B}}} . \: \bar c + \sqrt{N-1} \sqrt{M_1!} \:{\bar {\bar {\cal E}}} . {\bar d_3} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\end{aligned}$$ where ${\bar { \bar {\cal B}}}$ is a symmetric matrix, required by the unitarity conditions upon the time evolution. We have used the notation $\bar c = (c_1, c_2, ..., c_R)$ and $\bar d_2 = (d_{\vec n})$, etc. Note the structure of the equations; the coefficient of the term on the right side for every pair of parameters is the same. The structure of the above is $$\begin{aligned} i\frac{\partial}{\partial t} \left(\begin{array}{c} a \\ b \\ \bar c \\ \bar d \end{array} \right) = {\cal H}_{total} \left(\begin{array}{c} a \\ b \\ \bar c \\ \bar d \end{array} \right) \nonumber\end{aligned}$$ where ${\cal H}_{total}$ is the total Hamiltonian, a hermitian matrix. To simplify this set of equations, we operate on all of them with $i \frac{\partial}{\partial t}$ to yield the matrix equation $$\begin{aligned} - \frac{\partial^2}{\partial t^2} \left(\begin{array}{c} a \\ b \\ \bar c \\ \bar d \end{array} \right) = {\cal H}_{total}^2 \left(\begin{array}{c} a \\ b \\ \bar c \\ \bar d \end{array} \right) \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ = \left(\begin{array}{cccc} {\bf \frac{\Gamma_1^2}{2}} & \: \: \: 0 &\: \: \: \frac{\Gamma_1 {\bar \Gamma_2}}{\sqrt{2}} \sqrt{N!} &\: \: \: 0 \\ 0 & \: \: \: {\bf \frac{\Gamma_1^2}{2}+ N! \: {\bar \Gamma_2}.{\bar \Gamma_2} }& \: \: \: 0 & \: \: \: s \sqrt{N M!} \: {\bar {\bar {\cal B}}} \\ \frac{\Gamma_1 {\bar \Gamma_2}}{\sqrt{2}} \sqrt{N!} & \: \: \: 0 & \: \: \: {\bf N! {\bar \Gamma_2}{\bar \Gamma_2}+ s^2 (N M!) {\bar {\cal B}}.{\bar {\cal B}}} &\: \: \: 0 \\ 0 & \: \: \: s \sqrt{N M!} \: {\bar {\bar {\cal B}}} & \: \: \: 0 &\: \: \: {\bf s^2 (N M!) {\bar {\cal B}}.{\bar {\cal B}} } \end{array}\right) \left(\begin{array}{c} a \\ b \\ \bar c \\ \bar d \end{array} \right) \end{aligned}$$ The matrix ${\cal H}_{total}^2$ is a symmetric, explicitly time-independent matrix. Indeed, due to the physical separation of the successive interactions at parts of the photomultiplier, the matrix is structurally one-skip-tridiagonal and symmetric. Indeed, we can study the properties of a simple version of this set of equations, consistent with the symmetric nature of ${\bar {\bar {\cal B}}}$ etc., with new variable definitions, as $$\begin{aligned} - \frac{\partial^2}{\partial t^2} \left(\begin{array}{c} \alpha_1 \\ \alpha_2 \\ \alpha_3 \\ \alpha_4 \\ . \\ . \\ . \\ \alpha_i \\ . \\ . \\ . \\ . \\ \alpha_Q \end{array} \right) = \left(\begin{array}{cccccccccc} a_1^2 & \: \: \: 0 &\: \: \: a_1 a_2 & \: \: \: 0 \: \: \: & \: \: \: 0 \: \: \: & ... & ... & ... &... &\: \: \: 0 \: \: \: \\ 0 & \: \: \: a_1^2+a_2^2 & \: \: \: 0 & \: \: \: a_2 a_3 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ a_1 a_2 & \: \: \: 0 & \: \: \: a_2^2+a_3^2 & \: \: \: 0 \: \: \: &a_3 a_4 & ... & ... &... &... &0 \\ 0 & \: \: \: a_2 a_3 & \: \: \: 0 & \: \: \: a_3^2+a_4^2 \: \: \: &\: \: \: 0 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: a_3 a_4 & \: \: \: 0 \: \: \: &\: \: \: a_4^2+a_5^2 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: a_4 a_5 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: a_5 a_6 & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &a_{Q-2} a_{Q-1} \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: 0 & ... & ... & ... & ... &a_{Q-1}^2 \end{array}\right) \left(\begin{array}{c} \alpha_1 \\ \alpha_2 \\ \alpha_3 \\ \alpha_4 \\ . \\ . \\ . \\ \alpha_i \\ . \\ . \\ . \\ . \\ \alpha_Q \end{array} \right) \nonumber\end{aligned}$$ In the above matrix, we have, for instance, collapsed the set of components ${\underbar c}$ into one component $\alpha_3$. Hence $a_2 \propto R$ and is, among other things, a macroscopic multiple of $\Gamma_2$ etc. The same goes for $\alpha_5$ etc. We use $\alpha_1$ to represent $a$ in the calculation, so its initial value (at $t=0$) is $1$, while the other parameters and their first derivatives w.r.t time are $0$ at $t=0$. In addition, $Q$ is all the possible kinds of states, in fact $Q \sim e^R$ since there are an exponential number of ways of partitioning cascaded photons amongst $R$ states. The one-skip-tridiagonal structure of the matrix has some interesting properties, which are very similar to the tri-diagonal structure seen in typical coupled oscillator problems. The eigenvalues of this matrix are positive semi-definite (see Appendix 1). This means that the time evolution is purely unitary and no dissipation emerges from the mathematics of the problem. In addition, from an inspection of the equations we have obtained, the $a_i$’s increase by factors of roughly $R \sqrt{N}$ as $i$ increases. For simplicity, we consider $a_i$ chosen in a fairly simple manner to study the time evolution of a state that starts with $\alpha_1|_{t=0}=1, \frac{\partial \alpha_1}{\partial t}|_{t=0}=0$, i.e., the electron in a superposed state at both the slits. Accordingly, in Figures 2-5, we display the results of simulating the above equations for a simple form of the frequency matrix ${\cal G}$ chosen with $a_1=1, a_i=\sqrt{10} \: \: \forall \: i \ne 1$. However, when we study the time-dependence of $\alpha_1$, we will need to use a more realistic assumption for the $a_i$, namely, $a_1=a; a_i = a \:( R \: \sqrt{N})^{i-1} \: \forall i>1$. For clarity, $R$ is the macroscopic number of possible modes available at the first step (inside the photomultiplier) and $N$ is the relatively small (but greater than 1) number of photons produced at each stage of the photomultiplier cascade. We can add small random terms to the elements of the matrix in a manner consistent with leaving it real, symmetric and of the form we are studying. This does not amend the results we discuss below. In addition, use may be made of Gershgorin’s [@Gersh] theorem to understand why the eigenvalues are confined (with the numbers in our example) to the range $(0,..,40)$. For the purposes of graphing results, we have used $$\begin{aligned} {\cal G} = \left(\begin{array}{cccccccccc} 1 & \: \: \: 0 &\: \: \: \sqrt{10} & \: \: \: 0 \: \: \: & \: \: \: 0 \: \: \: & ... & ... & ... &... &\: \: \: 0 \: \: \: \\ 0 & \: \: \: 11 & \: \: \: 0 & \: \: \: 10 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ \sqrt{10} & \: \: \: 0 & \: \: \: 20 & \: \: \: 0 \: \: \: &10 & ... & ... &... &... &0 \\ 0 & \: \: \: 10 & \: \: \: 0 & \: \: \: 20 \: \: \: &\: \: \: 0 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 10 & \: \: \: 0 \: \: \: &\: \: \: 20 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 10 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: 10 & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &10 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: 0 & ... & ... & 10 & 0 &10 \end{array}\right) \end{aligned}$$     Obtaining decaying behavior from a collection of harmonic oscillators ===================================================================== To understand why we see behavior exhibited in Fig. 4 and Fig. 5, we analyze the problem as follows: the relevant eigenvalues and eigenvectors are those of the matrix $\cal G$. The normalized eigenvectors ($Q$ of them) of $\cal G$ are written as $\hat e_1, \hat e_2, ..., \hat e_Q$. We define the following matrices, in the original state vectors of the system as $$\begin{aligned} \hat {\cal V}^T = \left(\begin{array}{cccccc} ( . & . & . & \hat e_1 & . & . )\\ ( . & . & . & \hat e_2 & . & . ) \\ ( . & . & . & ... & . & . ) \\ ( . & . & . & ... & . & . ) \\ ( . & . & . & \hat e_Q & . & . ) \end{array}\right) \nonumber \\ \hat {\cal V} = \left(\begin{array}{cccccc} \Bigg( \hat e_1 \Bigg) & \Bigg( \hat e_2 \Bigg) & ... & . & . & \Bigg( \hat e_Q \Bigg) \end{array}\right) \nonumber \\ \hat {\cal V^T} \hat {\cal V} = {\cal I} \rightarrow \hat {\cal V}^T = \hat {\cal V}^{-1} \end{aligned}$$ Also, define the weight vector $w$ for the eigenstates to produce the initial system state, as well as the initial system state ${\cal B}(t=0)$ as $$\begin{aligned} w = \left(\begin{array}{c} w_1\\ w_2\\ .\\ .\\ w_Q \end{array}\right) \: \: \: \: \: {\cal B}(t=0) = \left(\begin{array}{c} 1\\ 0\\ .\\ .\\ 0 \end{array}\right) \end{aligned}$$ We note that since $\cal B$ represents the initial state of the system, we must have $$\begin{aligned} {\cal V} . w = {\cal B}(t=0) \end{aligned}$$ This yields, formally, the solution to the initial weights as $w = {\cal V}^T . {\cal B}(t=0)$. This implies, upon inspection, that $$\begin{aligned} w_i = {\hat e}_i^{(1)} \end{aligned}$$ where we have used the notation ${\hat e}_i^{(1)}$ for the first element of eigenvector ${\hat e}_i$. In this notation, we have, for the time-evolution of $\alpha_1(t)$, the amplitude of the initial state, $$\begin{aligned} \alpha_1(t) = \sum_{i=1}^Q {\hat e}_i^{(1)} {\hat e}_i^{(1)} \cos {\Omega_i t} = \sum_{i=1}^Q w_i^2 \cos {\Omega_i t} \end{aligned}$$ This is obtained from the general solution to the state of the system ${\cal B}(t)$, which is, $$\begin{aligned} {\cal B}(t) = e^{i\sqrt{{\cal G}} t } {\cal V} w = e^{i\sqrt{{\cal G}} t } {\cal B}(t=0) \end{aligned}$$ and applying the initial condition $\alpha_1(t=0)=1$. Let’s now study the dependence of the weights $w_i$ upon the eigenvalues $\Omega_i^2$ in this simplified model. In particular, we want the large $\Omega_i$ dependence of $w_i^2$. We begin with the eigenvalue equation for $\cal G$, i.e., ${\cal G} {\hat e}_i = \Omega_i^2 {\hat e}_i $. Writing down each term and noting that $w_i={\hat e}_i^{(1)}$, the first component of the normalized eigenvector, we get $$\begin{aligned} a_1^2 {\hat e}_i^{(1)} + a_1 a_2 {\hat e}_i^{(3)} = \Omega_i^2 {\hat e}_i^{(1)} \nonumber \\ a_1 a_2 {\hat e}_i^{(1)} + (a_2^2+a_3^2) {\hat e}_i^{(3)} + a_3 a_4 {\hat e}_i^{(5)}= \Omega_i^2 {\hat e}_i^{(3)} \nonumber \\ ... \end{aligned}$$ which can be solved generally as $$\begin{aligned} {\hat e}_i^{(1)} = {\hat e}_i^{(2n+1)} \frac{a_1 a_2...a_{2n}}{\Omega_i^{2n} - \Omega_i^{2n-2} (a_1^2+a_2^2+...+a_{2n-1}^2)+...+ (-1)^n a_1^2 a_3^2...a_{2n-1}^2} \nonumber \end{aligned}$$ Note that we can write down a similar condition for ${\hat e}_i^{(2)}$. There are two potential eigenvectors that solve the problem for one $\Omega_i$ - one series with ${\hat e}_i^{(1)} =0$ (and all subsequent odd-index eigenvector components equal to 0) and the other with ${\hat e}_i^{(2)}=0$ (and so for the subsequent even-index weights). The solution with ${\hat e}_i^{(1)} =0$ has zero weight and so doesn’t contribute to the value of $\alpha_1(t)$ at [*any*]{} time $t$ and can be ignored. We therefore need to only keep the odd-numbered components of the eigenvectors and the only eigenvectors that matter are the ones where the odd-numbered weights are non-zero. In what follows, we use the more general form, i.e., $a_1=a, \: a_i = a\: R N^{\frac{i-1}{2}} \: \forall i>1$. For the largest eigenvalues, the formula for the eigenvector components can be approximated by $$\begin{aligned} {\hat e}_i^{(2n+1)} \approx {\hat e}_i^{(1)} \frac{\Omega_i^{2n}}{a_1 a_2...a_{2n}} \end{aligned}$$ Hence, the normalization condition for the eigenvector ${\hat e}_i$ is $$\begin{aligned} ({\hat e}_i^{(1)})^2 \bigg( 1 + (\frac{\Omega_i^2}{a_1 a_2})^2 + (\frac{\Omega_i^4}{a_1 a_2 a_3 a_4})^2 + ...\bigg) = 1 \end{aligned}$$ Hence, approximating for the largest eigenvalues as $Q\rightarrow \infty$, $$\begin{aligned} ({\hat e}_i^{(1)})^2 \equiv w_i^2 \approx e^{- \frac{\Omega_i^4}{a_1^2 a_2^2}} = e^{- \frac{\Omega_i^4}{R^2 \: N a^4}} \end{aligned}$$ For small eigenvalues, the opposite limit can be applied, namely $$\begin{aligned} {\hat e}_i^{(2n+1)} \approx {\hat e}_i^{(1)} \frac{(-1)^{n-1} \Omega_i^2 (\sum_m^{`} A_m)+ (-1)^n a_1^2 a_3^2...a_{2n-1}^2}{a_1 a_2...a_{2n}} \nonumber \end{aligned}$$ where $A_m$ has $(n-1)$ multiples of coefficients. Specifically, for the first four components, we note the pattern $$\begin{aligned} {\hat e}_i^{3} = {\hat e}_i^{1} \: \frac{\Omega_i^2 - a_1^2}{a_1 a_2} = {\hat e}_i^{1} \: \frac{\Omega_i^2 - a^2}{a^2 R \sqrt{N}}\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ {\hat e}_i^{5} = - {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a_1^2a_3^2}{a_1^2+a_2^2+a_3^2}}{\frac{a_1 a_2 a_3 a_4}{a_1^2+a_2^2+a_3^2}} \approx - {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a^2}{3}}{\frac{R^2 N \: a^2}{3}} \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \nonumber \\ {\hat e}_i^{7} = {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a_1^2a_3^2 a_5^2}{a_1^2a_3^2+a_1^2 a_4^2+a_1^2 a_5^2+a_2^2a_4^2+a_2^2a_5^2+a_3^2a_5^2}}{\frac{a_1 a_2 a_3 a_4 a_5 a_6}{a_1^2a_3^2+a_1^2 a_4^2+a_1^2 a_5^2+a_2^2a_4^2+a_2^2a_5^2+a_3^2a_5^2}} \approx {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a^2}{6}}{ \frac{R^3 N^{3/2}a^2}{6}} \: \: \: \:\nonumber \\ {\hat e}_i^{9} = - {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a_1^2a_3^2 a_5^2 a_7^2}{a_1^2a_3^2(a_5^2+a_6^2+a_7^2) +(a_1^2 +a_2^2+a_3^2) a_5^2 a_7^2+a_1^2 a_4^2 (a_6^2+a_7^2) + a_2^2 a_4*2 (a_6^2+a_7^2)}}{\frac{a_1 a_2 a_3 a_4 a_5 a_6 a_7 a_8}{a_1^2a_3^2(a_5^2+a_6^2+a_7^2) +(a_1^2 +a_2^2+a_3^2) a_5^2 a_7^2+a_1^2 a_4^2 (a_6^2+a_7^2) + a_2^2 a_4*2 (a_6^2+a_7^2)}} \approx - {\hat e}_i^{1} \: \frac{\Omega_i^2 - \frac{a^2}{16}}{\frac{R^4 N^2 a^2}{16}} \end{aligned}$$ Note the $R\sqrt{N}$ factors that grow in the denominator and we have replaced terms in the denominator by the largest (in each case) in our approximations for ${\hat e}_i^{5}, {\hat e}_i^{7}, ...$. Again, using the normalization condition for the eigenvector ${\hat e}_i$, the small eigenvalue expansion leads to $$\begin{aligned} ({\hat e}_i^{(1)})^2 \bigg( 1 + \frac{1}{R^2N} \big(\frac{\Omega_i^2-a^2}{a^2 } \big)^2 + \frac{1}{R^4N^2} \big(\frac{\Omega_i^2-\frac{a^2}{3}}{\frac{a^2}{3}} \big)^2 + \frac{1}{R^6N^3} \big(\frac{\Omega_i^2-\frac{a^2}{6}}{\frac{a^2}{6}} \big)^2 + \frac{1}{R^8N^4} \big(\frac{\Omega_i^2-\frac{a^2}{16}}{\frac{ a^2}{16}} \big)^2 +..\bigg)=1 \nonumber \\ \rightarrow ({\hat e}_i^{(1)})^2 \equiv w_i^2 \approx \frac{1}{\bigg(1 + \frac{1}{R^2N} (\frac{ \Omega_i^2}{a^2} -1)^2\bigg) } \nonumber \end{aligned}$$ where we have kept only the first term in the set of $Q$ terms in the denominator. Again, to make some numerical estimates, we plot the squared weights and eigenvalues for the various finite-sized systems for our simple choice of the $\cal G$ matrix in Eqn 13, in Figures 6-13.         The time-dependence of $\alpha_1(t)$ is written finally as with the functional forms we have derived in the previous section. We write, in the limit of an infinite number of oscillators and using the chain rule to write the sum over each of the eigenvalues as an integral over the actual eigenvalues, $$\begin{aligned} \alpha_1(t) = \sum_{i=1}^Q w_i^2 \cos {\Omega_i t} \approx \int_0^{\infty} d \Omega \: {\cal D}(\Omega) \: \: w(\Omega)^2 \cos {\Omega t} \end{aligned}$$ From an inspection of the density of eigenvalues, we note that it is approximately constant (the slope of the $i$-vs-$\Omega_i$ curves), hence we can use ${\cal D}(\Omega) = {\cal C}$. For short times, we use the large eigenvalue limit of the weight/eigenvalue dependence, i.e., $$\begin{aligned} \alpha_1(t)|_{ST} \approx {\cal C} \int_0^{\infty} d \Omega \: \: e^{- \frac{\Omega^4}{R^2 N a^4} } \cos {\Omega t} \end{aligned}$$ where ${\cal C}$ must be set such that we recover $ \alpha_1(t=0)|_{ST} =1$ from the integral approximation. The time dependence is therefore, essentially, the Fourier transform of the squared weights as a function of the eigenvalues. To estimate the behavior of this integral, we note the two limits [@Boyd], namely, $$\begin{aligned} \int_0^{\infty} dx \: e^{-x^4} \cos{(xt)} |_{large \: t} = 2^{1/6} \sqrt{\frac{\pi}{3}}\: \frac{1}{t^{\frac{1}{3}}} e^{- \frac{2^{1/3} 3}{16} t^{4/3}} \cos{ \bigg( 3^{\frac{3}{2}} \frac{2^{1/3}}{16} t^{4/3} - \frac{\pi}{6} \bigg) } \nonumber \\ \int_0^{\infty} dx \: e^{-x^4} \cos{(xt)} |_{small \: t} = 0.906402 - 0.153177 t^2 \approx 0.906402 e^{- 0.168995 t^2} \end{aligned}$$ Then we rescale the integral in Equation (24) (using $B=\frac{1}{R^2 \:N a^4}$ and also ${\cal C} = \frac{B^{1/4}}{0.906402}$) $$\begin{aligned} \int_0^{\infty} dx \: e^{-B x^4} \cos{(xt)} = \frac{1}{B^{1/4}} \int_0^{\infty} dy \: e^{-y^4} \cos{(y \frac{t}{B^{1/4}})} \end{aligned}$$ which means the short-time dependence of $\alpha_1$ is $$\begin{aligned} \alpha_1(t)|_{ST} \approx e^{- 0.168995 \: R \: \sqrt{N} a^2 t^2} \end{aligned}$$ i.e., the parameter $\alpha_1$, the amplitude of the “unmeasured” state, decays with time like a Gaussian, with time scale $\tau^2_{ST}=\frac{1}{R \sqrt{N} a^2}$ in natural units. The dependence upon $R$ (the macroscopic number of possible end-states) and $a^2$ is identical to the results in [@Balian2], with a very different physical realization. To reiterate, the decay of $\alpha_1$ to $0$ represents the “collapse” of the superposed state into the other states where the electron is localized to slit $1$. For the long-time limit, we use the small eigenvalue limit for the squared weights to obtain (with $A = \frac{1}{R \sqrt{N}}$) and also take the upper limit of the rapidly convergent integral off to infinity, to get $$\begin{aligned} \alpha_1(t)|_{LT} \approx \int_0^{\infty} d \Omega \: \: \frac{{\cal D}(\Omega)}{1+A^2 (\frac{\Omega^2}{a^2}-1)^2} \cos {\Omega t} \approx {\cal C} \int_0^{\infty} d \Omega \: \: \frac{1}{1+A^2 (\frac{\Omega^2}{a^2}-1)^2} \cos {\Omega t} \end{aligned}$$ We use the relation (for small $A$), $$\begin{aligned} \int_0^{\infty} \frac{dx}{1 + A^2 (x^2-1)^2} \cos{x t} = \frac{ \pi e^{- \frac{t}{\sqrt{2 A}} }}{2 \sqrt{A}} \cos{ \left( \frac{t}{\sqrt{2A}}-\frac{\pi}{4} \right) } \end{aligned}$$ so the envelope of the graph of the parameter $\alpha_1$ is $$\begin{aligned} | \alpha_1(t)|_{LT}| \approx {\cal C} \pi \frac{e^{- \frac{t}{\sqrt{2A}} }}{2 \sqrt{A}} \end{aligned}$$ and obtain (upon suitably substituting for $A$) $$\begin{aligned} | \alpha_1(t)|_{LT}| \sim e^{- \sqrt{R \sqrt{N}} \: a \: t} \end{aligned}$$ The parameter $\alpha_1(t)$ falls off exponentially in the large $t$ limit. In addition, the collapse occurs over the time-scale $\tau_{LT}= \frac{1}{\sqrt{R \sqrt{N}}}$; this is unlike the situation in [@Balian2], undoubtedly owing to the difference in the process of measurement and the manner (in that calculation) of the connection of the phonon heat bath to the magnetic system. Here, in both cases, the scale of the time dependence includes the factor $R$, which is the macroscopic number of states accessible to the photons in the photo-multiplier. In the limit $R \rightarrow \infty$, we get instantaneous “collapse” of the amplitude to 0. Contrary to the analysis in [@Balian1], we make no distinction between the “registration” of the “collapsed” wave-function and the “collapse”. In our case, the “collapse” is simultaneous with “registration” - a macroscopic number of photons is produced in a continuum of states with corresponding localization of the electron to slit $1$. The analysis in [@Balian1] also relies on using a system in a metastable state near a critical point to make a measurement. The practical problem with this is that such a system would settle into one or the other stable state (all up or all down spin) even with statistical fluctuation. The interaction hence needs to be proportional to $R$ right away. That is not our approach, as we see the number of modes appear in a rather natural way as the number of accessible channels for cascade photons. Recurrence times ================ In the general case, where the elements of $\cal G$ are random, with the constraints that it is a symmetric matrix of the form in Eqn.12, the eigenvalues are likely distributed with the Wigner semi-circle distribution [@Wigner1] and the smallest separation is $\sim \frac{1}{R}$. The recurrence time for such a collection of coupled harmonic oscillators thus goes to $\infty$ and it is extremely unlikely that a collapsed state will revert to an un-collapsed form, as we have seen above. Connections to Mott’s thought experiment ======================================== Mott [@Mott] considered the theoretical explanation of why $\alpha$-particle decay into an $s$-wave state leads to tracks in a cloud chamber, rather than a spherically symmetric fuzz around the decay site. Employing a simple argument, he showed that the various possible exit states of the $\alpha$-particle should be straight rays emanating from the decay site. In our language, each such ray would be equivalent to a “detection at a slit”. From the above argument, there is an amplitude that any one of these rays is picked due to the interaction parameter with each of the macroscopic number of atoms in the cloud chamber. Once one of the interactions reaches and affects a macroscopic number of atoms, the mathematics of Schrodinger’s equation (and the density matrix) produces a collapse into one of the available ray-states, depending upon which of the interactions succeeds (with the cloud-chamber atoms), without any additional assumptions required. Conclusions =========== We have studied the measurement, using a macroscopic apparatus, of the double-slit experiment and deduced, from Schrodinger’s equation, that the wave-function collapse upon measurement is a straightforward result of the interaction Hamiltonian. The time scales for the short-term and long-term collapse are $\tau_{ST}^2=\frac{1}{R \sqrt{N} a^2}$ and $\tau_{LT}= \frac{1}{\sqrt{R \sqrt{N}}}$ respectively. In addition, the short-term collapse follows a Gaussian, while the long-term behavior is exponential. This approach therefore connects the quadratic time dependence of the transition probability from Fermi’s golden rule to the exponential behavior seen in radioactive decay, for instance. The calculation thus produces results parallel to [@Balian2], except that the “registration” process is much quicker here (related to $\tau_{LT} \sim \frac{1}{\sqrt{R}}$) and doesn’t behave as in the phonon bath in that reference. Some useful connections are made to the Mott experiment. Acknowledgments =============== Useful discussions with Scott Thomas, Dan Friedan and Thomas Banks are acknowledged and especially to the latter for pointing out the similarities to Allahverdyan [*et al*]{} [@Balian1; @Balian2; @Balian3]. The hospitality of the NHETC and the Department of Astronomy and Physics at Rutgers University are sincerely acknowledged. Appendix 1 ========== A matrix ${\cal G}$ is positive semi-definite if $x^T {\cal G} x \ge 0$ for any vector $x$. Consider the matrix defined as $$\begin{aligned} {\cal G} = \left(\begin{array}{cccccccccc} a_1^2 & \: \: \: 0 &\: \: \: a_1 a_2 & \: \: \: 0 \: \: \: & \: \: \: 0 \: \: \: & ... & ... & ... &... &\: \: \: 0 \: \: \: \\ 0 & \: \: \: a_1^2+a_2^2 & \: \: \: 0 & \: \: \: a_2 a_3 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ a_1 a_2 & \: \: \: 0 & \: \: \: a_2^2+a_3^2 & \: \: \: 0 \: \: \: &a_3 a_4 & ... & ... &... &... &0 \\ 0 & \: \: \: a_2 a_3 & \: \: \: 0 & \: \: \: a_3^2+a_4^2 \: \: \: &\: \: \: 0 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: a_3 a_4 & \: \: \: 0 \: \: \: &\: \: \: a_4^2+a_5^2 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: a_4 a_5 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: a_5 a_6 & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &a_{Q-2} a_{Q-1} \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: 0 & ... & ... & ... & ... &a_{Q-1}^2 \end{array}\right)\end{aligned}$$ For an arbitrary vector $x$, with components $x_1, x_2, ..., x_Q$ and the definitions $a_{-1}=a_{0}=a_{Q}=a_{Q+1}=0$ and $x_{-1}=x_0=x_{Q+1}=x_{Q+2}=0$, we write the scalar as $$\begin{aligned} \left(\begin{array}{cccc} x_1 & x_2 &\: \: \: . & x_Q \end{array}\right) \left(\begin{array}{cccccccccc} a_0^2+a_1^2 & \: \: \: 0 &\: \: \: a_1 a_2 & \: \: \: 0 \: \: \: & \: \: \: 0 \: \: \: & ... & ... & ... &... &\: \: \: 0 \: \: \: \\ 0 & \: \: \: a_1^2+a_2^2 & \: \: \: 0 & \: \: \: a_2 a_3 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ a_1 a_2 & \: \: \: 0 & \: \: \: a_2^2+a_3^2 & \: \: \: 0 \: \: \: &a_3 a_4 & ... & ... &... &... &0 \\ 0 & \: \: \: a_2 a_3 & \: \: \: 0 & \: \: \: a_3^2+a_4^2 \: \: \: &\: \: \: 0 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: a_3 a_4 & \: \: \: 0 \: \: \: &\: \: \: a_4^2+a_5^2 & ... & ... &... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: a_4 a_5 \: \: \: &\: \: \: 0 & ... & ... & ... &... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: a_5 a_6 & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &a_{Q-2} a_{Q-1} \\ 0 & \: \: \: ... & \: \: \: ... & \: \: \: ... \: \: \: &\: \: \: ... & ... & ... & ... & ... &0 \\ 0 & \: \: \: 0 & \: \: \: 0 & \: \: \: 0 \: \: \: &\: \: \: 0 & ... & ... & ... & ... &a_{Q-1}^2 + a_Q^2 \end{array}\right) \left(\begin{array}{c} x_1 \\ x_2 \\ .\\ .\\ .\\ .\\ .\\ .\\ .\\ .\\ .\\ .\\ x_Q \end{array}\right) \nonumber\end{aligned}$$ Hence $$\begin{aligned} x^T {\cal G} x = \bigg( a_{-1} a_0 x_{-1} x_1 + (a_0^2+a_1^2) x_1^2 + a_1 a_2 x_1 x_3 \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ + \: \: a_{0} a_1 x_{0} x_2 + (a_1^2+a_2^2) x_2^2 + a_2 a_3 x_2 x_4 \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ + \: \: a_{1} a_2 x_1 x_3 + (a_2^2+a_3^2) x_3^2 + a_3 a_4 x_3 x_5 \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ + \: ... \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \:\: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ + \: \: a_{Q-2} a_{Q-1} x_{Q-2} x_Q + (a_{Q-1}0^2+a_Q^2) x_1^2 + a_Q a_{Q+1} x_Q x_{Q+2} \bigg) \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \: \nonumber \\ = a_1^2 x_2^2 + a_{Q-1}^2 x_Q^2 + (a_1 x_1+a_2 x_3)^2 + (a_2 x_2+a_3 x_4)^2 + ... + (a_{Q-2} x_{Q-2}+a_{Q-1} x_Q)^2 \ge 0 \nonumber \end{aligned}$$ and is true $\forall x$. The matrix is, hence, positive-definite. [5]{} Borg, Frank “Quantum Profiles and Paradoxes”, arxiv.physics 0611164v1 16-Nov-2006 (1962). Allahverdyan, Armen E. ; Balian, Roger & NieuwenHuizen, Theo M. “The Quantum Measurement Process: Lessons from an exactly solvable model”, arxiv:quant-ph/0702135v2 (2007) Balian, Roger “Emergences in Quantum Measurement Processes”, KronoScope 13:1 85-95 (2013) Allahverdyan, Armen E. ; Balian, Roger & NieuwenHuizen, Theo M. “Curie-Weiss model of the quantum measurement process”, Europhysics Letters 61.4 452-458 (2003). Gerschgorin, S. , “Über die Abgrenzung der Eigenwerte einer Matrix”, Izv. Akad. Nauk. USSR Otd. Fiz.-Mat. Nauk (in German), 6: 749–754 (1931). Boyd, John P. “The Fourier Transform of the quartic Gaussian $e^{-Ax^4}$: Hypergeometric functions, power series, steepest descent asymptotics and hyperasymptotics and extensions to $e^{-Ax^{2n}}$”, Applied Mathematics and Computation 0096-3003 (2014). Also at http://dx.doi.org/10.1016/j.amc.2014.05.001 (2014). Wigner, E. “On the Distribution of the Roots of Certain Symmetric Matrices.” Ann. of Math. 67, 325-328, 1958. Mott, N. F. “The wave mechanics of $\alpha$-ray tracks.” Proc. Roy. Soc. London. A (1929). Coleman, S. “Quantum mechanics in your face.” APS New England Sectional Meeting (1994)

ArXiv

--- abstract: | [exponential asymptotics, asymptotics, hyperasymptotics, transseries, boundary value problems.]{} We introduce templates for exponential asymptotic expansions that, in contrast to matched asymptotic approaches, enable the simultaneous satisfaction of both boundary values in classes of linear and nonlinear equations that are singularly perturbed with an asymptotic parameter $\epsilon\rightarrow 0^+$ and have a single boundary layer at one end of the interval. For linear equations, the template is a transseries that takes the form of a sliding ladder of exponential scales. For nonlinear equations, the transseries template is a two-dimensional array of exponential scales that tilts and realigns asymptotic balances as the interval is traversed. An exponential asymptotic approach also reveals how boundary value problems force the surprising presence of transseries in the linear case and negative powers of $\epsilon$ terms in the series beyond all orders in the nonlinear case. We also demonstrate how these transseries can be resummed to generate multiple-scales-type approximations that can generate uniformly better approximations to the exact solution out to larger values of the perturbation parameter. Finally we show for a specific example how a reordering of the terms in the exponential asymptotics can lead to an acceleration of the accuracy of a truncated expansion. author: - 'C.J.Howls' title: 'Exponential asymptotics and boundary value problems: keeping both sides happy at all orders.' --- \[firstpage\] Introduction ============ In this paper we examine the role of exponential asymptotics in boundary value problems (BVPs) for classes of second-order ordinary differential equations with a small positive asymptotic parameter $\epsilon$ multiplying the highest derivative. In particular we consider both linear and nonlinear problems with boundary layers at one end of the finite domain. At first sight, this appears to be a well-trodden classical problem with little more to be revealed. Systems that cannot be solved exactly can be attacked using a vast toolbox of asymptotic approaches, most notably matched asymptotic expansions (MAE), see for example Nayfeh (1973) and references therein. Despite, or perhaps because of, the success of MAE approaches, BVPs have received less attention from asymptotic analysts than initial value problems. Results on existence and uniqueness of solutions of second-order nonlinear two-point BVPs predated MAE (Wasow 1956 and references therein). Formal theorems on the existence and character of asymptotic solutions of BVPs have been proved (Wasow 1956; Howes 1978). More recent asymptotic BVP work (e.g., O’Malley & Ward 1997; Ou & Wong 2003, 2004; Ward 2005) has considered shock fronts and multiple spike solutions. However, we shall show that, even for such a well-studied problem, the deployment of an exponential-asymptotic approach throws up subtle, surprising and generic asymptotic issues left unanswered by previous approaches. Exponential asymptotics (Berry 1989; Segur [*et al*]{} 1991) seeks to obtain a formal expansion of a function that incorporates all relevant exponential scales. These scales are manifested typically by the exponential prefactors of (often) divergent infinite series. Techniques exist for managing the divergence and analytical continuation of these series (Berry & Howls 1991; Olde Daalhuis 1993), which lead to exponentially improved numerical approximations that are uniformly valid in wider ranges of the underlying parameters (Olde Daalhuis 1998, Olde Daalhuis & Olver 1998). This approach can provide better error bounds (Boyd 1994) and algebraic methods to resolve complex geometric structures (Howls 1997; Delabaere & Howls 2002). In this paper we shall not be concerned with the implementation of the hyperasymptotics. Rather, we seek to identify the templates, or form for the expansions that are required as input to these hyperasymptotic approaches. The templates are the atomic building blocks which can be manipulated, using resummation techniques, to produce more sophisticated, ‘molecular asymptotic’ approximations. Exponential asymptotic approaches to solutions of linear and nonlinear ordinary differential equations have been well explored (Chapman [*et al*]{} 1998; Olde Daalhuis 1998, 2004, 2005[*a*]{}[*b*]{}; Howls & Olde Daalhuis 2003; Olafsdóttir [*et al*]{} 2005), but have concentrated on initial value problems. O’Malley and Ward (1997) applied exponential asymptotics to linear convection-diffusion PDE BVPs. Lee & Ward (1995) considered exponentially sensitive ill-conditioning in nonlinear ODE BVPs. Ward (2005) also provided an asymptotic description of multiple spike solutions and the associated interior boundary layers in Carrier-Pearson-type systems. However this paper is the first time that a detailed analysis of the satisfaction of singularly perturbed BVPs at each and every exponential order has taken place. The approach reveals the following surprising results. First, as is well known, MAE usually violate one of the boundary conditions. This error may be negligible at small values of the asymptotic parameter $\epsilon$, but as $\epsilon$ grows the range of validity (both numerical and analytical) of the expansion may decrease. Here we show how, for a similar effort, an exponential asymptotic approach satisfies both boundary conditions and hence obtains an expansion that is valid in a wider range with at least a comparable accuracy. Second, an exponential approach also reveals a generic, yet previously unknown, intricate and interlocking ladder (linear) or array (nonlinear) structure within the expansions that is essential to the satisfaction of boundary conditions. Third, for initial value problems arising from linear ordinary differential systems, usually only a finite number of asymptotic series play a role, each being a leading order (in the exponential-asymptotic sense) expansion of one of the solutions. The analysis here demonstrates explicitly how boundary values force the presence of “transseries" in linear systems here being infinite sums of exponentially-prefactored asymptotic series in the small parameter $\epsilon$. Transseries occur naturally in nonlinear differential systems (Olde Daalhuis 2005[*a*]{}[*b*]{}; Costin 1998; Costin & Costin 2001; Chapman [*et al*]{} 2007) due to the mixing of exponential scales by nonlinear terms, but their role in linear systems has not been highlighted explicitly before. They are significant, because the resummation of transseries can generate non-local information about the singularity structure of the problem and can be used to extend the range of validity of solutions. Fourth, for nonlinear BVPs, exponential asymptotics uncovers the presence of, at first sight, paradoxical [*negative*]{} powers of the small parameter $\epsilon$ in expansions. Fortunately these are contained within series that are premultiplied by decaying exponentials in $1/\epsilon$ and so behave regularly as $\epsilon \rightarrow 0^+$. Fifth, there is a choice in the order in which terms are resummed and this may be exploited, where possible, to allow for the derivation of numerically more efficient approximations to the solutions of nonlinear BVPs. At the outset, we stress that we are not proposing that an exponential-asymptotic approach should replace that of MAE (or any other) in boundary layer calculations. MAE remains a powerful and (where appropriately applied) numerically accurate approach. We aim to highlight the underlying asymptotic structure of BVPs to facilitate extensions to a matched asymptotic approach when needed. The outline of the paper is as follows. In section 2, we study a pedgagocial linear BVP to highlight the issues from an exponential asymptotic viewpoint. In section 3 we use general arguments to introduce the presence of transseries and the associated sliding ladder of exponential scales, demonstrating in section 4 how this will be a generic phenomenon in linear BVPs. In section 5 we introduce nonlinear BVPs, again with a pedagogical example, and the transseries array structure that allows boundary value satisfaction at both ends of the interval. The coefficients in the leading exponential orders are derived in sections 6 and 7 with a demonstration of the intricate realignment of balancing of terms, together with a demonstration of how it is possible to resum them to generate multiple-scales solutions. Moving to higher order terms in section 8 we uncover a contradiction that forces the presence of negative powers of the small parameter in the transseries. In section 9 we discuss how the reordering of terms within transseries can accelerate numerical agreement. We end with a discussion and suggestions for future work. A linear example of boundary failure ==================================== We take a pedagogical approach and first illustrate the issues we wish to address through a specific example. The general case will be explained later. We consider $$\begin{aligned} \label{1} \epsilon u''(x)+(2x+1)u'(x)+2u(x)=0, \qquad 0<x<1, \\ u(0)=\alpha, \ \ u(1)=\beta, \qquad 0<\epsilon<<1,\end{aligned}$$ where prime denotes differentiation with respect to $x$. Conventionally the argument for a matched asymptotic approach runs along the following lines. Substitution of the standard ansatz $$\label{simple} u(x)\sim\sum_{r=0}^{\infty}a_r(x)\epsilon^r$$ into (\[1\]) generates first order linear differential equations for $a_r(x)$. Hence $a_r(x)$ cannot in general satisfy the two boundary conditions simultaneously. A boundary layer exists and there is an apparent need for a second expansion that obeys a rescaled equation, valid in the boundary layer and which satisfies the inner boundary condition. The inner expansion is then matched to the outer expansion involving $a_r(x)$ that satisfies the other boundary condition. The inner and outer expansions are matched to determine the remaining unknown constants. Rescaling (\[1\]) with a change of variables $x=\epsilon X$ we may derive a composite expansion (at leading order) $$\label{leadm} u_{\rm matched}(x)= {3\beta}/{(1+2x})+(\alpha-3\beta)\re^{-x/\epsilon}.$$ Clearly we have $u_{\rm matched}(0)=\alpha$. However $u(1)=\beta+{\mathcal O}\left(\re^{-1/\epsilon}\right)$ and so fails to satisfy the boundary condition at $x=1$, by at least the size of the inner expansion. For small values of $\epsilon$, this may well be a numerically negligible error, but this error will grow with $\epsilon$, so reducing the range of validity of the solution. In this example, we learn more by following the approach of Latta (Nayfeh 1973, pp.145-154). We try the simple exponential asymptotics approach of a WKB ansatz $$\label{latta} u(x)\sim \sum_{r=0}^{\infty}a_r(x)\epsilon^r+\re^{-F(x)/\epsilon}\sum_{r=0}^{\infty}b_r(x)\epsilon^r.$$ We shall require $F(0)=0$ so that terms in the two series can balance at ${\mathcal O}(\epsilon^r)$ when $x=0$. However, for general $x\ne 0, 1$, substitution of (\[latta\]) into (\[1\]) and balancing at each order ${\mathcal O}(\epsilon^r)$ and ${\mathcal O}\left(\re^{-F(x)/\epsilon}\right)$, we obtain: $$\begin{aligned} {\mathcal O}(\epsilon^0): &\qquad& (2x+1)a_0'(x)+2a_0(x)=0, \\ {\mathcal O}(\epsilon^1): &\qquad& (2x+1)a_1'(x)+2a_1(x)=-a_0''(x), \\ {\mathcal O}\left(\re^{-F(x)/\epsilon}\epsilon^{-1}\right): &\qquad& F'(x)^2-(2x+1)F'(x)=0.\end{aligned}$$ The last equation coupled with $F(0)=0$ gives $$\label{Feqn} F(x)=x^2+x.$$ $F(x)$ is positive on $0<x<1$, as it should be, by assumption, so that (at least from an exponential-asymptotic viewpoint) no turning point occurs on that interval that can give rise to an interior boundary layer. The remaining equations for $a_r$ and $b_r$ can be solved order by order. The simplified equations for $b_r(x)$ are given below. The boundary conditions at $x=0$ are $$\label{bc0} a_r(0)+b_r(0)=\delta_{r0}\alpha,$$ where $\delta_{ij}$ is the Kronecker delta. These can be satisfied by design. However the boundary condition at $x=1$ can still only be satisfied up to exponential accuracy in $\epsilon$: $$\label{bc1} a_r(1)+{\mathcal O}\left(\re^{-F(1)/\epsilon}\right)=\delta_{r0}\beta.$$ Ignoring the exponential error at the right-hand boundary we generate the relations: $$\label{aterm} a_r(x)=(a_{r-1}'(1)-a_{r-1}'(x))/(2x+1), \qquad a_0(x)=3\beta/(2x+1),$$ $$\label{bterm} b_r'(x)={b_{r-1}''(x)}/({2x+1}), \qquad b_r(0)=-a_r(0), \qquad b_0(x)=\alpha-3\beta.$$ The latter equations show that the $b_r$ are here actually constants. Hence the leading order solution is $$\label{approx} u_{\rm WKB}(x)= \left({3\beta}/({2x+1})+(\alpha-3\beta)\re^{-(x^2+x)/\epsilon}\right)\left(1+{\mathcal O}\left(\epsilon\right) \right).$$ Note that for $x>0$, the exponentially small term could be absorbed into the ${\mathcal O}(\epsilon)$ term, but is kept separate to delineate the existence of the boundary layer. The matched and Latta-WKB approaches provide asymptotic “solutions" that satisfy the boundary condition near to the boundary layer at $x=0$. They both fail to satisfy the boundary condition at $x=1$. From an exponential asymptotic (and for $x(x+1)/\epsilon={\mathcal O}(1)$, a numerical) viewpoint, this is unsatisfactory. Comparison of (\[leadm\]) and (\[approx\]) shows that, for a similar amount of effort, we have $u_{\rm matched}(1)= \left(\beta+(\alpha-3\beta)\re^{-1/\epsilon}\right)\left(1+{\mathcal O}\left(\epsilon\right) \right),$ $u_{\rm WKB}(1)= \left(\beta+(\alpha-3\beta)\re^{-2/\epsilon}\right)\left(1+{\mathcal O}\left(\epsilon\right) \right).$ The order of the exponential error in the WKB approach at the right-hand boundary is here less than in the matched case. This suggests that an extension of the WKB approach may be worth pursuing. We develop this approach in the next section. A transseries approach ====================== The exponential prefactors of the $a_r$-series and $b_r$-series in (\[latta\]) balance at $x=0$ by design. At $x=1$, the $a_r$ satisfy the boundary data, but the $b_r$ do not. Now suppose there is an additional series present, satisfying the differential equation (\[1\]), but prefactored by $\re^{-F(1)/\epsilon}$. At $x=1$, this can be used to cancel off the contribution from the $b_r$-series. However, while addressing the satisfaction of the boundary condition at $x=1$, the presence of this new series means that the boundary condition at $x=0$ is now violated exponentially. This new series can still satisfy the boundary conditions at $x=0$ if there is another, additional, series prefactored by $\re^{-(F(x)+F(1))/\epsilon}$. For then, at $x=0$ with $F(0)=0$, this second new series balances the first new series at ${\mathcal O}(e^{-F(1)/\epsilon})$. In turn, this second series can satisfy the boundary condition at $x=1$ if there is another, third additional series prefactored by $\re^{-2F(1)/\epsilon}$. In turn this can satisfy the $x=0$ condition if there is another, fourth additional series prefactored by $\re^{-(F(x)+2F(1))/\epsilon}$ and so on and so forth. (Problems with a single boundary layer near $x=1$ would require $F(1)$=0 in (\[Feqn\]) and a reversal of roles of $F(0)$ and $F(1)$.) In summary, the boundary values generate a ladder of series with exponential prefactors defining the runs (see figure \[walk\]). The middle column in figure \[walk\] can be visualised as a sliding scale of prefactors that can be used to line up different exponential scales. At $x=0$ the scale lines up with both with the $0$-prefactor and also the scales at $pF(1)$, for integer $p\ge0$ (left hand panel). This ensures that the boundary conditions at $x=0$ are satisfied. As $x$ travels from $0$ to $1$ this scale slides down (middle panel) so as to line up with the exponential prefactors $pF(1)$, $p\ge1$ at $x=1$ (right-hand panel). The terms that are algebraic in $\epsilon$ in the re-balanced transseries can then be paired up so as to satisfy the boundary conditions, as we show below. With this ladder, the zero-prefactored series exactly satisfies the boundary condition at $x=1$, with no exponential error. The template for the full asymptotic expansion is thus: $$\label{template} u_{\rm trans}(x;\epsilon)\sim\sum_{p=0}^{\infty}\re^{-pF(1)/\epsilon}\sum_{r=0}^{\infty}a_r^{(p)}(x)\epsilon^r+\sum_{p=0}^{\infty}\re^{-\left(F(x)+pF(1)\right)/\epsilon}\sum_{r=0}^{\infty}b_r^{(p)}(x)\epsilon^r,$$ where $a_r^{(0)}(x)=a_r(x)$. Such a sum of exponentially-prefactored series is called a transseries (Olde Daalhuis 2005[*a*]{}, [*b*]{}; Costin 2001; Chapman [*et al*]{} 2007). The $p$-sums are transseries with prefactors that only depend on $\epsilon$, not $x$. When approached directly from the asymptotics, the presence of transseries in linear equations is, at first sight, surprising. We shall show below that this is the general form of the exponential-asymptotic template for the class of problems under consideration. More general templates might be derived for higher order equations. The template for BVPs with internal boundary layers must be modified to take account of the associated connection problems. Substituting (\[template\]) into (\[1\]) and hereafter using $\cdot_{,x}$ to denote $x$-derivatives, we arrive at the recurrence relations $$\begin{aligned} a_r^{(p)}(x)&=&\left(a_{r-1,x}^{(p)}(1)-3b_r^{(p-1)}(1)-a_{r-1,x}^{(p)}(x)\right)/\left(2x+1\right), \\ b_{r,x}^{(p)}(x)&=&b_{r-1,xx}^{(p)}(x)/(2x+1).\end{aligned}$$ with $b_r^{(-1)}(x)=0$. The BVP then becomes $$\label{pbcs0} a_r^{(p)}(0)+b_r^{(p)}(0)=\delta_{r0}\delta_{p0}\alpha, % \\ \label{pbcs1} \qquad a_r^{(p)}(1)+b_r^{(p-1)}(1)=\delta_{r0}\delta_{p0}\beta.$$ Simple calculations give the leading orders of each series as $$%a_0^{(0)}(x)=\frac{3\beta}{2x+1}, \qquad a_0^{(p)}(x)=-\frac{3^p(\alpha-3\beta)}{2x+1}, \qquad b_0^{(p)}(x)=3^p(\alpha-3\beta). a_0^{(0)}(x)=3\beta/(2x+1), \ \ \ a_0^{(p)}(x)=-3^p(\alpha-3\beta)/(2x+1), \ \ \ b_0^{(p)}(x)=3^p(\alpha-3\beta).$$ Hence we have $$\begin{aligned} u_{\rm trans}(x;\epsilon)&=&\left\{\frac{3\beta}{2x+1}-\frac{(\alpha-3\beta)}{2x+1}\sum_{p=1}^{\infty}3^p \re^{-pF(1)/\epsilon} \right. \nonumber \\ &\ & \left. + \ (\alpha-3\beta)\sum_{p=0}^{\infty}3^p \re^{-(F(x)+pF(1))/\epsilon}\right\}(1+{\mathcal O}(\epsilon)).\end{aligned}$$ The $p$-sums are convergent for $|3\re^{-F(1)/\epsilon}|<1$. Summing these is equivalent to changing the order of the $p$ and $r$ sums in transseries template (\[template\]), (Olde Daalhuis 2005[*a*]{}, [*b*]{}; Costin 2001; Chapman [*et al*]{} 2007). The result is $$\label{psum} u_{\rm trans}(x;\epsilon)=\left(\frac{3}{2x+1}\left\{\frac{\beta-\alpha \re^{-F(1)/\epsilon}}{1-3\re^{-F(1)/\epsilon}}\right\}+\left\{\frac{\alpha-3\beta}{1-3\re^{-F(1)/\epsilon}}\right\}\re^{-F(x)/\epsilon}\right)\left(1+{\mathcal O}(\epsilon)\right). %\\ %&=&\left(\frac{3}{2x+1}\left\{\frac{\beta-\alpha \re^{-2/\epsilon}}{1-3\re^{-2/\epsilon}}\right\}+\left\{\frac{\alpha-3\beta}{1-3\re^{-2/\epsilon}}\right\}\re^{-(x^2+x)/\epsilon}\right)\left(1+{\mathcal O}(\epsilon)\right).$$ The approximation (\[psum\]) satisfies both boundary conditions. In this sense (at least) approximation (\[psum\]) is better than either of the WKB or MAE approaches. If terms in $\re^{-F(1)/\epsilon}$ are neglected in (\[psum\]), we recover the WKB result (\[approx\]). Higher order approximations can in principle be derived by similar resummations of the transseries at each order of ${\mathcal O}(\epsilon^r)$. The results satisfy the boundary conditions at each order. In Appendix A we show how the $p$-series may be resummed by a multiple scales approach. In figure \[Asympplots\] we compare the leading order results of the exponential asymptotic, Latta-WKB and matched asymptotic approaches with the exact solution for $\alpha=1$, $\beta=0$ and the comparatively large small parameter $\epsilon=1$. Clearly the exponential asymptotic approach is the only one to satisfy both boundary conditions. For the large value of $\epsilon=1$, the other asymptotic approaches (especially MAE) agree noticeably less well with the exact result away from the boundary layer near $x=1$. In the right-hand graph in figure \[Asympplots\] we compare the integral of the relative errors over the entire range $0\le x\le 1$: $$\label{integerror} \Delta_{\rm approx}(\epsilon)=\int_0^1 |u_{\rm exact}(x;\epsilon)-u_{\rm approx}(x;\epsilon)|\rd x,$$ for values of $\epsilon$ between 0 and 1. Clearly the cumulated error in the exponential asymptotic approach is smaller over the entire range of $x$. Comparison with the exact solution and the general case ======================================================= A similar analysis can be carried out for more general solutions of second-order linear two-point boundary value problems. Consider the system: $$\begin{aligned} \label{gen} \epsilon u''(x)+c(x)u'(x)+d(x)u(x)=0, \qquad 0<x<1, \\ u(0)=\alpha, \qquad u(1)=\beta, \qquad 0<\epsilon <<1,\end{aligned}$$ where $c(x), d(x)>0$ for $0<x<1.$ Let the solutions of the equation be $$\label{us} u_1(x;\epsilon)=T_1^{(x)}(\epsilon), \ \ u_2(x;\epsilon)=\re^{-F(x)/\epsilon}T_2^{(x)}(\epsilon),$$ where the $T_j^{(\zeta)}(\epsilon)$ and $F(x)$ are given by $$T_j^{(\zeta)}(\epsilon)=\sum_{r=0}^{N-1}\tilde{T}_{j,r}^{(\zeta)}\epsilon^r+R_{j,N}(\zeta,\epsilon), \qquad F(x)=\int_0^xc(\eta)d\eta.$$ For a choice of $u_1$ and $u_2$ satisfying (\[gen\]), the self-consistent solution of the BVP is $$\label{exactu} u(x;\epsilon)=\frac{\left[\alpha u_2(1)-\beta u_2(0)\right]u_1(x)+\left[\beta u_1(0)-\alpha u_1(1)\right]u_2(x)}{\left[u_1(0)u_2(1)-u_1(1)u_2(0)\right]},$$ provided the denominator does not vanish. This can be written as $$\begin{aligned} \label{gen1} u(x;\epsilon)=&\ &\left\{\frac{\alpha-\beta T_1^{(0)}(\epsilon)/T_1^{(1)}(\epsilon)}{1-\re^{-F(1)/\epsilon}T_1^{(0)}(\epsilon)T_2^{(1)}(\epsilon)/T_1^{(1)}(\epsilon)T_2^{(0)}(\epsilon)}\right\}\frac{T_2^{(x)}(\epsilon)}{T_2^{(0)}(\epsilon)} \re^{-F(x)/\epsilon} \nonumber \\ &+& \left\{\frac{\beta-\alpha \re^{-F(1)/\epsilon}T^{(1)}_2(\epsilon)/T^{(0)}_2(\epsilon)}{1-\re^{-F(1)/\epsilon}T_1^{(0)}(\epsilon)T_2^{(1)}(\epsilon)/T_1^{(1)}(\epsilon)T_2^{(0)}(\epsilon)}\right\}\frac{T_1^{(x)}(\epsilon)}{T_1^{(1)}(\epsilon)}.\end{aligned}$$ Replacing the $T_j^{(\zeta)}(\epsilon)$ by their asymptotic expansions and expanding the quotient leads naturally to transseries (\[template\]). Hence for linear second-order boundary value systems of type (\[gen\]) the ladder transseries template (\[template\]) is the general form. Nonlinear second-order 2-point boundary value problems ====================================================== Given the ladder template for linear second-order 2-point BVP, it is natural to ask if a generalisation exists for analogous nonlinear boundary value systems. Such systems have been studied asymptotically for decades and form the raison-d’être for many matched asymptotic approaches to boundary layers. Significant results were provided by Wasow (1956), O’Malley (1969), Howes (1978). Wasow (1956) established conditions under which systems of the form $$\label{Wasoweq} \epsilon u''(x)=F_1(u,x)u'(x)+F_2(u,x), \qquad u(0)=\alpha, \ u(1)=\beta, \qquad 0<\epsilon<<1,$$ have absolutely convergent perturbative series on the interval $0<x<1$. O’Malley (1969) considered a systems approach to the problem, and Howes (1978) used the stability of the $\epsilon=0$ problem to study the existence of boundary, shock and corner layer solutions. We shall again restrict ourselves to a situation where only a single boundary layer exists, so follow Howes (1978) p.79 and consider his example (E3): $$\label{node} \epsilon u''(x)+u'(x)u(x)-u(x)=0, \qquad u(0)=\alpha, \ u(1)=\beta,$$ where $1<\beta<\alpha+1$. Under these conditions (Howes 1978, p.83) a boundary layer exists near to $x=0$. Generalisations of what follows could include, for example, larger regions of the $(\alpha, \beta)$ plane or $\epsilon$-dependent boundary data. The conventional MAE approach to (\[node\]) generates the composite expansion $$\label{unonmatched} u_c(x)=\left(x+(\beta-1)\frac{(\beta-1)\tanh\left(\frac{(\beta-1)x}{2\epsilon}\right)+\alpha}{\beta-1+\alpha\tanh\left(\frac{(\beta-1)x}{2\epsilon}\right)}\right)\left(1+{\mathcal O}\left(\epsilon\right)\right).$$ Note that while $u_c(0)=\alpha$, we again have $u_c(1)=\beta+{\mathcal O}\left(\re^{-(\beta-1)/\epsilon}\right)\ne \beta$. Clearly, when $\beta>1$, this error may diminish as $\epsilon \rightarrow 0^+$. However for larger values of $\epsilon$, or small positive $\beta-1$ this error may become numerically significant. The WKB approximation to the solution takes the form: $$\label{lattesoln} u_{\rm WKB}(x)=\left((x+\beta-1)+\frac{(1+\alpha-\beta)(\beta-1)}{x+\beta-1}\re^{-x\left(\beta-1+x/2\right)/\epsilon}\right)\left(1+{\mathcal O}\left(\epsilon\right)\right).$$ Clearly we have $u_{\rm WKB}(0)=\alpha$, but $u_{\rm WKB}(1)=\beta+{\mathcal O}\left(\re^{-\left(\beta-1/2\right)/\epsilon}\right)$. Hence, as with MAE, the WKB template also fails to satisfy the boundary data at $x=1$. The WKB approach is just the first two terms of a transseries expansion of the general solution of nonlinear ODE which takes the form $$u(x)\sim \sum_{n=0}^{\infty}C^nu_{n}(x,\epsilon), \qquad u_n(x, \epsilon)\sim \re^{-nF(x)/\epsilon}\sum_{r=0}^{\infty}a_r^{(n)}\epsilon^r, \qquad F(0)=0,$$ with the constant $C$ determined by the boundary conditions. It will here be absorbed into the leading order coefficient of $u_1$, namely $a_0^{(1)}$. The necessity to include integer powers of $e^{-F(x)/\epsilon}$ is caused by the nonlinear terms in the differential equation which generate successively higher order exponential scales that must be balanced (at each $x\ne 0$), at least in a formal solution. However, we see immediately that this template still suffers from the same problem as the previous approaches. Specifically the boundary conditions then require: $$\label{cap3} \sum_{n=0}^{\infty}a_r^{(n)}(0)=\delta_{r0}\alpha, \qquad a_{r}^{(0)}(1)+\sum_{n=1}^{\infty}\re^{-nF(1)/\epsilon}a_r^{(n)}(1)``="\delta_{r0}\beta,$$ at each $r=0, 1, 2, \dots$. For the time being, we put aside the issue of being able to satisfy the boundary conditions with an infinite sums of terms. The point is that, for $F(1)\ne 0$, the boundary conditions at $x=1$ still leave terms unbalanced at ${\mathcal O}\left(\re^{-nF(1)/\epsilon}\right)$ for any integer value of $n>0$. Inspired by the linear example above, we modify the transseries template as $$\label{cap4} u(x)\sim\sum_{n=0}^{\infty} \sum_{p=0}^\infty u_n^{(p)}(x,\epsilon),$$ $$\label{cap55} u_0(x)\sim \sum_{r=0}^{\infty}a_r^{(0,0)}(x)\epsilon^r, \qquad u_n^{(p)}(x,\epsilon)\sim \re^{-(nF(x)+pF(1))/\epsilon}\sum_{r=r_{\rm min}(n,p)}^{\infty}a_r^{(n,p)}(x)\epsilon^r. %\qquad F(0)=0.$$ The exponent $F(x)$ again slides between $F(0)\equiv0$ and $F(1)$ as $x$ runs between $0$ and $1$. The initial assumption of the minimum algebraic order in $\epsilon$, is $r_{\rm min}(n,p)=0$. However, we shall see later that even in this simple example, this is not always so. In contrast to the single exponent $\re^{-F(x)/\epsilon}$ in the linear case, we now have an infinite number of exponential scales $\re^{-nF(x)/\epsilon}$, $n=0,1,2 \dots$. At $x=0$ these all degenerate to $e^{-nF(0)/\epsilon}=1$ and so can be made to balance at each ${\mathcal O}(\epsilon^r)$ with the corresponding terms from $u_0$. At $x=1$ these scales become $\re^{-nF(1)/\epsilon}$, $n=1,2 \dots$. To satisfy the boundary condition at each exponential order thus again requires the inclusion of additional series prefactored with scales $\re^{-pF(1)/\epsilon}$, $p=1,2 \dots$. However, substitution shows that these boundary series will multiply series in $\re^{-nF(x)/\epsilon}$ in the $uu'(x)$ term to produce further contributions with prefactors $\re^{-(nF(x)+pF(1))/\epsilon}$, $n, p=0,1,2 \dots$. All of these series have to be balanced at $x=0,1$ to satisfy the boundary data. Fortunately this is possible, as explained in the diagrams below. From (\[cap55\]) the exponents labelled by integers $(n,p)$ generate not a ladder of exponential scales, but a planar lattice (see figure \[nonlineararray\]). The heights of the lattice points are at the exponential scales $\re^{-(nF(x)+pF(1))/\epsilon}$. At $x=0$, $nF(0)=0$, and so balances occur between $n=0,1,2,3\dots$ series prefactored by $\re^{-pF(1)/\epsilon}$ at each value of $p$. As $x$ increases to 1, $F(x)$ varies and the planar lattice tilts. When $x=1$, the prefactors become $\re^{-(n+p)F(1)/\epsilon}$ and the scales have reconfigured so that balancing now takes place between series prefactored by exponentials with $n+p=K$ for fixed $K$ (along diagonal rows on the right in figure 3). We shall demonstrate this reconfiguration in detail by explicit calculation for the system (\[node\]) in the following sections. Derviation of the coefficients ============================== Substitution of (\[cap4\]) into (\[node\]) and balancing at ${\mathcal O}\left(\re^{-(nF(x)+pF(1))/\epsilon}\right)$ gives $$\label{cap6} \epsilon u_{n,xx}^{(p)}(x)+\sum_{m=0}^{n}\sum_{q=0}^p u_m^{(q)}(x)u_{n-m,x}^{(p-q)}(x)-u_n^{(p)}(x)=0.$$ For fixed $(n,p)$ we can substitute (\[cap55\]) into (\[cap6\]) and balance at ${\mathcal O}(\epsilon^r)$. We obtain the following balances at $(n,p)=(0,0)$, ${\mathcal O}(\epsilon^r)$ and for $n\ne 0$, ${\mathcal O}(1/\epsilon)$, respectively $$\label{cap7aab} \qquad a_{r-1,xx}^{(0,0)}(x)+\sum_{s=0}^r a_s^{(0,0)}(x)a_{r-s,x}^{(0,0)}(x)-a_r^{(0,0)}(x)=0.$$ $$\label{cap7} n^2F'(x)a_0^{(n,p)}(x)-\sum_{m=0}^{n-1}\sum_{s=0}^p (n-m)a_0^{(m,s)}(x)a_0^{(n-m,p-s)}(x)=0.$$ These recurrence relations now generate the leading order asymptotic terms, with associated arbitrary constants $C_r^{(n,p)}$. From $(n,p)=(0,0)$ at ${\mathcal O}(\epsilon^0)$ we find: $$\label{cap10} %\left(a_{0,x}^{(0,0)}-1\right)a_0^{(0,0)}(x)=0 \qquad \Rightarrow \qquad a_0^{(0,0)}(x)=x+C_0^{(0,0)}.$$ At ${\mathcal O}(\epsilon)$ when $(n,p)=(0,0)$, we have $$\label{cap11} a_{0,xx}^{(0,0)}+a_0^{(0,0)}a_{1,x}^{(0,0)}+a_1^{(0,0)}a_{0,x}^{(0,0)}-a_1^{(0,0)}(x)=0 \qquad \Rightarrow \qquad a_1^{(0,0)}(x)=C_1^{(0,0)}.$$ It is then simple to prove from $(n,p)=(0,0)$, ${\mathcal O}(\epsilon^r)$, that $$\label{cap12} \qquad a_r^{(0,0)}(x)=C_r^{(0,0)}, \ r>0.$$ $F(x)$ is deduced from $(n,p)=(1,0)$ at ${\mathcal O}(1/\epsilon)$: $$\label{cap8} F'(x)a_0^{(1,0)}(x)-a_0^{(0,0)}(x)a_0^{(1,0)}(x)=0 \qquad \Rightarrow \qquad F'(x)=a_0^{(0,0)}(x).$$ This assumes that $a_0^{(1,0)}(x)\ne0$. This result allows us to recast (\[cap7\]) as a recurrence relation to generate $a_0^{(n,p)}(x)$. For nonzero values of $n$ at ${\mathcal O}(1/\epsilon)$ we have: $$\begin{aligned} \label{cap7a} n(n-1)a_0^{(0,0)}a_0^{(n,p)}(x)&=&\sum_{m=0}^{n-1}\sum_{s=1}^p (n-m)a_0^{(m,s)}(x)a_0^{(n-m,p-s)}(x) \nonumber \\ &\ & + \sum_{m=1}^{n-1}(n-m)a_0^{(m,0)}a_0^{(n-m,p)}.\end{aligned}$$ It is straightforward to show from (\[cap7a\]) that $$\label{cap7aa} %a_0^{(n,0)}=\frac{\left(a_0^{(1,0)}\right)^n}{2^{n-1}\left(a_0^{(0,0)}\right)^{n-1}}.\left(a_0^{(1,0)}/2a_0^{(0,0)}\right)^na_0^{(0,0)} a_0^{(n,0)}=2a_0^{(0,0)}\left(a_0^{(1,0)}/2a_0^{(0,0)}\right)^n.$$ The term $a_0^{(1,0)}(x)$ can be found, from $(n,p)=(1,0)$, ${\mathcal O}(\epsilon^0)$ and using (\[cap12\]): $$\begin{aligned} \label{cap7aaa} %-a_0^{(1,0)}(x)\left(1+a_0^{(0,0)}(x)a_1^{(0,0)}(x)\right)-a_0^{(0,0)}(x)\frac{da_0^{(1,0)}}{dx}&=&0\\ -a_0^{(1,0)}(x)\left(1+a_0^{(0,0)}(x)C_1^{(0,0)}\right)-a_0^{(0,0)}(x)a_{0,x}^{(1,0)}&=&0,\end{aligned}$$ $$\label{cap7aaaa} \Rightarrow \ a_0^{(1,0)}(x)=C_0^{(1,0)}\re^{-C_1^{(0,0)}(x+C_0^{(0,0)})}/(x+C_0^{(0,0)}).$$ When $n=0$, $p\ne0$, further terms can be deduced to be constants from the ${\mathcal O}(\epsilon^0)$ equations by first observing that, when $p=1$: $$\label{cap20} a_0^{(0,0)}(x)a_{0,x}^{(0,1)}+a_0^{(0,1)}(x)a_{0,x}^{(0,0)}(x)-a_0^{(0,1)}(x)=0, \qquad \Rightarrow \qquad a_0^{(0,1)}(x)=C_0^{(0,1)}.$$ Substitution of this result into the corresponding equation for $p>1$ gives: $$\label{cap19} \sum_{s=0}^p a_0^{(0,s)}(x)a_{0,x}^{(0,p-s)}(x)-a_0^{(0,p)}(x)=0.$$ It follows by iteration that: $$\label{cap21} a_0^{(0,p)}(x)=C_0^{(0,p)}, \qquad p=1,2,3, \dots.$$ We find these constants in the next section. Boundary conditions reconfiguring the exponential scales ======================================================== We now apply the boundary conditions, $u(0)=\alpha$, $u(1)=\beta,$ balanced first at each exponential order, and then at each algebraic order of $\epsilon$. The balancing of terms within the relevant exponential scales is illustrated graphically in figure \[termboundaryalign\]. ![Balancing of terms $a_r^{(n,p)}$ at fixed $r$ between transseries with varying $(n,p)$ (\[cap4\]-\[cap55\]) at $x=0$ and $x=1$ due to the realignment of the exponential scales. Terms joined by the grey lines balance at the values of $x$ shown. At $x=0$ at each fixed $p$, an infinite sum over all terms $n$ must satisfy the boundary conditions at the left hand end of the interval. At $x=1$ terms with $n+p=K$ for each fixed $K=0,1,2 \dots$ must be summed to satisfy the boundary conditions at the right-hand end of the interval.[]{data-label="termboundaryalign"}](figure4.eps) In what follows, we set ${a_r^{(n,p)}}(0)={C_r^{(n,p)}}$. At $x=0$, since $F(0)=0$ the exponential scales ${\mathcal O}\left(\re^{-nF(x)/\epsilon}\right)$ all collapse to the same degenerate level and the distinct exponential scales are given by ${\mathcal O}\left(\re^{-pF(1)/\epsilon}\right)$. So for fixed $p$, the terms $a_r^{(n,p)}$ balance at order ${\mathcal O}\left(\re^{-pF(1)/\epsilon}\epsilon^r\right)$ according to: $$\label{capbc1} %\sum_{n=0}^{\infty}{a_r^{(n,p)}}(0)=\sum_{n=0}^{\infty}{C_r^{(n,p)}}=\delta_{r0}\delta_{p0}\alpha, \ \ \Rightarrow \ \ \sum_{n=0}^{\infty}{C_0^{(n,0)}}=\alpha, \ \ \sum_{n=0}^{\infty}{C_r^{(n,p)}}=0, \ (p, r\ne0). \sum_{n=0}^{\infty}{a_r^{(n,p)}}(0)=\delta_{r0}\delta_{p0}\alpha, \ \ \Rightarrow \ \ \sum_{n=0}^{\infty}{a_0^{(n,0)}}(0)=\alpha, \ \ \sum_{n=0}^{\infty}{a_r^{(n,p)}}(0)=0, \ (p, r\ne0).$$ At $x=1$, the template (\[cap4\]-\[cap55\]) reconfigures the exponential scales. The orders ${\mathcal O}\left(\re^{-(n+p)F(1)/\epsilon}\right)$ with $n+p=K$, $K={\rm const}$, become degenerate. Within these scales the terms $a_r^{(n,p)}$, consequently align at each algebraic ${\mathcal O}(\epsilon^r)$ according to: $$\label{capbc2} \sum_{n=0}^{K}{a_r^{(n,K-n)}}(1)=\delta_{r0}\delta_{K0}\beta \ \ \Rightarrow \ \ {a_r^{(0,0)}}(1)=\delta_{r0}\beta, \ \ \sum_{n=0}^{K}{a_r^{(n,K-n)}}(1)=0, \ (n,r\ne0).$$ We make progress by first considering the conditions at $x=1$. Using (\[cap10\]), (\[cap7aaaa\]), (\[capbc2\]) we can identify immediately that $$\label{capbc3} a_0^{(0,0)}(1)= 1+C_0^{(0,0)} =\beta \qquad \Rightarrow \qquad a_0^{(0,0)}(x)= x+\beta-1,$$ $$\label{capbc3a} a_r^{(0,0)}(1)= \delta_{r0}\beta=C_r^{(0,0)} =0 \qquad r>0, \ \ \Rightarrow \ \ a_0^{(1,0)}(x)=C_0^{(1,0)}/(x+\beta-1).$$ Hence, returning to $x=0$ we have, for example, $$\label{capbc4} \sum_{n=1}^{\infty}{a_0^{(n,0)}}(0)=\alpha-\beta+1.$$ Using the form (\[cap7aa\]) we may sum this infinite series formally as a geometric progression and obtain $$\label{capbc4a} \frac{2a_0^{(0,0)}(0)a_0^{(1,0)}(0)}{2a_0^{(0,0)}(0)-a_0^{(1,0)}(0)}=\alpha-\beta+1,$$ $$\label{capbc4b} \frac{2(\beta-1)C_0^{(1,0)}(0)}{2(\beta-1)^2-C_0^{(1,0)}(0)}=\alpha-\beta+1 \qquad \Rightarrow \qquad C_0^{(1,0)}=\frac{2(\alpha-\beta+1)(\beta-1)^2}{(\alpha+\beta-1)}.$$ Returning back to $x=1$ we then have: $$a_0^{(0,1)}(1)+a_0^{(1,0)}(1)=0 \ \ \Rightarrow \ \ C_0^{(0,1)}+\frac{C_0^{(1,0)}}{\beta}=0 \ \ \Rightarrow \ \ C_0^{(0,1)}=-\frac{2(\alpha-\beta+1)(\beta-1)^2}{\beta(\alpha+\beta-1)}.$$ Hence we finally have the leading orders of the first three $(n,p)$ transseries as: $$\label{lead} a_0^{(0,0)}(x)= x+\beta-1, \ \ a_0^{(1,0)}(x)= \frac{2(\alpha-\beta+1)(\beta-1)^2}{(\alpha+\beta-1)(x+\beta -1)},\ \ a_0^{(0,1)}(x)=-a_0^{(1,0)}(1).%=-\frac{2(\alpha-\beta+1)(\beta-1)^2}{\beta(\alpha+\beta-1)}.$$ From (\[cap21\]) and (\[capbc2\]) we can also deduce that for this example $a_r^{(0,0)}(x)=0$, $r\ge1$. Note that this means that for the special case of (\[node\]), the expansion $u_0^{(0,0)}$ will not generate a Stokes phenomenon, since it truncates after the first term. This also reflects the fact that there are no internal boundary layers. Note also this means that the exponent $F(x)$ is not here directly connected to a factorial-over-power late-term ansatz (Dingle 1973; Berry 1989; Chapman [*et al*]{} 1998). We may resum the $n$-transseries for $p=0$ using (\[lead\]) and obtain the following: $$\begin{aligned} \label{nresum} u_{\rm trans}(x)&=&\left(a_0^{(0,0)}(x)+\sum_{n=1}^{\infty}\re^{-nF(x)/\epsilon}a_0^{(n,0)}(x)\right)\left(1+{\mathcal O}(\epsilon)\right) \\ \label{nresum1} &=&\left(a_0^{(0,0)}(x)+\frac{2a_0^{(0,0)}(x)a_0^{(1,0)}(x)\re^{-F(x)/\epsilon}}{2a_0^{(0,0)}(x)-a_0^{(1,0)}(x)\re^{-F(x)/\epsilon}}\right)\left(1+{\mathcal O}(\epsilon)\right).\end{aligned}$$ For most practical purposes, this result may be all that is needed. A comparison of the numerical errors for the MAE (\[unonmatched\]), WKB (\[lattesoln\]) and resummation (\[nresum\]) is shown in figure \[comp1\] for typical values of $\epsilon=1/10$ and $1$ with fixed $\alpha=3/2$ and $\beta=2$. First, for small $\epsilon$ (left hand graph), although the errors committed by the approximations at $x=1$ are too small to display on the scale of the graph, the transseries approximation is uniformly numerically better than either the matched or WKB solutions. Secondly, for larger values of $\epsilon$ both the WKB and transseries approximations are uniformly better approximations than the MAE. The latter violates the boundary condition at $x=1$ by a noticeably larger amount. Although it satisfies the boundary condition at $x=0$, the transseries approximation is not uniformly better than the WKB near there, agreeing more with the MAE. However the transseries approximation quickly does beat WKB as $x$ increases and is definitely better at $x=1$. Note that the transseries approximation (\[nresum\]) only contains terms that have $p=0$. If we summed the $n$-series with $p=1$ and added these in at ${\mathcal O}(\epsilon)$, the error at $x=1$ would diminish. In the next section we show how the higher order balances reveal unexpected behaviour in the nonlinear template. Higher order balancing ====================== Moving on to higher exponential orders, at $x=1$ we need to balance terms involving the $(n,p)=(2,0), (1,1), (0,2)$ series (cf. figure \[termboundaryalign\]): $$a_0^{(0,2)}(1)+a_0^{(1,1)}(1)+a_0^{(2,0)}(1)=0.$$ The behaviour of the $(n,p)=(2,0)$, $(0,2)$ terms have already been derived above: $$a_0^{(2,0)}(x)=\frac{\left(a_0^{(1,0)}(x)\right)^2}{2a_0^{(0,0)}(x)}=\frac{2 (\alpha -\beta +1)^2(\beta -1)^4 }{(x+\beta -1)^3 (\alpha +\beta -1)^2}, \qquad a_0^{(0,2)}(x)=C_0^{(0,2)}.$$ Seeking the $a_0^{(1,1)}(x)$ term we use (\[cap6\]) with $(n,p)=(1,1)$: $$\label{n1p1} \epsilon u_{1,xx}^{(1)}+u_0^{(0)}u_{1,x}^{(1)} +u_0^{(1)}u_{1,x}^{(0)} +u_1^{(0)}u_{0,x}^{(1)} +u_1^{(1)} u_{0,x}^{(0)}-u_1^{(1)}=0.$$ Assuming $r_{\rm min}(1,1)=0$ (cf. (\[cap55\])), at ${\mathcal O}(1/\epsilon)$ in (\[n1p1\]), after simplification we have $$%F'(x)^2a_0^{(1,1)}-F'(x)a_0^{(0,0)}a_0^{(1,1)}-F'(x)a_0^{(0,1)}a_0^{(1,0)}+0+0+0&=&0 \\ %F'(x)\left(a_0^{(0,0)}a_0^{(1,1)}-a_0^{(0,0)}a_0^{(1,1)}-a_0^{(0,1)}a_0^{(1,0)}\right)&=&0 \\ F'(x)a_0^{(0,1)}(x)a_0^{(1,0)}(x)=0. %F'(x)=0, \ {\rm and/or} \ a_0^{(0,1)}(x)=0, \ {\rm and/or} \ a_0^{(1,0)}(x)=0$$ This obviously leads to a contradiction, since prior calculations show that neither of the conditions $F'(x)=0$, $a_0^{(0,1)}(x)=0$, or $a_0^{(1,0)}(x)=0$ hold. We may resolve this impasse by taking $r_{\rm min}(1,1)=-1$. For then, at ${\mathcal O}(1/\epsilon^2)$, (\[n1p1\]) is satisfied by $F'(x)=a_0^{(0,0)}$ and at ${\mathcal O}(1/\epsilon)$ we have $$a_0^{(0,0)}(x)a_{-1,x}^{(1,1)}(x)+a_{-1}^{(1,1)}(x)=-a_0^{(0,0)}(x)a_0^{(0,1)}(x)a_0^{(1,0)}(x),$$ with general solution, $$\label{aminus1} a_{-1}^{(1,1)}(x)=\frac{C_{-1}^{(1,1)} \beta (\alpha +\beta -1)^2+4 x (\beta -1)^4 (\alpha -\beta +1)^2}{\beta (x+\beta -1) (\alpha +\beta -1)^2}.$$ Given that, from above, there is no term $a_{-1}^{(0,1)}(0)$, the boundary condition that this term must satisify at $x=0$ is (cf. figure \[termboundaryalign\]), $$\label{minus1bc} \sum_{n=1}^\infty a_{-1}^{(n,1)}(0)=0.$$ By examination of the ${\mathcal O}(1/\epsilon^2)$ balances in the $(n,p)=(1,p)$ equations (\[cap6\]), it is possible to show that $$a_{-1}^{(n,1)}(x)=n\left({a_0^{(1,0)}(x)/ 2 a_0^{(0,0)}(x)}\right)^{n-1}a_{-1}^{(1,1)}(x). %a_{-1}^{(n,1)}(x)=\frac{n a_{-1}^{(1,1)}(x)\left(a_0^{(1,0)}(x)\right)^{n-1}}{2^{n-1}\left(a_0^{(0,0)}(x)\right)^{n-1}}.$$ Hence we can sum (\[minus1bc\]) to obtain (cf. (\[aminus1\])) $$\frac{4 a_{-1}^{(1,1)}(0)\left(a_0^{(0,0)}(0)\right)^{2}}{\left(2a_0^{(0,0)}(0)-a_0^{(1,0)}(0)\right)^{2}}=0 \qquad \Rightarrow \qquad a_{-1}^{(1,1)}(x)=\frac{4 x (\beta -1)^4 (\alpha -\beta +1)^2}{\beta (x+\beta -1) (\alpha +\beta -1)^2}.$$ Consequently, since $a_{-1}^{(2,0)}(x)=0$ we can now determine the constant $a_{-1}^{(0,2)}\equiv C_{-1}^{(0,2)}$ from the boundary condition at $x=1$: $$\begin{aligned} &a_{-1}^{(2,0)}(1)&+a_{-1}^{(1,1)}(1)+a_{-1}^{(0,2)}(1)=0 \nonumber \\ \Rightarrow \ &a_{-1}^{(0,2)}(x)&=- a_{-1}^{(1,1)}(1)=-\frac{4 (\beta -1)^4 (\alpha -\beta +1)^2}{\beta^2}.\end{aligned}$$ Finally, it is then possible to resum the contributions from $a_{-1}^{(n,1)}$ to obtain $$\label{am1} \sum_{n=0}^\infty \frac{a_{-1}^{(n,1)}(x)}{\epsilon}\re^{-(nF(x)+F(1))/\epsilon}=\frac{4 a_{-1}^{(1,1)}(x)\left(a_0^{(1,0)}(x)\right)^{2}\re^{-(F(x)+F(1))/\epsilon}}{\epsilon\left(2a_0^{(0,0)}(x)-a_0^{(1,0)}(x)\re^{-F(x)/\epsilon}\right)^{2}}.$$ The exponential asymptotic approach has revealed the unexpected and counterintuitive presence of $1/\epsilon$ terms in the expansion as $\epsilon\rightarrow 0^+$. However, these $1/\epsilon$ terms are prefactored by $\re^{-F(1)/\epsilon}$ with $F(1)>0$. Hence they are not only beyond all algebraic orders, but also do vanish exponentially fast as $\epsilon\rightarrow 0^+$. The existence of an ${\mathcal O}(1/\epsilon)$ term in the $(n,p)=(1,1)$ series forces the presence of ${\mathcal O}(1/\epsilon^p)$ terms in series with $p>1$. The corrected form of the template (\[cap55\]) has: $$r_{\rm min}(n,p)=\left\{\begin{array}{cc}-p, \qquad n>p, \\-{\rm Floor}[(n+p)/2], & n\le p.\end{array}\right. %u\sim\sum_{n=0}^\infty\sum_{p=0}^\infty \sum_{r=-p}^\infty a_r^{(n,p)}\epsilon^r \re^{-(nF(x)+pF(1))/\epsilon},$$ As above, since the negative powers of $\epsilon$ are multiplied by $\re^{-pF(1)/\epsilon}$, they will remain technically beyond all algebraic orders and vanish as $\epsilon\rightarrow 0^+$. The presence of these terms means that in generic cases, as $p$ increases, additional work must be done to obtain resummed expansions accurate to\ ${\mathcal O}(\re^{-n(F(x)+F(1))/\epsilon}\epsilon)$. For example, it is a simple algebraic exercise to show that the calculation to derive the resummation analogous to (\[am1\]) of the ${\mathcal O}(\epsilon^0)$ terms with $p=1$, $$\label{am2} \sum_{n=0}^\infty \re^{-n(F(x)+F(1))/\epsilon}a_{0}^{(n,1)}(x)\epsilon^0,$$ actually requires first a complete derivation of all the terms in $a_1^{(n,0)}(x)$. A detailed set of calculations generates the following results for $n\geq 1$: $$\label{Keqn} a_{1}^{(n,0)}(x)=n\left({a_0^{(1,0)}/ 2 a_0^{(0,0)}}\right)^{n-1}a_{1}^{(1,0)}+{K_n/a_0^{(0,0)}} \left({a_0^{(1,0)}/ 2 a_0^{(0,0)}}\right)^{n},$$ $$\begin{aligned} \label{Leqn} a_{0}^{(n,1)}(x)&=&n\left({a_0^{(1,0)}/ 2 a_0^{(0,0)}}\right)^{n-1}a_{0}^{(1,1)} \nonumber \\ &\ &+ n(n-1){\left(a_0^{(1,0)}\right)^{n-2}/ \left(2a_0^{(0,0)}\right)^{n-1}}a_{1}^{(1,0)}a_{-1}^{(1,1)} \nonumber \\ &\ & -2(n-1)\left({a_0^{(1,0)}/ 2 a_0^{(0,0)}}\right)^{n}a_{0}^{(0,1)} \nonumber \\ & \ & +2L_n{\left(a_0^{(1,0)}\right)^{n-1}/ \left(2a_0^{(0,0)}\right)^{n+1}}a_{-1}^{(1,1)}.\end{aligned}$$ The coefficients $K_n$ in (\[Keqn\]) and $L_n$ in (\[Leqn\]) satisfy the recurrence relations $$\begin{aligned} {(n-1)}K_n/2&=&\sum_{m=2}^{n-1}K_m-{(2n+1)(n-1)/ n},\\ {(n-1)}L_n/2&=&\sum_{m=2}^{n-1}L_m+\sum_{m=2}^{n-1}(n-m)K_m-(2n+1)(n-1).\end{aligned}$$ A $z$-transform of these relations generates $$\label{Ksum} \sum_{n=2}^\infty K_nt^n=H(t)\equiv{2t(1-2{\rm Li}_2(t))\over(1-t)^2}+{2(1+t)\ln(1-t)\over 1-t},$$ $$\label{Lsum} \sum_{n=2}^\infty L_nt^n=G(t)\equiv{2t(1+t)(t-2{\rm Li}_2(t))\over(1-t)^3}+{8t\ln(1-t)\over (1-t)^2},$$ where ${\rm Li}_2(t)$ is the dilogarithm function (DLMF 2010). The sum over $n$ of (\[Keqn\]) at $x=0$ can be achieved using (\[Ksum\]) and according to the boundary value, be set equal to zero. It only depends on unknown $a_1^{(1,0)}$ and the already known $a_0^{(1,0)}$ and $a_0^{(0,0)}$. Coupled with the general solution of $$a_0^{(0,0)}a_{1,x}^{(1,0)}+a_{1}^{(1,0)}=a_{0,xx}^{(1,0)},$$ we can use this sum to determine $a_1^{(1,0)}(x)$. This can be substituted into (\[Leqn\]), which, in turn can be summed at $x=0$ using (\[Lsum\]) to find $a_0^{(1,1)}(x)$. In turn (\[Leqn\]) can then be multiplied by $\re^{-(nF(x)+F(1))/\epsilon}$ and summed over $n$ to obtain (\[am2\]). The resulting expression is large and so this is left as an exercise for the reader. In principle this scheme can be repeated for each $p$. For given $p$, the corresponding ${\mathcal O}(\epsilon)$ estimate will involve a sum over $n$ of $a_{0}^{(n,p)}(x)\re^{-(nF(x)+pF(1))}$. To determine the $a_{0}^{(n,p)}(x)$ will likely require the determination of all terms from $a_{r}^{(n,p)}(x)$, $0>r>r_{\min}(n,p)$ beforehand. This will become an increasingly arduous task. However, as the transseries template contains three sums (in $n$, $p$ and $r$) there is the potential to alter the order of resummation of the terms to improve the numerical accuracy of the approximation. Alternative resummations ======================== The above resummations are over $n$ for each $p$ and fixed order $\epsilon^r$. Alternatively it may be possible to reorder the resummations, for example resumming first over $p$ and obtain numerically better agreement with the exact solution. Due to the particular form of (\[node\]) we may proceed here down this route by reverting to a single transseries formulation of the form $$\begin{aligned} \label{pw1} u(x)=\sum_{k=0}^{\infty}\lambda^k \tilde{u}_k(x; \epsilon), \qquad \tilde{u}_0(x, \epsilon)=\sum_{r=0}^{\infty} \tilde{a}_r^{(0)}(x)\epsilon^r,\end{aligned}$$ where $\lambda$ is here just an ordering parameter that will be eventually set equal to one. In the absence of the boundary data, the $\tilde{u}_k(x; \epsilon)$ can be regarded as providing terms of order ${\mathcal O}\left(\exp(-kF(x)/\epsilon)\right)$. Substitution into (\[node\]) and balancing at order $\lambda^k$ generates at ${\mathcal O}(\lambda^0)$: $$\begin{aligned} \label{pw2} \epsilon \tilde{u}_0''(x)+\tilde{u}_0(x) \tilde{u}_0'(x)-\tilde{u}_0(x)=0\end{aligned}$$ and balancing at ${\mathcal O}(\epsilon^r)$ and setting $\tilde{a}_r^{(0)}(1)=\delta_{r0}\beta$ we have $$\begin{aligned} \label{pw4} \tilde{a}_r^{(0)}(x)=\delta_{r0}(x+\beta-1) \qquad \tilde{u}_0(x)=x+\beta-1,\end{aligned}$$ as above. Note that the truncated nature of $ \tilde{u}_0(x)$ is due to the particular system under study. In general it would be only expressible as a formal, divergent series. At order ${\mathcal O}(\lambda)$ we then have $$\begin{aligned} \label{pw5} %\epsilon \tilde{u}_1''(x)+\tilde{u}_1(x) \left(\tilde{u}_0'(x)-1\right)+\tilde{u}_0(x) \tilde{u}_1'(x)&=&0 \\ \qquad \epsilon \tilde{u}_1''(x)+(x+\beta -1)\tilde{u}_1'(x)&=&0.\end{aligned}$$ Applying the boundary condition, e.g., that (with $\lambda=1$) $$\label{pw6} \tilde{u}_0(0)+\tilde{u}_1(0)=\alpha, \qquad \tilde{u}_0(1)+\tilde{u}_1(1)=\beta ,$$ we find $$\label{pw7} \tilde{u}_1(x)=(\alpha -\beta +1) \left({\rm erf}\left(\frac{x+\beta -1}{\sqrt{2\epsilon } }\right)-{\rm erf}\left(\frac{\beta }{\sqrt{2\epsilon} }\right)\right) \slash \left({\rm erf}\left(\frac{\beta -1}{\sqrt{2\epsilon}}\right)-{\rm erf}\left(\frac{\beta }{\sqrt{2\epsilon }}\right)\right).$$ The numerical approximation of $\tilde{u}_0(x)+\tilde{u}_1(x)$ is better than any of the MAE or resummations so far (see figures \[comp1\], \[comp2\]). This is because the boundary conditions can here be satisfied at both ends by the leading two orders $\tilde{u}_0(x)$, $\tilde{u}_1(x)$. Note that the exact forms (\[pw4\]), (\[pw7\]) must incorporate resummations of the $r$ series in algebraic powers of $\epsilon$, and this is here actually the reason why both boundary conditions have here been able to be satisfied. That we can achieve this is due to the fortunate peculiar form of (\[pw5\]) due to the truncation of the expansion in (\[pw4\]). This serendipity would not extend to general systems and the $r$ sum would not have been automatically resummed. We can also solve the linear inhomogeneous equation at ${\mathcal O}(\lambda^2)$ for (\[node\]) with boundary conditions $u_2(0)=u_2(1)=0$. Addition of $\tilde{u}_2(x)$ to $\tilde{u}_0(x)+\tilde{u}_1(x)$ improves the numerical agreement even further (figure \[comp2\]). ![Comparison of relative percentage errors for $\epsilon=1/10$, $1$ with $\alpha=3/2, \beta=2$ for resummed approximations using only $\tilde{u}_0(x)$, $\tilde{u}_0(x)+\tilde{u}_1(x)$ and $\tilde{u}_0(x)+\tilde{u}_1(x)+\tilde{u}_2(x)$. Note the satisfaction of the boundary condition at both ends and the rapid apparent convergence.[]{data-label="comp2"}](figure6.eps) The form of $\tilde{u}_1(x)$ in relation to the double-transseries template requires an explanation. If we substitute the asymptotic expansion ${\rm erf}(s)\sim 1-\exp \left(-s^2\right)/\sqrt{\pi } s$, $s \rightarrow +\infty$, into $\tilde{u}_1(x)$ we find $$\label{pw10} \tilde{u}_1(x)\sim v(x)-v(1) \qquad v(x)=\frac{(\beta-1)(\alpha-\beta+1)\re^{-F(x)/\epsilon}}{(x+\beta-1)\left(1-(\beta-1)\re^{-F(1)/\epsilon}/\beta\right)},$$ where $F(x)$ is as defined in the sections above. At first sight, the form of $v(x)$ suggests that $\tilde{u}_1(x)$ might be interpreted as the ${\mathcal O}(\epsilon^0)$ term of the resummed $p$-series for $n=1$, (\[nresum\]). However a simple calculation shows this not to be the case. In fact $\tilde{u}_1(x)$ does not satisfy the same boundary condition as the resummed (\[nresum\]). The reordering is more complicated with terms of ${\mathcal O}(\re^{-F(x)/\epsilon})$ appearing also in the higher order $\tilde{u}_k$. Higher order $\tilde{u}_k$ satisfy inhomogeneous linear differential equations $$\label{pw15} \epsilon \tilde{u}_k''(x)+(x+\beta -1)\tilde{u}_k'(x)=f(\tilde{u}_0, \tilde{u}_0', \tilde{u}_1, \tilde{u}_1', \dots \tilde{u}_{k-1}), \qquad \tilde{u}_k(0)=\tilde{u}_k(1)=0.$$ The solutions of the homogeneous form equation will be of the same form for all $k$: $$\label{pw16} A_k(\epsilon)v_1(x)+B_k(\epsilon)v_2(x)\re^{-F(x)/\epsilon}, \qquad v_l(x)\sim\sum_{r_{\rm min}}^\infty d_r^{(l)}(x)\epsilon^r, \ l=1,2.$$ Terms of ${\mathcal O}(\exp(-jF(x)/\epsilon))$, $0\le j\le k$ will arise from the inhomogeneous terms and so the $A_k(\epsilon)$ and $B_k(\epsilon)$ will contain terms in ${\mathcal O}(\re^{-jF(1)/\epsilon}), \ 0\le j\le k$ to satisfy the boundary conditions. Hence the expansion (\[pw1\]) can be seen as non-trivial reordered sum over $p$ and $r$, ordered by the index $n$. For this particular system, due to the truncated nature of $\tilde{u}_0$, we are able to solve for the $\tilde{u}_k$ exactly and so effectively achieve this double resummation. For situations where $\tilde{u}_0$ does not truncate this will not be achievable in general. Note that the transseries template is forced by the presence of boundary values, but the form of the lambda series is driven by the differential equations. It is an open question as to what is the ordering of terms in the transseries that optimises the numerical agreement with the exact solution, when only a finite number of terms is taken in the approximation. Discussion ========== We conclude with a brief comparison of our nonlinear results with those of Wasow (1956) and a discussion of further work. In our notation, Wasow considered (\[Wasoweq\]) with $F_1(x,u,\epsilon)$ and $F_2(x,u,\epsilon)$ regular analytic with respect to $u$, $\epsilon$ and $C^{(2)}$ in $x$ in a region of $(x,u,\epsilon)$ space containing $u=u_0(x)$, $0\le x \le 1$ and $\epsilon=0$. Wasow’s expansion takes the form $$u(x,\epsilon)=\sum_{r=0}^{\infty}v_r(x,\epsilon)\mu^r, \qquad v_r(x,\epsilon)=\re^{-F(x)/\epsilon}w_r(x,\epsilon), \ r>0,$$ where $\mu=\alpha-v_0(0,\epsilon)=\alpha-\beta+1$, $v(1,\epsilon)=\beta$, the $w_r(x,\epsilon)$ are bounded and $F(x)$ is as defined above. He proved that the $v_r$ expansion is uniformly and absolutely convergent with respect to $\epsilon$, $\mu$ and $x$ for $0\le\epsilon\le\epsilon_1$, $|\alpha-\beta+1|\le \mu_1$, $0\le x\le 1.$ Wasow then reordered the terms to give $$u(x,\epsilon)=\sum_{m=0}^\infty c_m(x,\alpha,\epsilon)\re^{-mF(x)/\epsilon}$$ as uniformly and absolutely convergent. He stated that the coefficients $c_m(x,\alpha, \epsilon)$ are regular analytic and possess a power series in $\epsilon$. He omited the full proof as it was, in his own words, “somewhat detailed". Wasow’s results are existence proofs. He took the resummed route directly, his “reordering" is related to the transseries. Here we have taken the reverse route of solving for the transseries first and then resumming as we have been motivated by an exponential asymptotics approach. Our explicit calculation reveals the intricate and subtle nature of the form of Wasow’s $c_m$ or $w_r$ coefficients and reveals an interesting realignment of exponential scales that are required to satisfy the boundary values. The template has been dictated by the boundary values and is independent of the equation and so is expected to be more generally valid. Wasow’s statements that the coefficients “possess an asymptotic expansion in powers of $\epsilon$" is strictly incomplete, since it ignores the presence of series beyond all orders, and in inverse powers of the asymptotic parameter. Given all this, Wasow’s work suggests that our resummations may be convergent, at least for small values of $\mu=\alpha-\beta+1$. The preliminary work of this paper has opened up a multitude of potential follow-up problems in exponential asymptotics. Further work could include: rigorous justification of the exponential approach, with a proof (or otherwise) of convergence; derivation of proper error bounds for the reordered summations; extensions to higher order BVPs; adjustment of the template to include interior layers and shocks; a full hyperasymptotic treatment based on the fundamental transseries template; extensions of the transeries template to PDE BVPs. Some of these will be discussed elsewhere. The author acknowledges the hospitality of the Pacific Institute for the Mathematical Sciences and the Department of Mathematics at UBC, Vancouver, where some of this work was undertaken, and AB Olde Daalhuis for very helpful discussions. Motivated by the form of (\[psum\]) we seek, [*a priori*]{}, a multiple-scales solution of (\[1\]) that corresponds to the summation of all the $p$-exponentials in (\[template\]). We use the ansatz: $$\label{Wansatz} %&\sim&\sum_{r=0}^{\infty}a_r(x)\epsilon^r+\sum_{p=1}^{\infty}\re^{-pF(1)/\epsilon}\sum_{r=0}^{\infty}a_r^{(p)}(x)\epsilon^r+\sum_{p=0}^{\infty}\re^{-(F(x)+pF(1))/\epsilon}\sum_{r=0}^{\infty}b_r^{(p)}(x)\epsilon^r \\ u(x;\epsilon)\sim \sum_{r=0}^{\infty}W_r(x,X)\epsilon^r+\re^{-F(x)/\epsilon}\sum_{r=0}^{\infty}V_n(x,X)\epsilon^r,$$ where the scaled “variable" is the constant $X=F(1)/\epsilon$. We substitute(\[Wansatz\]) into (\[1\]) and balance at ${\mathcal O}(\epsilon^r)$ and ${\mathcal O}(\re^{-F(x)/\epsilon}\epsilon^r)$, ignoring the $\epsilon$-dependence in the $X$ terms. Since $X$ is a constant, it does not actually generate a derivative in $\partial/\partial X$. We thus obtain recurrence relations that are identical to those we obtained when we substituted (\[latta\]) into (\[1\]). Hence the recurrence relations for W and V are, for $r\ge 0$, $$\label{Wc} W_r(x,X)=(c_r-W_{r-1}'(x,X))/(2x+1), \ \ V'_r(x,X)=(V_{r-1}''(x,X))/(2x+1),$$ with $W_{-1}(x,X)=$ $V_{-1}(x,X)$ $=0$ and prime denoting $x$-differentiation. The constants $c_r$ are determined from a modified set of boundary conditions. The terms in the expansion in (\[bc0\]), (\[bc1\]) and (\[pbcs0\]) have been balanced at orders of ${\mathcal O}(\re^{-F(x)/\epsilon}\epsilon^r)$. Starting from (\[Wansatz\]) due to the apparent similarity of (\[latta\]) and the second row of (\[Wansatz\]), we find that we go around in a circle. The second sum in (\[Wansatz\]) cannot satisfy the boundary condition at $x=1$ exactly, without including further series, for the reasons outlined above. Thus the boundary conditions have to be modified. Instead of (\[bc0\]-\[bc1\]) we have: $$\begin{aligned} \label{Wbc0} W_r(0,X)+V_r(0,X)=\delta_{r0}\alpha, \ \ W_r(1,X)+\re^{-X}V_r(1,X)=\delta_{r0}\beta. \end{aligned}$$ Note that the exponential in (\[Wbc0\]) is actually $\re^{-F(1)/\epsilon}$, precisely the order of neglected terms in (\[bc1\]) which lead to the ansatz (\[template\]). That we can include this exponential term now is because we treat $X$ as varying on a different scale to $\epsilon$. From (\[Wbc0\]) the $c_r$ in (\[Wc\]) can now be found. A short calculation gives $$\begin{aligned} \label{Wr} \nonumber W_r(x,X)&=&\frac{(W_{r-1}'(1,X)-W_{r-1}'(x,X))-3\re^{-X}(W'_{r-1}(0,X)-W'_{r-1}(x,X))}{(1-3\re^{-X})(2x+1)}, \\ \label{Vr} V_r(x,X)&=&(W_{r-1}'(0)-W_{r-1}'(0))/(1-3\re^{-X}).\end{aligned}$$ with initial terms $W_0(x,X)=3(\beta-\alpha\re^{-X})/\left\{(1-3\re^{-X})(2x+1)\right\}$ and $V_0(x,X)=(\alpha-3\beta)/(1-3\re^{-X})$. Inserting the value of $\epsilon X=F(1)=2$, to leading order in $\epsilon$ we recover (\[psum\]). Note that by neglecting the terms in $\re^{-X}$ these relations reduce to those of (\[aterm\]-\[bterm\]), as they should. Berry, M.V. 1989 Uniform asymptotic smoothing of Stokes discontinuities. [*Proc. Roy. Soc. Lond. A*]{} [**422**]{}, 7-21. Berry M.V., & Howls C.J. 1991 Hyperasymptotics for integrals with saddles. [*Proc. Roy. Soc. Lond. A*]{} [**434**]{}, 657-675. Boyd W.G.C 1994 Gamma function asymptotics by an extension of the method of steepest descents. [*Proc. Roy. Soc. Lond. A*]{} [**447**]{}, 609-630. Chapman, S. J., Howls, C. J., King, J. R. & Olde Daalhuis, A. B. 2007 Why is a shock not a caustic? The higher order Stokes phenomenon and smoothed shock formation. [*Nonlinearity*]{} [**20**]{}, 2425-2452. Chapman, S. J., King J. R. & Adams K. L. 1998 Exponential asymptotics and Stokes line in nonlinear ordinary differential equations. [*Proc. Roy. Soc. Lond. A*]{} [**454**]{}, 2733-2755. Costin, O. 1998 On Borel summation and Stokes phenomena for rank-1 nonlinear systems of ordinary differential equations. [*Duke Math. J.*]{} [**93**]{}, 289-344. Costin, O. & Costin, R. D. 2001 On the formation of singularities of solutions of nonlinear differential systems in antistokes directions. [*Invent. Math.*]{} [**145**]{}, 425-485. Delabaere, E. & Howls, C. J. 2002 Global asymptotics for multiple integrals with boundaries. [*Duke Math. J.*]{} [**112**]{}, 199-264. DLMF to appear 2010 [*Digital library of mathematical functions*]{}. National Institute of Standards and Technology from [http://dlmf.nist.gov/25.12\#E1]{}. Dingle, R. B. 1973 [*Asymptotic expansions, their derivation and interpretation*]{}. London: Academic Press. Howes, F.  A. 1978 Boundary-interior layer interactions in nonlinear singular perturbation theory, [*Memoirs of the American Mathematics Society*]{} [**15**]{} No.203. Howls, C. J., 1992 Hyperasymptotics for integrals with finite endpoints. [*Proc. Roy. Soc. Lond. A*]{} [**439**]{}, 373-396. Howls, C. J., 1997 Hyperasymptotics for multidimensional integrals, exact remainder terms and the global connection problem, [*Proc. Roy. Soc. Lond. A.*]{} [**453**]{}, 2271-2294. Howls, C. J., Langman, P. J. & Olde Daalhuis, A. B. 2004 On the higher order Stokes phenomenon. [*Proc. R. Soc. Lond. A*]{} [**460**]{}, 2285-2303. Howls, C. J. & Olde Daalhuis, A. B. 2003 Hyperasymptotic solutions of inhomogeneous linear differential equations with a singularity of rank one. [*Proc. R. Soc. Lond. A*]{} [**459**]{}, 2599-2612. Lee, J. Y. & Ward, M. J. 1995 On the asymptotic and numerical analysis of exponentially ill-conditioned singularly perturbed boundary value problems. [*Studies in Appl. Math.*]{} [**94**]{}, 271-326. Nayfeh, A. H., 1973 [*Perturbation Methods*]{}. NewYork: John Wiley & Sons. Ólafsdóttir, E. I., Olde Daalhuis, A. B., & Vanneste J. 2005 Stokes-multiplier expansion in an inhomogeneous differential equation with a small parameter [*Proc. R. Soc. Lond. A*]{} [**461**]{}, 2243-2256. Olde Daalhuis, A. B. 1993 Hyperasymptotics and the Stokes phenomenon. [*Proc. R. Soc. Edin.*]{} [**123**]{}, 731-743. Olde Daalhuis, A. B. 1998 Hyperasymptotic solutions of higher order linear differential equations with a singularity of rank one. [*Proc. R. Soc. Lond. A*]{} [**454**]{}, 1-29. Olde Daalhuis, A. B. & Olver, F. W. J. 1998 On the asymptotic and numerical solution of linear ordinary differential equations. [*SIAM Rev.*]{} [**40**]{}, 463-495. Olde Daalhuis, A. B. 2004 On higher order Stokes phenomena of an inhomogeneous linear ordinary differential equation. [*J. Comput. Appl. Math.*]{} [**169**]{}, 235–246. Olde Daalhuis, A. B. 2005$a$ Hyperasymptotics for nonlinear ODEs I: A Riccati equation. [*Proc. R. Soc. Lond.A.*]{} [**461**]{}, 2503-2520. Olde Daalhuis, A. B., 2005$b$ Hyperasymptotics for nonlinear ODEs II: The first Painlevé equation and a second-order Ricatti equation. [*Proc. R. Soc. Lond. A.*]{}[**461**]{}, 3005-3021. O’Malley, R. E. 1969 On a boundary value problem for a nonlinear differential equation with a small parameter, [*SIAM  J.  App.  Math.*]{} [**17**]{}, 569-581. O’Malley, R.  E., & Ward, M.  J., 1997 Exponential asymptotics, boundary layer resonance, and dynamic metastability. In [*Mathematics is for Solving Problems*]{} (ed. P. Cook [*et al.*]{}). SIAM, pp. 189Ð203. Ou, C. H., & Wong, R., 2003 On a two-point boundary value problem with spurious solutions. [*Stud. App. Math.*]{} [**111**]{}, 377-408. Ou, C. H., & Wong R., 2004 Shooting method for nonlinear singularly perturbed boundary value problems. [*Stud. App. Math.*]{} [**112**]{}, 161-200. Segur, H., Tanveer, S., & Levine, J. (Eds.) [*Asymptotics beyond all orders*]{}. NATO Science Series B: Physics [**284**]{}. Ward, M. J. 2005 Spikes for singularly perturbed reaction-diffusion systems and Carrier’s Problem. In [*Differential equations and asymptotic theory in mathematical physics*]{} (ed. C. Hua & R. Wong). Series in Analysis Vol. 2, World Scientific, Singapore, pp. 100-188. Wong, R., & Yang H. 2002 On a boundary layer problem. [*Stud. App. Math.*]{} [**108**]{}, 369-398. Wong, R., & Zhao Y. 2008 On the number of solutions to Carrier’s problem. [*Stud. App. Math.*]{} [**120**]{}, 213-245. Wasow, W. 1956 Singular perturbations of boundary value problems for nonlinear differential equations of the second-order. [*Comm. Pure App. Math.*]{} [**9**]{}, 93-113. \[lastpage\]

ArXiv

--- abstract: | We show how analysis of a quasar high-magnification microlensing event may be used to construct a map of the frequency-dependent surface brightness of the quasar accretion disk. The same procedure also allows determination of the disk inclination angle, the black hole mass (modulo the caustic velocity), and possibly the black hole spin. This method depends on the validity of one assumption: that the optical and ultraviolet continuum of the quasar is produced on the surface of an azimuthally symmetric, flat equatorial disk, whose gas follows prograde circular orbits in a Kerr spacetime (and plunges inside the marginally stable orbit). Given this assumption, we advocate using a variant of first-order linear regularization to invert multi-frequency microlensing lightcurves to obtain the disk surface brightness as a function of radius and frequency. The other parameters can be found by minimizing $\chi^2$ in a fashion consistent with the regularized solution for the surface brightness. We present simulations for a disk model appropriate to the Einstein Cross quasar, an object uniquely well-suited to this approach. These simulations confirm that the surface brightness can be reconstructed quite well near its peak, and that there are no systematic errors in determining the other model parameters. We also discuss the observational requirements for successful implementation of this technique. author: - Eric Agol and Julian Krolik title: Imaging a Quasar Accretion Disk with Microlensing --- =cmbsy10 =cmmib10 Introduction ============ Due to their great distance and small intrinsic size, it is not possible to obtain a resolved optical image of a quasar with current technology. The angular size of a quasar optical emission region is of order ${10^{-8}}^{\prime\prime}$, a scale so small as to require a baseline of several thousand kilometers to resolve. Thus, until optical VLBI becomes practical, quasar structure will need to be probed by other, indirect, means. Reverberation mapping provides an instructive example of the difficulties of such indirect approaches, for it has proven difficult both to implement and to interpret. We believe that the subject of this paper, microlensing by stars in an intervening galaxy, is a more promising method. The circumstantial evidence that black holes power quasars is convincing: accretion onto black holes can be very efficient in converting rest mass energy to photons (up to 40%) or to bulk momentum (forming radio jets/lobes) in a compact region which has an effective temperature near where the quasar spectrum peaks. There is statistical evidence that quiescent black holes in the nuclei of galaxies at low redshift could be the remnants of quasars. And, quasars have properties very similar to Seyfert galaxies, for some of which there is good spectroscopic evidence for a central black hole. The nature of the accretion flow is quite uncertain, however, although it is probably geometrically thin since angular momentum will support the accretion flow against collapse and geometrically thick disks are by nature inefficient (if quasar disks were thick, there would be problems producing the huge luminosities observed within a reasonable mass budget). Attempts to constrain the character of the innermost accretion flow by means of spectral modeling have made little progress for several reasons. There are major systematic uncertainties about fundamental issues (e.g., the vertical distribution of energy dissipation, the physics to include in radiation transfer solutions). In addition, any particular model depends on a sizable number of free parameters (mass of the central black hole, accretion rate, viscosity parameter, inclination angle), so that parameter estimation is tricky. Analysis of a microlensing event is potentially a more powerful tool. Rather than guess a specific model for the accretion disk, the history of magnification in the event can be used to directly infer the disk surface brightness as a function of radius and frequency. The only assumptions required are that the continuum emission surface is geometrically flat, the material forming it follows circular orbits, and general relativity determines the dynamics of both the matter and the photons. To apply this technique requires study of a particular gravitationally-lensed quasar. Many are now known, and some appear to undergo occasional fluctuations due to individual stars in the lens galaxy magnifying the quasar, a phenomenon referred to as microlensing. The Einstein Cross is particularly well-suited to this problem for a number of reasons that we will detail in §1.3. Several authors have attempted to constrain the character of the accretion disk in this system by comparing predicted lightcurves to the data compiled during microlensing events (Jaroszyński et al. 1992, Rauch & Blandford 1991, Jaroszyński & Marck 1994, Czerny et al. 1994). However, the results have all been somewhat inconclusive due to the ordinary spectral modeling difficulties described above. The remainder of this paper is organized as follows: In §1, we describe how a caustic crossing can be used to perform the mapping, and discuss why the Einstein Cross is such a suitable target for this sort of study. In §2, we discuss the model in detail, considering our assumptions, the inversion technique, and error propagation. In §3, we present a variety of simulations of caustic-crossing events, and demonstrate that we can measure most model parameters. We show some examples of the accuracy with which the intensity at the accretion disk may be determined, and discuss how the reliability of the results depends on the quality of the observational data, on the character of the monitoring, and, to some degree, on the character of the regularization scheme. In the final section, we discuss the results and present our conclusions. Smooth thin disk ---------------- For a standard thin accretion disk with constant accretion rate and no advection of heat, the energy generation per unit area as a function of radius is given by $$\label{energy} Q = {3 \over 4\pi}{ GM_{BH} \dot M \over r^3} R_R (r).$$ where $M_{BH}$ is the mass of the black hole, $\dot M$ is the accretion rate, $r$ is the radius within the disk, and $R_R$ is a correction factor that combines outward advection of energy associated with the angular momentum flux and relativistic effects (Page & Thorne 1974, notation from Krolik 1998). $R_R$ is a function of the black hole spin $a_s = a/M_{BH}$. Though equation \[\[energy\]\] describes the functional dependence of the energy released with radius, it does not specify whether this energy is thermal or mechanical, and does not specify the dissipation as a function of height within the accretion disk. The appearance of a standard thin accretion disk can vary significantly depending on how and where the energy is released. To further our understanding, it would be very desirable to actually measure the local spectrum. One potential method to achieve this is to observe a quasar during the sort of high-magnification microlensing event that occurs when a caustic crosses the source (Grieger et al., 1988 and 1991, Gould & Gaudi 1997). Grieger et al. (1991) advocate inverting the microlensing lightcurve to obtain the one-dimensional surface brightness of the quasar, $P_\nu(x)=\int I_\nu(x,y) dy$, where $I_\nu(x,y)$ is the specific intensity of the quasar at sky coordinates $(x,y)$. Their method is quite elegant, but relies on first-order regularization (the assumption that the quasar profile is smooth). This assumption is problematic for black hole models since the energy release increases rapidly towards smaller radii, and relativistic effects can cause sharp peaks in the quasar profile. These sharp features are smoothed out when this method is used (see Figure 1). We believe a better approach is to solve instead for the surface brightness at the accretion disk, with the relatively benign assumptions that the disk is planar, axisymmetric, and isotropically emitting (in the fluid frame), and most importantly, likely to vary smoothly with radius. This approach provides a better inversion of the quasar profile if the disk assumptions are correct. This method recovers the spectrum as a function of radius at the accretion disk, which can then be compared with disk atmosphere models or other spectral modeling. In addition, it can help constrain disk parameters, such as the inclination angle, without relying on a specific accretion disk spectral model. Fig. 1.Plot of recovered one-dimensional disk profile as a function of impact parameter for a face-on accretion disk using the Grieger et al. (1991) technique. Solid line is the model profile; dotted line is recovered profile. The parameter $\alpha$ is distance across the source perpendicular to the caustic line in units of gravitational radii. High magnification microlensing event ------------------------------------- The spectrum of a quasar as a function of time is a convolution of the lensing magnification with the surface brightness of the quasar. Near a fold caustic, two bright images merge (Schneider et al. 1992). The resulting magnification has the specific form: $$\label{camp} A(x,t)=A_0 + {K \over \sqrt{x-v_c(t-t_0)}} \Theta\left[x-v_c(t-t_0)\right],$$ where $A_0$ is the magnification due to the additional images, $x$ is the position on the quasar plane perpendicular to the caustic, $K$ is the strength of the caustic \[units of $({\rm distance})^{1/2}$\], $v_c$ is the speed of the caustic (assumed to be positive), $\Theta$ is the step function \[$\Theta(x)=1$ for $x>0$, $0$ otherwise\], and $t_0$ is the time at which the caustic crosses the $x=0$ point measured relative to the center of the lightcurve (Schneider & Weiß 1986). This equation is valid whenever the source size is small compared to the Einstein ring radius of the microlensing star. If that ratio is small, there is only a small probability of any of a number of problems that might invalidate equation \[\[camp\]\]. The list of potential problems includes: 1) the possibility that the source is projected behind a cusp (where the caustic curve discontinuously changes direction) or behind the crossing of multiple fold caustics; 2) significant curvature in the caustic; 3) variation of $K$ along the caustic on the scale of the source. Grieger et al. (1988) demonstrate that a source of size smaller than $\sim 10$% of the Einstein radius of the typical lensing mass is necessary for the second and third assumptions to be valid. Microlensing Laboratory: The Einstein Cross ------------------------------------------- The Einstein Cross is a quasar (z=1.695) lensed into four images by a nearby (z=0.0394) barred spiral galaxy (Huchra et al. 1985). The luminosity distance to the quasar is $D_L=3\times 10^{28} h_{75}^{-1}$ cm, where $h_{75}=H_o$/(75 km/s/Mpc), assuming $\Omega_m=1,\Lambda=0$. This quasar has been observed to undergo fluctuations due to microlensing roughly once per year (Irwin et al. 1989; R. Webster, private communication). It is particularly well suited for studying microlensing because the lensing galaxy is nearby. This happy coincidence makes the time delays between the four images all less than a day, so that it is easy to distinguish intrinsic variability from microlensing variability. It also makes the stellar velocities projected onto the source plane quite high, so that microlensing is frequent, and also makes individual events comparatively brief (only about a month, in contrast to the years to decade timescales characteristic of more distant lenses). In addition, the Einstein radius projected onto the quasar plane is quite large compared to the quasar size, validating the assumptions made in the previous subsection, and also making the microlensing variations strong. The models of Witt et al. (1993) suggest that for image A, $$\langle K \rangle \simeq 8 \left({r_E\over 5.8\times 10^{16}{\rm cm}} \right)^{1\over 2} M_9^{-{1\over 2}},$$ where $r_E=5.8\times 10^{16}{\rm cm}(m/0.2)^{1/2} h_{75}^{-1/2}$ is the Einstein radius projected to the quasar plane (Schneider et al. 1992), $m$ is the typical mass (in Solar units) of a star causing microlensing, and $M_9=M_{BH}/10^9 M_\odot$. In equation (3), and in the rest of the paper, we adopt $r_g \equiv GM/c^2$ as the unit of distance. For concreteness, we will use $K=8$ in our simulations. The parameter $A_0$ can be approximated by $\langle A_0 \rangle = |(1-\sigma)^2-\gamma^2|^{-1}$ where $\gamma$ is the shear (Witt et al. 1993). For image A, the estimated range of the microlensing parameters is $\sigma=0.3-0.4$ and $\gamma=0.4-0.5$, giving $A_0=3-9$; we use $A_0=6$ in our simulations. Microlensing of an Accretion Disk ================================= Caustic crossing ---------------- As the caustic crosses the quasar, the observed lightcurve is a convolution of the magnification, equation \[\[camp\]\] and $P_\nu(x)$: $$\label{eqonedim} F_\nu(t)=\int_{-\infty}^{\infty} A(x,t) P_\nu(x) dx.$$ Note that if the flux within a waveband can be measured outside of the caustic and subtracted off, then the dependence on $A_0$ disappears, and the $K$ parameter becomes degenerate with an arbitrary scaling of $P_\nu(x)$. As discussed by Grieger et al. (1991), equation \[\[eqonedim\]\] can be inverted using regularization to find $P_\nu(x)$. Similar techniques have been used for measuring the limb-darkening of stars during galactic microlensing events (Gaudi & Gould 1998, Albrow et al. 1998). Black Hole geometry ------------------- Near a black hole, relativistic effects cause Doppler beaming of the emitted radiation, gravitational red shifts, and bending of photon trajectories. To image the surface of an accretion disk, these relativistic effects must be accounted for using a relativistic transfer function (defined in Cunningham 1975). To compute the transfer function (see equation \[\[fint\]\] below), we make several simplifying assumptions: (1) the accretion disk is thin, i.e. $h \ll r$; (2) the gas follows prograde circular orbits outside the marginally stable radius $r_{ms}$, and undergoes freefall within $r_{ms}$ with constant angular momentum and energy equal to those obtaining at $r_{ms}$; (3) the disk is flat and lies in the equatorial plane of the black hole; (4) the gas emits isotropically in its rest frame, i.e., there is no limb darkening in the accretion disk atmosphere. The first two assumptions are appropriate if pressure gradients cause forces much smaller than the gravitational force in the $z$ and $r$ directions, respectively. This condition is not met in advection-dominated accretion flows or slim accretion disks (Beloborodov 1998). In the case of slim accretion disks, the orbital frequency is nearly Keplerian, and deviates by less than 20% when $\dot M c^2/L_{Edd}\leq 1000$; however, the disk scale height can become a large fraction of the radius, which changes the emitted angle of radiation relative to the disk normal and can cause shadowing which we do not take into account. The third assumption is inappropriate if the disk is warped; however, Bardeen-Petterson precession (1975) can align the disk and black hole by the time the gas reaches the inner radii. The fourth assumption is a simplification for greater ease in the inversion computation since the disk can be viewed from only one angle and thus at most one emitted angle can be observed at each radius/azimuth of the disk. For each radius, there is a limited range of emitted angles which are observed, so our inversion will give some sort of average of the intensities within that range. Fig. 2. The upper curve shows the maximum $\mu_e=\cos{\theta}$, where $\theta$ is the angle between the disk normal and the direction of the photon in the fluid rest frame. The lower curve shows the minimum $\mu_e$. The disk parameters are $i=30^\circ$ and $a_s=0.998$. Figure 2 shows the range of emitted angles (in the fluid rest frame, $\mu_e$ is the cosine of the normal to the disk) for a disk inclined at 30$^\circ$. For a face-on disk, only one emitted angle is seen at each radius for all azimuths, so this assumption simply corresponds to mapping the specific intensity of the disk at $\mu_e(r)$. We could have assumed some limb-darkening law, but this is not warranted by the crudeness of the inversion technique. Figure 3 shows the disk geometry. The inclination angle of the accretion disk, $\mu=\cos{i}$, is $i=0^\circ$ when the disk is face-on, and 90$^\circ$ when the disk is edge-on. The caustic crossing angle $\phi_c$ is measured with respect to the $(\alpha,\beta)$ coordinates, which are defined so that the $\beta$ coordinate lies parallel to the projection of the disk axis onto the sky plane (in Figure 3, the disk spin axis is pointing out of the page), with the black hole at the origin. We will use units of $r_g$ for the $(\alpha,\beta)$ coordinates. Fig. 3. Geometry of the accretion disk. The disk axis points up out of the page, while the $\beta$ axis lies in the page. The line parallel to the $y$ axis is the caustic. The rotation of the accretion disk causes beaming of the radiation, leading to a hot spot on the approaching side, and a cold spot on the receding side. For an exactly face-on disk, the disk is symmetric, so no hot/cold spots exist, but a dip occurs inside the inner edge of the disk. In Figure 4 we show $P_\nu(x)$ for a blackbody accretion disk at two frequencies. As the disk becomes more edge-on, the profile becomes more asymmetric as the Doppler aberration becomes stronger. If the temperature of the disk decreases outwards, then the size of the hot spot will increase for smaller frequencies, and become less asymmetric, as can be seen by comparing Figure 4(a) and 4(b). Figures 4(c) and 4(d) show $P_\nu(x)$ for different caustic crossing angles, showing the hotspot is oblong. Fig. 1. Plots of the disk 1-D profile for a disk with $M_9=1$, $\dot m = 1$, and $a_s=0.998$. Figures (a) and (b) compare inclination angle: $i=$ 0$^\circ$ (solid), 30$^\circ$ (dotted) and 60$^\circ$ (dashed) for $\phi_c=0$. Figures (c) and (d) compare $\phi_c=$ 0 (solid), 30$^\circ$ (dotted), 60$^\circ$ (short dashed), 90$^\circ$ (long dashed) for $i=30^\circ$. The frequencies are in the quasar rest frame. Prediction of the Lightcurve ---------------------------- With these assumptions and parameter choices, the observed flux may be predicted from the run of intensity with radius: $$\label{fint} F(\nu_o,t)=\int d\alpha d\beta A\left[x(\alpha,\beta),t\right] g^3 I(\nu_o/g,r_e),$$ where $g = \nu_o/\nu_e$ is the redshift between the observer and the emitter and $I(\nu_e,r)$ is the specific intensity at the accretion disk (we have assumed it is independent of the emitted angle). The relation between the caustic coordinate and the black hole coordinates is $x(\alpha,\beta) =\alpha\cos{\phi_c}+\beta\sin{\phi_c}$ (see Figure 3). The physical variables $\nu_e$ and $r_e$ may be related via a Jacobian to $\alpha$ and $\beta$ through their functional dependence on redshift $g(\mu_o,a_s)$, the inclination angle of the disk $\mu_o$, and the black hole spin $a_s$. The magnification $A$ depends on $v_c$, $t_0$, $\phi_c$, $K$, and $A_0$, for a total of seven model parameters. We compute the transfer function by shooting rays from infinity at a grid in $(\alpha, \beta)$ until they cross the equatorial plane. The computational method is based on Rauch & Blandford (1994), and is described in Agol (1997). We use a nested grid of rays that is more finely sampled towards the center to resolve the inner parts of the accretion disk in greater detail. For a given observed frequency $\nu_o$, we compute $\nu_e$ as well as the emitted radius at each $(\alpha,\beta)$, and interpolate these on the (pre-specified) grid of radii and frequencies at the accretion disk. To compute a normal transfer function, we would then simply sum over the grid. In this case, we multiply each ray by a further factor that describes the magnification due to the microlensing at any particular time, and then sum over the grid. This procedure may be summarized $$\label{mateq} F_i = T_{ij} I_j,$$ where $i=1,N_F$ ($N_F= N_t\times N_\nu$ = \# observed frequencies $\times$ \# observed times) labels each measured observed frequency/time, and $j=1,N_I$ ($N_I=$ \# emitted frequencies $\times$ \# emitted radii) labels each emitted frequency/radius pair. The matrix ${\bf T}$ contains the integration and interpolation factors. Regularized Inversion --------------------- Attempting to directly invert equation \[\[mateq\]\] for ${\bf I}$ given an observed set of ${\bf F}$ is impossible since the matrix ${\bf T}$ is generally singular, so that noise in the lightcurve is magnified strongly during inversion. This fact requires the introduction of some sort of [*a priori*]{} knowledge in order to make inversion feasible. Regularization is a particularly useful way to do this, as discussed in Press et al. (1992), because the “prejudice" injected into the solution is usually relatively benign and also relatively controllable. The essence of the linear regularization method is to minimize both the deviation of the model from the data, and also the deviation of the model from “smoothness," as defined by some sort of differencing operator. In our case, at any given frequency, we expect the emitted intensity $I_\nu$ to be smooth as a function of radius, but not the one-dimensional profile $P_\nu$. In our specific implementation of the method, we also impose several other restrictions on the solution. We expect that the emitted intensity in the fluid frame diminishes as the black hole event horizon is approached; we therefore require $I_\nu(r)$ to approach zero as $r$ approaches $r_g$. In fact, because those regions are so strongly redshifted from almost any inclination angle, the intensity in the fluid frame is almost completely unconstrained by the data, so physical assumptions have a very strong impact on the solution in this region. Similarly, we also require $I_\nu$ to approach zero at very large radii, for there is little energy available there to dissipate. Particularly at low frequency, it may sometimes be desirable to relax this constraint. In addition, we would like the inverted intensities to be positive definite; to achieve this, we maximize $A_I=(\sum_i I_i w_i)/N_I$, where $w_i$ is a weighting factor. The most appropriate weighting is $w_i=r_i^2$ since the radiating area associated with each logarithmic radius interval scales as $r_i^2$. Finally, for any choice of grid, there will always be some radius/emitted frequency pairs that are not constrained by the data because Doppler shifts push $\nu_o$ outside the observed region (these are the intensities for which the corresponding column of ${\bf T}$ is all zeros). In order to prevent those frequencies from contributing to the smoothing condition, we require that the associated intensities be zero. Combining all these considerations leads to the following regularization operator: $$\begin{aligned} {\cal B} = {1\over A_I^2}\left\{ {\sum_{i=N_\nu+1}^{N_I} (I_i w_i-I_{i-N_\nu}w_{i-N_\nu})^2 + N_I(I_1w_1)^2+ }\right. \cr \left. {N_I(I_{N_I}w_{N_I})^2+N_I^4\sum_{\{j:{\rm if} T_{ij}=0 \forall i\}} (I_jw_j)^2 }\right\},\end{aligned}$$ where $N_\nu$ is the number of frequency grid points at the accretion disk. The first sum describes the smoothness as a function of radius at each frequency, the two isolated terms give the boundary conditions at the innermost and outermost radii, and the last sum is the factor encouraging minimization of those $I_j$ unconstrained by data. Note that the specific intensity vector is ordered with all the frequencies at one radius grouped together, so that $I_i$ and $I_{i-N_\nu}$ give the intensity at the same frequency, but adjacent radii. The smoothing operator ${\cal B}$ has the useful property of providing a model-independent measure of the “smoothness” of different solutions, due to the normalization by $A_I$. Other differencing schemes might also be used; in the examples we have explored, it makes little difference to the outcome. Although the regularization condition is designed to be relatively innocuous, no such injection of prejudice can be altogether free from consequences (we will discuss the effect of our particular choice in §3.2). We stress that the details of the regularization condition are always subject to “tuning" in the light of either theoretical expectations, or, better, the implications of real data. To solve for ${\bf I}$, we minimize the following function: $$\label{fmin} f({\bf I})={1\over N_F}\chi^2+\lambda{\cal B}$$ with $$\chi^2 = \sum_{i=1}^{N_F}\left({\sum_{j=1}^{N_I}T_{ij}I_j - F_i\over \sigma_i}\right)^2,$$ where $\lambda$ is a constant, and $\sigma_i$ is the error on $F_i$. We start with a direct solution of $\partial f / \partial {\bf I} = 0$, setting $A_I=1$, $$\label{isolve} {\bf I} = \left({\bf M}^T{\bf M} + {\lambda\over A_I} {\bf H}\right)^{-1}{\bf M}^T {\bf G} \equiv {\bf Q}{\bf M}^T {\bf G},$$ where $M_{ij} = T_{ij} /\sigma_j$, $G_i = F_i /\sigma_i$, and ${\bf H}$ is defined such that ${\cal B} = {\bf I}\cdot{\bf H}\cdot{\bf I}$. We then update $A_I$ from the solution, and iterate until $A_I$ converges. In some instances, $A_I$ can become negative after an iteration. If this happens, we re-set $A_I$ to be 0.1 times its value at the previous iteration, and recalculate the step. We assume that the magnification outside the caustic, $A_0$, can be measured from the lightcurve, and subtracted off. Then, the parameter $K$ (the caustic magnification factor) is completely degenerate with the disk surface brightness since a decrease in magnification corresponds to an increase in the surface brightness of the source. Consequently, we can determine the [*shape*]{} of the surface brightness profile, but not its absolute level. Several other parameters also remain to be determined after the direct inversion for the surface brightness profile. We call them collectively $\bzeta =(t_0, v_c, \mu, \phi_c, a_s)$. To find them, we fix $\lambda$ and compute $\chi^2$ over a coarse grid in this five-dimensional parameter space. Starting from the $\bzeta$ giving the smallest $\chi^2$ in this grid, we refine our estimate of these parameters using the Levenberg-Marquardt method. We compute the partial derivatives of $\chi^2$ with respect to the axes in $\bzeta$ space by finite differences. If a parameter with boundaries goes out of bounds, we fix it at the value where it went out of bounds, and keep it fixed throughout the rest of the minimization. This generally occurred with $a_s$ when it was near 0 or 1, and for $\mu$ and $\phi_c$ when the disk was face-on. Fixing this improved estimate for the best-fit $\bzeta$, we increase $\lambda$ and re-solve for the surface brightness profile until $\chi^2 = N_F$. The number of degrees of freedom against which to compare $\chi^2$ is not clearly defined for several reasons. One is that many of the model parameters are not entirely free; for several ($\mu$, $M_{BH}$, $v_c$) there are prejudices or constraints from other experiments. Another reason is the variable weight given the smoothing constraint [*vis-a-vis*]{} the data, as we are minimizing $\chi^2 + \lambda {\cal B}$ rather than $\chi^2$. In the limit of large $\lambda$, there is effectively only one free parameter for each frequency in the fit to the $I_i$; in the limit of $\lambda = 0$, there are as many free parameters as there are grid points. Given these considerations, $N_F$ is an upper bound to the true number of degrees of freedom; by raising $\lambda$ until $\chi^2=N_F$, we ensure that we do not overfit the data. Errors ------ The word “error" has several different meanings in this context, and it is important to distinguish them. First of all, the errors in the inferred intensities have different properties from the errors in the model parameters. Second, both are potentially subject to systematic error as well as random error. We will begin by estimating the random error in the intensities ${\bf I}$. Formally, we may say that $$\label{erri} \delta I_i^2 = \sum_j\left({\partial I_i \over \partial F_j}\right)^2 \sigma_j^2,$$ where ${\partial I_i \over \partial F_j}= \sum_k Q_{ik} T_{kj}/\sigma_j^2 $ (see equation \[\[isolve\]\]), and we assume the fluxes have uncorrelated errors. The $\lambda = 0$ case is of special interest because it reveals which $I_i$ (i.e., which frequency/radius pairs) are so constrained by the data that even without regularization they may be reliably determined. In this limit, ${\partial I_i \over \partial F_j} = T^{-1}_{ij}$, so the uncertainty in $I_i$ is given by $\delta I_i^2 = W^{-1}_{ii}$ where $W_{ij}= \sum_k T_{ki}T_{kj}/(\sigma_i \sigma_j)$. ${\bf W}^{-1}$ can be computed by singular value decomposition; in practice, we replace the singular values with a small number. We show an example of this procedure in Figure 9. The “formal accuracy" of our inversion is illustrated by a plot of $U_i \equiv I^r_i/\delta I_i$ (the superscript $r$ stands for recovered intensity) as a function of frequency and radius, computed from equation \[\[erri\]\] (Figure 9). The results for both $\lambda =0$ and the maximum $\lambda$ consistent with the data are shown. In the case $\lambda = 0$, we use the original $I_i$ instead of the recovered values to compute $U_i$. The formal accuracy depends on the true surface brightness: for a given radius, $U_i$ tends to peak where the flux is largest. Also, $U_i$ diminishes at large radii, since those radii aren’t monitored for long enough to truly determine $I_i$. In real solutions $\lambda \neq 0$, and the smoothing operator correlates the intensities at neighboring radii sharing the same frequency. The uncertainties in this case are most easily estimated by a Monte Carlo procedure in which the lightcurves are perturbed by random realizations of noise in the data. The distribution of $I_i$ after re-solving each of these realizations gives the random error in $I_i$. In evaluating these estimates, it is important to understand that points weakly constrained by the data have little sensitivity to measurement errors because they are primarily determined by the smoothness constraint. As a result, their random errors are artificially small. To check for systematic errors in the intensities, we will compute the difference between the original and recovered surface brightness, $\Delta I_i = I^o_i - I^r_i$, where superscript $o$ stands for “original." The systematic error is, of course, far more strongly model-dependent than the random error. It depends on the real intensity distribution, the character of the data (particularly the sampling), and the inversion scheme. These considerations will be discussed at greater length in §\[bestcase\]. We estimate the uncertainties in the model parameters $\bzeta$ two ways: through the same Monte Carlo procedure as for the $I_i$, and also through mapping out the $\chi^2$ found by direct solution of the original data (at fixed $\lambda$) for different choices of $\bzeta$. Examples will be shown in §\[bestcase\], \[varyparam\]. Simulations =========== Range of Parameters Examined ---------------------------- To determine how the inversion works in practice, we performed simulated inversions, varying the parameters describing the underlying model, the parameters describing the data set, and the parameters specifying the details of the solution technique. By varying the model parameters, we learn about whether the method is sensitive to the intrinsic nature of the quasar, or the microlensing event; by varying the observational parameters, we determine what the requirements will be for successful experiments; by varying the solution parameters, we learn how to tune the solution technique for optimum results. ### Model parameters $\bzeta$ In all of our simulations we assume that the intrinsic radiated intensity in the fluid frame is a black body at the local effective temperature, isotropic in the outer half-space. Detailed non-LTE spectra computed for our fiducial parameters (see the next several paragraphs) are consistent with both the observed optical/ultraviolet spectrum and the microlensing size constraint for the Einstein Cross, given the freedom to choose an extinction correction and macrolens magnification (Hubeny & Agol, in preparation). However, we examined a number of possibilities for the other parameters defining the intrinsic character of the quasar and the microlensing events. Although we do not know the mass of the black hole, we may set reasonable bounds on what it could be. If the intrinsic bolometric luminosity is $3\times 10^{46}$ erg/s (Rauch & Blandford 1991), the quasar would be at its Eddington limit if $M_{BH}\sim 2\times 10^8 M_\odot$. On the other hand, the size of the optical emitting region is limited to no more than $\sim 2\times 10^{15}$ cm at $\sim 10^{15}$ Hz (quasar rest frame) from microlensing (Wambsganss et al. 1990). If this equals the radius of maximum emission for an accretion disk ($\sim 10 r_g$), then $M_{BH}\sim 10^9 M_\odot$. On this basis we suppose that the true mass is between $2\times 10^8 M_\odot$ and $10^9 M_\odot$. Our choice for the fiducial model will be $10^9 M_{\odot}$. The units we used in the simulations are $r_g$ for length and $\Delta t$, the sampling rate, for time. The units of $v_c$ are then $r_g/\Delta t$: $$%from p. 82 of OUX microlensing notes v_c = 0.29 \left({V_c \over 5000 {\rm km/s}}\right)\left({M_{BH}\over 10^9 M_\odot}\right)^{-1} \left({\Delta t \over 1 {\rm day}}\right),$$ where $V_c$ is the caustic velocity in km/s with distance measured at the quasar plane, while time is measured at the observer. The caustic velocity, $V_c$, is quite uncertain, but is likely to be in the range $3000 - 5000$ km/s (Wyithe et al. 1999). For $\Delta t=3$ days, this corresponds to $0.5 \leq v_c \leq 4$, the range we span in our simulations. In the fiducial model, we choose $v_c = 1$. We try values of $\phi_c$ between $0$ and $2\pi$, with $\pi/2$ for our fiducial model. We choose $t_0=0$ for our fiducial model, and also try $t_0=5$ to see whether this technique works when the central time of the monitoring does not coincide with passage of the caustic line across the center of the black hole. There are no observational estimates of the inclination angle; however, unification arguments for radio-loud AGN suggest that quasars are less face-on than blazars, but closer to face-on than radio galaxies, so we choose a fiducial inclination of $30^\circ$. We also look at cases with $\theta = 0$ and $\theta = 60^{\circ}$. Because accretion can spin up black holes, and because Kerr holes permit more efficient accretion than Schwarzschild black holes, we choose $a_s=0.998$ for our fiducial model, but also study one example with $a = 0$. The last parameter is the accretion rate. With the fiducial choices for the other parameters, the observed spectrum is best reproduced with $\dot m \approx 1 = \dot M/(1 M_\odot/yr)$. ### Observational parameters We vary the number of observations, time sampling interval, SNR, and number of observed wave bands, as well as the model parameters. It is especially important to determine how the quality of the result depends on the number of observations because these observations must be targets of opportunity, and thus will impact other observations at a given telescope. We explore what happens for experiments with between 5 and 41 observations (in all cases, we assume uniform spacing). The ratio between $\Delta t$ and the duration of the microlensing event is implicitly given by $v_c$. We try two choices for the SNR (as measured outside the microlensing event): 100 for each image in the best case, and 50 in the worst case (these were chosen based on current ground-based errors, Rachel Webster, priv. comm.). Since disks are broad-band emitters, a broad range of observing frequencies is necessary. Observations in the four wave bands V, B, R, and I (or equivalent) should be routine; observations in U, J, H, K, or in the UV with HST will be much more difficult to obtain, but will yield much more information. To see just how important the additional bands are, we try using just ground-based data in 4 or 8 bands, or 8 ground + 3 HST bands in the best case. The short-wavelength bands are especially important for hot disks, since the deepest part of the potential well is seen at the shortest wavelengths. ### Solution parameters Several considerations determine the number of frequency and radius points at which we may solve for the surface brightness. The number of frequency points is not simply equal to the number of colors at which the quasar is monitored because of the extensive Doppler shifting. We found that in practice the best solution grids in both frequency and radius space were logarithmic. We solve for the intensity at frequencies equally spaced logarithmically between $3\times 10^{14}$ Hz and $10^{16}$ Hz, and radii equally spaced logarithmically from $r_g$ to $r_{out}$, with $r_{out}=500r_g$. In the initial testing of the inversion using the fiducial model, we found the smallest number of radii and emitted frequencies for which we could obtain $\chi^2 = N_F$ for some $\lambda$ was 15 radii and 10 frequencies, which we subsequently used for all the simulations. Best-case simulation {#bestcase} -------------------- For our “best-case” simulation (designated A1 in Table 1), we fixed $\bzeta$ at the fiducial parameter choices. The observational parameters were: 41 observations, $SNR = 100$, and 11 spectral bands. Fig. 5. Solid points are lightcurves with noise added. The top two curves have been shifted upwards by the amount indicated for clarity. The solid lines are the lightcurves from the reconstructed disk profile. Figure 5 shows the lightcurves for this example, with the observed bands de-redshifted for $z=1.695$. Note that the higher frequencies, which come from nearer to the black hole, are magnified more strongly than the lower frequencies. Figure 6 shows the original and reconstructed disk intensity, $I(\nu_e,r_e)$, as a function of frequency and radius for the best fit parameters for this best case (see Table 1). The overall shapes are reproduced quite well. Figure 7 shows the same results in a different format: we have multiplied surface brightness times $r_{e}^2$, and plotted the data differently for clarity. Note that the low and high frequencies and small and large radii are poorly constrained since the simulated lightcurve only covers $-20 r_g < r < 20 r_g$ and 1600 ${\rm \AA} < \lambda < 3~\mu$m. Consequently, at these points the regularization tries to make the flux per log radius constant as a function of radius. Figure 8 shows the ratio of the reconstructed to the original one-dimensional profile, $P_{\nu_o}(x)$ (computed from $I^r_i$) for run A1. Fig. 6. Specific intensity $I_\nu(r_e)$ versus emitted frequency and radius. The shaded surface is the original surface brightness, while the skeleton plot is the surface brightness recovered. Fig. 7. Specific intensity $I_\nu(r_e)$ times $r_e^2$ versus $r_e$ for each emitted frequency. The solid curves are the original $I_\nu r_e^2$, while the dashed curves are the minimum and maximum values of the recovered $I_\nu r_e^2$ for 20 simulations. Each curve is shifted upwards by 2 with respect to the curve below - the zero point is for the lowest curve. We have not plotted negative intensities. Fig. 8. Ratio of recovered to original one dimensional disk profile as a function of position and observed frequency. To discuss the reliability of this solution, we begin by contrasting the region in the $r_e$–$\nu_e$ plane where the random error is predicted to be relatively small with the region where the actual error is small. As can be seen in Figure 9, the region of large $U_i$ for $\lambda \neq 0$ largely, but not entirely, coincides with the region of large $I_i^o /|\Delta I_i|$. Moreover, both of these regions follow a track defined, not surprisingly, by the requirement that $r_e^2 I_{\nu_e}(r_e)$ is relatively large. Elsewhere in the plane, the contribution to the flux is so small that the intensity is virtually unconstrained by the data. Fig. 9. Plot of $I^r_i/\Delta_i$ (a) and formal accuracy, $U_i$, of reconstructed disk profile for $\lambda = 1.1$ (b) and $\lambda=0$ (c). The dotted line in panel (b) shows where the peak of $B_\nu$ occurs at each radius. Where $U_i \simeq I_i^o/|\Delta I_i|$, the error is predominantly random error, and $\delta I_i$ is a good predictor of its magnitude (in fact, in this region $\Delta I_i$ has a Gaussian distribution of the correct width). However, there is also a zone on the large radius side of the high-intensity track where the systematic error is as large or larger than the random error. The nature of this systematic error is revealed by studying Figure 7. The smoothing condition tends to raise the intensity in regions where it should be small, and diminish it where it is large. Because $U_i$ is rarely large enough to be interesting where $I_i$ is small, it is the latter effect that dominates in the region of the $r_e$–$\nu_e$ plane highlighted in figure 9. At least within the context of this model, this systematic error is not the result of the specific choice of smoothing constraint: we have tried a second-order linearization scheme to see if we could get rid of the systematic deviation; however, we still found that the recovered intensity was flatter than the original. If we relax the condition that the intensity should be zero at the last radial bin, then the intensity approaches a constant for each frequency at large radius. Fig. 10. Ratio of recovered to original one dimensional disk profile as a function of position and observed frequency, using the technique of Grieger et al. (1990). We performed a regularized inversion using the Grieger et al. (1991) technique for comparison. Figure 10 shows their inversion on a data set equivalent to run A1. Since the regularization constraint attempts to smooth $P_\nu$, the Doppler peaks are smoothed over, and the noise from the lightcurve appears to still be present in the $P_\nu$. After maximizing $\lambda$ consistent with $\chi^2 = N_F$, we then fix $\lambda$ and vary each component of $\bzeta$, minimizing $\chi^2$ with respect to the other parameters. This procedure shows how well each parameter can be constrained for a given simulation, or whether there are other local minima. In Figure 11 we show the $\Delta\chi^2$ for each model parameter. For this particular model, the physically interesting parameters, inclination angle ($\mu$) and caustic velocity ($v_c$), have well-defined minima. The time of origin crossing ($t_0$) and the caustic crossing angle ($\phi_c$) are also well-behaved. The black hole spin has a rather flat $\Delta\chi^2$ distribution. However, the minimum does lie at the correct value. Figure 11 also has a histogram of the parameters from each noise realization, showing that the minimum of the $\chi^2$ distribution has few outliers. Fig. 11. Change in $\chi^2$ versus each parameter for the A1 case. The vertical lines show the values of the original parameters. The histograms show the best-fit parameter results of 20 lightcurve realizations, with the right hand axis labelling the number in each bin. Varying model parameters {#varyparam} ------------------------ Table 1 shows the results of varying the model parameters, keeping the observational parameters fixed at the “best-case” values. Runs A1-A10 each have recovered surface brightness which look similar to A1, and the $U_i$ are quite similar. Runs A11a and A12 reproduce the intensities well for the lowest frequencies and for radii outside $r_{ms}$; however, the recovered intensities are non-zero inside $r_{ms}$, contrary to the input model. Table 1 shows the average and standard deviation of the recovered parameters, $\bzeta$, measured for 20 Monte Carlo realizations. In all the cases we examined, the distribution of recovered parameters is centered near the true model parameters, showing that there are no systematic offsets introduced by our inversion. This is encouraging since it means that this technique has the potential to measure important global properties of the accretion disk/black hole system. Run A2 shows that we can determine the time that the caustic crosses the black hole rather accurately. We have also tried cases with $t_0=\pm 10$, and we find that these are also measured quite well, and that the intensity is reproduced as well as in the A1 run. Runs A3 and A4 show that we can distinguish between different caustic velocities, which means that we can constrain the black hole mass in terms of the lens velocity. Runs A5 and A6 show that we can measure the disk inclination angle for a wide range of intrinsic angles. Runs A7, A8, A9, and A10 show that we can measure the angle at which the caustic crosses the accretion disk rather accurately. In some cases there is a degeneracy between $\phi_c$ and $2\pi-\phi_c$ when the disk inclination is small, but this should not affect the recovered surface brightness since the disk is approximately symmetric in this case. Figure 12 shows the $\chi^2$ topology for each parameter for Run A10, a somewhat special case in which $\phi_c=\pi$. The parameters $t_0$ and $\mu$ have local minima away from the correct minimum; however, these can be ruled out because some inferred intensities have large negative excursions in the false minimum. Fig. 12. Plot of $\Delta\chi^2$ (minimized over all other parameters) vs. each parameter for run A10. The solid vertical lines show the original parameters. The dotted lines show the absolute value of the sum of the inverted intensities which are negative (the scale is from 0 to 1). Constraining the spin can be difficult, particularly when $\phi_c \simeq \pi$. For example, in Run A10 $a_s$ is not constrained at all (see Figure 12). $\chi^2$ has as deep a minimum at $a_s = 0$ as it does at the correct value $a_s = 0.998$. Only if the coarse search in $\bzeta$ space is lucky enough to discover the true minimum will the Levenberg-Marquardt procedure home in on the correct value. It is not clear why some $\phi_c$ are more favorable for determining $a_s$. Whether $a_s$ can be constrained at all depends on how strongly one believes in the model. If no emission is permitted inside the marginally stable orbit (Run A11b), $a_s$ can be constrained because there is a sizable difference between the marginally stable orbit around a Schwarzschild black hole ($6 r_g$) and a maximal Kerr black hole ($\simeq r_g$). However, if one is unwilling to make this assumption, the distinction between the spins largely disappears (Runs A11a and A12). The reason for this indistinguishability is shown in Figure 13, which shows a contour plot for the redshift as a function of position for black holes with spins $a_s=0.01$ and $a_s = 0.99$, including the regions inside $r_{ms}$. The two plots are almost identical around $10 r_g$, where most of the observed radiation comes from in this model. Figure 14 shows the $\chi^2$ (minimized over all other parameters) vs. $a_s$ and recovered paramters for 25 simulations assuming that emission only occurs outside of $r_{ms}$. The $\chi^2$ has a clear minimum near the correct spin, and the simulations show that the spin can be rather accurately recovered. Since $r_{ms}$ increases with decreasing spin, we can only hope to obtain a lower limit on $a_s$ by assuming $r > r_{ms}$. Indeed, the $\chi^2$ vs. $a_s$ is flat in the case of zero spin (A11a). Fig. 13. Redshift as a function of position for black holes with spin $a_s=0.01$ (solid line) and $a_s=0.99$ (dashed line). Fig. 14. Plot of the $\chi^2$ vs. $a_s$ (minimized over all other parameters) for cases A1 and A10, but with the additional assumption that emission only occurs outside $r_{ms}$. The solid dots show $\chi^2$ for $\phi_c=\pi/2$ (case A1) while the open circles for A10. The solid histogram shows the resulting $a_s$ measured for 25 monte carlo simulations for A1; the dotted line for A10. If the disk is much hotter than we have assumed, then the parameters will not be as well constrained as those that we have used, as we would then sample only the outer regions of the disk. To illustrate this point, we have run a somewhat unrealistic model, A13, with the fiducial accretion rate, but a black hole mass of $2\times10^8 M_\odot$ (near the Eddington limit). A standard blackbody accretion disk around a black hole with this mass cannot fit the observations as its spectrum is too steep and the magnification must be much larger than in standard models of the lens galaxy. The error on the measured spin is much larger than for Run A1 (see Table 2). The $U_i$ for Run A13 are comparable to those in Run A1, so the intensities are recovered similarly well. To see how well we can perform the inversion when the assumption of smooth radial variation is incorrect, we multiplied the accretion disk intensity by $$\label{fluceq} 1+\sin{\left[6\pi {\log(r)-\log(r_{in})\over \log(r_{out})-\log(r_{in})}\right]},$$ which makes the disk three logarithmically spaced annuli (runs A14 and A15). Surprisingly, the recovered model parameters, $\bzeta$, are accurate. Whether the radial variations can be discovered depends on the number of observations. In Run A14 (15 observations), the correct overall shape is found, but the radial modulation not reproduced; in Run A15 (41 observations), the radial dependence of the recovered intensities is more nearly correct. This indicates that the inversion is only accurate if our smoothness model assumption is met on the smallest scale probed by the sampling. Figure 15 shows the results of run A14. Fig. 15. Plot of original (solid lines) and recovered (dotted lines) for blackbody disk with fluctuations added (model A14). Each curve shows a different frequency (from $3\times 10^{14}$ to $10^{16}$ Hz), shifted by 2 units for clear separation. The dotted lines show the maximum and minimum recovered intensity from 20 Monte Carlo realizations, with only positive intensities plotted. We have not tried breaking the assumption of azimuthal symmetry, as the transfer function is computed assuming it. Varying observation parameters ------------------------------ The question of how many observations are necessary is addressed with Runs M2a-d (Table 2). First, we compare fewer observations (21) at the same sampling rate (M2a) and for the same duration (M2b). In each case, the rms scatter of the model parameters is remarkably small compared to Run A1. Thus, [*if the underlying model is correct*]{}, it appears that we can determine the model parameters with a high degree of accuracy with relatively few observations. A smaller number of observations, however, impairs our ability to reconstruct the true surface brightness of the accretion disk, as the inversion relies more upon the smoothing constraint than on the actual data. This can be seen in Figure 16, which shows the derived formal accuracies for various numbers of observations. Fig. 16. This shows a plot of the $U_i$ as a function of radius for various numbers of observations at a single frequency: solid line is $N_t=41$ (A1), dotted line is $N_t=21$ (M2b), short-dashed line is $N_t=11$ (M2d), and long-dashed line is $N_t=5$ (M2e). The total amount of information about the intensities that may be gleaned also depends on the number of observations. It is obvious that the number of frequencies at which the intensity may be inferred scales in proportion to the number of bands whose lightcurves are measured. In addition, comparing A1 with M2b-d shows that the number of radial points with reliable solutions increases slowly with increasing number of observation times $N_t$. This is because each observation constrains most strongly the minimum radius where the caustic crosses. Since the observations are spaced linearly, while the radii are spaced logarithmically, the number of radial points constrained by the data is $\propto \log(r_{max}/r_{min}) \propto \log(T/\Delta t) \propto \log(N_t)$, where $T$ is the total duration and $r_{max}$ and $r_{min}$ are radii corresponding to the radial limits of the region with a reliable solution. In addition, $U_i \propto N_t^{1/2}$ (see Figure 16), in the usual fashion of signal-to-noise ratios. In addition to the number of observations, the SNR of each observation affects the quality of the inversion. Runs M3a-b with SNR=50 demonstrate this dependence. For Run M3a, the recovered parameters have similar errors to those in Run A1, except for the errors on the spin. For Run M3b, the errors on recovered parameters are roughly double those in Run M2d. For fixed model parameters and fixed number of observations, on average the errors on ${\bf I}$ are directly proportional to the errors on ${\bf F}$; in other words, $\langle U_i \rangle \propto SNR$, where the average is over 20 simulations. For individual simulations, the $U_i$ depend on $\lambda$, which causes scatter in $U_i$. Next, we look at the dependence of the solution on which frequencies are observed. Run M4a shows that the best fit model parameters, $\bzeta$ are not strongly dependent on the infrared frequencies. However, excluding infrared bands reduces $U_i$ at the lowest frequencies. Run M4b shows that the best fit $\bzeta$ are strongly dependent on observing the ultraviolet frequencies. This run has the largest errors on model parameters and poorest agreement with the true parameters. In addition, the $U_i$ at the highest frequencies are strongly reduced for this run. Comparing Runs A1, M4a,b, we find that the number of frequencies (at the accretion disk) constrained by the data $\propto N_\nu$, which shows why broad frequency coverage is crucial for mapping out the disk spectrum. Runs A1, A3, and A4 show that both the number of radial points with reliable recovered intensities and $U_i$ are insensitive to the total duration of the monitoring within the range of uncertainties in $v_c$. Runs A2 shows that the time at which we begin the observations does not strongly affect the derived parameters or surface intensity, as long as we catch the region near the peak. Discussion ========== Observational requirements -------------------------- Experimental design depends on which questions are the goal, and to what accuracy one aims to answer them. Consequently, this discussion, much like our earlier discussion of the experimental errors, divides according to whether the primary aim is to map the surface brightness of the disk, or to infer the model parameters. If the goal is to obtain a map of the the disk intensity as a function of radius and frequency over as large region in the $r_e$–$\nu_e$ plane as is possible, it pays most to invest in multiple monitoring bands because the number of radius-frequency pairs for which it is possible to find a solution is $\propto N_\nu$, but rises only logarithmically with $N_t$. We expect the intensity to vary as a power law in radius, so a better observation strategy might be to space observations logarithmically in time; however, this will be difficult to achieve in practice, which is why we have assumed equal time spacing. The accuracy of the solution is directly proportional to the SNR in the data. To map the regions nearest to the black hole, the highest UV frequencies are crucial (although this depends on the assumption that the spectral peak moves to higher frequencies at smaller radii); however, lower frequencies are required if one is interested in obtaining a broad-band spectrum. Note also that a higher sampling frequency increases the chance that one will obtain monitoring data during the time when the caustic line is near the black hole. To obtain high formal accuracy for a broad range in radius requires a lightcurve that is finely sampled for a long time period. To determine the intensities, the model parameters must also be well-defined. When the observations are too few or the SNR too small, the uncertainties in $\bzeta$ contribute to the uncertainty in $I_i$. If the goal is to simply constrain the model parameters, and one believes that the model is correct, then only a small number of observations may be required (5 observations were sufficient in our simulations to constrain all model parameters to better than 20% at $1 \sigma$; 11 observations for $\leq$ 10% $1\sigma$ accuracy). Since the black hole mass and caustic velocity are quite uncertain, it may be necessary to have more observations to be sure to obtain the few near the peak of the caustic crossing that are essential for success. The highest frequencies are most important for constraining model parameters. In our simulations we assumed that only statistical errors affected the fluxes, and that these errors can be estimated from the observational data. There might also be systematic errors in the fluxes due, for example, to emission line contributions, inaccurate calibration between bands, or cosmic rays. The quality of the result could also be affected by these problems. Under what circumstances might the model fail? ---------------------------------------------- The inversion scheme we propose assumes a model that is plausible—a geometrically thin relativistic accretion disk in which azimuthal variations are quickly smoothed out, microlensed by a caustic system whose basic length scale is much larger than the size of the bright region of the disk. However, we are by no means guaranteed that even this general framework is correct. In this section we discuss what would happen if our method is attempted, but one of these assumptions is invalid. The most basic of our assumptions is that the surface brightness varies smoothly as a function of radius. Given sufficiently dense sampling with good SNR, even quite sharp gradients could be recognized by our procedure. On the other hand, if the sampling is inadequate, the existence of such features would appear only as a troubling inability to find a solution with adequate $\chi^2$, unless $\lambda$ is taken to be very small. Another potential source of trouble is departures from azimuthal symmetry. In the absence of microlensing, “spots" can modulate the lightcurve on the orbital period if they are in the relativistic portion of the disk and the inclination is relatively large (Abramowicz et al. 1991). In this case, that would mean periods of $\simeq 31 M_9 (r/10r_g)^{3/2}(1+z)/2.7$ days (in the observed frame). Because the largest observed variations in the Einstein Cross are $\sim 10\%$ on year-long timescales, this effect cannot be too strong in this system. However, there might be a range of spot brightness in which they are too weak to show up in the ordinary lightcurve, yet strong enough to cause some periodic modulation of the lightcurve during a microlensing event (Gould & Miralda-Escudé 1997). Because the likely duration of a microlensing event (a few weeks) is several to ten times the orbital period for the brightest part of the disk, it is possible that this effect might be seen directly in the lightcurve. If not, they might still make it difficult to find a solution with acceptable $\chi^2$. Particularly if the break in azimuthal symmetry is approximately $\propto e^{i\phi}$ and the disk is nearly face-on, a strong spot might be confused with Doppler boosting, leading to a mistaken inference for the disk inclination. Several physical effects might make disks geometrically thick—radiation pressure support if the luminosity approaches Eddington (Abramowicz et al. 1988), gas pressure support if the ions retain most of their heat (Rees et al. 1982, Narayan & Yi 1995), or an optically thick outflow. If any of these mechanisms acts, the orbital velocity at the photosphere would no longer be that corresponding to circular free-fall in the equatorial plane of the black hole, so that the general relativistic transfer function we apply would no longer be valid. We would expect, then, difficulties in finding a solution with acceptable $\chi^2$, but the portion of our solution describing the outer regions of the disk should be only weakly affected. The surface brightness model we have used for the inversion simulations is roughly consistent with both the observed spectrum (given the uncertainty in reddenning) and the current microlensing size constraint, but it is not unique. If the actual spectrum has strong emission in the far ultraviolet (which could be true if the reddening is greater than our estimate), the observable portion of the spectrum will be dominated by emission far from the horizon, weakening all the relativistic effects. If so, the disk inclination and black hole spin will be poorly constrained. Uncertain reddening can have other effects, also. Because we can expect it to be uniform across the face of the disk, it should not affect the inferred radial profile of the disk, but it could well introduce additional uncertainty into our estimate of the intrinsic disk spectrum at any given radius. Our assumption that the bright part of the disk is small relative to the caustic length scale is unlikely to be broken, except in the outer regions of the accretion disk. In this case, for which the microlensing optical depth is $\sim 1$ (Witt & Mao 1994), the caustic scale is essentially the size of the Einstein ring due to a single star. Consequently, the ratio between the disk size and the caustic scale is only $\simeq 0.01 M_9 (r/10 r_g) m^{-1/2} (h/0.75)^{1/2}$, where $m$ is the mean mass of microlensing stars in Solar units, so according to the Grieger et al. (1988) criterion, the caustic assumption is likely to be valid out to 100 $r_g$. In our fiducial model, for example, the flux at $100 r_g$ peaks at an observed wavelength $1[(1+z)/2.7]\mu$m, where we have scaled to the redshift of the Einstein Cross. However, there might be difficulties in practice from a related problem: measuring $A_0 F_\nu$. Again, if the disk size is much smaller than the Einstein radius of a single star, then $A_0$ is approximately constant during a high amplification event; thus, the same criterion for success applies as in the previous paragraph. Of course, observations when the quasar is outside the caustic are still required to measure $A_0$. Given the expectation that the quasar is smaller at higher frequencies, observing at the highest frequencies will provide the best constraint on $A_0$. To determine how these difficulties and those listed in §1.2 will affect the inversion, we are currently running simulations of full microlensed lightcurves appropriate for the Einstein Cross, to which we will apply our inversion algorithm (Wyithe & Agol, in preparation). In any of these instances of model inappropriateness, the impact on specific inferred parameters depends somewhat on details of the inversion procedure. For example, if the procedure we have outlined is followed (i.e., minimizing $\chi^2$ by varying $\bzeta$ at fixed $\lambda$, then raising $\lambda$ until $\chi^2$ meets our definition of acceptability), difficulty in achieving satisfactory $\chi^2$ reduces the ultimate $\lambda$. This means that, in effect, more degrees of freedom are “spent" on fitting the $I(r,\nu)$, leaving fewer for defining $\bzeta$. If the problem is lack of smoothness in the radial profile, this transfer of effort is reasonable; if the problem is different, however, and if one cares about the accuracy of the $\bzeta$ parameters, one might choose to modify the procedure in a way that keeps $\lambda$ fixed at a relatively large value. One disadvantage of using linear regularization is that the intensities are not required to be positive definite, though it is impossible to emit a negative number of photons. We have tried to incorporate this by trying three other methods: maximum entropy; replacing $I_i$ by $log(I_i)$ in ${\cal B}$; and the method of projections onto convex sets. Each technique finds solutions that are local minima with large $\chi^2$. Another useful technique might be to make the further assumption that the spectrum at each radius can be described as a blackbody, and then solve for the temperature as a function of radius. These methods are all non-linear, and thus intrinsically slow. They therefore impede the exploration of parameter space, but might be interesting for future work. Multiple microlensing events ---------------------------- Of the five parameters in $\bzeta$, three ($t_o$, $v_c$, and $\phi_c$) will change from one event to the next, but the other two ($a_s$ and $i$) should remain fixed. Since we expect roughly one event per year, it should be possible to combine observations of several events in order to more tightly constrain $a_s$ and $i$. Connection to X-ray microlensing events --------------------------------------- The optical and ultraviolet continua of quasars are not the only portions of the spectrum radiated by the inner part of the accretion disk. The X-ray continuum must also come from somewhere near that region. It, too, should therefore be microlensed in much the same way as the optical and ultraviolet continuum we have discussed in this paper. Whether the same technique can be successfully applied to the X-ray continuum depends on the same considerations as discussed in §4.2, but several of them are more likely to present problems in the context of X-rays than for the optical/ultraviolet continuum. There have been numerous suggestions, for example, that the X-rays are produced in a relatively small number of compact active regions (e.g., Haardt et al. 1994) that might have substantial velocities relative to the disk (Beloborodov 1999). If so, the assumptions of azimuthal symmetry, radial smoothness, and also simple circular orbital motion might all be suspect. Nonetheless, it would certainly be worthwhile to monitor the X-ray flux during a microlensing event, in the hope that its emissivity distribution is sufficiently consistent with our assumptions that it, too, could be mapped. Combining this data set with the optical/ultraviolet data would also provide an independent constraint on the $\bzeta$ parameters, which should all be the same for the same event. Summary ------- We have demonstrated that monitoring microlensing events in the Einstein Cross quasar has great potential for both revealing the structure of its continuum emission with unprecedented resolution, and potentially constraining such basic parameters of the quasar as the spin of its black hole (if we assume emission occurs only outside of the marginally stable circular orbit), the mass of the black hole (modulo the caustic velocity), and the inclination angle of its disk relative to our line of sight. If this potential is realized, and the analytic method we have proposed is implemented successfully, we may be able to begin answering such fundamental questions as: What is the intrinsic local spectrum of the disk? How close is it to thermal? And, most fundamentally, does the dissipation distribution in accretion disks vary with radius in the fashion predicted (see equation 1) long-ago? We would like to thank Andy Gould and Chris Kochanek for useful discussions. We thank Casey Papovich, Keiichi Wada, Andrew Zirm, and Viktor Ziskin for time on their workstations. Thanks to David Heyrovsky and the referee for comments which improved the manuscript. This work was partially supported by NASA Grant NAG 5-3929 and NSF Grant AST-9616922. Abramowicz, M., Bao, G., Lanza, A. & Zhang, X. 1991, A & A 245, 454 Abramowicz, M., Czerny, B., Lasota, J.-P. & Szuszkiewicz, E. 1988,  332, 646 Agol, E., 1997, PhD Thesis, University of California, Santa Barbara Albrow, M. et al., the PLANET Collaboration, 1998, ApJ submitted, astro-ph/98114 79 Bardeen, J. M. & Petterson, J. A., 1975, ApJ, 195, L65 Beloborodov, A. M., 1998, MNRAS, 297, 739 Beloborodov, A. 1999, astro-ph/9809383 Corrigan, R.T. et al., 1991, AJ, 102, 304 Czerny, B., Jaroszyński, M., & Czerny, M., 1994, MNRAS, 268, 135 Cunningham, C. T., 1975, ApJ, 202, 788 Gaudi, B. S. & Gould, A., 1998, ApJ, submitted, astro-ph/9802205 Gould, A. & Miralda-Escude, J., 1997, ApJ, 483, L13 Gould, A. & Gaudi, B. S., 1997, ApJ, 486, 692 Grieger, B., Kayser, R., & Schramm, T., 1991, A & A, 252, 508 Grieger, B., Kayser, R., & Refsdal, S., 1988, A & A, 194, 54 Haardt, F., Maraschi, L. & Ghisellini, G. 1994,  432, L95 Huchra, J., et al., 1985, AJ, 90, 691a Irwin, M. J., et al., 1989, AJ, 98, 1989 Jaroszyński, M. & Marck, J.-A., 1994, A & A, 291, 731 Krolik, J. H., 1998, [Active Galactic Nuclei]{}, (Princeton: Princeton University Press) Jaroszyński, M., Wambsganss, J., Paczyński, B., 1992, ApJ, 396, L65 Narayan, R. & Yi, I., 1994, ApJ, 428, L13 Page, D. N. & Thorne, K. S., 1974, ApJ, 191, 499 Press, W., Teukolsky S., Vetterling W., & Flannery, B., 1992, [*Numerical Recipes in FORTRAN*]{}, 2nd edition, (Cambridge: Cambridge University Press) Pringle, J. E., 1981, ARAA, 19, 137 Rauch, K. P., & Blandford, R. D., 1991, ApJ, 395, L65 Rauch, K. P., & Blandford, R. D., 1994, ApJ, 421, 46 Rees, M. J., Phinney, E. S., Begelman, M. C., & Blandford, R. D., 1982, Nature, 295, 17 Schneider, P., Ehlers, J., & Falco, E., 1992, [*Gravitational Lenses*]{} (New York: Springer-Verlag) Schneider, P. & Weiß, A., 1986, A & A, 164, 237 Wambsganss, J., Paczyński, B., & Schneider, P., 1990, ApJ, 358, L33 Wisotzki, L., Köhler, T., Kayser, R., and Reimers, D., 1993, A & A, 278, L15 Witt, H. J., Kayser, R., & Refsdal, S., 1993, A & A, 268, 501 Witt, H. J. & Mao, S., 1994, ApJ, 429, 66 Witt, H. J., Mao, S., & Schechter, P. L., 1995, ApJ, 443, 18 Wyithe, J. S. B., Webster, R. L., & Turner, E. L., 1999, astro-ph/9901341 [ccccccc]{} A1 & 0 & 1 & 0.866 &$\pi/2$& 0.998 & &-.1$\pm$.2&.96$\pm$.06&.84$\pm$.02&1.56$\pm$.03&.96$\pm$.05&.01-1.9 A2 & 5 & 1 & 0.866 & $\pi/2$ & 0.998 & &-4.8$\pm$.2&.96$\pm$.06&.84$\pm$.02&1.58$\pm$.02&.97$\pm$.03&.002-1.7 A3 & 0 & 2 & 0.866 & $\pi/2$ & 0.998 & &-.10$\pm$.13&1.95$\pm$.095&.83$\pm$.03&1.58$\pm$.03&.93$\pm$.08&.01-1.6 A4 & 0 & .5 & 0.866 & $\pi/2$ & 0.998 & &-.18$\pm$.2&.50$\pm$.02&.85$\pm$.02&1.57$\pm$.02&.97$\pm$.03&.001-2.3 A5 & 0 & 1 & 1.0 & n/a & 0.998 & &-.06$\pm$.07&1.02$\pm$.04&.999$\pm$.001&1.53$\pm$.08&.99$\pm$.02&.0001-1.8A6 & 0 & 1 & 0.5 & $\pi/2$ & 0.998 & &-.03$\pm$.08&.95$\pm$.05&.50$\pm$.02&1.57$\pm$.01&.92$\pm$.06&.005-1.8 A7 & 0 & 1 & 0.866 & $\pi/4$ & 0.998 & &-.02$\pm$.13&.99$\pm$.04&.84$\pm$.01&.83$\pm$.05&.95$\pm$.11&.01-1.7 A8 & 0 & 1 & 0.866 & $3\pi/4$& 0.998 & &-.03$\pm$.1&1.01$\pm$.03&.86$\pm$.01&2.33$\pm$.04&.99$\pm$.02&.003-1.9 A9 & 0 & 1 & 0.866 & $3\pi/2$ & 0.998 & &.1$\pm$.1&.98$\pm$.06&.84$\pm$.02&4.72$\pm$.02&.95$\pm$.05&.006-1.9 A10 & 0 & 1 & 0.866 & $\pi$ & 0.998 & &.03$\pm$.07&1.03$\pm$.03&.86$\pm$.01&3.16$\pm$.08&.99$\pm$.03&.005-1.8 A11a & 0 & 1 & 0.866 & $\pi/2$ & 0. & &-.46$\pm$.26&.95$\pm$.07&.78$\pm$.06&1.57$\pm$.02&.17$\pm$.32&.005-.8 A11b&-.22$\pm$.24&.87$\pm$.07&.87$\pm$.02&1.57$\pm$.02&.09$\pm$.12&.001-4.4 A12 & 0 & 1 & 0.866 & $\pi/2$ & 0.5 & &-.29$\pm$.20&.91$\pm$.09&.83$\pm$.07&1.57$\pm$.02&.55$\pm$.25&.03-.7 \[tab1\] [ccccccccccc]{} A13 & 41 & 100 & 11 & 1 &.93$\pm$.07&-.2$\pm$.4&.85$\pm$.04&1.58$\pm$.06&.89$\pm$.19&.01-2 A14 & 15 & 100 & 11 & 2.9 &2.6$\pm$.3&-.04$\pm$.06&.90$\pm$.04&1.57$\pm$.03&.87$\pm$.19&.07-.9 A15 & 41 & 100 & 11 & 1 &.98$\pm$.06&.05$\pm$.20&.88$\pm$.05&1.57$\pm$.02&.79$\pm$.20&.002-.18 M2a & 21 & 100 & 11 & 1 &.99$\pm$.04& -.13$\pm$.15&.84$\pm$.01&1.58$\pm$.02&.98$\pm$.05&.5-5.2M2b & 21 & 100 & 11 & 2 &1.9$\pm$.13&-.12$\pm$.07&.82$\pm$.02&1.59$\pm$.03&.93$\pm$.07&.2-4.8 M2c & 15 & 100 & 11 & 2.9 &2.7$\pm$.3&-.09$\pm$.1&.83$\pm$.03&1.59$\pm$.03&.95$\pm$.06&.6-6.6 M2d & 11 & 100 & 11 & 4 &3.9$\pm$.3&-.03$\pm$.07&.83$\pm$.03&1.56$\pm$.03&.9$\pm$.1&1.7-11 M2e & 5 & 100 & 11 & 10 &$8\pm2$&0.$\pm$.15&.75$\pm$.1&1.48$\pm$.28&.93$\pm$.11&9-38 M3a & 41 & 50 & 11 & 1 &.94$\pm$.07&-.25$\pm$.22&.83$\pm$.03&1.58$\pm$.03&.94$\pm$.14&.006-1.1M3b & 11 & 50 & 11 &4 &3.9$\pm$.3&-.03$\pm$.14&.81$\pm$.06&1.55$\pm$.1&.88$\pm$.17&.8-8.6 M4a & 15 & 100 & 8 & 2.9 &2.8$\pm$.24&-.04$\pm$.08&.84$\pm$.03&1.57$\pm$.03&.9$\pm$.1&.004-5.M4b & 41 & 100 & 4 & 1&.88$\pm$.09&-.6$\pm$.4&.80$\pm$.06&1.61$\pm$.05&.85$\pm$.14&.25-3.3 \[tab2\]

ArXiv

--- abstract: 'In this paper, we consider a 3d cubic focusing nonlinear schrödinger equation (NLS) with slowing decaying potentials. Adopting the variational method of Ibrahim-Masmoudi-Nakanishi [@IMN], we obtain a condition for scattering. It is actually sharp in some sense since the solution will blow up if it’s false. The proof of blow-up part relies on the method of Du-Wu-Zhang [@DWZ].' address: - 'Qing Guo, College of Science, Minzu University of China, Beijing, 100081, P.R. China' - ' Hua Wang, School of Mathematics and Statistics and Hubei Province Key Laboratory of Mathematical Physics, Central China Normal University, Wuhan, 430079, P.R. China' - 'Xiaohua Yao, School of Mathematics and Statistics and Hubei Province Key Laboratory of Mathematical Physics, Central China Normal University, Wuhan, 430079, P.R. China' author: - 'Qing Guo,  Hua Wang  and Xiaohua Yao' title: Dynamics of the focusing 3D cubic NLS with slowly decaying potential --- Introduction ============ In this paper, we consider a 3d cubic focusing NLS with slowly decaying potentials ($\rm{NLS_{k}}$) $$\label{1.1} \left\{ \begin{aligned} i&\partial_{t}u-H_{\alpha}u+|u|^{2}u=0,\;\;(t,x) \in {{\bf{R}}\times{\bf{R}}^{3}}, \\ u&(0, x)=u_{0}(x)\in H^{1}({\bf{R}}^{3}), \end{aligned}\right.$$ where $u: {\bf R}\times {\bf R}^{3}\rightarrow {\bf C}$ is a complex-valued function, $H_{\alpha}=-\Delta+V(x)$ and $V(x)=\frac{k}{|x|^{\alpha}}$ with $k>0$ and $1<\alpha\leq 2$. Throughout this paper, we use the symbol $V(x)$ instead of $\frac{k}{|x|^{\alpha}}$ since we frequently use the general property of $V$: $V\geq 0$, $x\cdot\nabla V\leq 0$, $2V+x\cdot V\geq 0$ and $3x\cdot V+x\nabla^{2}V x^{T}\leq 0$. As $\frac{k}{|x|^{\alpha}}>0$ and $\frac{k}{|x|^{\alpha}} \in L_{loc}^{1}$, $H_{\alpha}$ is defined as a unique self-adjoint operator associated with the non-negative quadratic form $<(-\Delta+\frac{k}{|x|^{\alpha}})f, f>$ on $C_{0}^{\infty}({\bf R}^{3})$. Moreover, $H_{\alpha}$ is purely absolutely continuous and has no eigenvalue. Since $k>0$, the kernel $e^{-tH_{\alpha}}(x, y)$ of $e^{-tH_{\alpha}}$ satisfies the upper Gaussian estimate [@S] i.e., for $\forall t>0$, $\forall x$, $y\in {\bf R}^{3}$, $$\begin{aligned} \label{heatkernel} 0\leq e^{-tH_{\alpha}}(x, y)\leq e^{t\Delta}(x, y)=(4\pi t)^{-\frac{3}{2}}e^{-\frac{|x-y|^{2}}{4t}},\end{aligned}$$ which implies that Hardy inequality , Mikhlin multiplier theorem and Littlewood-Paley theory (Bernstein inequalities Lemma \[Bernstein\], Littlewood-Paley decomposition Lemma \[LPdecompostion\] and square function estimates Lemma \[square\]) associated with $H_{\alpha}$. Hence it follows from Hardy inequality and Stein complex interpolation that the standard Sobolev norms and the Sobolev norms associated with $H_{\alpha}$ are equivalent (see Lemma \[Sobolev\]). Recently, Mizutani [@M] showed that $e^{-itH_{\alpha}}$ satisfies global-in-time Strichartz estimates for any admissible pairs. Combining the Sobolev norm equivalence and the Strichartz estimates and following the same line of the proof of Theorem 2.15 and Remark 2.16 of [@KMVZ] yield that ($\rm{NLS_{k}}$) is locally well-posed and scatters in $H^{1}({\bf R}^{3})$. [@KMVZ] \[localwellposedness\] Let $u_{0}\in H^{1}({\bf R}^{3})$. Then the following are true. \(i) There exist $T=T(\|u_{0}\|_{H^{1}})>0$ and a unique solution $u\in C((-T, T), H^{1}({\bf R}^{3}))$ of ($\rm{NLS_{k}}$). \(ii) There exists $\epsilon_{0}>0$ such that if for $0<\epsilon<\epsilon_{0}$, $$\|e^{-itH_{\alpha}}u_{0}\|_{L_{t,x}^{5}({\bf R}^{+}\times{\bf R}^{3})}<\epsilon,$$ then the solution $u$ of ($\rm{NLS_{k}}$) is global in the positive time direction and satisfies $$\begin{aligned} \|u\|_{L_{t,x}^{5}({\bf R}^{+}\times{\bf R}^{3})}\lesssim\epsilon.\end{aligned}$$ The similar result holds in the negative time direction. \(iii) For any $\phi\in H^{1}({\bf R}^{3})$, then there exist $T>0$ and a solution $u\in C((T, +\infty), H^{1}({\bf R}^{3}))$ of ($\rm{NLS_{k}}$) such that $$\lim_{t\rightarrow+\infty}\|u(t)-e^{-itH_{\alpha}}\phi\|_{H^{1}({\bf R}^{3})}=0.$$ The similar result holds in the negative time direction. \(iv) If $u:{\bf R}\times{\bf R}^{3}\rightarrow {\bf C}$ is a global solution of ($\rm{NLS_{k}}$) with $$\begin{aligned} \label{scatteringbound} \|u\|_{L_{t,x}^{5}({\bf R}\times{\bf R}^{3})}<+\infty,\end{aligned}$$ then the solution $u(t)$ scatters in $H^{1}$. That is, there exists $\phi_{\pm}\in H^{1}({\bf R}^{3})$ such that $$\lim_{t\rightarrow\pm\infty}\|u(t)-e^{-itH_{\alpha}}\phi_{\pm}\|_{H^{1}({\bf R}^{3})}=0.$$ Moreover, the $H^{1}$ solution $u$ obeys the mass and energy conservation laws: $$\begin{aligned} \label{mass} M(u)=\displaystyle\int_{{\bf R}^{3}}|u(t, x)|^{2}dx=M(u_{0}),\end{aligned}$$ and $$\begin{aligned} \label{energy} E(u)=E_{k}(u)=\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla u(x)|^{2}dx +\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}V(x)|u(x)|^{2}dx -\frac{1}{4}\displaystyle\int_{{\bf R}^{3}}|u(x)|^{4}dx=E(u_{0}).\end{aligned}$$ In the case $k=0$, Holmer-Roudenko [@HR] and Duyckaerts-Holmer-Roudenko [@DHR] employed the concentration-compactness approach of Kenig-Merle [@KM] to obtain sharp criteria between scattering and blow up for ($\rm{NLS_{0}}$) in terms of conservation laws ( and ) and the ground state $Q$, which is the unique positive radial exponential decaying solution of the elliptic equation $$\begin{aligned} \label{ellipticequation} \Delta Q-Q+Q^{3}=0.\end{aligned}$$ Fang-Xie-Canzenave [@FXC] and Akahor-Nawa [@AN] extended the result in [@HR; @DHR] to the general power and dimensions. Subsequently, Killip-Murphy-Visan-Zhang [@KMVZ] established a corresponding sharp threshold between scattering and blow up for ($\rm{NLS_{k}}$) with $k>-\frac{1}{4}$ and $\alpha=2$. Recently, Miao-Zhang-Zheng [@MZZ] used the interaction Morawetz-type estimates and the equivalence of Sobolev norms to prove all solutions scatter for ($\rm{NLS_{k}}$) with $k>0$, $\alpha=1$ and $-|u|^{p-1}u$ $(\frac{7}{3}<p<5)$ in place of $|u|^{2}u$ (i.e., nonlinear Schrödinger equation with repulsive Coulomb potential in the defocusing). The goal of this paper is to extend the sharp scattering criterion in [@KMVZ] from $\alpha=2$ to $1<\alpha\leq 2$ when $k>0$ in some sense. Obviously, for $1<\alpha<2$, the equation ($\rm{NLS_{k}}$) doesn’t enjoy scaling invariant. Therefore, we cannot apply scaling as indicated in [@KMVZ] to get a critical element ( a minimal blow up solution). Hence, we shall adopt the variational argument based on the work of Ibrahim-Masmoudi-Nakanishi [@IMN] to overcome the difficulty. Recently, the same argument have been applied to the focusing mass-supercritical nonlinear Schrödinger equation with repulsive Dirac delta potential on the real line (see [@II]). To state our main result, we introduce some notation now. We define the functional $S_{k}$ as $$\begin{aligned} \label{action} S_{k}(\varphi):=E(\varphi)+\frac{1}{2}M(\varphi)=\frac{1}{2}\|\varphi\|_{{\mathcal H}_{k}^{1}}-\frac{1}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx,\end{aligned}$$ where $$\begin{aligned} \label{Sobolev1} \|\varphi\|_{{\mathcal H}_{k}^{1}}^{2}=\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx +\displaystyle\int_{{\bf R}^{3}}V(x)|\varphi(x)|^{2}dx +\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{2}dx,\end{aligned}$$ which is equivalent to $\|\varphi\|_{H^{1}}$ by $k>0$ by Hardy’s inequality. Denote the scaling quantity $\varphi_{\lambda}^{a, b}$ by $$\begin{aligned} \label{scalingquantity} \varphi_{\lambda}^{a, b}:=e^{a\lambda}\varphi(e^{-b\lambda}x),\end{aligned}$$ where $(a, b)$ satisfies the condition $$\begin{aligned} \label{parameter} a>0, \;\; b\leq 0,\;\; 2a+b> 0\;\; 2a+3b\geq 0,\;\; (a,b)\neq (0,0).\end{aligned}$$ We define the scaling derivative of $S_{k}(\varphi_{\lambda}^{a, b})$ at $\lambda=0$ by $K_{k}^{a, b}(\varphi)$. $$\begin{aligned} \label{functionalK} K_{k}^{a, b}(\varphi)&:={\mathcal{L}}^{a,b}S_{k}(\varphi)=\frac{d}{d\lambda}\Big{|}_{\lambda=0}S_{k}(\varphi_{\lambda}^{a, b})\nonumber\\ &=\frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx +\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}V|\varphi(x)|^{2}dx +\frac{b}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx\\ &\;\;+\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{2}dx -\frac{4a+3b}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\nonumber\end{aligned}$$ In particular, when $(a, b)=(3, -2)$, $$\begin{aligned} \label{functionalP} P_{k}(\varphi):=K_{k}^{3,-2}(\varphi)=2\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx -\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx -\frac{3}{2}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx,\end{aligned}$$ which is related with the Virial identity of ($\rm{NLS_{k}}$) with $\phi(x)=|x|^{2}$, and when $(a, b)=(3, 0)$, $$\begin{aligned} \label{functionalI} I_{k}(\varphi):=\frac{1}{3}K_{k}^{3,0}(\varphi)=\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx +\displaystyle\int_{{\bf R}^{3}}V|\varphi(x)|^{2}dx +\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{2}dx -\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx.\end{aligned}$$ We note that, to get existence of minimal blow-up solutions, we need to use the functional $I_{k}$ instead of $P_{k}$. so that we can apply the linear profile decomposition Lemma \[linearprofile\]. The sharp threshold quantity $n_{k}$ are determined by the following minimizing problem $$\begin{aligned} \label{threshold} n_{k}=\inf\{S_{k}(\varphi):\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}, P_{k}(\varphi)=0\}.\end{aligned}$$ When $k=0$, $n_{0}$ is positive and is achieved by $Q$, which is a unique radial solution of (see [@AN]). The sharp criteria between scattering and blow up as mentioned above can be described by $n_{0}$. We state the results of [@HR; @DHR; @FXC; @AN] in terms of $n_{0}$ as follows. [@HR; @DHR; @FXC; @AN] \[scattering0\] Let $u_{0}\in H^{1}({\bf R}^{3})$ satisfy $S_{0}(u_{0})<n_{0}$. \(i) If $P_{0}\geq 0$, then the solution $u$ of ($\rm{NLS_{0}}$) is global and scatters. \(ii) If $P_{0}< 0$ and $u_{0}$ is radial or $xu_{0}\in L^{2}({\bf R}^{3})$, the the solution $u$ of ($\rm{NLS_{0}}$) blows up in finite time in both time directions. Furthermore, if $\psi\in H^{1}({\bf R}^{3})$ satisfying $\frac{1}{2}\|\psi\|_{H^{1}}^{2}<n_{0}$, then there exists a global solution of ($\rm{INLS_{0}}$) that scatters to $\psi$ in the positive time direction. The analogous statement holds in the negative time direction. When $k>0$, we prove that $n_{k}=n_{0}$ and $n_{k}$ is never attained (see Lemma \[attain\]). For succinctness, we next define two subsets of $H^{1}({\bf R}^{3})$ as follows: $$\begin{aligned} \label{n+} {\mathcal{N}}^{+}:=\{\varphi\in H^{1}({\bf R}^{3}): S_{k}(\varphi)<n_{0}, P_{k}(\varphi)\geq 0\}\end{aligned}$$ and $$\begin{aligned} \label{n-} {\mathcal{N}}^{-}:=\{\varphi\in H^{1}({\bf R}^{3}): S_{k}(\varphi)<n_{0}, P_{k}(\varphi)< 0\}.\end{aligned}$$ Now we state our main result. \[scattering1\] Let $u$ be the solution of ($\rm{NLS_{k}}$) on $(-T_{min}, T_{max})$, where $(-T_{min}, T_{max})$ is the maximal life-span. \(i) If $u_{0}\in {\mathcal{N}}^{+}$, then $u$ is global well-posedness, $u(t)\in {\mathcal{N}}^{+}$ for any $t\in {\bf R}$ and scatters. \(ii) If $u_{0}\in {\mathcal{N}}^{-}$, then $u(t)\in {\mathcal{N}}^{-}$ for any $t\in (-T_{min}, T_{max})$ and one of the following four statements holds true: \(a) $T_{max}<+\infty$ and $\lim_{t\uparrow T_{max}}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(b) $T_{min}<+\infty$ and $\lim_{t\downarrow -T_{min}}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(c) $T_{max}=+\infty$ and there exists a sequence $\{t_{n}\}_{n=1}^{+\infty}$ such that $t_{n}\rightarrow+\infty$ and $\lim_{t_{n}\uparrow +\infty}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(d) $T_{min}=+\infty$ and there exists a sequence $\{t_{n}\}_{n=1}^{+\infty}$ such that $t_{n}\rightarrow-\infty$ and $\lim_{t_{n}\downarrow -\infty}\|\nabla u(t)\|_{L^{2}}=+\infty$. Here the blow-up result is proved by the method of Du-Wu-Zhang [@DWZ]. This present paper is organized as follows. We fix notations at the end of Section 1. In Section 2, as preliminaries, we state some our required lemmas, including Sobolev norm equivalence, Strichartz estimates, stability theory, Littlewood-Paley theory, some limit lemmas between $H_{\alpha}^{n}$ and $H_{\alpha}^{\infty}$, linear profile decomposition and nonlinear profiles for $|x_{n}|\rightarrow+\infty$. In Section 3, using the variational idea of Ibrahim-Masmoudi-Nakanishi [@IMN], we obtain that if $\psi\in \mathcal{N}^{+}$, $P_{k}(\psi)$ and $I_{k}(\psi)$ have the same sign and $S_{k}(\psi)$ is equivalent to $\|\psi\|_{H^{1}}$ and if $\psi\in \mathcal{N}^{\pm}$, $P_{k}(\psi)$ has the uniform bounds, which play a vital role to get blow-up and scattering results. In Section 4, using the upper bound of $P_{k}(\psi)$ for $\psi\in \mathcal{N}^{-}$ and adopting the method of Du-Wu-Zhang [@DWZ], we establish blow-up part of Theorem \[scattering1\]. Global part of Theorem \[scattering1\] can be obtained by the lower bound of $P_{k}(\psi)$ for $\psi\in \mathcal{N}^{+}$ and local well-posedness $i$ of Theorem \[localwellposedness\]. In the last section, we show the scattering part of Theorem \[scattering1\] in two steps. In Step 1, by contradiction, if scattering fails, then a critical element must exist. In Step 2, we utilize lower bound of $P_{k}(\psi)$ for $\psi\in \mathcal{N}^{+}$ to preclude the critical element. Putting the last two sections together completes the proof of Theorem \[scattering1\]. **Notations:**: We fix notations used throughout the paper. In what follows, we write $A\lesssim B$ to signify that there exists a constant $c$ such that $A\leq cB$, while we denote $A\sim B$ when $A\lesssim B\lesssim A$. Given a real number $\alpha$, $\alpha-=\alpha-\epsilon$ for $0<\epsilon\ll 1$. Let $L_{I}^{q}L_{x}^{r}$ be the space of measurable functions from an interval $I\subset {\bf R}$ to $L_{x}^{r}$ whose $L_{I}^{q}L_{x}^{r}$- norm $ \|\cdot\|_{L_{I}^{q}L_{x}^{r}}$ is finite, where $$\begin{aligned} \|u\|_{L_{I}^{q}L_{x}^{r}}=\Big(\displaystyle\int_{I}\|u(t)\|_{L_{x}^{r}}^{q}dt\Big)^{\frac{1}{r}}.\end{aligned}$$ When $I={\bf R}$, we may use $L_{t}^{q}L_{x}^{r}$ instead of $L_{I}^{q}L_{x}^{r}$, respectively. In particular, when $q=r$, we may simply write them as $L_{t,x}^{q}$, respectively. Moreover, the Fourier transform on ${\bf R}^{3}$ is defined by $\hat{f}(\xi)=(2\pi)^{-\frac{3}{2}}\displaystyle\int_{{\bf R}^{3}} e^{-ix\cdot\xi}f(x)dx$. Define the inhomogeneous Sobolev space $H^{s}({\bf R}^{3})$ and and the homogeneous Sobolev space $\dot{H}^{s}({\bf R}^{3})$, respectively, by norms $$\|f\|_{H^{s}({\bf R}^{3})}=\|(1+|\xi|^{2})^{\frac{s}{2}}\hat{f}(\xi)\|_{L^{2}({\bf R}^{3})}=\|(1+\Delta)^{\frac{s}{2}} f\|_{L^{2}({\bf R}^{3})}$$ and $$\|f\|_{\dot{H}^{s}({\bf R}^{3})}=\||\xi|^{s}\hat{f}(\xi)\|_{L^{2}({\bf R}^{3})}=\|\Delta^{\frac{s}{2}} f\|_{L^{2}({\bf R}^{3})}.$$ Denote the inhomogeneous Sobolev space and homogeneous Sobolev space adapted to $H_{\alpha}$ by ${\mathcal{H}}_{k}^{s, p}({\bf R}^{3})$ and ${\mathcal{\dot H}}_{k}^{s, p}({\bf R}^{3})$, respectively, with norms $$\|f\|_{{\mathcal{H}}_{k}^{s, p}({\bf R}^{3})}=\|(1+H_{\alpha})^{\frac{s}{2}} f\|_{L^{p}({\bf R}^{3})}$$ and $$\|f\|_{{\mathcal{\dot H}}_{k}^{s, p}({\bf R}^{3})}=\|H_{\alpha}^{\frac{s}{2}} f\|_{L^{p}({\bf R}^{3})}.$$ ${\mathcal{H}}_{k}^{s}({\bf R}^{3})$ and ${\mathcal{\dot H}}_{k}^{s}({\bf R}^{3})$ are shorthand by ${\mathcal{H}}_{k}^{s, 2}({\bf R}^{3})$ and ${\mathcal{\dot H}}_{k}^{s, 2}({\bf R}^{3})$, respectively. Given $p\geq 1$, let $p'$ be the conjugate of $p$, that is $\frac{1}{p}+\frac{1}{p'}=1$. [**Acknowledgement**]{}  The first author is financially supported by the China National Science Foundation (No.11301564, 11771469), the second author is financially supported by the China National Science Foundation ( No. 11771165 and 11571131), and the third author is financially supported by the China National Science Foundation( No. 11771165). Preliminaries {#sec-2} ============= As we’ve mentioned in the introduction, the heat kernel associated with $H_{\alpha}$ satisfies , so Mikhlin multiplier theorem holds, which implies that for $\forall 1<p<\infty$, $$\begin{aligned} \label{Lpbound} \|f\|_{L^{p}}\lesssim \|(1+H_{\alpha})f\|_{L^{p}}.\end{aligned}$$ And we have the following Hardy type inequality for $H_{\alpha}$ (e.g., see [@KMVZS] for $\alpha=2$). $$\begin{aligned} \label{Hardy} \Big\||x|^{-s}f\Big\|_{L^{p}({\bf R}^{3})}\lesssim \|H_{\alpha}^{\frac{s}{2}}f\|_{L^{p}({\bf R}^{3})}\lesssim \|(1+H_{\alpha})^{\frac{s}{2}}f\|_{L^{p}({\bf R}^{3})},\end{aligned}$$ where $0<s<3$ and $1<p<\frac{3}{s}$. Using and Stein complex interpolation yields the following Sobolev norm equivalence (see [@H] for $V\geq 0$ and $V\in L^{\frac{3}{2}}$ and [@ZZ; @KMVZS; @MZZ] for $\alpha=2$). \[Sobolev\] Let $k>0$ and $0<\alpha< 2$, $1<p<\frac{3}{s}$ and $0\leq s\leq 2$, then $$\begin{aligned} \label{ihomogeneous} \|(1+H_{\alpha})^{\frac{s}{2}}f\|_{L^{p}({\bf R}^{3})}\sim \|(1-\Delta)^{\frac{s}{2}}f\|_{L^{p}({\bf R}^{3})}.\end{aligned}$$ As the heat kernel associated with $H_{\alpha}$ satisfies , the kernel of Riesz potentials $(1+H_{\alpha})^{-\frac{s}{2}}$ $$(1+H_{\alpha})^{-\frac{s}{2}}(x, y)=\frac{1}{\Gamma(\frac{s}{2})}\displaystyle\int_{0}^{+\infty}e^{-t(1+H_{\alpha})}(x, y)t^{\frac{s}{2}-1}dt$$ satisfies $$|(1+H_{\alpha})^{-\frac{s}{2}}(x, y)|\lesssim |x-y|^{s-3},$$ which, by Hardy-Littlewood-Sobolev inequality, implies that $$\begin{aligned} \label{HLS} \|(1+H_{\alpha})^{-\frac{s}{2}}f\|_{L^{\frac{3p}{3-ps}}}\lesssim \|f\|_{L^{p}}.\end{aligned}$$ It suffices to prove that with $s=2$ and $1<p<\frac{3}{2}$ holds. Indeed, if it is true, then follows from Stein complex interpolation and the $L^{p}$-boundedness of $(1+H_{\alpha})^{iy}$ with $\forall y\in {\bf R}$ and $1<p<+\infty$ (which can be obtained by and Sikora-Wright [@SW]). Let $\chi(x)$ be a smooth compact supported function such that $\chi(x)=1$ for $|x|\leq 1$ and $\chi(x)=0$ for $|x|\geq 2$. On one hand, using Hölder inequality and Sobolev embedding yields that $$\begin{aligned} \|(1+H_{\alpha})f\|_{L^{p}}&\leq \|(1-\Delta)f\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}f\Big\|_{L^{p}}\nonumber\\ &\leq \|(1-\Delta)f\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}\chi f\Big\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}(1-\chi) f\Big\|_{L^{p}}\nonumber\\ &\lesssim \|(1-\Delta)f\|_{L^{p}}+\Big\||x|^{-\alpha}\chi\Big\|_{L^{\frac{3}{2}}}\|f\|_{L^\frac{3p}{3-2p}}+\|f\|_{L^{p}}\nonumber\\ &\lesssim \|(1-\Delta)f\|_{L^{p}}.\end{aligned}$$ On the other hand, using Hölder inequality, and with $s=2$ gives that $$\begin{aligned} \|(1-\Delta)f\|_{L^{p}}&\leq \|(1+H_{\alpha})f\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}f\Big\|_{L^{p}}\nonumber\\ &\leq \|(1+H_{\alpha})f\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}\chi f\Big\|_{L^{p}}+k\Big\|\frac{1}{|x|^{\alpha}}(1-\chi) f\Big\|_{L^{p}}\nonumber\\ &\lesssim \|(1+H_{\alpha})f\|_{L^{p}}+\Big\||x|^{-\alpha}\chi\Big\|_{L^{\frac{3}{2}}}\|f\|_{L^\frac{3p}{3-2p}}+\|f\|_{L^{p}}\nonumber\\ &\lesssim \|(1+H_{\alpha})f\|_{L^{p}}.\end{aligned}$$ Thus, we get with $s=2$ and $1<p<\frac{3}{2}$ and then conclude the proof. Recently, Mizutani [@M] proved that the solution to free Schrödinger equations with a class of slowly decaying repulsive potentials including $k|x|^{-\alpha}$ with $k>0$ and $0<\alpha<2$ satisfies global-in-time Strichartz estimates for any admissible pairs. Besides, it is well known that Strichartz estimates for free Schrödinger equation with inverse-square potentials were established by Burq-Planchon-Stalker-Tahvildar-Zadeh [@BPSTZ]. Hence, we have the following global-in-time Strichartz estimate. [@BPSTZ; @M] \[Strichartz\] Let $k>0$ and $0<\alpha\leq 2$. Then the solution $u$ of $iu_{t}-H_{\alpha}u=F$ with initial data $u_{0}$ obeys $$\begin{aligned} \label{Strichartzestimates} \|u\|_{L_{t}^{q}L_{x}^{r}}\lesssim \|u_{0}\|_{L_{x}^{2}}+\|F\|_{L_{t}^{\tilde{q}'}L_{x}^{\tilde{r}'}}\end{aligned}$$ for any $2\leq q, \tilde{q}\leq \infty$ with $\frac{2}{q}+\frac{3}{r}=\frac{2}{\tilde{q}}+\frac{3}{\tilde{r}}=\frac{3}{2}$. Once we have Strichartz estimates Lemma \[Strichartz\] and Sobolev norm equivalence Lemma \[Sobolev\], the local well-posedness Theorem \[localwellposedness\] and stability result Lemma \[stability\] for ($\rm{NLS_{k}}$) can be obtained by the same proofs as in Theorem 2.15 and Theorem 2.17 of [@KMVZ], respectively. [@KMVZ] \[stability\] For $k>0$ and $0<\alpha\leq 2$. Let $\tilde{u}$ be the solution of $$\label{perturbation} \left\{ \begin{aligned} i&\tilde{u}_{t}-H_{\alpha}\tilde{u}+|\tilde{u}|^{2}\tilde{u}=e,\;\;(t,x) \in {I\times{\bf{R}}^{3}}, \\ \tilde{u}&(0, x)=\tilde{u}_{0}(x)\in H^{1}({\bf{R}}^{3}), \end{aligned}\right.$$ for some ’error’ $e$. Given $u_{0}\in H^{1}({\bf{R}}^{3})$ and assume $$\begin{aligned} \label{twoupperbound} \|u_{0}\|_{H^{1}}+\|\tilde{u}_{0}\|_{H^{1}}\leq A \;\;\text{and}\;\; \|\tilde{u}\|_{L_{t,x}^{5}}\leq M\end{aligned}$$ for some $A$, $M>0$. For any given $\frac{1}{2}\leq s<1$, there exists $\epsilon_{0}=\epsilon_{0}(A, M)$ such that if $0<\epsilon<\epsilon_{0}$ and $$\begin{aligned} \label{error} \|u_{0}-\tilde{u}_{0}\|_{H^{s}}+\Big\|(1-\Delta)^{\frac{s}{2}}e\Big\|_{N(I)}<\epsilon,\end{aligned}$$ where $$N(I):= L_{I,x}^{\frac{10}{7}}+L_{I}^{\frac{5}{3}}L_{x}^{\frac{30}{23}}+L_{I}^{1}L_{x}^{2},$$ then there exists a solution $u$ of ($\rm{NLS_{k}}$) such that $$\begin{aligned} \label{difference} \|u-\tilde{u}\|_{S_{\alpha}^{s}(I)}\lesssim_{A, M}\epsilon,\end{aligned}$$ $$\begin{aligned} \label{oneupperbound} \|u\|_{S_{\alpha}^{1}(I)}\lesssim_{A, M}1,\end{aligned}$$ where $$S_{\alpha}^{s}(I)=L_{I}^{2}{\mathcal{H}}_{\alpha}^{s,6}\cap L_{I}^{\infty}{\mathcal{H}}_{\alpha}^{s}.$$ Since Mikhlin multiplier theorem for $H_{\alpha}$ holds, naturally, we have the Littlewood-Paley theory associated with $H_{\alpha}$ (see [@KMVZS] for $\alpha=2$). We first give the definition of Littlewood-Paley projection via the heat kernel as follows: For $N\in 2^{\bf{Z}}$, $$\begin{aligned} \label{projection} P_{N}:=e^{-\frac{1}{N^{2}}H_{\alpha}}-e^{-\frac{4}{N^{2}}H_{\alpha}}.\end{aligned}$$ We next state Littlewood-Paley decomposition, square function estimate and Bernstein estimates (see [@KMVZS] for $\alpha=2$). \[LPdecompostion\] Let $1<p<\infty$. If $k>0$ and $0<\alpha\leq 2$, then $$\begin{aligned} \label{decompostioneq} f=\sum_{N\in 2^{{\bf{Z}}}}P_{N}f\end{aligned}$$ as elements of $L^{p}({\bf R}^{3})$. In particular, the sum converges in $L^{p}({\bf R}^{3})$. \[square\] Let $0\leq s<2$ and $1<p<+\infty$. If $k>0$ and $0<\alpha\leq 2$, then $$\begin{aligned} \label{squareestimate} \Big\|(\sum_{N\in 2^{{\bf{Z}}}}N^{2s}|P_{N}f|^{2})^{\frac{1}{2}}\Big\|_{L^{p}({\bf R}^{3})}\sim \|(H_{\alpha})^{\frac{s}{2}}\|_{L^{p}({\bf R}^{3})}.\end{aligned}$$ \[Bernstein\] Let $1<p\leq q<+\infty$ and $s\in {\bf R}$. If $k>0$ and $0<\alpha\leq 2$, then \(i) $$\begin{aligned} \label{Bernsteinpp} \|P_{N}f\|_{L^{p}({\bf R}^{3})}\lesssim \|f\|_{L^{p}({\bf R}^{3})}.\end{aligned}$$ \(ii) $$\begin{aligned} \label{Bernsteinpq} \|P_{N}f\|_{L^{q}({\bf R}^{3})}\lesssim N^{3(\frac{1}{p}-\frac{1}{q})}\|f\|_{L^{p}({\bf R}^{3})}.\end{aligned}$$ \(iii) $$\begin{aligned} \label{Bernsteinsim} \|(H_{\alpha})^{\frac{s}{2}}P_{N}f\|_{L^{p}({\bf R}^{3})}\sim N^{s}\|f\|_{L^{p}({\bf R}^{3})}.\end{aligned}$$ In order to establish linear profile decomposition associated with $e^{-itH_{\alpha}}$ and find a critical element, we apply the argument of [@KMVZE] to get some convergence results. For convenience, define two operators: Let $\{x_{n}\}_{n=1}^{+\infty}\subset {\bf R}^{3}$, $$\begin{aligned} \label{operatortranslation} H_{\alpha}^{n}:=-\Delta +V(x+x_{n})\;\;\text{and}\;\; H_{\alpha}^{\infty}:=\left\{ \begin{array}{rcl} -\Delta+V(x+\bar{x})& &{ x_{n}\rightarrow \bar{x}\in {\bf{R}}^{3}}\\ -\Delta\quad\quad\quad& &{ |x_{n}|\rightarrow +\infty} \end{array}\right.\end{aligned}$$ Obviously, $\tau_{x_{n}}H_{\alpha}^{n}\psi=H_{\alpha}\tau_{x_{n}}\psi$, where $\tau_{y}\psi(x)=\psi(x-y)$. So $H_{\alpha}$ doesn’t commute with $\tau_{x}$. $H_{\alpha}^{\infty}$ can be regarded as limits of $H_{\alpha}^{n}$ in the following sense (see [@KMVZE] for $\alpha=2$). \[limit\] Let $k>0$ and $1<\alpha< 2$. Assume $t_{n}\rightarrow\bar{t}\in {\bf R}$. and $\{x_{n}\}_{n=1}^{+\infty}\rightarrow\bar{x}\in{\bf R}^{3}$ or $|x_{n}|\rightarrow+\infty$. Then $$\begin{aligned} \label{operatorlimit} \lim_{n\rightarrow+\infty}\|(H_{\alpha}^{n}-H_{\alpha}^{\infty})\psi\|_{H^{-1}}=0,\;\;\forall \psi\in H^{1}.\end{aligned}$$ $$\begin{aligned} \label{grouplimit} \lim_{n\rightarrow+\infty}\|(e^{-it_{n}H_{\alpha}^{n}}-e^{-i\bar{t}H_{\alpha}^{\infty}})\psi\|_{H^{-1}}=0,\;\;\forall \psi\in H^{-1}.\end{aligned}$$ $$\begin{aligned} \label{fractionaloperatorlimit} \lim_{n\rightarrow+\infty}\|((1+H_{\alpha}^{n})^{\frac{1}{2}}-(1+H_{\alpha}^{\infty})^{\frac{1}{2}})\psi\|_{L^{2}}=0,\;\;\forall \psi\in H^{1}.\end{aligned}$$ For any $2<q\leq+\infty$ and $\frac{2}{q}+\frac{3}{r}=\frac{3}{2}$. $$\begin{aligned} \label{Strichartzlimit} \lim_{n\rightarrow+\infty}\|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{q}L_{x}^{r}}=0,\;\;\forall \psi\in L^{2}.\end{aligned}$$ If $\bar{x}\neq 0$, then for any $t>0$, $$\begin{aligned} \label{heatlimit} \lim_{n\rightarrow+\infty}\|(e^{-tH_{\alpha}^{n}}-e^{-tH_{\alpha}^{\infty}})\delta_{0}\|_{H^{-1}}=0.\end{aligned}$$ Here we only give the proof of because the others can be obtained by using the same method of Lemma 3.3 in [@KMVZE], where in the proof of - we need to replace $\dot{H}^{1}$, $\dot{H}^{-1}$, $H_{\alpha}^{n}$, $H_{\alpha}^{\infty}$ and homogeneous Sobolev equivalence Theorem 2.2 in [@KMVZE] with $H^{1}$, $H^{-1}$, $1+H_{\alpha}^{n}$, $1+H_{\alpha}^{\infty}$ and inhomogeneous Sobolev equivalence Lemma \[Sobolev\], respectively and in the proof of we need to replace $\dot{H}^{-1}$ by $H^{-1}$. It only suffices to prove in the case $(q, r)=(\infty, 2)$, since the general case can be obtained by interpolating with the end-point Strichartz estimates (i.e., $(q, r)=(2, 6)$). By density and Strichartz estimates, let $\psi$ be a smooth function with compact support, so that for $\forall M>0$, $$\begin{aligned} \label{dispersive} |(e^{-it\Delta}\psi)(x)|\lesssim_{\psi}\langle t\rangle^{-\frac{3}{2}}\Big(1+\frac{|x|}{\langle t\rangle}\Big)^{-M}.\end{aligned}$$ By the definition of $H_{\alpha}^{\infty}$, we consider two cases: $|x_{n}|\rightarrow+\infty$ and $x_{n}\rightarrow\bar{x}$. For the first case $|x_{n}|\rightarrow+\infty$, we have $$\begin{aligned} e^{it\Delta}\psi&=e^{-itH_{\alpha}^{n}}\psi+i\displaystyle\int_{0}^{t}e^{-i(t-s)H_{\alpha}^{n}}V(x+x_{n})e^{is\Delta}\psi ds\\ &=e^{-itH_{\alpha}^{n}}\psi+i\displaystyle\int_{0}^{t}e^{-i(t-s)H_{\alpha}^{n}}(\chi_{|x+x_{n}|\leq R}+\chi_{|x+x_{n}|> R})V(x+x_{n})e^{is\Delta}\psi ds,\end{aligned}$$ where $R$ is a sufficiently large number that will be chosen later and $\chi_{A}$ is a characteristic function on a set $A$. Hence by Strichartz estimates and dispersive estimates , we find that $$\begin{aligned} \label{twocases} \|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{\infty}L_{x}^{2}}&\lesssim \|\chi_{|x+x_{n}|\leq R}V(x+x_{n})e^{it\Delta}\psi\|_{L_{t,x}^{\frac{10}{7}}} +\|\chi_{|x+x_{n}|> R}V(x+x_{n})e^{it\Delta}\psi\|_{L_{t}^{q}L_{x}^{r}}\nonumber\\ &\lesssim \Big\|\chi_{|x+x_{n}|\leq R}V(x+x_{n})\langle t\rangle^{-\frac{3}{2}}\Big(1+\frac{|x|}{\langle t\rangle}\Big)^{-M}\Big\|_{L_{t,x}^{\frac{10}{7}}}\nonumber\\ &\quad+\Big\|\chi_{|x+x_{n}|> R}V(x+x_{n})\langle t\rangle^{-\frac{3}{2}}\Big(1+\frac{|x|}{\langle t\rangle}\Big)^{-M}\Big\|_{L_{t}^{q}L_{x}^{r}}\nonumber\\ &:=I_{1}+I_{2},\end{aligned}$$ where $(q,r)=(1,2)$ if $1<\alpha\leq \frac{3}{2}$ and $(q, r)=(\frac{10}{7}, \frac{10}{7})$ if $\frac{3}{2}<\alpha\leq 2$. We note that $|x_{n}|\geq R$ for $n$ large enough. So for the first part $I_{1}$, $R<|x|\sim |x_{n}|$. If $|x|\leq \langle t\rangle$, then $|x_{n}|\lesssim \langle t\rangle$. Therefore, $$\begin{aligned} \label{nlarge1} I_{1}\lesssim \|V(x+x_{n})\|_{L_{x}^{\frac{10}{7}}(|x+x_{n}|\leq R)}\Big\||t|^{-\frac{3}{2}}\Big\|_{L_{t}^{\frac{10}{7}}(|t|\gtrsim |x_{n}|)}\lesssim R^{-\alpha+\frac{21}{10}}|x_{n}|^{-\frac{4}{5}}.\end{aligned}$$ If $|x|> \langle t\rangle$, $$I_{1}\lesssim \|\chi_{|x+x_{n}|\leq R}V(x+x_{n})\langle t\rangle^{-\frac{3}{2}+M}|x|^{-M}\|_{L_{t,x}^{\frac{10}{7}}}.$$ When $|t|\leq 1$, it is easy to get $$\begin{aligned} \label{nlarge2} I_{1}\lesssim R^{-\alpha+\frac{21}{10}}|x_{n}|^{-M}.\end{aligned}$$ When $|t|>1$ and $M$ sufficiently large, $$\|\langle t\rangle^{-\frac{3}{2}+M}\|_{L_{t}^{\frac{10}{7}}(|t|\lesssim |x|)}\lesssim |x|^{-\frac{3}{2}+M+\frac{7}{10}},$$ so $$\begin{aligned} \label{nlarge3} I_{1}\lesssim R^{-\alpha+\frac{21}{10}}|x_{n}|^{-\frac{4}{5}}.\end{aligned}$$ Putting , and together gives that $I_{1}$ tends to 0 as $n$ approaches $+\infty$, regardless of $R>0$. For the second part $I_{2}$, we consider two subcases: $1<\alpha\leq\frac{3}{2}$ and $\frac{3}{2}<\alpha\leq 2$. If $1<\alpha\leq \frac{3}{2}$, $(q,r)=(1,2)$. When $|x|\leq \langle t\rangle$, using Hölder inequality, we have $$\begin{aligned} \label{Rlarge1} I_{2}\lesssim \|V(x+x_{n})\|_{L_{x}^{\frac{3}{\alpha-}}(|x+x_{n}|> R)}\Big\||t|^{-\frac{3}{2}}\Big\|_{L_{t}^{1}L_{x}^{\frac{6}{3-2(\alpha-)}}(|x|\leq\langle t\rangle)}\lesssim R^{1-\frac{\alpha}{\alpha-}}.\end{aligned}$$ When $|x|> \langle t\rangle$, $$\begin{aligned} \label{Rlarge2} I_{2}\lesssim \|V(x+x_{n})\|_{L_{x}^{\frac{3}{\alpha-}}(|x+x_{n}|> R)}\Big\||t|^{-\frac{3}{2}}|x|^{-M}\langle t\rangle^{M}\Big\|_{L_{t}^{1}L_{x}^{\frac{6}{3-2(\alpha-)}}(|x|>\langle t\rangle)}\lesssim R^{1-\frac{\alpha}{\alpha-}}.\end{aligned}$$ It follows from and that $I_{2}$ can be chosen arbitrarily small if we take $R$ sufficiently large. If $\frac{3}{2}<\alpha\leq 2$, $(q, r)=(\frac{10}{7}, \frac{10}{7})$. When $|x|\leq \langle t\rangle$, using Hölder inequality, we have $$\begin{aligned} \label{Rlarge3} I_{2}\lesssim \|V(x+x_{n})\|_{L_{x}^{2}(|x+x_{n}|> R)}\Big\||t|^{-\frac{3}{2}}\Big\|_{L_{t}^{\frac{10}{7}}L_{x}^{5}(|x|\leq\langle t\rangle)}\lesssim R^{-\alpha+\frac{3}{2}}.\end{aligned}$$ When $|x|> \langle t\rangle$, $$\begin{aligned} \label{Rlarge4} I_{2}\lesssim \|V(x+x_{n})\|_{L_{x}^{2}(|x+x_{n}|> R)}\Big\||t|^{-\frac{3}{2}}|x|^{-M}\langle t\rangle^{M}\Big\|_{L_{t}^{\frac{10}{7}}L_{x}^{5}(|x|>\langle t\rangle)}\lesssim R^{-\alpha+\frac{3}{2}}.\end{aligned}$$ It follows from and that $I_{2}$ can be chosen arbitrarily small if we take $R$ sufficiently large. So we get in the case $|x_{n}|\rightarrow+\infty$. Now we turn to the other case $x_{n}\rightarrow\bar{x}$. Here we obtain that $$\begin{aligned} e^{-itH_{\alpha}^{\infty}}\psi&=e^{-itH_{\alpha}^{n}}\psi+i\displaystyle\int_{0}^{t}e^{-i(t-s)H_{\alpha}^{n}}(V(x+x_{n})-V(x+\bar{x}))e^{-isH_{\alpha}^{\infty}}\psi ds\end{aligned}$$ Hence by Strichartz estimates , we have $$\begin{aligned} \|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{\infty}L_{x}^{2}}&\lesssim \|(V(x+x_{n})-V(x+\bar{x}))e^{-itH_{\alpha}^{\infty}}\psi\|_{L_{t}^{2}L_{x}^{\frac{6}{5}}}\nonumber\\ &=\|(V(x+x_{n}-\bar{x})-V(x))e^{-itH_{\alpha}}\tau_{\bar{x}}\psi\|_{L_{t}^{2}L_{x}^{\frac{6}{5}}}\end{aligned}$$ Replacing $x_{n}-\bar{x}$ by $x_{n}$ and $\tau_{\bar{x}}\psi$ by $\psi$, so we can suppose that $\bar{x}=0$. Hence, for $\forall \epsilon>0$, $|x_{n}|<\epsilon$ when $n$ is large enough. Besides, by Newton-Leibniz formula, $$|V(x+x_{n})-V(x)|\lesssim |x_{n}|\displaystyle\int_{0}^{1}|x+(1-t)x_{n}|^{-\alpha-1}dt.$$ Using Hölder inequality and Strichartz estimates yields that $$\begin{aligned} \|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{\infty}L_{x}^{2}} &\lesssim \|(V(x+x_{n})-V(x))\|_{L_{x}^{\frac{15}{11}}(|x|\leq 2\epsilon)}\|e^{-itH_{\alpha}}\psi\|_{L_{t}^{2}L_{x}^{10}}\nonumber\\ &\quad+\|(V(x+x_{n})-V(x))\|_{L_{x}^{\frac{3}{2}}(|x|> 2\epsilon)}\|e^{-itH_{\alpha}}\psi\|_{L_{t}^{2}L_{x}^{6}}\nonumber\\ &\lesssim\epsilon^{\frac{1}{5}}\|\psi\|_{{\mathcal H}_{k}^{\frac{1}{10}}}+|x_{n}|\epsilon^{-(\alpha-1)},\end{aligned}$$ which implies that $$\begin{aligned} \lim_{n\rightarrow+\infty}\|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{\infty}L_{x}^{2}}\lesssim \epsilon^{\frac{1}{5}}.\end{aligned}$$ Since $\epsilon$ can be chosen arbitrarily, $$\begin{aligned} \lim_{n\rightarrow+\infty}\|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t}^{\infty}L_{x}^{2}}=0.\end{aligned}$$ Thus, we get in the case $x_{n}\rightarrow\bar{x}$. Once getting and , we follow the same proof of Corollary 3.4 for $\alpha=2$ in [@KMVZE] and replace $\dot{H}^{1}$ and $H_{\alpha}^{n}$ with $H^{1}$ and $1+H_{\alpha}^{n}$ in the procedure of the proof, respectively, to get the following decaying estimates. \[decay\] Let $k>0$, $1<\alpha<2$. Given $\psi\in H^{1}$, $t_{n}\rightarrow\pm\infty$ and any sequence $\{x_{n}\}_{n=1}^{+\infty}\subset {\bf R}^{3}$. We have $$\begin{aligned} \label{decay6} \lim_{n\rightarrow+\infty}\|e^{-it_{n}H_{\alpha}^{n}}\psi\|_{L_{x}^{6}}=0.\end{aligned}$$ Moreover, if $\psi\in H^{1}$, then for $2<p<\infty$, $$\begin{aligned} \label{decayp} \lim_{n\rightarrow+\infty}\|e^{-it_{n}H_{\alpha}^{n}}\psi\|_{L_{x}^{p}}=0.\end{aligned}$$ Using , Sobolev equivalence Lemma \[Sobolev\] and interpolation yields the following convergence (see the same result and the detailed proof for $\alpha=2$ in [@KMVZ]). \[Strichartzlimitlemma\] Let $x_{n}\rightarrow\bar{x}\in {\bf R}^{3}$ or $x_{n}\rightarrow\pm\infty$. Then for $\forall\psi\in H^{1}$, $k>0$, $1<\alpha<2$, $$\begin{aligned} \label{Strichartzlimit5} \lim_{n\rightarrow+\infty}\|(e^{-itH_{\alpha}^{n}}-e^{-itH_{\alpha}^{\infty}})\psi\|_{L_{t,x}^{5}}=0.\end{aligned}$$ To get the parameters of linear profile decomposition are asymptotically orthogonal, we finally need two weak convergence results Lemma \[weak1\] and Lemma \[weak2\]. Since they are the direct consequences of Lemma \[limit\] and the detailed proof can be found in Lemma 3.8 and Lemma 3.9 for $\alpha=2$ in [@KMVZE] with a small modification by replacing $\dot{H}^{1}$ and $\Delta$ with $H^{1}$ and $1+\Delta$, respectively, and using the inequality $\|\sqrt{H_{\alpha}^{n}}\psi_{n}\|_{L^{2}}\lesssim \|\psi_{n}\|_{H^{1}}$, where the implicit constant is independent of $n$, we omit their proof here. \[weak1\] Let $\psi_{n}\in H^{1}({\bf R}^{3})$ satisfy $\psi_{n}\rightharpoonup 0$ in ${H}^{1}({\bf R}^{3})$ and let $t_{n}\rightarrow\bar{t}\in {\bf R}$. Then for $k>0$, $1<\alpha\leq2$, $$\begin{aligned} \label{twiceweak} e^{-it_{n}H_{\alpha}^{n}}\psi_{n}\rightharpoonup 0 \;\;\text{in}\;\; {H}^{1}({\bf R}^{3}).\end{aligned}$$ \[weak2\] Let $\psi\in H^{1}({\bf R}^{3})$ and let $\{(t_{n}, y_{n})\}\subset {\bf R}\times{\bf R}^{3}$, $|t_{n}|\rightarrow\infty$ or $|x_{n}|\rightarrow\infty$ . Then for $k>0$, $1<\alpha\leq2$, $$\begin{aligned} \label{onceweak} (e^{-it_{n}H_{\alpha}^{n}}\psi)(\cdot+y_{n})\rightharpoonup 0 \;\;\text{in}\;\; {H}^{1}({\bf R}^{3}).\end{aligned}$$ Use Lemma \[limit\], Lemma \[decay\], Lemma \[weak1\], Lemma \[weak2\] and Littlewood-Paley theory Lemma \[LPdecompostion\]- Lemma \[Bernstein\] and follow the proof of Proposition 5.1 for $\alpha=2$ in [@KMVZ] to give the following linear profile decomposition. Similar to the above, we need to use ${\mathcal{H}}_{k}^{1}$, $1+H_{\alpha}^{n}$, $1+H_{\alpha}^{\infty}$, and $H^{-1}$ to replace ${\mathcal{\dot{H}}}_{k}^{1}$, $H_{\alpha}^{n}$, $H_{\alpha}^{\infty}$ and $\dot{H}^{-1}$, respectively, in some appropriate places (e.g. in the proof of ). \[linearprofile\] Let $\{\phi_{n}\}_{n=1}^{+\infty}$ be a uniformly bounded sequence in $H^{1}({\bf R}^{3})$. Then there exist $M^{*}\in {\bf N}\cup\{+\infty\}$, a subsequence of $\{\phi_{n}\}_{n=1}^{M^{*}}$, which is denoted by itself, such that for $k>0$, $1<\alpha\leq2$, the following statements hold. \(1) For each $1\leq j\leq M\leq M^{*}$, there exist (fixed in $n$) a profile $\psi^{j}$ in $H^{1}({\bf R}^{3})$, a sequence (in $n$) of time shifts $t_{n}^{j}$ and a sequence (in $n$) of space shifts $x_{n}^{j}$, and there exists a sequence (in $n$) of remainder $W_{n}^{M}$ in $H^{1}({\bf R}^{3})$ such that $$\begin{aligned} \label{decomposition} \phi_{n}=\sum_{j=1}^{M}e^{it_{n}^{j}H_{\alpha}}\tau_{x_{n}^{j}}\psi^{j}+W_{n}^{M}:=\sum_{j=1}^{M}\psi_{n}^{j}+W_{n}^{M}.\end{aligned}$$ \(2) For each $1\leq j\leq M$, $$\begin{aligned} \label{zeroinfty} \text{either}\; t_{n}^{j}=0\; \text{for any}\; n\in {\bf N}\;\;\; \text{or}\; \lim_{n\rightarrow+\infty}t_{n}^{j}=\pm\infty\\ \text{either}\; x_{n}^{j}=0\; \text{for any}\; n\in {\bf N}\;\;\; \text{or}\; \lim_{n\rightarrow+\infty}|x_{n}^{j}|=+\infty.\end{aligned}$$ \(3) The time and space sequence have a pairwise divergence property. Namely, for $1\leq j\neq k\leq M$, $$\begin{aligned} \label{divergence} \lim_{n\rightarrow+\infty}(|t_{n}^{j}-t_{n}^{k}|+|x_{n}^{j}-x_{n}^{k}|)=+\infty.\end{aligned}$$ \(4) The remainder sequence has the following asymptotic smallness property and weak convergence property: $$\begin{aligned} \label{smallness} \lim_{M\rightarrow M^{*}}\overline{\lim}_{n\rightarrow+\infty}\|e^{-itH_{\alpha}}W_{n}^{M}\|_{L_{t,x}^{5}}=0\end{aligned}$$ and $$\begin{aligned} \label{remainderweak} \tau_{-x_{n}^{M}}e^{-it_{n}^{M}H_{\alpha}}W_{n}^{M}\rightharpoonup 0 \;\;\text{in}\;\; H^{1},\;\;\text{as}\;\; n\rightarrow+\infty.\end{aligned}$$ \(5) For each fixed $M$, we have the asymptotic Pythagorean expansion as follows: $$\begin{aligned} \label{expansion0} \|\phi_{n}\|_{L^{2}}^{2}=\sum_{j=1}^{M}\|\psi^{j}\|_{L^{2}}^{2}+\|W_{n}^{M}\|_{L^{2}}^{2}+o_{n}(1),\end{aligned}$$ $$\begin{aligned} \label{expansion1} \|\phi_{n}\|_{{\mathcal{\dot{H}}}_{k}^{1}}^{2}=\sum_{j=1}^{M}\|\tau_{x_{n}^{j}}\psi^{j}\|_{{\mathcal{\dot{H}}}_{k}^{1}}^{2}+\|W_{n}^{M}\|_{{\mathcal{\dot{H}}}_{k}^{1}}^{2}+o_{n}(1)\end{aligned}$$ and $$\begin{aligned} \label{expansion4} \|\phi_{n}\|_{L^{4}}^{4}=\sum_{j=1}^{M}\|\psi_{n}^{j}\|_{L^{4}}^{4}+\|W_{n}^{M}\|_{L^{4}}^{4}+o_{n}(1),\end{aligned}$$ where $o_{n}(1)\rightarrow 0$ as $n\rightarrow+\infty$. Specially, $$\begin{aligned} \label{expansions} S_{k}(\phi_{n})=\sum_{j=1}^{M}S_{k}(\psi_{n}^{j})+S_{k}(W_{n}^{M})+o_{n}(1)\end{aligned}$$ and $$\begin{aligned} \label{expansioni} I_{k}(\phi_{n})=\sum_{j=1}^{M}I_{k}(\psi_{n}^{j})+I_{k}(W_{n}^{M})+o_{n}(1).\end{aligned}$$ Following the proof of Theorem 6.1 for $\alpha=2$ in [@KMVZ] and using Theorem \[scattering0\], Lemma \[stability\], Lemma \[limit\] and Lemma \[Strichartzlimitlemma\], replacing $\frac{1}{|x|^{2}}$ and homogenous Sobolev spaces with $\frac{1}{|x|^{\alpha}}$ inhomogeneous Sobolev spaces, respectively, in some appropriate places gives the following lemma on nonlinear profiles when $|x_{n}|\rightarrow+\infty$. \[nonlinearprofile\] Let $k>0$, $1<\alpha\leq2$, $\psi\in H^{1}({\bf R}^{3})$ and the time sequence $t_{n}\equiv 0$ or $t_{n}\rightarrow\pm\infty$ such that $$\begin{aligned} \label{time0} \text{if}\;\; t_{n}\equiv 0,\;\;S_{0}(\psi)<n_{0}\;\;\text{ and}\;\; P_{0}(\psi)\geq 0\end{aligned}$$ and $$\begin{aligned} \label{timeinfty} \text{if}\;\; t_{n}\rightarrow\pm\infty,\;\;\frac{1}{2}\|\psi\|_{H^{1}}^{2}<n_{0}.\end{aligned}$$ Let $$\psi_{n}=e^{-it_{n}H_{\alpha}}\tau_{x_{n}}\psi,$$ where the space sequence $x_{n}$ satisfies $|x_{n}|\rightarrow+\infty$. Then taking $n$ large enough, we have that the solution $u(t):=NLS_{k}(t)\psi_{n}$ of ($\rm{NLS_{k}}$) with initial data $u_{0}=\psi_{n}$ is global and satisfies $$\|NLS_{k}(t)\psi_{n}\|_{S_{\alpha}^{1}(I)}\lesssim_{\|\psi\|_{H^{1}}}1.$$ Moreover, for $\forall \epsilon >0$, there exist a positive number $N=N(\epsilon)$ and a smooth compact supported function $\chi_{\epsilon}$ on ${\bf R}\times {\bf R}^{3}$ satisfying $$\begin{aligned} \label{dense} \|NLS_{k}(t)\psi_{n}(x)-\chi_{\epsilon}(t+t_{n}, x-x_{n})\|_{Z}<\epsilon\;\;\text{for}\;\;n\geq N,\end{aligned}$$ where the norm $$\|f\|_{Z}:=\|f\|_{L_{t,x}^{5}}+\|f\|_{L_{t,x}^{\frac{10}{3}}}+\|f\|_{L_{t}^{\frac{30}{7}}L_{x}^{\frac{90}{31}}}+\|f||_{L_{t}^{\frac{30}{7}}{\mathcal{H}}_{k}^{\frac{31}{60},\frac{90}{31}}}$$ Variational characterization {#section3} ============================ We start with proving the positivity of $K_{k}^{a, b}$ near $0$ in the $H^{1}({\bf R}^{3})$. \[positive\] If the uniform $L^{2}$-bounded sequence $\varphi_{n}\in H^{1}({\bf R}^{3})\setminus \{0\}$ satisfies $\lim_{n\rightarrow+\infty}\|\nabla \varphi_{n}\|_{L^{2}}=0$, then for sufficiently large $n\in {\bf N}$, $K_{k}^{a, b}(\varphi_{n})>0$. By the fact that $-2V\leq x\cdot\nabla V\leq 0$ and $V\geq 0$, we always have that for large enough $n$ $$\begin{aligned} K_{k}^{a, b}(\varphi_{n}) &=\frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi_{n}(x)|^{2}dx +\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}V|\varphi_{n}(x)|^{2}dx +\frac{b}{2}\displaystyle\int_{{\bf R}^{3}}x\cdot\nabla V |\varphi_{n}(x)|^{2}dx\\ &\;\;+\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{2}dx -\frac{4a+3b}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{4}dx\\ &\geq \frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi_{n}(x)|^{2}dx +\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{2}dx -\frac{4a+3b}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{4}dx\\ &\geq \frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi_{n}(x)|^{2}dx -\frac{4a+3b}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{4}dx\\ &\geq \frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi_{n}(x)|^{2}dx -\frac{4a+3b}{4}c\Big(\displaystyle\int_{{\bf R}^{3}}|\varphi_{n}(x)|^{2}dx\Big)\Big(\displaystyle\int_{{\bf R}^{3}}|\nabla\varphi_{n}(x)|^{2}dx\Big)^{\frac{3}{2}}\\ &\geq \frac{2a+b}{4}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi_{n}(x)|^{2}dx>0,\end{aligned}$$ where we have used the Gagliardo-Nirenberg inequality in the line before last and the assumptions $\lim_{n\rightarrow+\infty}\|\nabla \varphi_{n}\|_{L^{2}}=0$ and uniform boundedness of $\|\varphi_{n}\|_{L^{2}}$ in the last line. Simple computation gives $$\|\nabla \varphi_{\lambda}^{a, b}\|_{L^{2}}=e^{\frac{1}{2}(2a+b)\lambda}\|\nabla \varphi\|_{L^{2}},$$ which implies that $$\begin{aligned} \label{gradient0} \lim_{\lambda\rightarrow-\infty}\|\nabla \varphi_{\lambda}^{a, b}\|_{L^{2}}=0.\end{aligned}$$ So it follows from Lemma \[positive\] that for sufficiently small $\lambda<0$, $$\begin{aligned} \label{positiveK} K_{k}^{a, b}(\varphi_{\lambda}^{a, b})>0.\end{aligned}$$ For brevity, let $\overline{\mu}=2a+b$ and $\underline{\mu}=2a+3b$. Next we introduce the functional $$J_{k}^{a,b}(\varphi):=\frac{1}{\overline{\mu}}(\overline{\mu}-{\mathcal{L}}^{a,b})S_{k}(\varphi) =\frac{1}{\overline{\mu}}(\overline{\mu}S_{k}(\varphi)-K_{k}^{a, b}(\varphi)).$$ The lemma shows that the positivity of $J_{k}^{a,b}(\varphi)$ and the monotonicity of $J_{k}^{a,b}(\varphi_{\lambda}^{a, b})$ in $\lambda$. \[monotonicity\] For any $\varphi\in H^{1}({\bf R}^{3})$, we have the following two identities: $$\begin{aligned} \label{positiveJ} \overline{\mu}J_{k}^{a,b}(\varphi)&=(\overline{\mu}-{\mathcal{L}}^{a,b})S_{k}^{a,b}(\varphi)\nonumber\\ &=-b\displaystyle\int_{{\bf R}^{3}}V|\varphi|^{2}dx-\frac{b}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V)|\varphi|^{2}dx-b\|\varphi\|_{L^{2}}^{2}+\frac{a+b}{2}\|\varphi\|_{L^{4}}^{4}.\end{aligned}$$ and $$\begin{aligned} \label{monotonicityJ} ({\mathcal{L}}^{a,b}-\overline{\mu})(\underline{\mu}-{\mathcal{L}}^{a,b})S_{k}^{a,b}(\varphi)=b^{2}\displaystyle\int_{{\bf R}^{3}}(-3x\cdot\nabla V-x\nabla^{2} Vx^{T})|\varphi|^{2}dx, +(a+b)a\|\varphi\|_{L^{4}}^{4}\end{aligned}$$ where $\nabla^{2} V$ is Hessian matric of $V$. By simple computations, we have $${\mathcal{L}}^{a,b}\|\nabla\varphi\|_{L^{2}}^{2}=\overline{\mu}\|\nabla\varphi\|_{L^{2}}^{2},\;\; {\mathcal{L}}^{a,b}\|\varphi\|_{L^{2}}^{2}=\underline{\mu}\|\varphi\|_{L^{2}}^{2},\;\; {\mathcal{L}}^{a,b}\|\varphi\|_{L^{4}}^{4}=(4a+3b)\|\varphi\|_{L^{4}}^{4},$$ $${\mathcal{L}}^{a,b}\Big(\displaystyle\int_{{\bf R}^{3}}V|\varphi|^{2}dx\Big)=\underline{\mu}\displaystyle\int_{{\bf R}^{3}}V|\varphi|^{2}dx+b\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V)|\varphi|^{2}dx,$$ and $${\mathcal{L}}^{a,b}\Big(\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V)|\varphi|^{2}dx\Big)=2(a+b)\displaystyle\int_{{\bf R}^{3}}x\cdot\nabla V|\varphi|^{2}dx +b\displaystyle\int_{{\bf R}^{3}}x\nabla^{2} Vx^{T}|\varphi|^{2}dx,$$ which imply that and . We conclude the proof. \[positive-monotonicity\] By the definition of $V$, we have $2V+x\cdot V\geq 0$ and $3x\cdot\nabla V+x\nabla^{2}Vx^{T}\leq 0$, which together with and yield that $J_{k}^{a,b}(\varphi)> 0$ and ${\mathcal{L}}^{a,b}J_{k}^{a,b}(\varphi)\geq 0$ for any $\varphi\in H^{1}({\bf R}^{3})\setminus\{0\}$ which implies that the function $\lambda \mapsto J_{k}^{a,b}(\varphi_{\lambda}^{a, b})$ is increasing. Using the functional $K_{k}^{a,b}$ with $(a, b)$ satisfying , we introduce a general minimizing problem: $$\begin{aligned} \label{thresholdab} n_{k}^{a,b}=\inf\{S_{k}(\varphi):\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}, K_{k}^{a,b}(\varphi)=0\}.\end{aligned}$$ In particular, when $(a,b)=(3,-2)$, it is namely $n_{k}$ in . In fact, we shall show that $$\begin{aligned} \label{threshold4} n_{k}^{a,b}=n_{0}^{a,b}=n_{k}=n_{0}.\end{aligned}$$ To this end, we introduce another one general minimizing problem in terms of the positive functional $J_{0}^{a,b}$ for any ${a, b}$ satisfying : $$\begin{aligned} \label{thresholdJ} j^{a,b}=\inf\{J_{0}^{a,b}(\varphi):\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}, K_{0}^{a,b}(\varphi)\leq 0\}.\end{aligned}$$ In the following lemma, we shall show a relation between the two minimizers $n_{0}^{a, b}$ in (i.e., $k=0$) and $j^{a,b}$ in . \[j=n0\] $$\begin{aligned} \label{jab=n0} j^{a,b}=n_{0}^{a, b}.\end{aligned}$$ On one hand, for any $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$ with $K_{0}^{a,b}(\varphi)\leq 0$, there are two possibilities: $K_{0}^{a,b}(\varphi)= 0$ and $K_{0}^{a,b}(\varphi)< 0$. When $K_{0}^{a,b}(\varphi)=0$, $S_{0}(\varphi)=J_{0}^{a,b}(\varphi)$ by the definition of $J_{k}^{a, b}$. Hence, $$\begin{aligned} \label{onehand} n_{0}^{a, b}\leq J_{0}^{a,b}(\varphi).\end{aligned}$$ When $K_{0}^{a,b}(\varphi)=K_{0}^{a,b}(\varphi_{0}^{a, b})< 0$, using the continuity of $K_{0}^{a, b}(\varphi_{\lambda}^{a, b})$ in $\lambda$ and the fact that for sufficiently small $\lambda<0$ such that $K_{0}^{a, b}(\varphi_{\lambda}^{a, b})>0$ holds by yields that there exists a $\lambda_{0}<0$ such that $K_{0}^{a, b}(\varphi_{\lambda_{0}}^{a, b})=0$. Using Remark \[positive-monotonicity\], we get $$S_{0}(\varphi_{\lambda_{0}}^{a, b})=J_{0}^{a, b}(\varphi_{\lambda_{0}}^{a, b})\leq J_{0}^{a,b}(\varphi_{0}^{a, b})=J_{0}^{a,b}(\varphi)$$ Hence, we still have . Altogether, for any $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$ with $K_{0}^{a,b}(\varphi)\leq 0$, we have . By the definition of $j^{a, b}$, we have $n_{0}^{a,b}\leq j^{a,b}$. On the other hand, for any $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$ with $K_{0}^{a,b}(\varphi)=0$. Of course, $K_{0}^{a,b}(\varphi)\leq 0$ and $S_{0}(\varphi)=J_{0}^{a,b}(\varphi)$, which implies that $j^{a, b}\leq n_{0}^{a,b}$. Thus, we complete the proof. Next we shall apply Lemma \[j=n0\] to get $n_{k}^{a,b}=n_{0}^{a,b}$. \[n=n0\] $$\begin{aligned} \label{nab=n0} n_{k}^{a,b}=n_{0}^{a, b}.\end{aligned}$$ On one hand, for any $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$ with $K_{k}^{a,b}(\varphi)= 0$, we have $S(\varphi)=J^{a,b}(\varphi)$ by the definition of $J_{k}^{a, b}$. Since $V\geq 0$ and $2V+x\cdot\nabla V\geq 0$, we get $K_{0}^{a,b}(\varphi)\leq K_{k}^{a,b}(\varphi)=0$ and $J_{0}^{a,b}(\varphi)\leq J_{k}^{a,b}(\varphi)$, which together with implies that $$n_{0}^{a, b}=j^{a,b}\leq J_{k}^{a,b}(\varphi).$$ Also as $K^{a,b}(\varphi)= 0$, taking infimum on both sides of the above inequality yields that $$n_{0}^{a, b}\leq n_{k}^{a, b}.$$ On the other hand, we review that $Q$ is the positive radial exponential decaying solution of the elliptic equation , so there exists a constant $c$ such that $Q(x)\lesssim e^{-c|x|}$ for any $x\in {\bf R}^{3}$ (e.g. See Theorem 8.1.1 in [@C]. Here we only need the decaying property of $Q$, not necessarily exponential decaying). Let $x_{n}$ be a sequence satisfying $|x_{n}|\rightarrow+\infty$. Hence, for any given $R>0$, there exists $N=N(R)>0$, for any $n\geq N$, we have $|x_{n}|\geq 2R$. we claim that $$\begin{aligned} \label{V01} \displaystyle\int_{{\bf R}^{3}}V(x)|Q(x-x_{n})|^{2}dx\rightarrow 0\;\;\text{as}\;\; n\rightarrow +\infty.\end{aligned}$$ Indeed, $$\begin{aligned} \displaystyle\int_{{\bf R}^{3}}V(x)|Q(x-x_{n})|^{2}dx &\leq \displaystyle\int_{|x|\leq R}V(x)|Q(x-x_{n})|^{2}dx +\displaystyle\int_{|x|> R}V(x)|Q(x-x_{n})|^{2}dx\\ &:=I_{1}+I_{2}.\end{aligned}$$ For the first part $I_{1}$, when $n\geq N$, $|x-x_{n}|\geq \frac{|x_{n}|}{2}$ and then $$|Q(x-x_{n})|\leq \sup_{|x|\geq\frac{|x_{n}|}{2}}|Q(x)|:=C(|x_{n}|)\rightarrow 0$$ as $n$ tends to $+\infty$. Therefore, $$\begin{aligned} I_{1}\lesssim C(|x_{n}|)\displaystyle\int_{|x|\leq R}V(x)dx\lesssim R^{(3-\alpha)}C(|x_{n}|).\end{aligned}$$ For the second part $I_{2}$, $V(x)\lesssim R^{-\alpha}$, which implies that $$I_{2}\lesssim R^{-\alpha}\displaystyle\int_{{\bf R}^{3}}|Q(x)|^{2}dx.$$ Combining the above two parts yields that the claim holds true. Noticing that $x\cdot\nabla V=(-\alpha) V$, thus we also have $$\begin{aligned} \label{nablaV0} \displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V)|Q(x-x_{n})|^{2}dx\rightarrow 0\;\;\text{as}\;\; n\rightarrow +\infty.\end{aligned}$$ Using the definition of $K_{k}^{a, b}$ and $K_{0}^{a, b}(Q)=0$ and taking $\theta$ sufficiently large, we have $$K_{k}^{a, b}(\tau_{x_{n}}Q)>K_{0}^{a, b}(\tau_{x_{n}}Q)=K_{0}^{a, b}(Q)=0$$ and $$K_{k}^{a, b}(\theta\tau_{x_{n}}Q)<0,$$ from which it follows that there must be $\theta_{n}>1$ satisfying $K_{k}^{a, b}(\theta_{n}\tau_{x_{n}}Q)=0$. We claim that $$\begin{aligned} \label{theta1} \lim_{n\rightarrow +\infty}\theta_{n}=1.\end{aligned}$$ In fact, by $K_{0}^{a, b}(Q)=0$, we have $$\frac{2a+b}{2}\displaystyle\int_{{\bf R}^{3}}|\nabla Q(x)|^{2}dx +\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}|Q(x)|^{2}dx =\frac{4a+3b}{4}\displaystyle\int_{{\bf R}^{3}}|Q(x)|^{4}dx,$$ which implies that $$\begin{aligned} K_{k}^{a, b}(\theta_{n}\tau_{x_{n}}Q)&=\theta_{n}^{2}\Big[\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}V|Q(x-x_{n})|^{2}dx +\frac{b}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |Q(x-x_{n})|^{2}dx\\ &\;\;+\frac{4a+3b}{4}(1-\theta_{n}^{2})\displaystyle\int_{{\bf R}^{3}}|Q(x)|^{4}dx\Big]=0.\end{aligned}$$ Hence, $$\frac{2a+3b}{2}\displaystyle\int_{{\bf R}^{3}}V|Q(x-x_{n})|^{2}dx +\frac{b}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |Q(x-x_{n})|^{2}dx +\frac{4a+3b}{4}(1-\theta_{n}^{2})\displaystyle\int_{{\bf R}^{3}}|Q(x)|^{4}dx=0.$$ Taking limit in the above quality and using and gives the claim . Using and yields that $$\lim_{n\rightarrow+\infty}S_{k}(\theta_{n}\tau_{x_{n}}Q)=S_{0}(Q)=n_{0}^{a,b},$$ which together with $K_{k}^{a, b}(\theta_{n}\tau_{x_{n}}Q)=0$ implies that $$n_{k}^{a, b}\leq n_{0}^{a,b}.$$ We conclude the proof. Lemma \[n=n0\] shows that $n_{k}^{a,b}=n_{0}^{a,b}=S_{0}(Q)$, so $n_{k}^{a,b}$ and $n_{0}^{a,b}$ don’t depend on the parameters $a$ and $b$. Therefore, holds true. It is known that $n_{0}$ is attained by $Q$. However, the following lemma shows that $n_{k}$ is never attained for any $k>0$. \[attain\] $n_{k}$ is never attained for any $k>0$. Suppose that there exists a $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$ such that $n_{k}$ is attained by $\varphi$. Namely, $P_{k}(\varphi)=0$ and $S_{k}(\varphi)=n_{k}$. As $\varphi\in H^{1}({\bf R}^{3})\setminus \{0\}$, $\lim_{|x|\rightarrow +\infty}\varphi(x)=0$. Following the proof of , we have $$\begin{aligned} \label{V0} \displaystyle\int_{{\bf R}^{3}}V(x)|\varphi(x-x_{n})|^{2}dx\rightarrow 0\;\;\text{as}\;\; n\rightarrow +\infty.\end{aligned}$$ where $|x_{n}|\rightarrow+\infty$ as $n\rightarrow +\infty$. Hence, we have $$\begin{aligned} -\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\tau_{x_{n}}\varphi|^{2}dx\; \text{ is positive and tends to zero as}\;\; n\rightarrow +\infty\end{aligned}$$ and $$\begin{aligned} 2\displaystyle\int_{{\bf R}^{3}}V|\tau_{x_{n}}\varphi|^{2}dx +\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V )|\tau_{x_{n}}\varphi|^{2}dx\; \text{ is positive and tends to zero as}\;\; n\rightarrow +\infty.\end{aligned}$$ Therefore, for $n$ sufficiently large, we have $$\begin{aligned} \label{K0} P_{k}(\tau_{x_{n}}\varphi)<P_{k}(\varphi)=0\end{aligned}$$ and $$\begin{aligned} \label{J} J_{k}^{3, -2}(\tau_{x_{n}}\varphi)<J_{k}^{3,-2}(\varphi)=S_{k}(\varphi)=n_{k}.\end{aligned}$$ By , for sufficiently small $\lambda<0$, we have $P_{k}((\tau_{x_{n}}\varphi)_{\lambda}^{3, -2})>0$, which combined with implies that there exists a $\lambda_{0}<0$ such that $P_{k}((\tau_{x_{n}}\varphi)_{\lambda_{0}}^{3, -2})=0$. Using Remark \[positive-monotonicity\] and , we get $$n_{k}=S_{k}((\tau_{x_{n}}\varphi)_{\lambda_{0}}^{3, -2})=J_{k}^{3, -2}((\tau_{x_{n}}\varphi)_{\lambda_{0}}^{3, -2}) < J_{k}^{3,-2}((\tau_{x_{n}}\varphi)_{0}^{3, -2})=J_{k}^{3,-2}(\tau_{x_{n}}\varphi)<n_{k},$$ which is impossible. Thus we complete the proof. To get the fact that $P_{k}(\varphi)$ in and $I_{k}(\varphi)$ in have the same sign under the condition $S_{k}(\varphi)<n_{0}$ with $\varphi\in H^{1}({\bf R}^{3})$, we introduce ${\mathcal{N}}_{a,b}^{\pm}\subset H^{1}({\bf R}^{3})$ defined by $$\begin{aligned} \label{nab+} {\mathcal{N}}_{a,b}^{+}:=\{\varphi\in H^{1}({\bf R}^{3}): S_{k}(\varphi)<n_{0}, K_{k}^{a,b}(\varphi)\geq 0\}\end{aligned}$$ and $$\begin{aligned} \label{nab-} {\mathcal{N}}_{a,b}^{-}:=\{\varphi\in H^{1}({\bf R}^{3}): S_{k}(\varphi)<n_{0}, K_{k}^{a,b}(\varphi)< 0\}.\end{aligned}$$ It is easy to see that ${\mathcal{N}}_{3,-2}^{\pm}={\mathcal{N}}^{\pm}$ in and . The above fact can be obtained if we show that ${\mathcal{N}}_{a,b}^{\pm}$ are independent of $(a, b)$ (i.e., ${\mathcal{N}}_{a,b}^{\pm}={\mathcal{N}}^{\pm}$), which follows from the contractivity of ${\mathcal{N}}_{a,b}^{+}$. \[independentofab\] Let $(a, b)$ satisfy , then ${\mathcal{N}}_{a,b}^{\pm}$ are independent of $(a, b)$. The proof is similar to the one of Lemma 2.9 in [@IMN] (see also Lemma 2.15 in [@II]), so we omit it. The following lemma shows that for any element $\varphi$ in ${\mathcal{N}}^{+}$, $S_{k}(\varphi)\sim \|\varphi\|_{{\mathcal{H}}_{k}^{1}}\sim \|\varphi\|_{H^{1}}$. \[equivalentSH\] Let $\varphi\in {\mathcal{N}}^{+}$, then $$\begin{aligned} \label{equivalenceSH} \frac{1}{4}\|\varphi\|_{{\mathcal{H}}_{k}^{1}}^{2}\leq S_{k}(\varphi)\leq\frac{1}{2}\|\varphi\|_{{\mathcal{H}}_{k}^{1}}^{2}.\end{aligned}$$ By Lemma \[independentofab\], we have $I_{k}(\varphi)\geq 0$, which implies that $$\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\leq \|\varphi\|_{{\mathcal{H}}_{k}^{1}}^{2}.$$ Thus, we have $$\frac{1}{2}\|\varphi\|_{{\mathcal{H}}_{k}^{1}}^{2}\geq S_{k}(\varphi)=\frac{1}{2}\|\varphi\|_{{\mathcal H}_{k}^{1}}-\frac{1}{4}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx \geq \frac{1}{4}\|\varphi\|_{{\mathcal{H}}_{k}^{1}}^{2},$$ which is namely . We complete the proof. Finally, we obtain the corresponding uniform bounds on $P_{k}(\varphi)$ when $\varphi\in {\mathcal{N}}^{\pm}$, which plays a vital role in the proof of Theorem \[scattering1\]. \[uniformbounds\] 1\. Let $\varphi\in {\mathcal{N}}^{-}$, then $$\begin{aligned} \label{upperbound} P_{k}(\varphi)\leq -4\Big(n_{0}-S_{k}(\varphi)\Big)\end{aligned}$$ 2\. Let $\varphi\in {\mathcal{N}}^{+}$, then $$\begin{aligned} \label{lowerbound} P_{k}(\varphi)\geq \min\Big\{4\Big(n_{0}-S_{k}(\varphi)\Big), \frac{2}{5}\Big(\|\nabla\varphi\|_{L^{2}}^{2}-\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx)\Big)\Big\}.\end{aligned}$$ For any $\varphi\in H^{1}({\bf R}^{3})$, define $s(\lambda):=S_{k}(\varphi_{\lambda}^{3,-2})$, then $$\begin{aligned} &s(\lambda)=\frac{1}{2}e^{4\lambda}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx +\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}V(e^{-2\lambda}x)|\varphi(x)|^{2}dx\nonumber\\ &\quad\quad\quad +\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{2}dx -\frac{1}{4}e^{6\lambda}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx,\nonumber\\ &s'(\lambda)=P_{k}(\varphi_{\lambda}^{3,-2})=2e^{4\lambda}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx -e^{-2\lambda}\displaystyle\int_{{\bf R}^{3}}\Big[x\cdot (\nabla V)(e^{-2\lambda}x)\Big]|\varphi(x)|^{2}dx\label{s'}\\ &\quad\quad\quad-\frac{3}{2}e^{6\lambda}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\nonumber\\ &\quad\quad\geq 2e^{4\lambda}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx -\frac{3}{2}e^{6\lambda}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx,\nonumber\\ &s''(\lambda)=8e^{4\lambda}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx +2e^{-2\lambda}\displaystyle\int_{{\bf R}^{3}}\Big[x\cdot (\nabla V)(e^{-2\lambda}x)\Big]|\varphi(x)|^{2}dx \label{s''}\\ &\quad\quad\quad+2e^{-4\lambda}\displaystyle\int_{{\bf R}^{3}}\Big[x(\nabla^{2} V)(e^{-2\lambda}x)x^{T}\Big]|\varphi(x)|^{2}dx -9e^{6\lambda}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\nonumber\\ &\quad\quad\leq 8e^{4\lambda}\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx -4e^{-2\lambda}\displaystyle\int_{{\bf R}^{3}}\Big[x\cdot (\nabla V)(e^{-2\lambda}x)\Big]|\varphi(x)|^{2}dx\nonumber\\ &\quad\quad\quad-9e^{6\lambda}\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\nonumber\\ &\quad\quad= 4s'(\lambda)-3\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx\leq 4s'(\lambda)\nonumber,\end{aligned}$$ where we have used the inequalities $x\cdot\nabla V$ and $3x\cdot\nabla V+x\nabla^{2} Vx^{T}\leq 0$. 1\. If $\varphi\in {\mathcal{N}}^{-}$, then it follows from that $s'(0)=P_{k}(\varphi)<0$ and $s'(\lambda)>0$ for sufficiently small $\lambda<0$. Thus, by the continuity of $s'(\lambda)=P_{k}(\varphi_{\lambda}^{3,-2})$ in $\lambda$, there exists a negative $\lambda_{0}<0$ such that $$\begin{aligned} s'(\lambda_{0})=P_{k}(\varphi_{\lambda_{0}}^{3,-2})=0\;\;\text{and}\;\;s'(\lambda)<0,\;\;\text{for}\;\;\forall \lambda\in (\lambda_{0},0].\end{aligned}$$ By the definition of $n_{0}$ , $s(\lambda_{0})=S_{k}(\varphi_{\lambda_{0}}^{3,-2})\geq n_{0}$. Integrating the inequality over $[\lambda_{0}, 0]$ yields that $$P_{k}(\varphi)=s'(0)=s'(0)-s'(\lambda_{0})\leq 4(s(0)-s(\lambda_{0}))\leq 4(S_{k}(\varphi)-n_{0})=-4(n_{0}-S_{k}(\varphi)).$$ Thus, we complete the proof of . 2\. If $\varphi\in {\mathcal{N}}^{+}$, we consider two cases: one is $8P_{k}(\varphi)\geq 3\|\varphi\|_{L^{4}}^{4}$ and the other is $8P_{k}(\varphi)< 3\|\varphi\|_{L^{4}}^{4}$. For the case $8P_{k}(\varphi)\geq 3\|\varphi\|_{L^{4}}^{4}$, it follows from the definition of $P_{k}$ that $$2P_{k}(\varphi)=4\displaystyle\int_{{\bf R}^{3}}|\nabla \varphi(x)|^{2}dx -2\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx -3\displaystyle\int_{{\bf R}^{3}}|\varphi(x)|^{4}dx,$$ and then we have $$10P_{k}(\varphi)\geq 4(\|\nabla\varphi\|_{L^{2}}^{2}-\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx),$$ that is, $$\begin{aligned} \label{lowerbound1} P_{k}(\varphi)\geq \frac{2}{5}\Big(\|\nabla\varphi\|_{L^{2}}^{2}-\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |\varphi(x)|^{2}dx\Big).\end{aligned}$$ For the other case $8P_{k}(\varphi)< 3\|\varphi\|_{L^{4}}^{4}$, by , we have $$\begin{aligned} \label{s'decreasing} 0<8s'(\lambda)<3e^{6\lambda}\|\varphi\|_{L^{4}}^{4}\;\;\text{and then}\;\;s''(\lambda)\leq 4s'(\lambda)-3e^{6\lambda}\|\varphi\|_{L^{4}}^{4}< -4s'(\lambda)\end{aligned}$$ at $\lambda=0$. Also as $s'(\lambda)$ and $s''(\lambda)$ are continuous, $s'(\lambda)$ decreases as $s$ increases until $s'(\lambda_{1})=0$ for some $0<\lambda_{1}<+\infty$ and is true over $[0, \lambda_{1}]$. Since $P_{k}(\varphi_{\lambda_{1}}^{3,-2})=s'(\lambda_{1})=0$, by the definition of $n_{0}$ , $s(\lambda_{1})=S_{k}(\varphi_{\lambda_{1}}^{3,-2})\geq n_{0}$. Integrating the second inequality in over $[0, \lambda_{1}]$, we have $$\begin{aligned} -P_{k}(\varphi)=s'(\lambda_{1})-s'(0)<-4(s(\lambda_{1})-s(0))\leq -4\Big(n_{0}-S_{k}(\varphi)\Big),\end{aligned}$$ which is $$\begin{aligned} \label{lowerbound2} P_{k}(\varphi)\geq 4\Big(n_{0}-S_{k}(\varphi)\Big).\end{aligned}$$ Putting and together yields . Criteria for global well-posedness and blow-up ============================================== In this section, we will give the criteria for global well-posedness and blow-up for the solution $u$ of ($\rm{NLS_{k}}$), which are partial results of Theorem \[scattering1\]. The proof of blow-up part is based on the argument of [@DWZ]. \[globalvsblowup\] Let $u$ be the solution of ($\rm{NLS_{k}}$) on $(-T_{min}, T_{max})$, where $(-T_{min}, T_{max})$ is the maximal life-span. \(i) If $u_{0}\in {\mathcal{N}}^{+}$, then $u$ is global well-posedness and $u(t)\in {\mathcal{N}}^{+}$ for any $t\in {\bf R}$. \(ii) If $u_{0}\in {\mathcal{N}}^{-}$, then $u(t)\in {\mathcal{N}}^{-}$ for any $t\in (-T_{min}, T_{max})$ and one of the following four statements holds true: \(a) $T_{max}<+\infty$ and $\lim_{t\uparrow T_{max}}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(b) $T_{min}<+\infty$ and $\lim_{t\downarrow -T_{min}}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(c) $T_{max}=+\infty$ and there exists a sequence $\{t_{n}\}_{n=1}^{+\infty}$ such that $t_{n}\rightarrow+\infty$ and $\lim_{t_{n}\uparrow +\infty}\|\nabla u(t)\|_{L^{2}}=+\infty$. \(d) $T_{min}=+\infty$ and there exists a sequence $\{t_{n}\}_{n=1}^{+\infty}$ such that $t_{n}\rightarrow-\infty$ and $\lim_{t_{n}\downarrow -\infty}\|\nabla u(t)\|_{L^{2}}=+\infty$. $(i)$ Define $$I^{+}=\{t\in (-T_{min}, T_{max}): u(t)\in {\mathcal{N}}^{+}\}.$$ Obviously, $0\in I^{+}\neq \Phi$. On one hand, since $S_{k}(u(t))=S_{k}(u_{0})<n_{0}$ and $P_{k}(u(t))$ is continuous in $t$, $I^{+}$ is a closed subset of $(-T_{min}, T_{max})$. On the other hand, by , we further obtain that, $I^{+}$ is a open subset of $(-T_{min}, T_{max})$. Therefore, $I^{+}=(-T_{min}, T_{max})$. Namely, for any $t\in (-T_{min}, T_{max})$, $u(t)\in {\mathcal{N}}^{+}$. It follows form that for any $t\in (-T_{min}, T_{max})$, $$\|u(t)\|_{H^{1}}^{2}\leq \|u(t)\|_{{\mathcal{H}}_{k}^{1}}^{2}\leq 4S_{k}(u(t))\leq 4n_{0}.$$ So by local well-posedness result $(i)$ of Theorem \[localwellposedness\], we have $(-T_{min}, T_{max})={\bf R}$, which implies that $u$ is global well-posedness and $u(t)\in {\mathcal{N}}^{+}$ for any $t\in {\bf R}$. Thus, we complete the proof of $(i)$. $(ii)$ Similarly above, we can show that $u(t)\in {\mathcal{N}}^{-}$ for any $t\in (-T_{min}, T_{max})$ by replacing with . In the sequel, we only consider the positive time because the negative time can be dealt with similarly. If $T_{max}<+\infty$, we naturally have $\lim_{t\uparrow T_{max}}\|\nabla u(t)\|_{L^{2}}=+\infty$. If $T_{max}=+\infty$, we shall prove $\lim_{t\uparrow +\infty}\|\nabla u(t)\|_{L^{2}}=+\infty$ by contradiction. Assume we have $$C_0=\sup_{t\in\mathbb R^+}\|\nabla u(t)\|_{L^2}<\infty.$$ Consider the localized Virial identity and define $$\label{vf} I(t):=\int_{{\bf R}^{3}}\phi(x)|u(t,x)|^2dx,$$ then by straight computations, we obtain that for any $\phi\in C^4(\mathbb R^3)$ (e.g., see Proposition 7.1 in [@H]) $$I'(t)=2{\operatorname{Im}}\int_{{\bf R}^{3}}\nabla\phi\cdot\nabla u\bar udx;$$ $$\begin{aligned} I''(t)=\int_{{\bf R}^{3}}4{\operatorname{Re}}\nabla\bar u\nabla^2\phi\nabla udx -\int_{{\bf R}^{3}}2\nabla\phi\cdot\nabla V|u|^2+\Delta\phi |u|^4dx -\int_{{\bf R}^{3}}\Delta^2\phi|u|^2dx.\end{aligned}$$ In particular, if $\phi$ is a radial function , $$\begin{aligned} \label{I'} I'(t)=2{\operatorname{Im}}\int_{{\bf R}^{3}}\phi'(r)\frac{x\cdot\nabla u}r\bar udx,\end{aligned}$$ $$\begin{aligned} \label{I''0} &I''(t)=4\int_{{\bf R}^{3}}\frac{\phi'}r|\nabla u|^2dx+4\int_{{\bf R}^{3}}\left(\frac{\phi''}{r^2}-\frac{\phi'}{r^3}\right)|x\cdot\nabla u|^2dx\\ &-2\int_{{\bf R}^{3}}\frac{\phi'}{r}x\cdot\nabla V|u|^2dx-\int_{{\bf R}^{3}}\left(\phi''(r)+\frac{2}r\phi'(r)\right) |u|^4dx\nonumber -\int_{{\bf R}^{3}}\Delta^2\phi|u|^2dx.\end{aligned}$$ [**$L^2$ estimate in the exterior ball**]{} Given $R\gg 1$, which will be determined later. Take $\phi$ in such that $$\phi=\begin{cases}0,&0\leq r\leq\frac R2;\\1,&r\geq R,\end{cases}$$ and $$0\leq\phi\leq1,\ \ 0\leq\phi'\leq\frac4R.$$ By and Hölder inequality, there holds that $$\begin{aligned} I(t)=&I(0)+\int_0^tI'(\tau)d\tau \leq I(0)+t\|\phi'\|_{L^\infty}M(u_{0}) C_0\\ \leq&\int_{|x|\geq\frac R2}|u_0|^2dx+\frac{4M(u_{0}) C_0t}R.\end{aligned}$$ Note that $$\int_{|x|\geq\frac R2}|u_0|^2dx=o_R(1),$$ and $$\int_{|x|\geq R}|u(t,x)|^2dx\leq I(t).$$ So for given $\eta_0>0$, if $$t\leq\frac{\eta_0R}{4M(u_{0}) C_0},$$ then we have that $$\begin{aligned} \label{outermass} \int_{|x|\geq\frac R2}|u(t,x)|^2dx\leq\eta_0+o_R(1).\end{aligned}$$ [**Localized Virial identity**]{} $I''(t)$ can be rewritten as $$\begin{aligned} \label{I''} I''(t)=4P_{k}(u)+R_1+R_2+R_3+R_4,\end{aligned}$$ where $$R_1=4\int_{{\bf R}^{3}}\left(\frac{\phi'}r-2\right)|\nabla u|^2dx+4\int_{{\bf R}^{3}} \left(\frac{\phi''}{r^2}-\frac{\phi'}{r^3}\right)|x\cdot\nabla u|^2dx,$$ $$R_2=-\int_{{\bf R}^{3}}\left(\phi''+\frac{2}r\phi'(r)-6\right)| u|^4dx,$$ $$R_3=-2\int_{{\bf R}^{3}}\left(\frac{\phi'}{r}-2\right)(x\cdot\nabla V)|u|^2dx,$$ $$R_4=-\int_{{\bf R}^{3}}\Delta^2\phi|u|^2dx.$$ At this stage, we choose another radial function $\phi$ such that $$\begin{aligned} \label{radialfunction} 0\leq\phi\leq r^2,\ \ \phi''\leq2,\ \ \phi^{(4)}\leq\frac4{R^2},\;\; \text{and}\;\; \phi=\begin{cases}r^2,&0\leq r\leq R;\\0,&r\geq 2R\end{cases}.\end{aligned}$$ First, note that $R_1\leq0$. If $\phi''\leq r^{-1}\phi'\leq0$, by $\phi'\leq2r$, it is easy to see that $R_1\leq0$. If $\phi''\leq r^{-1}\phi'\leq0$, by $\phi''\leq2$, it holds that $$R_1\leq 4\int_{{\bf R}^{3}}\left(\frac{\phi'}r-2\right)|\nabla u|^2dx+4\int_{{\bf R}^{3}} \left(\phi''-\frac{\phi'}r\right)|\nabla u|^2dx=4\int_{{\bf R}^{3}}\left(\phi''-2\right)|\nabla u|^2dx\leq 0.$$ Secondly, let $\chi^{4}(r)=|\phi''+\frac{2}r\phi'(r)-6|$. it is easy to see that ${\rm supp}\chi\subset[R,\infty).$ So by Gagliardo-Nirenberg inequality $$R_2\leq \displaystyle\int_{{\bf R}^{3}}\chi^{4}|u|^{4}dx\lesssim \|\nabla(\chi u)\|_{L^{2}}^{3}\|\chi u\|_{L^{2}} \leq C(M(u_{0}), C_0) \|u\|_{L^2(|x|>R)}^{\frac{1}{4}}.$$ By the properties of $\phi$, $$R_4\leq CR^{-2}\|u\|_{L^2(|x|>R)}^2.$$ Finally, by $x\cdot\nabla V\leq0$, and we obtain $R_3\leq0$. Putting all the above estimates together, there holds that for $R\gg 1$, $$\begin{aligned} \label{I''upperbound} I''(t)\leq 4P_{k}(u)+\tilde C\|u\|_{L^2(|x|>R)},\end{aligned}$$ where $\tilde C>0$ depending on $M(u_{0})$ and $C_0$. Using and yields that $t\leq T:=\eta_0R/(4M(u_{0}) C_0)$, $$I''(t)\leq 4P_{k}(u)+\tilde C(\eta_0^{1/2}+o_R(1)).$$ As $u(t)\in {\mathcal{N}}^{-}$, it follows from that there exists $$\beta_{0}:=-4(n_{0}-S_{k}(u(t)))=-4(n_{0}-S_{k}(u_{0}))<0$$ such that $$\begin{aligned} \label{I''upperbound1} I''(t)\leq 4\beta_{0}+\tilde C(\eta_0^{1/2}+o_R(1)).\end{aligned}$$ Choose $\eta_0$ sufficiently small and take $R$ sufficiently large such that $$4\beta_{0}+\tilde C\eta_0^{1/2}+o_R(1)<\beta_{0}.$$ Integrating over $[0, T]$ twice, we obtain that $$\begin{aligned} I(T)&\leq I(0)+I'(0)T+\int_0^T\int_0^t\beta_{0}\\ &\leq I(0)+I'(0)T+\beta_{0}\frac{T^2}2.\end{aligned}$$ Hence for $T=\eta_0R/(4M(u_{0})C_0)$, we obtain that $$\begin{aligned} \label{IT} I(T)\leq I(0)+I'(0)\eta_0R/(4M(u_{0}) C_0)+\alpha_0R^2,\end{aligned}$$ where the constant $$\alpha_0=\beta_0\eta_0^2/(4M(u_{0}) C_0)^2<0$$ is independent of $R$. At the same time, we note that $$\begin{aligned} \label{I0} I(0)=o_R(1)R^2,\ \ \ I'(0)=o_R(1)R.\end{aligned}$$ In fact, $$\begin{aligned} I(0)&\leq\int_{|x|<\sqrt{R}}|x|^2|u_0|^2dx+\int_{\sqrt{R}<|x|<2R}|x|^2|u_0|^2dx\\ &\leq RM(u_{0})+R^2\int_{|x|>\sqrt{R}}|u_0|^2dx=o_R(1)R^2.\end{aligned}$$ Similarly, we obtain the second estimate and then prove . Putting with together and choosing $R$ sufficiently enough, we find that $$I(T)\leq (o_R(1)+\alpha_0)R^2\leq\frac14\alpha_0R^2<0,$$ which is impossible since $I\geq 0$. Thus, we conclude the proof of blow-up part. Scattering result ================= In this section, we shall show the remaining part of Theorem \[scattering1\]. In the previous section, we have obtained that the solution $u(t)$ of ($\rm{NLS_{k}}$) is global and belongs to $\mathcal{N}^{+}$ if $u_{0}\in \mathcal{N}^{+}$. To get scattering result, by $(iv)$ of Theorem \[localwellposedness\], it’s enough to get . To this end, we introduce a definition. \[SC\] We say that $SC(u_0)$ holds if for $u_0\in H^1({\bf R}^{3})$ satisfying $u_{0}\in \mathcal{N}^{+}$, the corresponding global solution $u$ of ($\rm{NLS_{k}}$) satisfies . We first note that for $u_{0}\in \mathcal{N}^{+}$, there exists $\delta>0$ such that if $S_{k}(u_{0})<\delta$, then holds. In fact, by , $\|u_{0}\|_{H^{1}}\lesssim S_{k}(u_{0})$. Therefore, by $(ii)$ of Theorem \[localwellposedness\], taking $\delta>0$ sufficiently small gives . Now for each $\delta>0$, we define the set $S_\delta$ as follows: $$\begin{aligned} \label{Sdelta} S_\delta=\{u_0\in H^1({\bf R}^{3}):\ \ S_{k}(u_{0})<\delta \ \ and \ \ u_{0}\in{\mathcal{N}^{+}} \Rightarrow \ \eqref{scatteringbound}\ holds\}.\end{aligned}$$ We also define $$\begin{aligned} \label{nc} n_c=\sup\{\delta:\ \ u_0\in S_\delta\Rightarrow SC(u_0)\ \ holds \}.\end{aligned}$$ Hence, $0<n_{c}\leq n_{0}$. Next we shall prove that $n_{c}<n_{0}$ is impossible, which implies that $n_{c}=n_{0}$. Thus, we assume now $$n_{c}<n_{0}.$$ By the definition of $n_{c}$, we can find a sequence of solutions $u_{n}$ of ($\rm{NLS_{k}}$) with initial data $\phi_{n}\in {\mathcal{N}^{+}}$ such that $S_{k}(\phi_{n})\rightarrow n_{c}$ and $$\begin{aligned} \label{L5} \|u_{n}\|_{L_{{\bf R}^{+},x}^{5}}=+\infty\;\;\text{and}\;\;\|u_{n}\|_{L_{{\bf R}^{-},x}^{5}}=+\infty.\end{aligned}$$ In the subsequent subsection, our goal is to prove the existence of critical element $u_{c}\in H^1({\bf R}^{3})$, which is a global solution of ($\rm{NLS_{k}}$) with initial data $u_{c,0}$ such that $S_{k}(u_{c,0})=n_{c}$, $u_{c,0}\in {\mathcal{N}^{+}}$ and $SC(u_{c,0})$ does not hold. Moreover, we prove that if $\|u_{c}\|_{L_{t,x}^{5}}=+\infty$, then $K:=\{u_c(t): t\in {\bf R}\}$ is precompact in $H^1({\bf R}^{3})$. Before showing the existence and compactness of critical element $u_{c}$, we need a lemma related with the linear profile decomposition Lemma \[linearprofile\]. \[whole-partial\] Let $M\in {\bf N}$ and $\psi^{j}\in H^{1}({\bf R}^{3})$ for any $0\leq j\leq M$. Suppose that there exist some $\delta>0$ and $\epsilon>0$ with $2\epsilon<\delta$ such that $$\begin{aligned} \sum_{j=0}^{M}S_{k}(\psi^{j})-\epsilon\leq S_{k}\Big(\sum_{j=0}^{M}\psi^{j}\Big)\leq n_{0}-\delta,\:\:\;\; -\epsilon\leq I_{k}\Big(\sum_{j=0}^{M}\psi^{j}\Big)\leq \sum_{j=0}^{M}I_{k}(\psi^{j})+\epsilon.\end{aligned}$$ Then $\psi^{j}\in \mathcal{N}^{+}$ for any $0\leq j\leq M$. Suppose that for some $0\leq l\leq M$, $I_{k}(\psi^{l})<0$. By with $(a, b)=(3,0)$, we have $I_{k}\Big((\psi^{l})_{\lambda}^{3,0}\Big)>0$ for sufficiently small $\lambda<0$. Thus, by continuity of $I_{k}\Big((\psi^{l})_{\lambda}^{3,0}\Big)$ in $\lambda$, there exists $\lambda_{2}<0$ such that $I_{k}\Big((\psi^{l})_{\lambda_{2}}^{3,0}\Big)=0$. As $n_{k}^{3.0}=n_{0}$, by the increasing property of $J_{k}^{3,0}\Big((\psi^{l})_{\lambda}^{3,0}\Big)$ in $\lambda$, we have $$J_{k}^{3,0}(\psi^{l})\geq J_{k}^{3,0}\Big((\psi^{l})_{\lambda_{2}}^{3,0}\Big)=S_{k}\Big((\psi^{l})_{\lambda_{2}}^{3,0}\Big)\geq n_{0}.$$ By the nonnegativity of $J_{k}^{3,0}(\psi^{j})$ for any $0\leq j\leq M$ and $2\epsilon<\delta$, we have $$\begin{aligned} n_{0}&\leq J_{k}^{3,0}(\psi^{l})\leq \sum_{j=0}^{M}J_{k}^{3,0}(\psi^{j})=\sum_{j=0}^{M}\Big(S_{k}(\psi^{j})-\frac{1}{2}I_{k}(\psi^{j})\Big)\\ &\leq S_{k}\Big(\sum_{j=0}^{M}\psi^{j}\Big)+\epsilon-\frac{1}{2}I_{k}\Big(\sum_{j=0}^{M}\psi^{j}\Big)+\frac{1}{2}\epsilon\\ &\leq n_{0}-\delta+2\epsilon<n_{0},\end{aligned}$$ which is impossible. Hence, for each $0\leq j\leq M$, we obtain $$I_{k}(\psi^{j})\geq 0.$$ So $$S_{k}(\psi^{j})=J_{k}^{3,0}(\psi^{j})+\frac{1}{2}I_{k}(\psi^{j})\geq 0,$$ which together with $$\sum_{j=0}^{M}S_{k}(\psi^{j})\leq S_{k}\Big(\sum_{j=0}^{M}\psi^{j}\Big)+\epsilon\leq n_{0}-\delta+\epsilon<n_{0}$$ yields that $S_{k}(\psi^{j})<n_{0}$ for each $0\leq j\leq M$. By Lemma \[independentofab\], $\psi^{j}\in \mathcal{N}^{+}$ for any $0\leq j\leq M$. Existence and compactness of critical element --------------------------------------------- \[criticalelement\] There exists a $u_{c,0}$ in $H^1({\bf R}^{3})$ with $S_{k}(u_{c,0})=n_{c}$, $u_{c,0}\in {\mathcal{N}^{+}}$ such that if $u_c$ is the corresponding global solution of ($\rm{NLS_{k}}$) with the initial data $u_{c,0}$, then $\|u_c\|_{L_{t,x}^{5}({\bf R}\times{\bf R}^{3})}=+\infty$ and $K$ is precompact in $H^1({\bf R}^{3})$. We first note that $\{\phi_{n}\}_{n=1}^{+\infty}$ be a uniformly bounded sequence in $H^{1}({\bf R}^{3})$. In fact, since $\phi_{n}\in \mathcal{N}^{+}$ for any $n\in {\bf N}$, by , $$\|\phi_{n}\|_{H^{1}}\leq \|\phi_{n}\|_{{\mathcal{H}}_{k}^{1}}^{2}\leq 4S_{k}(\phi_{n})<4n_{0}.$$ We apply Lemma \[linearprofile\] to $\phi_{n}$ to get that for each $M\leq M^{*}$, $$\begin{aligned} \label{M} \phi_{n}=\sum_{j=1}^{M}\psi_{n}^{j}+W_{n}^{M},\end{aligned}$$ $$\begin{aligned} S_{k}(\phi_{n})=\sum_{j=1}^{M}S_{k}(\psi_{n}^{j})+S_{k}(W_{n}^{M})+o_{n}(1),\end{aligned}$$ and $$\begin{aligned} I_{k}(\phi_{n})=\sum_{j=1}^{M}I_{k}(\psi_{n}^{j})+I_{k}(W_{n}^{M})+o_{n}(1),\end{aligned}$$ which together with $\phi_{n}\in \mathcal{N}^{+}$ yield that there exist some $\delta>0$ and $\epsilon>0$ with $2\epsilon<\delta$ such that $$\begin{aligned} \sum_{j=1}^{M}S_{k}(\psi_{n}^{j})+S_{k}(W_{n}^{M})-\epsilon\leq S_{k}(\phi_{n})\leq n_{0}-\delta,\:\:\;\; -\epsilon\leq I_{k}(\phi_{n})\leq \sum_{j=1}^{M}I_{k}(\psi_{n}^{j})+I_{k}(W_{n}^{M})+\epsilon.\end{aligned}$$ According to Lemma \[whole-partial\], we have that for large $n$ and each $1\leq j\leq M$, $\psi_{n}^{j}$, $W_{n}^{M}\in {\mathcal{N}^{+}}$, and then $$\begin{aligned} \label{M=1} 0\leq \overline{\lim_{n\rightarrow+\infty}}S_{k}(\psi_{n}^{j}) \leq \overline{\lim_{n\rightarrow+\infty}}S_{k}(\phi_{n})=n_{c},\end{aligned}$$ where if equality holds in the last inequality for some $j$, we must have $M^{*}=1$ and $W_{n}^{1}\rightarrow 0$ in $H^{1}$. We claim that if equality holds in the last inequality of for some $j$ (w.l.g. let $j=1$), $u_{c,0}$ is namely $\psi^{1}$. Indeed, at this time we have $$\begin{aligned} \label{psi1} \phi_{n}=\psi_{n}^{1}+W_{n}^{1}=e^{it_{n}^{1}H_{\alpha}}\tau_{x_{n}^{1}}\psi^{1}+W_{n}^{1},\end{aligned}$$ $$\begin{aligned} \overline{\lim_{n\rightarrow+\infty}}S_{k}(\psi_{n}^{1})=n_{c}\end{aligned}$$ and $$\begin{aligned} \label{wn1} W_{n}^{1}\rightarrow 0 \;\;\text{in}\;\; H^{1}.\end{aligned}$$ Our target is to prove that $$\begin{aligned} \label{xt} x_{n}^{1}\equiv 0\;\;\text{ and}\;\; t_{n}^{1}\equiv 0.\end{aligned}$$ If is true, then we have $$\begin{aligned} \phi_{n}=\psi^{1}+W_{n}^{1},\;\; S_{k}(\psi^{1})=n_{c},\;\; \psi^{1}\in\mathcal{N}^{+}\end{aligned}$$ and $$\begin{aligned} \lim_{n\rightarrow+\infty}\|\phi_{n}-\psi^{1}\|_{H^{1}}=0.\end{aligned}$$ Here $\psi^{1}$ is namely our required $u_{c,0}$. Let $u_c$ be the solution of ($\rm{NLS_{k}}$) with the initial data $u_{c,0}=\psi^{1}$, then $u_{c}$ is global and $S_{k}(u_{c})=S_{k}(u_{c,0})=n_{c}$. Using Lemma \[stability\], it holds that $\|u_c\|_{L_{t,x}^{5}({\bf R}\times{\bf R}^{3})}=+\infty$. Otherwise, $\|u_n\|_{L_{t,x}^{5}}<+\infty$, which contradicts with . If is false, then either $|x_{n}^{1}|\rightarrow+\infty$ or $t_{n}^{1}\rightarrow\pm\infty$. I will see that it leads to $\|u_n\|_{L_{t,x}^{5}}<+\infty$ or $\|u_{n}\|_{L_{{\bf R}^{\pm},x}^{5}}<+\infty$ contradicting with . For the case $|x_{n}^{1}|\rightarrow+\infty$, by , we have $$\begin{aligned} \label{H1} \lim_{n\rightarrow +\infty}\|\psi_{n}^{1}\|_{{\mathcal H}_{k}^{1}}=\|\psi^{1}\|_{H^{1}}>0,\end{aligned}$$ which implies that when $t_{n}^{1}\equiv 0$, $$S_{0}(\psi^{1})=\overline{\lim_{n\rightarrow+\infty}}S_{k}(\psi_{n}^{1})=n_{c}<n_{0}\;\;\text{and} \;\;I_{0}(\psi^{1})=\overline{\lim_{n\rightarrow+\infty}}I_{k}(\psi_{n}^{1})\geq 0.$$ By , $P_{0}(\psi^{1})\geq 0$. Hence, when $t_{n}^{1}\equiv 0$, $\psi^{1}$ satisfies the condition . When $t_{n}^{1}\rightarrow\pm\infty$, apply and to get $$\frac{1}{2}\|\psi^{1}\|_{H^{1}}^{2}=\overline{\lim_{n\rightarrow+\infty}}S_{k}(\psi_{n}^{1})=n_{c}<n_{0},$$ that is, $\psi^{1}$ satisfies the condition . Using Theorem \[nonlinearprofile\] yields that the solution $NLS_{k}(t)\psi_{n}^{1}$ of ($\rm{NLS_{k}}$) with initial data $\psi_{n}^{1}$ is global and satisfies $$\|NLS_{k}(t)\psi_{n}^{1}\|_{S_{\alpha}^{1}(I)}\lesssim_{\|\psi^{1}\|_{H^{1}}}1.$$ We know that $W_{n}^{1}\rightarrow 0$ in $H^{1}$, which is $$\begin{aligned} \lim_{n\rightarrow+\infty}\|\phi_{n}-\psi_{n}^{1}\|_{H^{1}}=0.\end{aligned}$$ Using Lemma \[stability\] again, we obtain $\|u_n\|_{L_{t,x}^{5}}<+\infty$. For the other case $t_{n}^{1}\rightarrow\pm\infty$, we only cope with $t_{n}^{1}\rightarrow-\infty$ since $t_{n}^{1}\rightarrow+\infty$ can be dealt with similarly. Apply with $x_{n}^{1}\equiv 0$, , Strichartz estimates Lemma \[Strichartz\] and norm equivalence Lemma \[Sobolev\] to get that $$\begin{aligned} \lim_{n\rightarrow+\infty}\|e^{-itH_{\alpha}}\phi_{n}\|_{L_{{\bf R}^{+},x}^{5}} &\leq \lim_{n\rightarrow+\infty}\|e^{-i(t-t_{n}^{1})H_{\alpha}}\psi^{1}\|_{L_{{\bf R}^{+},x}^{5}} +\lim_{n\rightarrow+\infty}\|e^{-itH_{\alpha}}W_{n}^{1}\|_{L_{{\bf R}^{+},x}^{5}}\\ &\lesssim \lim_{n\rightarrow+\infty}\|W_{n}^{1}\|_{H^{1}} +\lim_{n\rightarrow+\infty}\|e^{-itH_{\alpha}}\psi^{1}\|_{L_{(-t_{n}^{1}, +\infty),x}^{5}}=0,\end{aligned}$$ which immediately implies that $\lim_{n\rightarrow+\infty}\|u_{n}\|_{L_{{\bf R}^{+},x}^{5}}=0$ by $(ii)$ of Theorem \[localwellposedness\]. Thus, we obtain that holds true. Next we turn to the other situation that equality doesn’t hold in the last inequality of for any $1\leq j\leq M$. So for each $1\leq j\leq M$ and $\psi_{n}^{j}\in \mathcal{N}^{+}$, there exists $\delta=\delta_{j}>0$ such that $$\begin{aligned} \overline{\lim_{n\rightarrow+\infty}}S_{k}(\psi_{n}^{j}) \leq n_{c}-2\delta,\;\; P_{k}(\psi_{n}^{j})\geq 0\;\;\text{and}\;\;I_{k}(\psi_{n}^{j})\geq 0.\end{aligned}$$ We shall use $\psi_{n}^{j}$ to constitute approximate solutions of $u_{n}$ under three cases: $|x_{n}^{j}|\rightarrow+\infty$; $x_{n}^{j}\equiv 0$ and $t_{n}^{j}\equiv 0$; $x_{n}^{j}\equiv 0$ and $t_{n}^{j}\rightarrow\pm\infty$ and then apply Lemma \[stability\] to get a contradiction. For some $j$ such that $|x_{n}^{j}|\rightarrow+\infty$, still holds for $\psi_{n}^{j}$. Using the same argument after , we obtain that $\psi^{j}$ satisfies or . Therefore, using Theorem \[nonlinearprofile\], we can constitute a global solution $v_{n}^{j}(t):=NLS_{k}(t)\psi_{n}^{j}$ of ($\rm{NLS_{k}}$) with initial data $\psi_{n}^{j}$ such that $$\|v_{n}^{j}\|_{L_{t,x}^{5}}\leq \|NLS_{k}(t)\psi_{n}^{j}\|_{S_{\alpha}^{1}(I)}\lesssim_{\|\psi^{j}\|_{H^{1}}}1.$$ For some $j$ such that $x_{n}^{j}\equiv 0$ and $t_{n}^{j}\equiv 0$, we apply $\psi^{j}\in \mathcal{N}^{+}$ to constitute a global solution $v_{n}^{j}(t):=NLS_{k}(t)\psi^{j}$ of ($\rm{NLS_{k}}$) with initial data $\psi^{j}$. For some $j$ such that $x_{n}^{j}\equiv 0$ and $t_{n}^{j}\rightarrow\pm\infty$, by $(iii)$ of Theorem \[localwellposedness\], there exists $\tilde{\psi}^{j}\in H^{1}$ such that $$\begin{aligned} \label{initialdata} \|NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}-e^{it_{n}^{j}H_{\alpha}}\psi^{j}\|_{{\mathcal{H}}_{k}^{1}} \sim \|NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}-e^{it_{n}^{j}H_{\alpha}}\psi^{j}\|_{H^{1}}\rightarrow 0\;\;\text{as}\;\;n\rightarrow+\infty,\end{aligned}$$ which implies that for each $1\leq j\leq M$ and $n$ large enough, $$\begin{aligned} S_{k}\Big(NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}\Big) \leq n_{c}-\delta,\;\; P_{k}\Big(NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}\Big)\geq 0\;\;\text{and}\;\;I_{k}\Big(NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}\Big)\geq 0.\end{aligned}$$ We set $v_{n}^{j}(0)=NLS_{k}(t_{n}^{j})\tilde{\psi}^{j}$. Then according to the definition of $n_{c}$ and $v_{n}^{j}(0)\in\mathcal{N}^{+} $, we obtain that the solution $v_{n}^{j}(t):=NLS_{k}(t+t_{n}^{j})\tilde{\psi}^{j}$ of ($\rm{NLS_{k}}$) with initial data $v_{n}^{j}(0)$ is global and satisfies uniform space-time bounds: $\|v_{n}^{j}\|_{L_{t,x}^{5}}<+\infty$. As a result, we can construct approximate solutions of ($\rm{NLS_{k}}$): $$\tilde{u}_{n}(t):=\sum_{j=1}^{M}v_{n}^{j}+e^{-itH_{\alpha}}W_{n}^{M}$$ and set $$e:=(i\partial_{t}-H_{\alpha})\tilde{u}_{n}+|\tilde{u}_{n}|^{2}\tilde{u}_{n}.$$ By and , we have $$\begin{aligned} \label{approximate1} \|\phi_{n}-\tilde{u}_{n}(0)\|_{H^{1}}=\|u_{n}(0)-\tilde{u}_{n}(0)\|_{H^{1}}\rightarrow 0\;\;\text{as}\;\;n\rightarrow+\infty,\end{aligned}$$ which implies that $$\begin{aligned} \label{approximate2} \overline{\lim}_{n\rightarrow+\infty}\|\tilde{u}_{n}(0)\|_{H^{1}}\;\;\text{ has a uniform bound independent of}\;\; M.\end{aligned}$$ Using the same argument of Lemma 7.3 in [@KMVZ] and replacing $H_{\alpha}$ and homogeneous fractional operator (e.g., $|\nabla|^{\frac{1}{2}}$) with $1+H_{\alpha}$ and inhomogeneous fractional operator (e.g., $(1+\Delta)^{\frac{1}{4}}$), respectively, we also obtain the same results as there for $\tilde{u}_{n}(t)$, that is, $$\begin{aligned} \label{approximate3} \overline{\lim}_{n\rightarrow+\infty}\|\tilde{u}_{n}\|_{L_{t,x}^{5}} \;\;\text{has a uniform bound independent of }\;\;M\end{aligned}$$ and $$\begin{aligned} \label{approximate4} \lim_{M\rightarrow M^{*}}\overline{\lim}_{n\rightarrow+\infty} \|(1+\Delta)^{\frac{1}{4}}e\|_{L_{t,x}^{\frac{10}{7}}}=0\end{aligned}$$ Applying - to Theorem \[stability\] gives $\|u_n\|_{L_{t,x}^{5}}<+\infty$, which is a contradiction with . Thus, we have completed the proof of existence of critical element $u_{c}$. Finally, we consider precompactness of $K$ in $H^{1}$. We recall that $u_{c}$ satisfies the following properties: $$\begin{aligned} S_{k}(u_{c}(t))=n_{c}\;\;,\;\; u_{c}(t)\in {\mathcal{N}^{+}}\;\;\text{ for}\;\; \forall t\in {\bf R}\;\;\text{ and}\;\; \|u_c\|_{L_{t,x}^{5}}=+\infty.\end{aligned}$$ In particular, for any time sequence $\{t_{n}\}_{n=1}^{+\infty}$, the sequence $\{u_{c}(t_{n})\}_{n=1}^{+\infty}$ also satisfies $$\begin{aligned} S_{k}(u_{c}(t_{n}))=n_{c}\;\;,\;\; u_{c}(t_{n})\in {\mathcal{N}^{+}}\;\;\text{ for}\;\; \forall t\in {\bf R}\;\;\text{ and}\;\; \|u_c\|_{L_{(-\infty, t_{n}),x}^{5}}=\|u_c\|_{L_{( t_{n}, +\infty),x}^{5}}=+\infty.\end{aligned}$$ Hence, regarding $u_{c}(t_{n})$ as the foregoing $\phi_{n}$ and noting that the fact $\phi_{n}$ converges $\psi^{1}$ in $H^{1}$ yieds that $K$ is precompact in $H^{1}$. Thus, we complete the whole proof. Precluding the critical element ------------------------------- In this subsection, we shall apply the localized Virial identities and to preclude the critical element $u_{c}$. First by the precompactness of $K$, we have uniform localization of $u_{c}$: For each $\epsilon>0,$ there exists $R=R(\epsilon)>0$ independent of $t$ such that $$\begin{aligned} \label{localization} \int_{|x|>R}\Big(|\nabla u_{c}(t,x)|^2+|u_{c}(t,x)|^2+|u_{c}(t,x)|^4\Big)dx\leq\epsilon.\end{aligned}$$ We next claim that there exists a constant $c$ such that for any $t\in {\bf R}$, $$\begin{aligned} \label{control} \|\nabla u_{c}(t)\|_{L^{2}}\geq c\|u_{c}(t)\|_{L^{2}}.\end{aligned}$$ Indeed, if it’s false, then there exists a time sequence $\{t_{n}\}_{n=1}^{+\infty}$ such that $$\|\nabla u_{c}(t_{n})\|_{L^{2}}\leq \frac{1}{n}\|u_{c}(t_{n})\|_{L^{2}}=\frac{1}{n}\|u_{c}(0)\|_{L^{2}},$$ which means that $u_{c}(t_{n}\rightarrow 0$ in $\dot{H}^{1}$. However, $\{u_{c}(t_{n})\}_{n=1}^{+\infty}$ is precompact in $H^{1}$. Hence, there exists a subsequence (still denoted by itself) $u_{c}(t_{n})\rightarrow 0$ in $H^{1}$. As $u_{c}(t_{n})\in {\mathcal{N}^{+}}$, by , $n_{c}=\lim_{n\rightarrow+\infty}S_{k}(u_{c}(t_{n}))=0$, which is impossible. Now we use localized Virial identities and only with $u_{c}$ in place of $u$ again, where we still choose the radial function $\phi$ satisfying . For $R_{1}$, $R_{2}$, $R_{3}$ and $R_{4}$ in , by , we have $$\begin{aligned} \label{importanterror} |R_{1}+R_{2}+R_{3}+R_{4}|\lesssim &\displaystyle\int_{|x|\geq R}\Big(|\nabla u_{c}(t,x)|^{2}+|u_{c}(t,x)|^{4}+\frac{1}{R^{2}}|u_{c}(t,x)|^{2} +\frac{1}{R^{\alpha}}|u_{c}(t,x)|^{2}\Big)dx\nonumber\\ &\rightarrow 0\;\;\text{as}\;\; R\rightarrow+\infty.\end{aligned}$$ And for $P_{k}(u_{c})$ in , by , $x\cdot \nabla V\leq 0$, and , we have $$\begin{aligned} \label{importantlowerbound} P_{k}(u_{c}(t))&\geq \min\Big\{4\Big(n_{0}-S_{k}(u_{c}(t))\Big), \frac{2}{5}\Big(\|\nabla u_{c}(t)\|_{L^{2}}^{2}-\frac{1}{2}\displaystyle\int_{{\bf R}^{3}}(x\cdot\nabla V) |u_{c}(t)|^{2}dx)\Big)\Big\}\nonumber\\ &\gtrsim \|\nabla u_{c}(t)\|_{L^{2}}^{2}\gtrsim \|u_{c}(t)\|_{H^{1}}^{2}\gtrsim S_{k}(u_{c}(t))=n_{c}.\end{aligned}$$ It follows from and that there exists $\delta_{0}>0$ such that for $R$ large enough, $$I''(t)\geq \delta_{0},$$ which means that $\lim_{t\rightarrow+\infty}I'(t)=+\infty$. However, it is impossible since $I'(t)$ is bounded. In fact, by , $$|I'(t)|\lesssim R.$$ Thus, we complete the proof of scattering part of Theorem \[scattering1\]. [99]{} T. Akahori and H. Nawa, Blowup and scattering problems for the nonlinear Schrödinger equations, *Kyoto J. Math.*, 53(2013), pp. 629-672. N. Burq, F. Planchon, J. Stalker and A. S. Tahvildar-Zadeh, Strichartz estimates for the wave and Schödinger equations with the inverse-square potential, *J. Funct. Anal.*, 203(2003), pp. 519-549. T. Cazenave, *Semilinear Schrödinger equations*, Courant Lecture Notes in Mathematics, 10. New York University, Courant Institute of Mathematical Sciences, New York; American Mathematical Society, Providence, RI, 2003. D. Du, Y. Wu and K. Zhang, On blow-up criterion for the nonlinear Schrödinger equation, *Discrete Contin. Dyn. Syst.*, 36(2016), pp. 3639-3650. T. Duyckaerts, J. Holmer and S. Roudenko, Scattering for the non-radial 3D cubic nonlinear Schrödinger equation, *Math. Res. Lett.*, 15(2008), pp. 1233-1250. D. Fang, J. Xie and T. Cazenave, Scattering for the focusing energy-subcritical nonlinear Schrödinger equaiton, *Sci. China Math.*, 54(2011), pp. 2037-2062. J. Holmer and S. Roudenko, A sharp condition for scattering of the radial 3D cubic nonlinear Schrödiger equation, *Comm. Math. Phy.*, 282(2008), pp. 435-467. Y. Hong, Scattering for a nonlinear Schrödinger equation with a potential, *Commun. Pure Appl. Anal.*, 15(2016), pp. 1571-1601. S. Ibrahim, N. Masmoudi and K. Nakanishi, Scattering threshold for the focusing nonlinear Klein-Gordon equation, *Anal. PDE*, 4(2011), pp. 405-460. M. Ikeda and T. Inui, Global dynamics below the standing waves for the focusing semilinear Schrödinger equation with a repulsive Dirac delta potential. C. Kenig and F. Merle, Global well-posedness, scattering and blow-up for the energy-critical, focusing, non-linear Schrödinger equaiton in the radial case, *Invent. Math.*, 166(2006), pp. 645-675. R. Killip, C. Miao, M. Visan, J. Zhang and J. Zheng, Multipliers and Riesz transforms for the Schrödinger operator with inverse-square potential, arXiv:1503.02716. R. Killip, C. Miao, M. Visan, J, Zhang and J. Zheng, The energy-critical NLS with inverse-square potential, arXiv: 1509.05822. R. Killip, J. Murphy, M. Visan and J. Zheng, The focusing cubic NLS with inverse-square potential in three space dimensions, arXiv: 1603.08912. C. Miao, J. Zhang and J. Zheng, Nonlinear Schrödinger equation with Coulomb potential, arXiv:1809.06685. H. Mizutani, Strichartz estimates for Schrödinger equations with slowly decaying potential, arXiv:1808.06987. A. Sikora and J. Wright, Imaginary powers of Laplace operators, *Proc. Amer. Math. Soc.*, 129(2001), pp. 1745-1754. B. Simon, Schrödinger semigroups, *Bull. Amer. Math. Soc.*, 7(1982), pp.447-526. J. Zhang and J. Zheng, Scattering theory for nonlinear Schrödinger with inverse-square potential, *J. Funct. Anal.*, 267(2014), pp.2907-2932.

ArXiv

--- abstract: 'In this paper we obtain precise estimates for the $L^2$ norm of the $s$-dimensional Riesz transforms on very general measures supported on Cantor sets in ${{\mathbb R}}^d$, with $d-1<s<d$. From these estimates we infer that, for the so called uniformly disconnected compact sets, the capacity $\gamma_s$ associated with the Riesz kernel $x/|x|^{s+1}$ is comparable to the capacity $\dot{C}_{\frac{2}{3}(d-s),\frac{3}{2}}$ from non-linear potential theory.' address: - 'Maria Carmen Reguera. Departament de Matemàtiques, Universitat Autònoma de Barcelona, Catalonia, Spain and School of Mathematics, University of Birmingham, Birmingham, UK.' - 'Xavier Tolsa. Institució Catalana de Recerca i Estudis Avançats (ICREA) and Departament de Matemàtiques, Universitat Autònoma de Barcelona, Catalonia' author: - Maria Carmen Reguera and Xavier Tolsa title: 'Riesz transforms of non-integer homogeneity on uniformly disconnected sets' --- [^1] Introduction ============ In this paper we provide estimates from below for the $L^{2}$ norm of the $s$-dimensional Riesz transform of measures supported on very general Cantor sets in ${{\mathbb R}}^d$, for $s\in(d-1,d)$. The bounds obtained are written in terms of the densities of the cubes appearing in the construction of the Cantor sets. Our estimates allow us to establish an equivalence between the capacity $\gamma_{s}$ associated with the s-dimensional Riesz kernel and the capacity $\dot{C}_{\frac{2}{3}(d-s),\frac{3}{2}}$ from non-linear potential theory for the so called uniformly disconnected compact sets. In this paper we combine some of the techniques of previous works on the subject by Mateu and Tolsa [@MT], [@T], with others from Eiderman, Nazarov, Tolsa, and Volberg [@ENV12], [@NToV]. The general case for arbitrary compact sets still remains open. To state our results precisely, we need to introduce some notation. For $0<s<d$ and $x\in{{\mathbb R}}^d\setminus\{0\}$, we denote $$K^s(x)=\frac x{|x|^{s+1}},$$ and we let ${{\mathcal R}}^s$ be the associated Riesz transform, so that for a measure $\mu$ in ${{\mathbb R}}^d$ and $x\in{{\mathbb R}}^d$, $${{\mathcal R}}^s \mu(x) =\int K^s(x-y)\,d\mu(y),$$ whenever the integral makes sense. To avoid delicate issues with convergence, we will work with the truncated Riesz transform $${{\mathcal R}}^s_{{{\varepsilon}}} \mu(x) =\int_{|x-y|>{{\varepsilon}}} K^s(x-y)\,d\mu(y).$$ We say that ${{\mathcal R}}^s \mu$ is bounded in $L^{2}(\mu)$ if the truncated Riesz transforms ${{\mathcal R}}^s_{{{\varepsilon}}} \mu$ are bounded in $L^{2}(\mu)$ uniformly in ${{\varepsilon}}$. We now construct the Cantor sets $E$ that we will study, by the following algorithm. Let $Q^0\subset {{\mathbb R}}^d$ be a compact set. Take now disjoint closed subsets $Q^1_1,\ldots,Q^1_{N_1}\subset Q^0$ and set $E_1= \bigcup_{i=1}^{N_1} Q_i^1$. In the general step $k$ of the construction, we are given a family of closed sets $Q_1^k,\ldots,Q^k_{N_k}$. Then for each set $Q_i^k$ we take a finite family of closed sets $Q^{k+1}_j$, $j\in I_{Q_i^k}$, contained in $Q_i^k$ and we denote $${{\mathcal Ch}}(Q_i^k)=\{Q^{k+1}_j\}_{j\in I_{Q_i^k}}$$ (the notation ${{\mathcal Ch}}(Q_i^k)$ stands for “children” of $Q_i^k$). Renumbering these cubes if necessary, we set $$\{Q_1^{k+1},\ldots,Q_{N_{k+1}}^{k+1}\} := \bigcup_{j=1}^{N_k}{{\mathcal Ch}}(Q_i^k).$$ Then we denote $$E_k = \bigcup_{i=1}^{N_{k+1}} Q_i^{k+1},\qquad E=\bigcap_{k=1}^\infty E_k,$$ so that $E$ is a compact set. The sets $Q^k_i$ in this construction will be called “cubes”, although they need not be “true” cubes. The side length of $Q_i^k$ is $\ell(Q_i^k):={\operatorname{diam}}(Q_i^k)$. In this construction we assume that, for all $i,k$, $$\label{eq000} \frac18\,\ell(Q_i^k)\leq \ell(Q_j^{k+1})\leq \frac13\,\ell(Q_i^k)\quad\mbox{ for $Q_j^{k+1}\in{{\mathcal Ch}}(Q_i^k)$,}$$ and the separation condition $$\label{eqsepara} {\operatorname{dist}}(Q_j^{k+1},Q_{j'}^{k+1})\geq c_{sep}\,\ell(Q_i^k)\quad \mbox{for all $Q_j^{k+1}, Q_{j'}^{k+1}\in{{\mathcal Ch}}(Q_i^k)$},$$ for some fixed constant $c_{sep}>0$. These conditions guaranty that $E$ is a totally disconnected set. In particular, we infer that $${\operatorname{dist}}(Q_i^k,\,E\setminus Q_i^k)\geq c^{-1}\,\ell(Q_i^k).$$ Notice that we allow the family of children of $Q_i^k$ to be formed by a single cube $Q_j^{k+1}$. On the other hand, the assumptions and imply that the number of children is bounded above uniformly. We denote by ${{\mathcal D}}$ the family of all the cubes $Q_i^k$ in the construction above. That is, $${{\mathcal D}}= \{Q_i^k\}_{\substack{1\leq k \leq\infty \\ 1\leq i\leq N_k}}.$$ Given a measure $\mu$ supported on $E$ and a cube $Q\in{{\mathcal D}}$, we consider the $s$-dimensional density of $\mu$ on $Q$: $$\Theta^{s}_{\mu}(Q) = \frac{\mu(Q)}{\ell(Q)^s}.$$ We can now state our main result. \[teopri\] Let $s\in (d-1,d)$. Let $\mu$ be a finite Borel measure supported on the Cantor set $E\subset{{\mathbb R}}^d$ described above. Suppose that $$\sup_{Q\in {{\mathcal D}}} \Theta_\mu^s(Q)\leq C.$$ Then, $$\label{eqdyad1} \|{{\mathcal R}}^s \mu\|_{L^2(\mu)}^2 \approx \sum_{Q\in{{\mathcal D}}} \Theta_\mu^s(Q)^2\,\mu(Q),$$ where the comparability constant depends only on $s$, $d$ and $c_{sep}$. We remark that the estimate from above is already known for any $0<s<d$ due to the work of Eiderman, Nazarov and Volberg [@ENV10] (extending previous arguments by Mateu, Prat and Verdera [@MPV] for the case $0<s<1$). So the novelty lies in the converse inequality. This is a very delicate estimate, in particular because we are not in the realm where the notion of Menger curvature is useful, namely the case $0<s\leq1$. In the case $0<s<1$, Mateu, Prat and Verdera [@MPV] have shown that holds for any arbitrary compact set $E$ (with ${{\mathcal D}}$ being the family of all dyadic cubes). Mateu and Tolsa [@MT] have studied the general case $0<s<d$ when $\mu$ is the probability measure on a Cantor set in which the densities of the cubes in the construction decrease with side length. This unnecessary restriction is removed by Tolsa in [@T]. The Cantor sets considered in [@MT] and [@T] are some kind of high-dimensional variants of the well known $1/4$ planar Cantor set. The advantage of these Cantor sets is that, in a given stage $k$ of the construction, all the cubes $Q_i^k$ have the same side lengths and densities. In this paper, we go a step forward by considering much more general Cantor sets and measures where we no longer have the aforementioned properties. On the down side, we have to restrict ourselves to $d-1<s<d$ due to the use of a maximum principle for the $s$-dimensional Riesz transforms which, apparently, fails for $s<d-1$. Theorem \[teopri\] can also be understood as a refined quantitative version of the results of Prat [@Pr], Vihtilä [@Vih] and Eiderman, Nazarov and Volberg [@ENV12], in the particular case of the preceding Cantor sets. In these works it is shown that, given $F\subset{{\mathbb R}}^d$ with $0<{{\mathcal H}}^s(F)<\infty$ and $\mu = \mathcal H^{s}|_F$, the $s$-dimensional Riesz transform with respect to $\mu$ is unbounded in $L^2(\mu)$, in the cases $0<s<1$ [@Pr]; $s\in (0,d)\setminus{{\mathbb Z}}$, $\mu$ with positive lower $s$-dimensional density [@Vih]; and for $d-1<s<d$, $\mu$ with zero lower $s$-dimensional density [@ENV12]. Finally, there is another quantitative result worth mentioning in the work of Jaye, Nazarov and Volberg [@JNV] that relates the boundedness of the fractional $s$-dimensional Riesz transform $d-1<s<d$ with a weak type estimate for a Wolff potential of exponential type. Theorem \[teopri\] has an important collorary regarding the capacities $\gamma_s$ and $ \dot C_{\frac23(d-s),\frac32}$. To present the corollary we need some extra definitions. Given a compact set $F\subset {{\mathbb R}}^d$, the capacity $\gamma_s$ of $F$ is defined by $$\gamma_s(F) = \sup|\langle T,1\rangle|,$$ where the supremum runs over all distributions $T$ supported on $F$ such that $\|{{\mathcal R}}^{s}(T)\|_{L^\infty({{\mathbb R}}^d)} \leq 1$. Denote by $\Sigma(F)$ the family of measures $\mu$ supported on $F$ such that $\mu(B(x,r))\leq r^n$ for all $x\in{{\mathbb R}}^d$ and $r>0$. It turns out that $$\gamma_s(F) \approx \sup\{\mu(F):\,\mu\in\Sigma(F),\,\|{{\mathcal R}}^s\mu\|_{L^2(\mu)}^2\leq \mu(F)\}.$$ This was first shown for $s=1,d=2$ by Tolsa [@Tolsa-sem], and it was extended to the case $s=d-1$ by Volberg [@Vo], and to the other values of $s$ and $d$ by Prat [@Laura]. Now we turn to non linear potential theory. Given $\alpha>0$ and $1<p<\infty$ with $0<\alpha p <d$, the capacity $\dot C_{\alpha,p}$ of $F\subset{{\mathbb R}}^d$ is defined as $$\dot C_{\alpha,p}(F) = \sup_\mu \mu(F)^p,$$ where the supremum is taken over all positive measures $\mu$ supported on $F$ such that $$I_\alpha(\mu)(x) = \int \frac1{|x-y|^{d-\alpha}}\,d\mu(x)$$ satisfies $\|I_\alpha(\mu)\|_{p'}\leq 1$, where as usual $p'=p/(p-1)$. We are interested in the characterization of $\dot C_{\alpha,p}$ in terms of Wolff potentials. Consider $$\dot W^\mu_{\alpha,p}(x) = \int_0^\infty \biggl(\frac{\mu(B(x,r))}{r^{d-\alpha p}}\biggr)^{p'-1}\,\frac{dr}r.$$ A classical theorem of Wolff establishes $$\dot C_{\alpha,p}(F) \approx \sup_\mu \mu(F),$$ where the supremum is taken over all measures $\mu$ supported on $F$ such that $\int \dot W_{\alpha,p}^\mu(x)\,d\mu\leq \mu(F)$ (see [@Adams-Hedberg Chapter 4], for instance). Finally we wish to remark that the class of Cantor sets $E$ considered in Theorem \[teopri\] coincides with the class of compact [*uniformly disconnected sets*]{}. According to [@David-Semmes-fractured p.156], a set $F\subset{{\mathbb R}}^d$ is called uniformly disconnected if there exists a constant $c_F>0$ so that for each $x\in F$ and $r>0$ one can find a closed (with respect to $F$) subset $A\subset F$ such that $A\subset B(x,r)$, $A\supset F\cap B(x,c_F^{-1}r)$, and ${\operatorname{dist}}(A,F\setminus A)\geq c_F^{-1}r$. One can check that any uniformly disconnected set can be constructed as one of the Cantor sets $E$ considered in Theorem \[teopri\], for a suitable separation constant $c_{sep}$, and replacing the constants $1/8$ and $1/3$ in by others if necessary. Conversely, it is immediate that any such set $E$ is uniformly disconnected. We are now ready to formulate the corollary to Theorem \[teopri\]. Let $F\subset{{\mathbb R}}^d$ be compact and uniformly disconnected, with constant $c_F$, and let $d-1<s<d$. Then $$\label{e.capacities} \gamma_s(F) \approx \dot C_{\frac23(d-s),\frac32}(F),$$ with the comparability constant depending only on $d$, $s$, and $c_F$. The estimate was proved by Mateu, Prat and Verdera when $0<s<1$ in [@MPV]. By using the upper estimate in the inequality , Eiderman, Nazarov and Volberg [@ENV10], showed that for all indices $0<s<d$ $$\gamma_s(F) \gtrsim \dot C_{\frac23(d-s),\frac32}(F),$$ for any compact set $F\subset{{\mathbb R}}^d$. It is known that the opposite inequality is false when $s$ is integer, see [@ENV10]. On the other hand, in the works of Tolsa and Mateu [@MT] and [@T] mentioned above, it is shown that the comparability holds for the Cantor sets studied in these papers for $0<s<d$. Although our corollary extends the result to a more general family of compact sets when $d-1<s<d$, the general case when $0<s<d$ is non integer and $F$ is a general compact set remains open. From Theorem \[teopri\] it follows easily that the comparability $\gamma_s(E) \approx \dot C_{\frac23(d-s),\frac32}(E)$ holds for the Cantor sets defined above, with the constant in the comparability depending only on $d$, $s,$ and the separation constant $c_{sep}$. Indeed, recall that one just has to show that $\gamma_s(E) \lesssim \dot C_{\frac23(d-s),\frac32}(E)$. To this end, take $\mu\in\Sigma(E)$ with $\|R^s\mu\|_{L^2(\mu)}^2\leq \mu(E)$ such that $\gamma_s(E) \approx \mu(E)$. Then, by Theorem \[teopri\], $$\sum_{Q\in{{\mathcal D}}} \Theta_\mu^s(Q)^2\,\mu(Q)\approx \|{{\mathcal R}}^s \mu\|_{L^2(\mu)}^2 \leq\mu(E).$$ It is easy to check that the above sum on the left had side is comparable to $\int \dot W_{\frac23(d-s),\frac32}^\mu(x)\,d\mu$. So one infers that $\gamma_s(E) \lesssim \dot C_{\frac23(d-s),\frac32}(E)$, as wished. In order to prove Theorem \[teopri\] we will consider a stopping time argument. The stopping conditions will take into account the oscillations of the densities on the different cubes from ${{\mathcal D}}$ and the possible large values of the $s$-dimensional Riesz transform on each cube. In this way we will split ${{\mathcal D}}$ into different families of cubes, which we will call “trees”, following an approach similar to the one in [@T]. One the main differences of the present work with respect to the latter reference is that, in some key steps, our work paper implements a variational argument borrowed from work of Eiderman, Nazarov, Volberg [@ENV12] and Nazarov, Tolsa and Volberg [@NToV], suitably adapted to our setting. This variational argument requires the $s$-dimensional Riesz transforms to satisfy the maximum principle mentioned above. The paper is organized as follows. Section 2 contains the basic background. Section 3 is devoted to the description of the stopping time argument and the properties associated to it. In Section 4 we prove that the sizes of trees obtained by the stopping time argument must be small. We will use a touching point argument for this purpose, where fact that $s$ is fractional will play an important role. In Section 5 we describe four relevant families of enlarged trees and we start analysing the easier ones. Section 6 contains some Fourier analysis that will be necessary for the development of Section 7. The analysis of the most difficult family of trees is included in Section 7. And at last, Section 8 puts together all the estimates obtained in previous sections to provide the proof of the Main Theorem \[teopri\]. Preliminaries ============= About cubes, trees ------------------ Below, to simplify notation we will write ${{\mathcal R}}$, $K$ and $\Theta$ instead of ${{\mathcal R}}^s$, $K^s$ and $\Theta_\mu^s$, respectively. Notice that if we replace the cubes $Q\in{{\mathcal D}}$ by $${{\widehat{Q}}} = \Bigl\{x\in {{\mathbb R}}^d: {\operatorname{dist}}(x,Q)\leq \frac1{10}c_{sep}\,\ell(Q)\Bigr\},$$ the separation condition still holds with a slightly worse constant. Analogously, is still satisfied, possibly after modifying suitably the constants $1/8$ and $1/3$. The advantage of ${{\widehat{Q}}}$ over $Q$ is that the Lebesgue measure of ${{\widehat{Q}}}$ is comparable to ${\operatorname{diam}}({{\widehat{Q}}})^d$, which is not guaranteed for $Q$. To avoid some technicalities, in the proof of Theorem \[teopri\] we will assume that the original cubes $Q\in{{\mathcal D}}$ satisfy ${{\mathcal L}}^d(Q)\approx\ell(Q)^d$, where ${{\mathcal L}}^d$ stands for the Lebesgue measure on ${{\mathbb R}}^d$. Moreover, we will assume that the separation constant $c_{sep}$ does not exceed $1/10$, say. For $j\geq0$, we denote by ${{\mathcal D}}_j$ the family of cubes of generation $j$ that appear in the construction of $E$, and we set ${{\mathcal D}}=\bigcup_{j\geq0} {{\mathcal D}}_j$. We assume that $\mu(Q)>0$ for all $Q\in{{\mathcal D}}$. Otherwise we eliminate $Q$ from the construction of $E$. If $R\in{{\mathcal D}}_j$, we denote by ${{\mathcal D}}_k(R)$ the family of the cubes from ${{\mathcal D}}_{j+k}$ which are contained in $R$. Notice that if $Q\in{{\mathcal D}}_k(R)$, then $$8^{-k}\leq\frac{\ell(Q)}{\ell(R)}\leq 3^{-k}.$$ Given a cube $Q\in{{\mathcal D}}$, for any constant $a>1$ we denote $$aQ = \{x\in{{\mathbb R}}^d:{\operatorname{dist}}(x,Q)\leq (a-1)\,\ell(Q)\}.$$ Also, we set $$p(Q) =\sum_{P\in{{\mathcal D}}:P\supset Q} \frac{\ell(Q)}{\ell(P)}\,\Theta(P).$$ So $p(Q)$ should be understood as a smoothened version of $\Theta(Q)$. Also, given $Q,R\in{{\mathcal D}}$ with $Q\subset R$, we set $$p(Q,R) =\sum_{P\in{{\mathcal D}}:Q\subset P\subset R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P).$$ We say that a cube $Q$ is $p$-doubling (with constant $c_{db}$) if $$\label{eqdob} p(Q)\leq c_{db}\,\Theta(Q).$$ Given a family of cubes ${{\mathcal T}}\subset {{\mathcal D}}$, we denote $$\sigma({{\mathcal T}}) =\sum_{Q\in{{\mathcal T}}} \Theta(Q)^2\,\mu(Q).$$ So Theorem \[teopri\] asserts that $\|{{\mathcal R}}\mu\|_{L^2(\mu)}^2 \approx\sigma({{\mathcal D}})$ under the assumption $\sup_{Q\in {{\mathcal D}}} \Theta_\mu^s(Q)\leq C$. \[lemdob0\] Suppose that $c_{db}$ is big enough, depending on $a,s,d$. Let $Q_0,Q_1,\ldots,Q_n$ be a family of cubes from ${{\mathcal D}}$ such that $Q_j$ is son of $Q_{j-1}$ for $1\leq j\leq n$. Suppose that $Q_j$ is not $p$-doubling for $1\leq j\leq n$. Then $$\label{eqcad35} \Theta(Q_j)\leq 2^{-j/2}\,p(Q_0).$$ For $1\leq j \leq n$, the fact that $Q_j$ is not $p$-doubling implies that $$\label{eqsak33} \Theta(Q_j) \leq \frac1{c_{db}}\,p(Q_j) = \frac1{c_{db}}\Biggl (\sum_{k=0}^{j-1} \frac{\ell(Q_j)}{\ell(Q_{j-k})}\, \Theta(Q_{j-k})+ \frac{\ell(Q_j)}{\ell(Q_0)}\,p(Q_0)\Biggr).$$ We will prove by induction on $j$. For $j=0$ this is in an immediate consequence of the definition of $p(Q)$. Suppose that holds for $0\leq h\leq j$, with $j\leq n-1$, and let us consider the case $j+1$. From and the induction hypothesis we get $$\begin{aligned} \Theta(Q_{j+1}) & \leq \frac1{c_{db}}\Biggl (\Theta(Q_{j+1}) + \sum_{k=1}^j \frac{\ell(Q_{j+1})}{\ell(Q_{j+1-k})}\, \Theta(Q_{j+1-k})+ \frac{\ell(Q_{j+1})}{\ell(Q_0)}\,p(Q_0)\Biggr)\\ &\leq \frac1{c_{db}}\Biggl (\Theta(Q_{j+1}) + \sum_{k=1}^j 3^{-k}\, \Theta(Q_{j+1-k})+ 3^{-j-1}\,p(Q_0)\Biggr)\\ &\leq \frac1{c_{db}}\Biggl (\Theta(Q_{j+1}) + \sum_{k=1}^j 3^{-k}\,2^{(-j-1+k)/2}p(Q_0) + 3^{-j-1}\,p(Q_0)\Biggr)\end{aligned}$$ Since $\sum_{k=1}^j 3^{-k}\,2^{(-j-1+k)/2}\lesssim 2^{-j/2}$, we obtain $$\begin{aligned} \Theta(Q_{j+1}) &\leq \frac1{c_{db}}\bigl (\Theta(Q_{j+1}) + C\,2^{-j/2}\,p(Q_0)+ 3^{-j-1}\,p(Q_0)\bigr)\\ &\leq \frac1{c_{db}}\bigl (\Theta(Q_{j+1}) + \tilde{C}\,2^{-j/2}\,p(Q_0)\bigr) \\\end{aligned}$$ It is straightforward to check that yields $\Theta(Q_{j+1})\leq 2^{-(j+1)/2}\,p(Q_0)$ if $c_{db}$ is big enough. For the rest of the paper we will fix $c_{db}$ so that holds. \[lemdob\] For a fixed $Q\in{{\mathcal D}}$, let $J\subset {{\mathcal D}}$ be a family of cubes contained in $Q$ such that for every $P\in J$, every cube $P'\in{{\mathcal D}}$ such that $P\subset P'\subset Q$ is not $p$-doubling. Then $$\sigma(J)\leq 2\,p(Q)^2\,\mu(Q).$$ For $j\geq0$, let $J_j$ be the subfamily of the cubes from $J$ which are $j$ generations below $Q$. That is, $P$ is from $J_j$ if it belongs to $J$ and it is an $j$-th descendant of $Q$. By the preceding lemma, it turns out that $$\Theta(P)\leq 2^{-j/2}\,p(Q)\qquad\mbox{ if $P\in J_j$.}$$ Taking also into account that the cubes from $J_j$ are pairwise disjoint, we get $$\sigma(J_j)\leq 2^{-j}\,p(Q)^2\sum_{P\in J_j:\subset Q}\mu(P) \leq 2^{-j}\,p(Q)^2\,\mu(Q).$$ Therefore, $$\sigma(J) = \sum_{j=0}^n \sigma(J_j) \leq \sum_{j=0}^\infty 2^{-j}\,p(Q)^2\,\mu(Q) = 2\,p(Q)^2\,\mu(Q),$$ where $n$ is the maximum number of generations between $Q$ and the cubes belonging to $J$. The operators $D_Q$ ------------------- Given a cube $Q\in{{\mathcal D}}$ and a function $f\in L^1(\mu)$, we denote by $m_Qf$ the mean of $f$ on $Q$ with respect to $\mu$. That is, $m_Qf =\frac1{\mu(Q)}\int_{Q} f\,d\mu$. Then we define $$D_Q f = \sum_{P\in{{\mathcal Ch}}(Q)} \chi_P\,(m_Pf - m_Q f).$$ The functions $D_Qf$, $Q\in{{\mathcal D}}$, are orthogonal, and it is well known that $$\|f\|_{L^2(\mu)}^2 = \sum_{Q\in{{\mathcal D}}}\|D_Q f\|_{L^2(\mu)}^2.$$ About the Riesz transform ------------------------- In the following lemma we collect a pair of useful estimates about the Riesz transform. \[lemcomp\] Let $Q,R\in{{\mathcal D}}$ with $Q\subset R$, and $x,y\in Q$. Then $$\label{eqfacc00} |{{\mathcal R}}(\chi_{R\setminus Q}\mu)(x)|\lesssim \sum_{P\in{{\mathcal D}}:Q\subset P\subset R} \frac{\mu(P)}{\ell(P)^s}$$ and $$\label{eqfacc11} |{{\mathcal R}}(\chi_{R\setminus Q}\mu)(x)- {{\mathcal R}}(\chi_{R\setminus Q}\mu)(y)| \lesssim \frac{|x-y|}{\ell(Q)}\,p(Q,R).$$ Also, $$\label{eqfacc12} \bigl|{{\mathcal R}}(\chi_{R\setminus Q}\mu)(x)- \bigl (m_Q({{\mathcal R}}\mu) - m_R({{\mathcal R}}\mu)\bigr)\bigr|\lesssim p(Q,R) + p(R).$$ The first and second inequalities follow by very standard methods, taking into account the separation property . Regarding , by the antisymmetry of the Riesz transform, we have$$\begin{aligned} m_Q({{\mathcal R}}\mu) - m_R({{\mathcal R}}\mu) & = m_Q({{\mathcal R}}(\chi_{Q^c}\mu)) - m_R({{\mathcal R}}(\chi_{R^c}\mu))\\ & = m_Q({{\mathcal R}}(\chi_{R\setminus Q})) + m_Q({{\mathcal R}}(\chi_{R^c}\mu)) - m_R({{\mathcal R}}(\chi_{R^c}\mu)).\end{aligned}$$ From the estimate , we deduce that for all $x'\in Q\subset R$, $y'\in R$, $$|{{\mathcal R}}(\chi_{R^c}\mu)(x')- {{\mathcal R}}(\chi_{R^c}\mu)(y')| \lesssim p(R).$$ Averaging, we get $$|m_Q({{\mathcal R}}(\chi_{R^c}\mu)) - m_R({{\mathcal R}}(\chi_{R^c}\mu))|\lesssim p(R).$$ Analogously, for $x\in Q$, we have $$\bigl|{{\mathcal R}}(\chi_{R\setminus Q})(x) - m_Q({{\mathcal R}}(\chi_{R\setminus Q})\bigr|\lesssim p(Q,R).$$ Therefore, $$\begin{aligned} \bigl|{{\mathcal R}}(\chi_{R\setminus Q})(x)- \bigl (m_Q(R\mu) - m_R(R\mu)\bigr)\bigr| & \leq \bigl|{{\mathcal R}}(\chi_{R\setminus Q})(x)- m_Q({{\mathcal R}}(\chi_{R\setminus Q}))\bigr| \\&\quad + \bigl|m_Q({{\mathcal R}}(\chi_{R^c}\mu)) - m_R({{\mathcal R}}(\chi_{R^c}\mu))\bigr|\\ & \lesssim p(Q,R) + p(R).\end{aligned}$$ The operators ------------- Let $\Phi:{{\mathbb R}}^d\to[0,\infty)$ be a $1$-Lipschitz function. Below we will need to work with the suppressed kernel $$\label{eqsuppressed} K_\Phi(x,y) = \frac{x-y}{\bigl (|x-y|^2+\Phi(x)\Phi(y)\bigr)^{(s+1)/2}}$$ and the associated operator $${{\mathcal R}}_\Phi\nu(x) =\int K_\Phi(x,y)\,d\nu(y),$$ for a signed measure $\nu$ in ${{\mathbb R}}^d$. This kernel (or a variant of this) appeared for the first time in the work of Nazarov, Treil and Volberg in connection with Vitushkin’s conjecture (see [@Vo]). For $f\in L^1_{loc}(\mu)$, one denotes ${{\mathcal R}}_{\Phi,\mu} f = {{\mathcal R}}_\Phi (f\,\mu)$. If ${{\varepsilon}}\approx\Phi(x)$, then it follows that, for any signed measure $\nu$ in ${{\mathbb R}}^d$, $$\label{e.compsup''} \bigl|{{\mathcal R}}_{{{\varepsilon}}}\nu(x) - {{\mathcal R}}_{\Phi}\nu(x)\bigr|\lesssim \sup_{r> \Phi(x)}\frac{|\nu|(B(x,r))}{r^s}.$$ See Lemmas 5.4 and 5.5 in [@To], for example. The following result is an easy consequence of a $Tb$ theorem of Nazarov, Treil and Volberg. See Chapter 5 of [@To], for example. \[teontv\] Let $\mu$ be a Radon measure in ${{\mathbb R}}^d$ and let $\Phi:{{\mathbb R}}^d\to[0,\infty)$ be a $1$-Lipschitz function. Suppose that - $\mu(B(x,r))\leq c_0\,r^s$ for all $r\geq \Phi(x)$, and - $\sup_{{{\varepsilon}}>\Phi(x)}|{{\mathcal R}}_{{\varepsilon}}\mu(x)|\leq c_1$. Then ${{\mathcal R}}_{\Phi,\mu}$ is bounded in $L^p(\mu)$, for $1<p<\infty$, with a bound on its norm depending only on $p$, $c_0$ and $c_1$. In particular, ${{\mathcal R}}_\mu$ is bounded in $L^p(\mu)$ on the set $\{x:\Phi(x)=0\}$. The corona decomposition {#seccorona} ======================== Recall that the starting cube from the construction of $E$ is denoted by $Q^0$. Below we choose constants $B\gg M\gg1\gg \delta_0$. For convenience we assume $B$ to be a power of $2$. Given a cube $Q$, we define its dyadic density $\Theta_{d}(Q)$, $$\Theta_{d}(Q):= 2^{j},$$ where $j\in{{\mathbb Z}}$ is such that $ 2^{j}\leq \Theta(Q) < 2^{j+1}$. Given a cube $R\in{{\mathcal D}}$, we define families $HD_0(R)$, $LD_0(R)$ and $BR_0(R)$ of cubes from ${{\mathcal D}}$ as follows: - We say that a cube $Q\subset{{\mathcal D}}$ belongs to $HD_0(R)$ if it is contained in $R$, $\Theta_{d}(Q)=B\Theta_{d}(R)$ and has maximal side length. Recall that $B$ is a power of $2$. - A cube $Q\subset{{\mathcal D}}$ belongs to $LD_0(R)$ if it is contained in $R$, $\Theta(Q)\leq \delta_0\,\Theta(R)$, it is not contained in any cube from $HD_0(R)$, and has maximal side length. - A cube $Q\subset{{\mathcal D}}$ belongs to $BR_0(R)$ if it is contained in $R$, it satisfies $$|m_Q {{\mathcal R}}\mu - m_R {{\mathcal R}}\mu|\geq M\,\bigr(\Theta(R) + p(Q)\bigl),$$ it is not contained in any cube from $HD_0(R)\cup LD_0(R)$, and has maximal side length. We consider a “doubling constant” $c_{db}>10$ in . Then we denote by $HD(R)$, $LD(R)$ and $BR(R)$ the families of maximal, and thus disjoint, $p$-doubling cubes (with constant $c_{db}$) which are contained in $HD_0(R)$, $LD_0(R)$ and $BR_0(R)$, respectively. We denote $${\operatorname{Stop}}_0(R)= HD_0(R)\cup LD_0(R)\cup BR_0(R)$$ and $${\operatorname{Stop}}(R)= HD(R)\cup LD(R)\cup BR(R).$$ For $Q\in{\operatorname{Stop}}_0(R)$, let $J_Q$ be the family of cubes from ${{\mathcal D}}$ which are contained in $Q$ and are not contained in any cube from ${\operatorname{Stop}}(R)$. Notice that, by Lemma \[lemdob\], $$\label{eqsigma1} \sigma(J_Q)\leq c\,p(Q)^2\,\mu(Q).$$ For $k\geq 1$, we define ${\operatorname{Stop}}^k(R)$ inductively: we set ${\operatorname{Stop}}^1(R)={\operatorname{Stop}}(R)$, and for $k>1$ a cube belongs to ${\operatorname{Stop}}^k(R)$ if it belongs to ${\operatorname{Stop}}(Q)$ for some $Q\in{\operatorname{Stop}}^{k-1}(R)$. Now we construct the family ${\operatorname{Top}}\subset{{\mathcal D}}$ as follows: $${\operatorname{Top}}= \{Q_0\} \bigcup_{k\geq1} {\operatorname{Stop}}^k(Q_0).$$ Given a cube $R\in{\operatorname{Top}}$, we denote by ${\operatorname{Tree}}(R)$ (and ${\operatorname{Tree}}_0(R)$) the family of cubes contained in $R$ which are not contained in any cube from ${\operatorname{Stop}}(R)$ (and ${\operatorname{Stop}}_0(R)$, respectively). Observe that the cubes from ${\operatorname{Stop}}(R)$ do not belong to ${\operatorname{Tree}}(R)$. Notice also that $${{\mathcal D}}= \bigcup_{R\in{\operatorname{Top}}} {\operatorname{Tree}}(R).$$ Moreover, the union is disjoint. The following lemma is an easy consequence of our construction. \[lemfac1\] For every $R\in{\operatorname{Top}}$, we have: - $R$ is $p$-doubling (with constant $c_{db}$). - Every cube $Q\in{\operatorname{Tree}}(R)$ satisfies $\Theta(Q)\leq 2B\,\Theta(R)$ and $p(Q)\leq (2B + c_{db})\Theta(R)$. - Every cube $Q\in{\operatorname{Tree}}_0(R)$ satisfies $$\label{eqc391} |m_R {{\mathcal R}}\mu - m_Q {{\mathcal R}}\mu|\leq M\,\bigr(\Theta(R) + p(Q)\bigl),$$ and $$\label{eqc392} \bigl|{{\mathcal R}}(\chi_{R\setminus Q} \mu)(x)\bigr| \leq 2 M \,\bigl (\Theta(R)+ p(Q)\bigr) \qquad \mbox{for all $x\in Q$.}$$ The statement (a) is an immediate consequence of the definition above. Let us turn our attention to (b). The estimate $\Theta(Q)\leq 2B\,\Theta(R)$ is clear if $Q\in{\operatorname{Tree}}_0(R)$. Otherwise, let $Q_0\in{\operatorname{Stop}}_0(R)$ such that $Q\subset Q_0$. Reasoning by contradiction let us suppose that $\Theta(Q)> 2B\,\Theta(R)$, and let $Q'\in{{\mathcal D}}$ be a cube such that $Q\subset Q'\subset Q_0$ with $\Theta(Q')\geq 2B\,\Theta(R)$ with maximal side length. Then, $\Theta(P)\leq 2B\,\Theta(R)$ for any $P\in {\operatorname{Tree}}(R)$ such that $Q'\subset P$. $$\begin{aligned} p(Q') & = \Theta(Q') + \sum_{P:Q'\subsetneq P\subset R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) + \sum_{P\supsetneq R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) \\ & \leq \Theta(Q') + c_2\,B\,\Theta(R) + p(R) \\ & \leq \Theta(Q') + c_2\,\Theta(Q') + c_{db}\,\Theta(R)\leq \bigl (1 + c_2 + \frac{c_{db}}B\bigr)\Theta(Q').\end{aligned}$$ Since $B>2$, if we choose $c_{db}>2(1+c_1)$, we get $$p(Q')\leq c_{db}\,\Theta(Q').$$ That is, $Q'$ is $p$-doubling, which implies either that $Q'\in{\operatorname{Stop}}(R)$ or there exists another cube $Q''\in{\operatorname{Stop}}(R)$ which contains $Q'$. This contradicts the fact that $Q\in{\operatorname{Tree}}(R)$. To prove that $p(Q)\leq (2B+c_{db})\Theta(R)$ we write $$\begin{aligned} \label{eqalg17} p(Q) & = \sum_{P:Q\subset P\subset R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) + \sum_{P\supsetneq R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) \\ & \leq 2B\Theta(R) + p(R) \leq 2B\Theta(R) + c_{db}\,\Theta(R).\nonumber\end{aligned}$$ The estimate in (c) is a direct consequence of the construction of ${\operatorname{Tree}}_0(R)$. On the other hand, follows from and the inequality : $$\begin{aligned} \bigl|{{\mathcal R}}(\chi_{R\setminus Q}\mu)(x)\bigr| & \leq \bigl|m_Q({{\mathcal R}}\mu) - m_R({{\mathcal R}}\mu)\bigr| + p(Q,R) + p(R)\\ &\leq M\,\bigr(\Theta(R) + p(Q)\bigl) + p(Q)+ c_{db}\,\Theta(R) \leq 2M\,\bigr(\Theta(R) + p(Q)\bigl),\end{aligned}$$ assuming that $M\geq c_{db}$. \[lemt0\] Let $R\in{\operatorname{Top}}$. We have (a) If $Q\in HD_0(R)$, then $Q$ is $p$-doubling with constant $c_{db}$. Thus $HD(R)=HD_0(R)$. (b) If $Q\in BR(R)$, then $$|m_Q {{\mathcal R}}\mu - m_R {{\mathcal R}}\mu|\geq \frac{M}2\,\Theta(R),$$ assuming $M\geq c\,c_{db}$. (c) If $Q\in LD(R)$, then $$p(Q) \leq c\,\delta_0^{\frac1{s+1}}B^{\frac s{s+1}}\,\Theta(R).$$ We will assume that $\delta_0\ll B^{-1}$, so that $$\label{eqld41} p(Q) \leq \delta_0^{\frac1{s+2}}\Theta(R)\leq \frac1B \Theta(R)\quad \mbox{ if $Q\in LD(R)$.}$$ We will assume that $M=C\, B$, where $C$ is some absolute constant such that the statement in (b) of the preceding lemma holds. The proof of (a) is already implicit in the proof of part (b) from Lemma \[lemfac1\]. Indeed, similarly to , $$\begin{aligned} p(Q) &= \sum_{P:Q\subset P\subset R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) + \sum_{P\supsetneq R} \frac{\ell(Q)}{\ell(P)}\,\Theta(P) \\ & \leq c_2 \max_{P:Q\subset P\subset R}\Theta(P) + p(R) \leq c_2\,\Theta(Q) + c_{db}\,\Theta(R)\\&\leq (c_2+ c_{db}\,B^{-1})\Theta(Q) \leq c_{db}\,\Theta(Q),\nonumber\end{aligned}$$ assuming $c_{db}\geq2c_2$ and $B$ big enough in the last inequality. For the proof of (b), we know that there exists $Q_{0}\in BR_{0}(R)$ such that $Q\subset Q_{0}$. The following estimate follow from the definition of $ BR_{0}(R)$: $$\begin{aligned} |m_Q {{\mathcal R}}\mu - m_R {{\mathcal R}}\mu| & \geq |m_{Q_{0}} {{\mathcal R}}\mu - m_R {{\mathcal R}}\mu|- |m_Q {{\mathcal R}}\mu - m_{Q_{0}} {{\mathcal R}}\mu|\\ & \geq M\,\bigr(\Theta(R) + p(Q_{0})\bigl) -|m_Q {{\mathcal R}}\mu - m_{Q_{0}} {{\mathcal R}}\mu|.\end{aligned}$$ So it suffices to prove $|m_Q {{\mathcal R}}\mu - m_{Q_{0}} {{\mathcal R}}\mu|\lesssim p(Q_{0})$, since we will choose $M \gg1$. From , and , we get $$\begin{aligned} |m_Q {{\mathcal R}}\mu - m_{Q_{0}} {{\mathcal R}}\mu|& \leq |m_Q {{\mathcal R}}(\chi_{Q_{0}\setminus Q}\mu)| + |m_Q {{\mathcal R}}(\chi_{Q_{0}^{c}}\mu) - m_{Q_{0}} {{\mathcal R}}(\chi_{Q_{0}^{c}}\mu)|\\ & \lesssim \sum_{\substack{P\in \mathcal D \\ Q\subsetneq P \subset Q_{0}}}\frac{\mu(P)}{\ell(P)^{s}} + p(Q_0) \\ & \lesssim \sum_{j=1}^{\infty} 2^{-j/2}p(Q_{0}) \lesssim p(Q_{0}),\end{aligned}$$ as wished. Let us see (c). Let $Q_0$ be the cube from $LD_0(R)$ that contains $Q$. Recall that $\Theta(Q_0)\leq \delta_0\,\Theta(R)$. Suppose that $$\Theta(Q) \geq \tau\,\Theta(R),$$ for some constant $\tau\gg\delta_0$ to be determined below. Consider the largest cube $Q'\in{{\mathcal D}}$ with $Q\subset Q'\subset Q_0$ such that $\Theta(Q')\geq\tau\,\Theta(R)$. Then it is immediate to check that $$\label{eqpp99} \Theta(Q')\approx\tau\,\Theta(R).$$ Indeed, it is enough to show that $\Theta(Q')\lesssim\tau\,\Theta(R)$, since the opposite inequality is clear from the choice on $Q'$. To prove this, consider the parent $Q''$ of $Q'$. From the fact that $\ell(Q'')\approx\ell(Q')$ we get $$\Theta(Q') = \frac{\mu(Q')}{\ell(Q')^s}\leq \frac{\mu(Q'')}{\ell(Q')^s} \approx \Theta(Q'')<\tau\,\Theta(R),$$ as wished. We are going to show that $p(Q')\leq c\,\Theta(Q')$. To this end we write $$\label{eqpp100} p(Q') \leq p(Q',Q_0) + \frac{\ell(Q')}{\ell(Q_0)}\,p(Q_0).$$ Since $\Theta(P)\leq \Theta(Q')$ for all $P$ with $Q'\subset P\subset Q_0$, it follows that $$\label{eqpp101} p(Q',Q_0)\lesssim \Theta(Q').$$ On the other hand, using the fact that $\mu(Q')\leq \mu(Q_0)$ we derive $$\Theta(Q')\leq\frac{\ell(Q_0)^s}{\ell(Q')^s}\,\Theta(Q_0) \leq \delta_0 \,\frac{\ell(Q_0)^s}{\ell(Q')^s}\,\Theta(R),$$ and so, taking int account that $\Theta(Q')\approx\tau\,\Theta(R)$ we deduce that $\dfrac{\ell(Q')^s}{\ell(Q_0)^s} \lesssim \dfrac{\delta_0}\tau.$ Therefore, recalling that $p(Q_0)\lesssim B\,\Theta(R)$, we get $$\frac{\ell(Q')}{\ell(Q_0)}\,p(Q_0) \lesssim\frac{\delta_0^{1/s}}{\tau^{1/s}}\,B\,\Theta(R).$$ Together with and , this gives that $$p(Q')\lesssim \Theta(Q') + \frac{\delta_0^{1/s}}{\tau^{1/s}}\,B\,\Theta(R).$$ Choosing $\tau = \delta_0^{1/(s+1)}\,B^{s/(s+1)}$, we get $p(Q')\leq c\,\Theta(Q')$, as wished. So if the constant $c_{db}$ has been taken big enough, $Q'$ is a $p$-doubling cube and then, by construction, $Q=Q'$, and so the statement (c) follows from . Controlling the size of the trees ================================= The objective of this section consists in proving the following. \[lemdif1\] For every $R\in{\operatorname{Top}}$, we have $$\sum_{Q\in{\operatorname{Tree}}_0(R)}\mu(Q)\leq c(B,\delta_0,M)\,\mu(R).\label{eqdif1}$$ As a consequence, $$\label{eqdif2} \sigma({\operatorname{Tree}}(R))\leq c'(B,\delta_0,M)\,\Theta(R)^2\,\mu(R).$$ This result is an easy consequence of the next one. \[lemtouch\] There are constants $\delta,\eta>0$ depending on $B,M,\delta_0$ such the following holds. Given $R\in{\operatorname{Top}}$, for any $R'\in{\operatorname{Tree}}(R)$, let ${{\mathcal S}}(R')$ denote the subfamily of the cubes $Q\in{\operatorname{Stop}}_0(R)$ which are contained in $R'$ and satisfy $\ell(Q)\geq\delta\,\ell(R')$. Then, ${{\mathcal S}}(R')$ is non-empty and $$\mu\biggl (\,\bigcup_{Q\in {{\mathcal S}}(R')} Q\biggr)\geq \eta\,\mu(R').$$ The estimate follows by applying the preceding result iteratively. Indeed, for each $k\geq0$ denote by ${{\mathcal A}}_k$ the collection of the cubes $R'\in{\operatorname{Tree}}_0(R)$ such that $$\delta^{k+1}\ell(R)< \ell(R') \leq \delta^{k}\,\ell(R)$$ and have maximal side length. Notice that this is a pairwise disjoint family. From Lemma \[lemtouch\] it follows that for, any $R'\in{{\mathcal A}}_k$, $$\mu\biggl (R'\cap \bigcup_{Q\in{{\mathcal A}}_{k+2}:\,Q\subset R'}Q\biggr) \leq (1-\eta)\,\mu(R').$$ Summing over all $R'\in{{\mathcal A}}_k$ we obtain $$\mu\biggl (\bigcup_{Q\in{{\mathcal A}}_{k+2}}Q\biggr) \leq (1-\eta)\,\mu\biggl (\bigcup_{Q\in{{\mathcal A}}_{k}}Q\biggr) .$$ By iterating this estimate we get $$\mu\biggl (\bigcup_{Q\in{{\mathcal A}}_{k}}Q\biggr)\leq (1-\eta)^{(k-1)/2}\,\mu(R),$$ and thus $$\sum_{Q\in{\operatorname{Tree}}_0(R)}\mu(Q)\leq c(\delta)\,\sum_{k\geq0} \sum_{Q\in{{\mathcal A}}_k}\mu(Q) \leq c(\delta)\,\sum_{k\geq0}(1-\eta)^{(k-1)/2}\mu(R) \leq c(\delta,\eta)\,\mu(R),$$ which yields . The inequality follows from , Lemma \[lemfac1\] (b) and the fact that $$\sigma({\operatorname{Tree}}_0(R))\leq B^2\,\Theta(R)^2\sum_{Q\in{\operatorname{Tree}}_0(R)}\mu(Q)\leq B^2c(B,M,m_0)\,\Theta(R)^2\,\mu(R).$$ To prove Lemma \[lemtouch\] we need to introduce some notation. In the situation of the lemma, given $Q\in {\operatorname{Tree}}_0(R)$ and an integer $N>1$, we denote $${{\mathcal F}}_N(Q)={\operatorname{Tree}}_0(R)\cap{{\mathcal D}}_N(Q)$$ and we consider the set $$F_N(Q)= \bigcup_{P\in {{\mathcal F}}_N(Q)}P.$$ Also, we consider the following families of cubes, for $N_0\geq N>1$, $$I_{N_0}(Q) = \biggl\{P\in{\operatorname{Tree}}_0(R):\, P\subset Q, \,\mu(P\cap F_{N_0}(Q))\leq \frac14 \mu(P)\biggr\}$$ and $${{\widetilde{{{\mathcal F}}}}}_{N,N_0}(Q) = \Bigl\{P\in{{\mathcal F}}_N(Q):\nexists P'\in I_{N_0}(Q) \text{ such that }P'\supset P\Bigr\}.$$ Then we denote $${{\widetilde{F}}}_{N,N_0}(Q)= \bigcup_{P\in{{\widetilde{{{\mathcal F}}}}}_{N,N_0}(Q)} P = F_N(Q) \setminus \bigcup_{P\in I_{N_0}(Q)}P .$$ \[lemfnt\] Let $N_0\geq N>1$ be integers. Let $Q\in{\operatorname{Tree}}_0(R)$ be such that $\mu(F_{N_0}(Q))\geq\frac12\,\mu(Q)$. Then, $$\mu({{\widetilde{F}}}_{N,N_0}(Q))\geq \mu({{\widetilde{F}}}_{N_0,N_0}(Q))\geq\frac14 \,\mu(Q).$$ The first inequality is trivial. To prove the second one, denote by ${{\widetilde{I}}}_{N_0}(Q)$ the subfamily of maximal cubes from $I_{N_0}(Q)$. Notice that $$F_{N_0}(Q)\setminus{{\widetilde{F}}}_{N_0,N_0}(Q)\subset \bigcup_{P\in {{\widetilde{I}}}_{N_0}(Q)} P.$$ Since the cubes from ${{\widetilde{I}}}_{N_0}(Q)$ are disjoint, we have $$\mu\bigl (F_{N_0}(Q)\setminus{{\widetilde{F}}}_{N_0,N_0}(Q)\bigr) \leq \sum_{P\in {{\widetilde{I}}}_{N_0}(Q)} \mu(F_{N_0}(Q)\cap P)\ \leq \frac14 \sum_{P\in {{\widetilde{I}}}_{N_0}(Q)} \mu(P) \leq \frac14\,\mu(Q).$$ Therefore, $$\mu({{\widetilde{F}}}_{N_0,N_0}(Q)) = \mu(F_{N_0}(Q)) - \mu\bigl (F_{N_0}(Q)\setminus{{\widetilde{F}}}_{N_0,N_0}(Q)\bigr) \geq \frac12 \,\mu(Q) - \frac14\,\mu(Q) = \frac14\,\mu(Q).$$ \[lembola\] Let $N\geq 1$. Let $R'\in{\operatorname{Tree}}_0(R)$ be such that $F_N(R')\neq\varnothing$. Then there exists an open ball $B_{R'}\subset\frac{10}9R'$ which satisfies $$B_{R'}\cap F_N(R')=\varnothing\quad\mbox{ and }\quad \partial B_{R'}\cap F_N(R')\neq\varnothing.$$ Let $Q_i$, $i=1,\ldots,m$ be the sons of $R'$. By the separation condition , we know that there exists some constant $\tau>0$ such that $${\operatorname{dist}}\biggl (Q_1,\bigcup_{i=2}^m Q_i\biggr) \geq \tau\,\ell(Q_1).$$ Thus $(1+\tau)Q_1 \setminus Q_1$ does not intersect any cube from ${{\mathcal D}}(R')$ different from $R'$. So we may take an open ball $B_0$ contained in $R'\cap \bigl ((1+\tau)Q_1 \setminus Q_1\bigr)$ with $r(B_0) \approx \ell(Q_1)\approx \ell(R')$, which clearly satisfies $B_0\cap F_N(R')=\varnothing$. Let $x_1$ be a point from $F_N(R')$ which is closest to the center $x_0$ of $B_0$. We move $B_0$ keeping its center in the segment $[x_0,x_1]$ till its boundary touches $F_N(R')$, and then we call it $B_{R'}$. Assuming $r\leq \ell(R')/20$, it follows that $B_{R'}\subset \frac{10}9R'$. Given a hyperplane $H\subset {{\mathbb R}}^d$, $x\in H$, $r>0$ and $0<\alpha\leq1/2$, we consider the cone $$X(x,H,\alpha) = \bigl\{y\in {{\mathbb R}}^d:\, {\operatorname{dist}}(y,H)\leq \alpha\,|x-y|\bigr\}.$$ For $0\leq r_1< r_2$, we also set $$X(x,H,\alpha,r_1,r_2) = \bigl\{y\in {{\mathbb R}}^d:\, {\operatorname{dist}}(y,H)\leq \alpha\,|x-y|,\,r_1\leq|x-y|\leq r_2\bigr\}.$$ \[lemcon\] Let $R'\in{\operatorname{Tree}}_0(R)$, $P\in{{\widetilde{F}}}_N(R')$, and $x\in P$. Given $0<\alpha\leq1/2$, there exists $m_0=m_0(\alpha)\geq 1$ such that for all $r\geq \ell(P)$ and $m\geq m_0$, $$\mu(X(x,H,\alpha,\tfrac1{10}\ell(P),r) \cap F_{mN}(R'))\leq c_3\,B\,\alpha^{s+1-d}\,r^s.$$ We assume for simplicity that $x=0$. It is clear that we may also assume that $r\leq{\operatorname{diam}}R'$. For $k\geq0$, let $r_k = (1-\alpha)^k\,r$ and consider the closed annulus $$A_k = A(0,r_{k+1},r_k),$$ so that $$X(0,H,\alpha,r_{k+1},r_k) = A_k \cap X(0,H,\alpha,0,r).$$ Let $n_0$ be the largest integer such that $r_{n_0}\leq \tfrac1{10}\ell(P)$. We set $$\mu\bigl(X(0,H,\alpha,\tfrac1{10}\ell(P),r) \cap F_{mN}(R')\bigr)\leq \sum_{k=0}^{n_0-1} \mu\bigl(X(0,H,\alpha,r_{k+1},r_k)\cap F_{mN}(R')\bigr).$$ We are going to estimate each summand on the right side. To this end, consider the $(d-1)$-dimensional annulus $A_k\cap H$. This can be covered by a family of closed balls $B_1^k,\ldots,B_{n_k}^k$ of radius $\alpha\,r_k$ centered in $A_k\cap H$ with $$n_k\approx \frac1{\alpha^{d-2}}.$$ To prove this notice that the $(d-1)$-dimensional volume of $A_k\cap H$ is comparable to $$\bigl(r_k-r_{k+1}\bigr)\,r_k^{d-2} = \alpha\,r_k^{d-1},$$ while the one of $B_i^k\cap H$ is $c\,\bigl(\alpha\,r_k\bigr)^{d-1}$ for each $i$. So $$n_k\approx \frac{\alpha\,r_k^{d-1}}{\bigl(\alpha\,r_k\bigr)^{d-1}} = \frac1{\alpha^{d-2}}.$$ We claim now that $$\label{eqbik10} X(0,H,\alpha,r_{k+1},r_k) \subset \bigcup_{i=1}^{n_k} 3B_i^k.$$ Indeed, given $y\in X(0,H,\alpha,r_{k+1},r_k)$, let $y'$ be its orthogonal projection on $H$, so that $|y-y'|={\operatorname{dist}}(y,H)$. Take now $$y'' = \frac{|y|}{|y'|}\,y'.$$ Observe that $y''\in H$ and $|y''|=|y|$, and so $y''\in H\cap X(0,H,\alpha,r_{k+1},r_k)$. Also, we have $$|y'-y''| = |y'|\,\biggl(1-\frac{|y|}{|y'|}\biggr) = \bigl| |y'|- |y|\bigr| \leq |y'-y|.$$ Therefore, $$|y-y''|\leq |y-y'| + |y'-y''|\leq 2|y-y'|=2{\operatorname{dist}}(y,H)\leq2\alpha r_k.$$ As a consequence, if $y''\in B_i^k$, then ${\operatorname{dist}}(y,B_i^k)\leq 2\,r(B_i^k)$ and so $y\in 3B_i^k$, which proves our claim. Next we are going to check that $$\mu\bigl(3B_i^k\cap F_{mN}(R')\bigr)\leq c\,B\,\Theta(R)\,r(B_i^k)^s = c\,B\,\Theta(R)\,\alpha^s\, (1-\alpha)^{sk}\,r^s.$$ for all $0\leq k\leq n_0$ and $1\leq i \leq n_k$. To this end notice that for these indices $k,i$, $$\label{eqmes19} r(B_i^k) = \alpha\,r_k\geq \alpha\,r_{n_0} \geq \alpha\,(1-\alpha)\,r_{n_0-1}\geq \frac\alpha{20}\,\ell(P)\geq \frac\alpha{20}\,8^{-N}\,\ell(R'),$$ recalling that $P\in{{\mathcal F}}_N(R')$ in the last inequality. On the other hand, if $\mu\bigl(3B_i^k\cap F_{mN}(R')\bigr)\neq0$, then there exists some cube $Q\in F_{mN}(R')$ which intersects $3B_i^k$ and we have $$\label{eqmes20} \ell(Q)\leq 2^{-mN}\ell(R').$$ If $m_0$ is big enough and $m\geq m_0$, from and we infer that $\ell(Q)\leq r(B_i^k)$. Thus there exists some ancestor $Q'$ of $Q$ such that $3B_i^k\subset (1+c_{sep})Q'$ which satisfies $\ell(Q')\approx r(B_i^k)$. In particular, either $Q'\in{\operatorname{Tree}}_0(R)$ or $Q'$ contains $R$ and $\ell(Q')\approx\ell(R)$. In any case (using that $R$ is p-doubling in the second one), we deduce that $\mu(Q')\leq c\,B\,\Theta(R)\,\ell(Q')^s$. So, by the separation property $$\mu\bigl(3B_i^k\cap F_{mN}(R')\bigr)\leq \mu((1+c_{sep})Q')= \mu(Q') \lesssim B\,\Theta(R)\,\ell(Q')^s\approx B\,\Theta(R)\,r(B_i^k)^s,$$ as wished. From the latter estimate we infer that $\mu\bigl(3B_i^k\cap F_{mN}(R')\bigr)\lesssim\,B\,\Theta(R)\,\alpha^s\, (1-\alpha)^{sk}\,r^s$ for all $i,k$. Together with and the fact that $n_k\approx\alpha^{2-d}$, this implies that $$\mu\bigl(X(0,H,\alpha,r_{k+1},r_k)\cap F_{mN}(R')\bigr)\lesssim B\,\Theta(R)\alpha^{2+s-d}\,(1-\alpha)^{sk}\,r^s.$$ As a consequence, $$\begin{aligned} \mu\bigl(X(0,H,\alpha,\tfrac1{10}\ell(P),r) \cap F_{mN}(R')\bigr)&\lesssim B\,\Theta(R)\,r^s\,\alpha^{2+s-d}\sum_{k=0}^{n_0-1}(1-\alpha)^{sk}\\ & \leq \frac{B\,\Theta(R)\,r^s\,\alpha^{2+s-d}}{1-(1-\alpha)^s} \approx B\,\Theta(R)\alpha^{1+s-d}\,r^s.\end{aligned}$$ \[lemtouchingpoint\] Suppose that $N$ and $m_0$ are big enough, depending on $\delta_0$ and $B$, and let $N_0$ be such that $N_0\geq m_0\,N$. Given $R'\in{\operatorname{Tree}}_0(R)$, let $B_{R'}$ be an open ball centered at some point from $R'$ with $r(B_{R'})\geq c^{-1}\ell(R')$ such that $$B_{R'}\cap F_N(R')=\varnothing \quad\mbox{and}\quad \partial B_{R'}\cap {{\widetilde{F}}}_{N,N_0}(R')\neq\varnothing.$$ Let $P\in {{\widetilde{{{\mathcal F}}}}}_{N,N_0}(R')$ be such that $P\cap \partial B_{R'}\neq \varnothing$. We have $$\label{eqtou3} |{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P}\mu)(x)|\geq C_1(B, \delta_{0})\, \Theta(R)\,N\quad\mbox{for all $x\in P$.}$$ Using Lemma \[lemcomp\], it is easy to check that it is enough to prove the estimate for $x\in P\cap \partial B_{R'}$. Let $\alpha$ be a parameter to be chosen later in the proof. Let $H$ be the hyperplane which is tangent to $\partial B_{R'}$ at $x$. For simplicity we assume that $H=\{x\in{{\mathbb R}}^d:x_d=0\}$ and we suppose that $B_{R'}$ is contained in the half plane $\{y\in{{\mathbb R}}^d:y_d\leq0\}$. Consider the cone $X_\alpha = X(x,H,\alpha)$ and let $V_\alpha = {{\mathbb R}}^d\setminus X_\alpha$. Also set $$V_\alpha^+ = \{y\in{{\mathbb R}}^d:y_d>\alpha\,|x-y|\},\qquad V_\alpha^- = \{y\in{{\mathbb R}}^d:y_d<-\alpha\,|x-y|\} ,$$ so that $V_\alpha = V_\alpha^+\cup V_\alpha^-$. Let $B_0$ be an open ball centered at $x$ such that $B_0\cap V_\alpha^-\subset B_{R'}$ with radius comparable to $r(B_{R'})$ with some constant depending on $\alpha$. See Figure \[f.uno\]. (-6.72,-3.94) rectangle (9.12,3.48); (0,0) circle (1.42cm); (-2,-1)– (0,0); (0,0)– (2,1); (-2,1)– (0,0); (0,0)– (2,-1); (0,-1.59) circle (1.59cm); (-0.18,-0.1) node\[anchor=north west\] [$x$]{}; (-2.52,0)– (2.58,0); (2.12,0.5) node\[anchor=north west\] [$\mathcal H$]{}; (1.5,0.5) node\[anchor=north west\] [$\alpha$]{}; (-2.42,0.82) node\[anchor=north west\] [$X_{\alpha}$]{}; (0,-0.44) node\[anchor=north west\] [$ V_\alpha^-$]{}; (0,1.18) node\[anchor=north west\] [$ V_\alpha^+$]{}; (-1.16,2.0) node\[anchor=north west\] [$B_0$]{}; (-1.56,-2.8) node\[anchor=north west\] [$B_{R'}$]{}; (0,0) circle (1.5pt); (-1.27,-0.64) circle (1.5pt); (1.27,-0.64) circle (1.5pt); Write $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P}\mu)(x)|\geq |{{\mathcal R}}(\chi_{(B_0\cap F_{N_0}(R')) \setminus P}\mu)(x)|-|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus B_0}\mu)(x)|$$ We can bound the second term on the right hand side using the size condition on the fractional Riesz transform: $$\label{e.outside} |{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus B_0}\mu)(x)|\leq \frac{\mu(R')}{r(B_0)^{s}} \leq c(\alpha) \frac{\mu(R')}{\ell(R')^{s}} \leq c(\alpha) B\,\Theta(R).$$ We focus on the term $|{{\mathcal R}}(\chi_{(B_0\cap F_{N_0}(R'))\setminus P}\mu)(x)|$. We will use that $(B_0\cap F_{N_0}(R'))\cap V_\alpha^+ =\varnothing$, by the construction of $B_0$, and that $x_{d}-y_{d}>0$ in $V_\alpha^-$. Denoting by ${{\mathcal R}}^d$ the $d$-th component of ${{\mathcal R}}$, we have $$\begin{aligned} {{\mathcal R}}^d(\chi_{B_0\setminus P}\mu)(x) & = {{\mathcal R}}^{d}(\chi_{(B_0\cap F_{N_0}(R')\setminus P)\cap V_\alpha^+}\mu)(x) + {{\mathcal R}}^{d}(\chi_{(B_0\cap F_{N_0}(R')\setminus P)\cap V_{\alpha}^-}\mu)(x) \\&\quad+{{\mathcal R}}^{d}(\chi_{(B_0\cap F_{N_0}(R')\setminus P)\cap X_{\alpha}}\mu)(x) \\ & \geq {{\mathcal R}}^{d}(\chi_{(B_0\cap F_{N_0}(R')\setminus P)\cap V_{\alpha}^-}\mu)(x) - |{{\mathcal R}}^{d}(\chi_{(B_0\cap F_{N_0}(R')\setminus P)\cap X_{\alpha}}\mu)(x)| \\ &\geq \int_{B_0\cap F_{N_0}(R')\cap V_\alpha^-\setminus P}\!\frac{x_{d}-y_{d}}{|x-y|^{s+1}}d\mu(y)- \!\!\int_{B_0\cap F_{N_0}(R')\cap X_{\alpha}\setminus P}\! \frac{|x_{d}-y_{d}|}{|x-y|^{s+1}}d\mu(y)\\ &\geq \int_{B_0\cap F_{N_0}(R')\cap V_{\alpha}^-\setminus P}\frac{\alpha}{|x-y|^s}d\mu(y)- \int_{B_0\cap F_{N_0}(R')\cap X_{\alpha}\setminus P}\frac{\alpha}{|x-y|^s}d\mu(y)\\ &=\int_{B_0\cap F_{N_0}(R')\setminus P}\frac{\alpha}{|x-y|^s}d\mu(y)- 2\int_{B_0\cap F_{N_0}(R')\cap X_{\alpha} \setminus P}\frac{\alpha}{|x-y|^s}d\mu(y)\\ & = (I)-(II).\end{aligned}$$ We will study each of $(I)$ and $(II)$ in turn. First, we will bound $(II)$ from above. To this end let $P_0\in{{\mathcal D}}$ be the smallest cube such that $1.1P_0$ contains $B_0$. It is clear that $\ell(P_0)\approx r(B_0)$. Given any $Q\in{{\mathcal D}}$, $Q^1$ stands for the parent of $Q$. Using Lemma \[lemcon\] then we get $$\begin{aligned} (II) & \leq 2\sum_{Q:P\subset Q\subsetneq P_0}\int_{( Q^1\setminus Q)\cap F_{N_0}(R')\cap X_\alpha} \frac{\alpha}{|x-y|^s}d\mu(y)\\ & \leq c\sum_{Q:P\subset Q\subsetneq P_0} \frac{\alpha\,\mu\bigl(( Q^1\setminus Q)\cap F_{N_0}(R')\cap X_\alpha\bigr)}{\ell(Q)^s}\nonumber \\ & \leq c \sum_{Q:P\subset Q\subsetneq P_0} \frac{B\,\Theta(R)\,\alpha^{2+s-d}\ell( Q^1)^s}{\ell(Q)^s} \nonumber\\ &\approx B\,\Theta(R)\,\alpha^{2+s-d}\,n_0,\nonumber\end{aligned}$$ where $n_0= \#\{Q\in{{\mathcal D}}:P\subset Q\subsetneq P_0\}$. Notice that in the third inequality we assumed $m_0(\alpha)$ big enough, so that the assumptions of Lemma \[lemcon\] are satisfied. We now look for a lower bound for $(I)$. First we fix a integer $k_{0}$ to be chosen later in the proof, depending on $B$ and $\delta_0$. For $j\geq1$, we denote by $P^{j}$, the $j$-th ancestor of $P$ in the Cantor construction. Then we have $$\begin{aligned} \label{eqI} (I) & \geq \sum_{k=1}^{n_1}\int_{(P^{k_{0}(k+1)}\setminus P^{k_{0}k}) \cap F_{N_0}(R')} \frac{\alpha}{|x-y|^s}d\mu(y)\\ & \geq c(k_0)^{-1} \sum_{k=1}^{n_1}\frac{\alpha\,\mu\bigl(P^{k_{0}(k+1)} \cap F_{N_0}(R')\setminus P^{k_{0}k}\bigr)}{\ell(P^{k_{0}(k+1)})^s},\nonumber\end{aligned}$$ where $n_1$ is the largest integer such that $P^{k_{0}(n_1+1)}\subset B_0$. To estimate $\mu\bigl(P^{k_{0}(k+1)} \cap F_{N_0}(R')\setminus P^{k_{0}k}\bigr)$ from below notice that $P^{k_{0}(k+1)}\not\in I_{N_0}(R')$, because otherwise $P\not\in {{\widetilde{{{\mathcal F}}}}}_{N,N_0}(R')$. This tells us that $$\mu\bigl(P^{k_{0}(k+1)}\cap F_{N_0}(R')\bigr)>\frac14\,\mu\bigl(P^{k_{0}(k+1)}\bigr) \geq \frac14\,\delta_0\,\Theta(R)\,\ell(P^{k_{0}(k+1)}\bigr)^s.$$ Taking into account that $$\mu\bigl(P^{k_{0}k}\bigr)\leq B\Theta(R)\,\ell(P^{k_{0}k}\bigr)^s \leq B\Theta(R)\,2^{-k_0s}\,\ell(P^{k_{0}(k+1)}\bigr)^s,$$ we deduce that for $k_0$ big enough, depending on $B$ and $\delta_0$, $$\mu\bigl(P^{k_{0}(k+1)}\cap F_{N_0}(R')\bigr)\geq 2\,\mu\bigl(P^{k_{0}k}\bigr),$$ and thus $$\mu\bigl(P^{k_{0}(k+1)} \cap F_{N_0}(R')\setminus P^{k_{0}k}\bigr)\geq \frac12\,\mu\bigl(P^{k_{0}(k+1)}\cap F_{N_0}(R')\bigr)\geq \frac18\,\delta_0\,\Theta(R)\,\ell(P^{k_{0}(k+1)}\bigr)^s.$$ Plugging this estimate into we obtain $$(I)\geq c(k_0,\delta_0)^{-1}\,\alpha\,\Theta(R)\,n_1.$$ Taking into account that $n_1\approx n_0/k_0$, we get $$(I)\geq c(B,\delta_0)^{-1}\,\alpha\,\Theta(R)\,n_0.$$ Together with the upper bound we found for (II), this gives $${{\mathcal R}}^d(\chi_{B_0\setminus P}\mu)(x)\geq \bigl[ c(B,\delta_0)^{-1}\,\alpha\,- c\,B\,\alpha^{2+s-d}\bigr]\Theta(R)\,n_0.$$ For $\alpha$ small enough depending on $B$ and $\delta_0$, the first factor on the right side is positive, and then we get ${{\mathcal R}}^d(\chi_{B_0\setminus P}\mu)(x)\geq c(\delta_0,B)^{-1}\Theta(R)n_0$. This implies , taking into account that $n_0$ is comparable to $N$ (with some constant that may depend on $\alpha$, and so on $\delta_0$ and $B$). \[lemcub2\] Let $M'>M$ be some large constant. Suppose that $N_0$ is a sufficiently large integer, depending on $B$ and $M'$. For all $R'\in{\operatorname{Tree}}_0(R)$, one of the two following conditions holds: - The union of the cubes from ${\operatorname{Stop}}_0(R)$ which belong to $\bigcup_{j=1}^{N_0+1} {{\mathcal D}}_j(R')$ has $\mu$-measure larger that $\mu(R')/4$. - There exists a family of pairwise disjoint cubes ${{\mathcal T}}(R')\subset \bigcup_{j=1}^{N_0+1} {{\widetilde{{{\mathcal F}}}}}_{j,N_0}(R')$ with $$\label{eqrest53} \mu\biggl (\,\bigcup_{P\in{{\mathcal T}}(R')} P\biggr)\geq \frac3{20}\,\mu(R')$$ such that, for each $P\in{{\mathcal T}}(R')$, $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P}\mu)(x)|\geq M'\, \Theta(R)\quad\mbox{ for all $x\in P$.}$$ Suppose that (a) does not hold. This means that $$\mu\bigl (F_{N_0}(R')\bigr) \geq \frac34\,\mu(R').$$ To show that (b) holds we choose $N_0=2n\, N$, where $n$ and $N$ are some big integers to be fixed below. In particular, we will require $n\geq m_0$, where $m_0=m_0(B,\delta_0)$ is as in Lemma \[lemtouchingpoint\]. Also we will assume that $N$ is big enough so that the same lemma holds and, moreover, $$C_1(B, \delta_{0})\,N \geq 3M',$$ where $C_1(B,\delta_0)$ is the constant in . We define ${{\mathcal T}}(R')$ as the subfamily of cubes $P$ from $\bigcup_{j=1}^{N_0+1} {{\widetilde{{{\mathcal F}}}}}_{j,N_0}(R')$ such that $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P}\mu)(x)|\geq M'\, \Theta(R)\quad\mbox{ for all $x\in P$}$$ and moreover have maximal side length. To prove we define an auxiliary family ${{\mathcal T}}_{aux}(R')$ of pairwise disjoint cubes from $\bigcup_{j=1}^{N_0} {{\widetilde{{{\mathcal F}}}}}_{j,N_0}(R')$ by a repeated application of a touching point argument, as follows. We will inductively construct families ${{\mathcal T}}_{aux}^j(R')$, $j=1,\ldots,n$. In the case $j=1$, by Lemma \[lembola\], there exists a ball $B_{R'}$ contained in $\frac{10}9R'$, with radius $\geq c^{-1}\,\ell(R')$ which satisfies $$B_{R'}\cap F_N(R')=\varnothing\quad\mbox{ and }\quad \partial B_{R'}\cap F_N(R')\neq\varnothing.$$ Let $P\in {{\mathcal F}}_{N}(R')$ be such that $\partial B_{R'}\cap P\neq\varnothing$. We set $${{\mathcal T}}_{aux}^1(R') = \{P\}.$$ Assume now that ${{\mathcal T}}_{aux}^1(R'),\ldots,{{\mathcal T}}_{aux}^{j}(R')$ have already been defined and let us construct ${{\mathcal T}}_{aux}^{j+1}(R')$. For each $Q\in F_{jN}(R')$ which intersects $F_{N_0}(R')$ and is not contained in any cube from ${{\mathcal T}}_{aux}^1(R')\cup\ldots\cup{{\mathcal T}}_{aux}^{j}(R')$, consider a ball $B_Q$ contained in $\frac{10}9Q$, with radius $\geq c^{-1}\,\ell(Q)$ which satisfies $$B_Q\cap F_N(Q)=\varnothing\quad\mbox{ and }\quad \partial B_Q\cap F_N(Q)\neq\varnothing.$$ Let $P_Q\in {{\mathcal F}}_N(Q)\subset {{\mathcal F}}_{(j+1)N}(R')$ be such that $\partial B_Q\cap P_Q\neq\varnothing$. We set $${{\mathcal T}}_{aux}^{j+1}(R') = \{P_Q\}_Q,$$ where $Q$ ranges over all the cubes described above. Notice that the cubes from ${{\mathcal T}}_{aux}(R')= {{\mathcal T}}_{aux}^1(R')\cup\ldots\cup{{\mathcal T}}_{aux}^{n}(R')$ are pairwise disjoint, by construction. We claim that if $m_0$ is sufficiently big, then $$\label{eqrest55} \mu\biggl (\bigcup_{P\in{{\mathcal T}}_{aux}(R')} P\biggr)\geq \frac9{10}\,\mu(F_{N_0}(R')).$$ To check this, observe first that there is some constant $\delta>0$ depending on $N$, $B$ and $\delta_{0}$ such that $$\mu(P_Q)\geq \delta\mu(Q),$$ for $Q$ and $P_Q$ as in the previous paragraph. Indeed, $\ell(P_Q)\geq 8^{-N}\ell(Q)$, and thus $$\mu(P_Q)\geq \delta_{0}\,\ell(P_Q)^s \Theta(R)\geq 8^{-Ns}\delta_{0}\,\ell(Q)^s \,\Theta(R)\geq 8^{-Ns}\delta_{0}B^{-1}\,\mu(Q) =:\delta\,\mu(Q).$$ Then, from the above construction, it turns out that $$\mu\biggl (\,\bigcup_{P\in{{\mathcal T}}_{aux}^j(R')}P\biggr) \geq \delta\,\biggl[\mu(F_{N_0}(R')) - \mu\biggl (\,\bigcup_{k=1}^{j-1}\bigcup_{P\in{{\mathcal T}}_{aux}^k(R')}P\biggr)\biggr].$$ Summing this estimate over $1\leq j\leq n$, we obtain $$\begin{aligned} \mu\biggl (\bigcup_{P\in{{\mathcal T}}_{aux}(R')}P\biggr) & \geq \delta n\,\mu(F_{N_0}(R')) - \delta \sum_{j=1}^n\mu\biggl (\bigcup_{k=1}^{j-1}\bigcup_{P\in{{\mathcal T}}_{aux}^k(R')}P\biggr)\\ & \geq \delta n\,\mu(F_{N_0}(R')) - \delta n \,\mu\biggl (\bigcup_{P\in{{\mathcal T}}_{aux}(R')}P\biggr).\end{aligned}$$ Thus, $$\mu\biggl (\bigcup_{P\in{{\mathcal T}}_{aux}(R')}P\biggr) \geq \frac{\delta n}{1+\delta n}\,\mu(F_{N_0}(R')).$$ So, if $n$ is big enough, the claim follows. To prove , denote now by ${{\mathcal T}}_0(R')$ the subfamily of the cubes $Q\in{{\mathcal T}}_{aux}(R')$ which contain some cube $P\in{{\widetilde{{{\mathcal F}}}}}_{N_0,N_0}(R')$. Notice that such cubes $Q$ belong to ${{\widetilde{{{\mathcal F}}}}}_{jN,N_0}(R')$ for some $j\in[1,n]$. So if $Q\in {{\mathcal T}}_{aux}(R')\setminus {{\mathcal T}}_0(R')$, then $${\operatorname{supp}}\mu\cap Q\subset F_{N_0}(R')\setminus {{\widetilde{F}}}_{N_0,N_0}(R').$$ Together with Lemma \[lemfnt\] this yields $$\begin{aligned} \mu\biggl (\,\bigcup_{Q\in {{\mathcal T}}_{aux}(R')\setminus {{\mathcal T}}_0(R')}Q\biggr) & \leq \mu\bigl (F_{N_0}(R')\setminus {{\widetilde{F}}}_{N_0,N_0}(R')\bigr) \\ &\leq \mu(F_{N_0}(R')) - \frac14 \mu(R')\leq \frac34\,\mu(F_{N_0}(R')).\end{aligned}$$ Thus, $$\begin{aligned} \mu\biggl (\,\bigcup_{Q\in {{\mathcal T}}_0(R')}Q\biggr) & = \mu\biggl (\bigcup_{Q\in{{\mathcal T}}_{aux}(R')} Q\biggr) - \mu\biggl (\,\bigcup_{Q\in {{\mathcal T}}_{aux}(R')\setminus {{\mathcal T}}_0(R')}Q\biggr)\\ & \geq \frac9{10}\,\mu(F_{N_0}(R'))- \frac34\,\mu(F_{N_0}(R')) = \frac3{20}\,\mu(F_{N_0}(R')).\end{aligned}$$ On the other hand, from the construction above and Lemma \[lemtouchingpoint\], it turns out that for each $P\in{{\mathcal T}}_0(R')\cap {{\mathcal F}}_{jN}(R')$, with $1\leq j\leq n$, there exists some cube $P'\in{\operatorname{Tree}}_0(R')\cap{{\mathcal F}}_{(j-1)N}(R')$ with $P\subset P'\subset R'$ such that $$|{{\mathcal R}}(\chi_{P'\cap F_{N_0}(R')\setminus P}\mu)(x)|\geq 3M'\, \Theta(R)\quad\mbox{ for all $x\in P$.}$$ Indeed, notice that $P\in{{\widetilde{{{\mathcal F}}}}}_{N,N_0-(j-1)N}(P')$ because $P\in{{\mathcal T}}_0(R')$ and we have $$N_0-(j-1)N\geq N_0-nN=nN\geq m_0\,N,$$ so that the assumptions in Lemma \[lemtouchingpoint\] hold. We distinguish now two cases. First, if $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P'}\mu)(x)|\leq 2M'\, \Theta(R)\quad\mbox{ for all $x\in P'$,}$$ then $$\begin{aligned} |{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P}\mu)(x)| & \geq |{{\mathcal R}}(\chi_{P'\cap F_{N_0}(R')\setminus P}\mu)(x)| - |{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P'}\mu)(x)|\\ & \geq M'\, \Theta(R)\quad\mbox{ for all $x\in P$,}\end{aligned}$$ so that $P$ belongs to ${{\mathcal T}}(R')$. In the second case, if $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P'}\mu)(x)|\geq 2M'\, \Theta(R)\quad\mbox{ for some $x\in P'$,}$$ then we deduce that, for all $y\in P'$, $$|{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P'}\mu)(y)|\geq 2M'\, \Theta(R) - c\,p(P') \geq 2M'\, \Theta(R) - C(B)\,\Theta(R) \geq M'\,\Theta(R),$$ assuming that $M'$ is big enough. So $P'\in{{\mathcal T}}(R')$, and thus $P\subset \bigcup_{Q\in{{\mathcal T}}(R')} Q$. Then we infer that $$\mu\biggl (\,\bigcup_{P\in{{\mathcal T}}(R')} P\biggr)\geq \mu\biggl (\,\bigcup_{P\in{{\mathcal T}}_0(R')} P\biggr)\geq \frac3{20}\,\mu(R').$$ \[lemrieszmax\] We have $$\|{{\mathcal R}}(\chi_{R'\setminus F_{N_0}(R')}\mu)\|_{L^2(\mu{{\lfloor}}F_{N_0}(R'))} \leq C(B,M)\,\Theta(R)\,\mu(R')^{1/2}.$$ We will use the technique of the suppressed kernels of Nazarov, Treil and Volberg. Consider the set $$H = \bigcup_{Q\in{\operatorname{Stop}}_0(R)} (1+\tfrac14c_{sep})Q.$$ For $x\in{{\mathbb R}}^d$, denote $\Phi(x) = {\operatorname{dist}}(x,H^c)$. Observe that $\Phi$ is a $1$-Lipschitz function, and if $x\in Q\in{\operatorname{Stop}}_0(R)$, then $\Phi(x)\approx\ell(Q)$. Recall that every $Q'\in{\operatorname{Tree}}_0(R)$ satisfies $\Theta(Q')\leq B\Theta(R)$. Then it easily follows that all the balls $B(x,r)$ with $x\in R$, $\Phi(x)\leq r\leq\ell(R)$, such that $$\mu(B(x,r))> C_3\,B\,\Theta(R)\,r^s$$ are contained in $H$ if $C_3$ is some sufficiently big constant. For $x, y\in{{\mathbb R}}^d$ we consider the kernel $K_\Phi(x,y)$ defined in and, given the measure $\sigma =\mu{{\lfloor}}R$, we consider the operator ${{\mathcal R}}_{\Phi,\sigma}$. By we know that given $x\in R$, for all $Q'\in{\operatorname{Tree}}_0(R)\cup {\operatorname{Stop}}_0(R)$ such that $x\in Q'$, $\bigl|{{\mathcal R}}(\chi_{R\setminus Q'} \mu)(x)\bigr| \leq C(B,M) \,\Theta(R)$. Using the separation condition , then we infer that $$|{{\mathcal R}}_{{{\varepsilon}}} \sigma(x)| \leq C(B,M)\,\Theta(R) \quad\mbox{for $x\in Q'\in{\operatorname{Stop}}_0(R)$ and for all ${{\varepsilon}}>c^{-1}\ell(Q)$}.$$ And also $$|{{\mathcal R}}_{{{\varepsilon}}} \sigma(x)| \leq C(B,M)\,\Theta(R) \quad\mbox{for $x\in Q'\in R\setminus{\operatorname{Stop}}_0(R)$ and for all ${{\varepsilon}}>0$}.$$ Notice that if $x$ belongs to some stopping cube $Q$ from ${\operatorname{Stop}}_0(R)$, then $\Phi(x)\approx \ell(Q)$. As a consequence, it follows that $$\sup_{{{\varepsilon}}\geq \Phi(x)}|{{\mathcal R}}_{{\varepsilon}}\sigma(x)| \leq C(B,M)\,\Theta(R) \quad\mbox{for $x\in R$.}$$ By Theorem \[teontv\] we deduce that ${{\mathcal R}}_{\Phi,\sigma}$ is bounded in $L^2(\sigma)$, with its norm not exceeding $C(B,M)\,\Theta(R)$. Therefore, $$\label{eqfaj21} \|{{\mathcal R}}_{\Phi,\sigma}\chi_{R'\setminus F_N(R')} \|_{L^2(\sigma{{\lfloor}}F_N(R'))} \leq c(B,M)\,\Theta(R)\,\mu(R')^{1/2}.$$ To estimate $\|{{\mathcal R}}(\chi_{R'\setminus F_{N_0}(R')}\sigma)\|_{L^2(\sigma{{\lfloor}}F_{N_0}(R'))}$, observe that if $y\in {\operatorname{supp}}\,\sigma\cap R'\setminus F_{N_0}(R')$ and $x\in {\operatorname{supp}}\,\sigma\cap F_{N_0}(R')$, then $y$ belongs to some cube $Q \in{\operatorname{Stop}}_0(R)$, while $x\not\in Q$. Then it follows that $$|x-y|\gtrsim \bigl (\Phi(x) + \Phi(y)\bigr).$$ So, $$\begin{aligned} |{{\mathcal R}}(\chi_{R'\setminus F_{N_0}(R')}\sigma)(x)|&\leq |{{\mathcal R}}_{\Phi,\sigma}(\chi_{R'\setminus F_{N_0}(R')})(x)| + c\,\sup_{r\geq \Phi(x)}\frac{\mu(B(x,r)\cap F_{N_0}(R'))}{r^s}\\ &\leq |{{\mathcal R}}_{\Phi,\sigma}(\chi_{R'\setminus F_{N_0}(R')})(x)| + C(B)\,\Theta(R).\end{aligned}$$ Together with this implies that $$\|{{\mathcal R}}(\chi_{R'\setminus F_N(R')}\sigma)\|_{L^2(\sigma{{\lfloor}}F_N(R'))} \leq C(B,M)\,\Theta(R)\,\mu(R')^{1/2}.$$ If the condition (a) from Lemma \[lemcub2\] holds we are done. Otherwise, we consider the family ${{\mathcal T}}(R')$ in the statement (b) of that lemma and we set $$G=\bigcup_{P\in{{\mathcal T}}(R')} P.$$ By Lemma \[lemrieszmax\], there exists some constant $c_5=c_5(B,M)$ such that the set $$V=\bigl\{x\in F_{N_0}(R'):|{{\mathcal R}}(\chi_{R'\setminus F_{N_0}(R')}\mu)(x)|>c_5\Theta(R)\bigr\}$$ satisfies $$\mu(V)\leq \frac18\mu(R').$$ Then, $$\mu(G\setminus V)\geq \biggl (\frac14 - \frac1{8}\biggr)\mu(R')= \frac1{8}\,\mu(R').$$ Moreover, for $x\in G\setminus V$, if we denote by $P_x$ the cube from ${{\mathcal T}}(R')$ that contains $x$, $$|{{\mathcal R}}(\chi_{R'\setminus P_x}\mu)(x)| \geq |{{\mathcal R}}(\chi_{F_{N_0}(R')\setminus P_x}\mu)(x)| - {{\mathcal R}}(\chi_{R'\setminus F_{N_0}(R')}\mu)(x) \geq (M'-c_3) \, \Theta(R).$$ From Lemma \[lemcomp\] we deduce that $$\begin{aligned} |m_{P_x} {{\mathcal R}}\mu - m_{R'}{{\mathcal R}}\mu| & \geq |{{\mathcal R}}(\chi_{R'\setminus P_x}\mu)(x)| - c\,p(P_x) - c\,p(R') \\ & \geq (M' - c_5- c\,B)\,\Theta(R)\geq 3M\,\Theta(R),\end{aligned}$$ assuming $M'$ big enough in the last inequality. However, since $R',P_x\in{\operatorname{Tree}}_0(R)$, we have $$|m_{P_x} {{\mathcal R}}\mu - m_{R'}{{\mathcal R}}\mu|\leq |m_{P_x} {{\mathcal R}}\mu - m_{R}{{\mathcal R}}\mu| + |m_{R} {{\mathcal R}}\mu - m_{R'}{{\mathcal R}}\mu| \leq 2M\Theta(R),$$ which is a contradiction. Thus (a) from Lemma \[lemcub2\] holds. Types of trees ============== Recall that given a collection of cubes ${{\mathcal A}}\subset{{\mathcal D}}$, $\sigma({{\mathcal A}})$ stands for $$\sigma({{\mathcal A}}) = \sum_{Q\in{{\mathcal A}}}\Theta(Q)^2\mu(Q).$$ So Theorem \[teopri\] consists in proving that $$\|{{\mathcal R}}\mu\|_{L^2(\mu)}^2 \approx \sigma({{\mathcal D}}),$$ under the assumption that $$\sup_{Q\in {{\mathcal D}}} \Theta(Q)\lesssim1.$$ For the proof we need to consider different types of trees. We say that ${{\mathcal T}}$ is a simple tree if it is of the form ${{\mathcal T}}={\operatorname{Tree}}(R)$, with $R\in {\operatorname{Top}}$, (defined in Section \[seccorona\]). Given a small constant $\delta_W>0$, we say that a simple tree ${{\mathcal T}}={\operatorname{Tree}}(R)$, $R\in{\operatorname{Top}}$, is of type $W$ (wonderful) if $$\mu\biggl (\,\bigcup_{Q\in BR(R)} Q\biggr)\geq \delta_W\,\mu(R).$$ If ${{\mathcal T}}$ is not of type $W$, we say that it is of type $NW$. Let $A>10$ to be fixed below. For the reader’s convenience, let us remark that we will choose $$\delta_0,\delta_{W}\ll 1\ll M\ll A\ll B.$$ For instance, we may choose $A=B^{1/100}$. We say that a simple tree ${{\mathcal T}}$ is of type $I\sigma$ (increasing $\sigma$) if it is type $NW$ and $$\sigma({\operatorname{Stop}}(R))>A\,\sigma(\{R\}).$$ We say that ${{\mathcal T}}$ is of type $D\sigma$ (decreasing $\sigma$) if it is of type $NW$ and $$\sigma({\operatorname{Stop}}(R))< A^{-1}\,\sigma(\{R\}).$$ On the other hand, we say that ${{\mathcal T}}$ is of type $S\sigma$ (stable $\sigma$) if it is of type $NW$ and $$A^{-1}\,\sigma(\{R\})\leq\sigma({\operatorname{Stop}}(R))\leq A\,\sigma(\{R\}).$$ Our analysis is going to need more complex trees, in fact, we will use three families of trees: the family of maximal decreasing trees $MDec$, the family of large wonderful trees denoted by $LW$ and the family of tame increasing trees $TInc$. Let us describe these three families in turn. [**(1)**]{} Given a $D\sigma$ simple tree ${{\mathcal T}}={\operatorname{Tree}}(R)$ we construct an $MDec-enlarged$ tree $\widetilde{{{\mathcal T}}}$ iteratively as follows. We set $$\widetilde{{{\mathcal T}}}_{1}:= {{\mathcal T}}\cup \bigcup_{\substack{P \in {\operatorname{Stop}}({{\mathcal T}})\\ {\operatorname{Tree}}(P) \text{ is } D\sigma }} {\operatorname{Tree}}(P).$$ We also define the stopping collection of the new tree as $${\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{1}):= \{P\in{\operatorname{Stop}}({{\mathcal T}}):\text{${\operatorname{Tree}}(P)$ is not $D\sigma$}\} \cup \!\!\bigcup_{\substack{P \in {\operatorname{Stop}}({{\mathcal T}})\\ {\operatorname{Tree}}(P) \text{ is } D\sigma }}\!\! {\operatorname{Stop}}(P),$$ where ${\operatorname{Stop}}({{\mathcal T}})\equiv{\operatorname{Stop}}(R)$. In general, for an integer $k\geq 2$, we define $$\widetilde{{{\mathcal T}}}_{k}:=\widetilde{{{\mathcal T}}}_{k-1}\cup \bigcup_{\substack{P \in {\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1})\\ {\operatorname{Tree}}(P) \text{ is } D\sigma }} {\operatorname{Tree}}(P),$$ and $${\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k}):= \{P\in{\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1}):\text{${\operatorname{Tree}}(P)$ is not $D\sigma$}\} \cup \!\!\bigcup_{\substack{P \in {\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1})\\ {\operatorname{Tree}}(P) \text{ is } D\sigma }} \!\!{\operatorname{Stop}}(P).$$ Our $MDec-enlarged$ tree $\widetilde{{{\mathcal T}}}$ will be the maximal union of $D\sigma$ trees, that is, $$\widetilde{{{\mathcal T}}}:=\bigcup_{k\geq 1} \widetilde{{{\mathcal T}}}_{k}$$ [**(2)**]{} Given a simple tree ${{\mathcal T}}={\operatorname{Tree}}(R)$ that is either $I\sigma$ or $S\sigma$, the following algorithm will return the desired $LW$ or $TInc$ tree. We first define the tree ${{\widetilde{{{\mathcal T}}}}}_{2}$ as $${{\widetilde{{{\mathcal T}}}}}_{2}:= {{\mathcal T}}\cup \bigcup_{P\in HD(R)}{\operatorname{Tree}}(P).$$ We define the stopping collection of the new tree ${{\widetilde{{{\mathcal T}}}}}_{2}$ as $${\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{2}):=\bigl({\operatorname{Stop}}({{\mathcal T}})\setminus HD(R)\bigr) \cup \bigcup_{P\in HD(R)} {\operatorname{Stop}}(P).$$ The collections of HD, LD and BR cubes are, respectively, $$HD_{2}:=HD({{\widetilde{{{\mathcal T}}}}}_{2}):= \bigcup_{P\in HD(R)} HD(P),$$ $$LD_{2}:=\bigcup_{P\in HD(R)} LD(P) \quad \text{ and }\quad BR_{2}:=\bigcup_{P\in HD(R)} BR(P).$$ In general, for an integer $k\geq 2$, if ${{\widetilde{{{\mathcal T}}}}}_{k-1}$ has already been defined and moreover the following two conditions hold: $$\sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{k}))> A \sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{k-1})) \quad\text{ and }\quad \mu(BR_{k})\leq \delta_{W}' \mu(HD_{k-1}),$$ then we define $$\widetilde{{{\mathcal T}}}_{k}:=\widetilde{{{\mathcal T}}}_{k-1}\cup \bigcup_{P\in HD_{k-1}} {\operatorname{Tree}}(P),$$ and $$\begin{aligned} {\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k})& := \{P\in{\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1})\setminus HD_{k-1}\} \cup \!\!\bigcup_{P\in HD_{k-1}} \!\!{\operatorname{Stop}}(P).\end{aligned}$$ Moreover, we denote $$HD_{k}:=\!\bigcup_{P\in HD_{k-1}} \!\!HD(P), \quad LD_{k}:=\!\bigcup_{P\in HD_{k-1}}\!\!LD(P)\quad \text{ and }\quad BR_{k}:=\!\bigcup_{P\in HD_{k-1}}\!\! BR(P).$$ We stop the algorithm when we reach one of the following conditions for some $k=N_0$: 1. If $$\label{e.stoplw} \mu(BR_{N_{0}})> \delta_{W}' \mu(HD_{N_{0}-1})$$ for certain $N_{0}\geq 2$, in which case we say that ${{\widetilde{{{\mathcal T}}}}}_{N_{0}}$ belongs to the $LW$ family and we say that $N_{0}$ is its order. 2. Or if $$\label{e.tame} \sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{N_{0}}))\leq A \sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{N_{0}-1}))$$ for certain $N_{0}\geq 2$. The boundedness of densities ensures that such condition must be reached at some $N_{0}$. Then ${{\widetilde{{{\mathcal T}}}}}_{N_{0}}$ belongs to the $TInc$ family and we say that $N_{0}$ is its order. The starting cube $R$ in the construction of a tree ${{\widetilde{{{\mathcal T}}}}}$ of type $MDec$, $TInc$, $LW$, or $W$ is called [*root of ${{\widetilde{{{\mathcal T}}}}}$*]{}, and we write $R={\operatorname{Root}}({{\widetilde{{{\mathcal T}}}}})$. We say that the trees of type $MDec$, $TInc$, $LW$, and $W$ are [*maximal trees*]{}. Note that the trees of type $W$ are simple. Maximal trees of type $MDec$ can be simple or non-simple. On the other hand, trees of type $TInc$ and $LW$ are non-simple. Estimates for the trees of type $W$ and $LW$ -------------------------------------------- Throughout this section we will use that the dyadic density $\Theta_{d}(Q)$ is constant for all cubes $Q\in HD_{k}$ for a fixed $k\in\mathbb N$ and we will denote such constant by $\Theta(HD_{k})$. In order to prove the desired estimates for trees of type $LW$ and $TInc$ we need the following lemma: \[l.sigmatinc\] Let ${{\widetilde{{{\mathcal T}}}}} \in TInc\cup LW$ be of order $N_{0}$. Then $$\sigma({{\widetilde{{{\mathcal T}}}}})\leq C(A,B, \delta_{0}, \delta_{W}) \sigma(HD_{N_{0}-1}).$$ First, we observe that by Lemma \[lemdif1\], $$\sigma({{\widetilde{{{\mathcal T}}}}}) \leq C(A,B, \delta_{0}) \sum_{k=0}^{N_{0}-1}\,\sigma(HD_{k}),$$ where, by convenience, we set $HD_{0}:=R$, where $R$ is the root of ${{\widetilde{{{\mathcal T}}}}}$. The lemma follows now trivially by the geometrically increasing nature of $\sigma(HD_{k})$. By iteration, it is enough to prove that for $2\leq k\leq N_{0}-1$ $$\label{e.dechd} \sigma(HD_{k-1})\leq \frac2A\, \sigma(HD_{k}),$$ recalling that $2A^{-1} <1$. Notice also that for $k=1$ $$\label{e.uglycontrol} \sigma(R)\leq 2A\sigma(HD_{1}).$$ Since $2\leq k\leq N_{0}-1$, by the definition of $TInc$ and $LW$ tree ${{\widetilde{{{\mathcal T}}}}}$, we have that $\sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{k}))>A\,\sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{k-1}))$. Thus $$\begin{aligned} \label{eqfa520} \sigma(HD_{k})+ \sigma(BR_{k}) +\sigma(LD_{k}) &= \sigma({\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k})\setminus {\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1}))\\ &\geq (A-1)\,\sigma({\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{k-1}))\geq (A-1)\,\sigma(HD_{k-1}).\nonumber\end{aligned}$$ On the other hand for $k=1$, ${{\widetilde{{{\mathcal T}}}}}_{1}$ is of either type $I\sigma$ or $S\sigma$, then we have $$\label{e.hd1} \sigma({\operatorname{Stop}}(\widetilde{{{\mathcal T}}}_{1}))\geq A^{-1}\sigma(R).$$ We also have that for all $k\geq 1$ $$\sigma(LD_{k})\leq \delta_{0}^{2}\,2\Theta(HD_{k-1})^{2}\,\mu(LD_{k})\leq 2 \delta_{0}^{2}\sigma(HD_{k-1})$$ and using the fact that all the $BR$ cubes in our tree ${{\widetilde{{{\mathcal T}}}}}$ have little mass, we obtain $$\sigma(BR_{k}) \leq 4B^{2}\Theta(HD_{k-1})^{2}\mu(BR_{k}) \leq 4\delta_{W}B^{2}\sigma(HD_{k-1}),$$ assuming $\delta_{W}\ll (2B)^{-2}$. So $\sigma(BR_{k}) +\sigma(LD_{k})\ll\sigma(HD_{k-1})$. If moreover we assume $\delta_{W}\ll (2B)^{-2}A^{-1}$ and $\delta_{0}\ll A^{-1}$, together with and , this implies that for all $k\geq 1$ $$\label{eqqq234} \sigma(BR_{k}) +\sigma(LD_{k})\leq \frac1{100} \sigma(HD_{k})$$ assuming $A$ big enough, and again by and , one deduces and . This concludes the proof. We are now ready to estimate $W$ and $LW$ trees. \[lemtw\] If ${{\mathcal T}}$ is of type $W$ or $LW$, then $$\sigma({{\mathcal T}})\leq C(A,B,\delta_0,M,\delta_W)\, \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.$$ We prove the case of a $LW$ tree of order $N_{0}\geq 1$, as the case of a $W$ follows with $N_{0}=1$. We claim that it is enough to prove the following estimate: $$\label{e.wond} \sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)}\sum_{Q\subsetneq P \subset R} \|D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 \geq C(A,B, \delta_0,M, \delta_{W}) \sigma(HD_{N_{0}-1}).$$ In fact, from and Lemma \[l.sigmatinc\] we see that $$\begin{aligned} \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 &\geq \sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)}\sum_{Q\subsetneq P \subset R} \|D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2\\ & \geq C(A, B,\delta_0, M, \delta_{W})\sigma(HD_{N_{0}-1}) \geq \widetilde{C}(A, B, M, \delta_{0}, \delta_{W}) \sigma({{\mathcal T}}),\end{aligned}$$ which gives exactly the desired estimate. Let us proceed with the proof of : $$\sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)}\sum_{Q\subsetneq P \subset R} \|D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 = \sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)} \| \sum_{Q\subsetneq P \subset R} D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.$$ For $R,Q$ as in the last sum we have $$\Bigl|\chi_Q\sum_{Q\subsetneq P \subset R} D_P({{\mathcal R}}\mu)\Bigr| = \bigl|m_{Q}({{\mathcal R}}\mu) - m_{R}({{\mathcal R}}\mu)\bigr| \geq \frac M2\,\Theta(R),$$ by Lemma \[lemt0\] (b). So we get $$\begin{aligned} \sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)}\sum_{Q\subsetneq P \subset R} \|D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 & \geq \sum_{R\in HD_{N_{0}-1}}\sum_{Q\in BR(R)} \frac{M^2}{4} \Theta(R)^2 \mu(Q)\\ & \gtrsim \sum_{R\in HD_{N_{0}-1}} \delta_{W}\frac{M^2}{4} \Theta(HD_{N_{0}-1})^2 \mu(R) \\ &\gtrsim \delta_{W}\frac{M^2}{4} \sigma(HD_{N_{0}-1}),\end{aligned}$$ using the condition in the second inequality. Initial estimates for $TInc$ trees ---------------------------------- We want to prove the following. \[lemtame\] Let ${{\mathcal T}}$ be a $TInc$ tree. Then $$\sigma({{\mathcal T}})\leq C(A,B,M,\delta_{0}, \delta_{W}) \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.$$ The proof requires a deep analysis that will be mostly carried in section 7. We prove some preliminary estimates below. Let us consider a $TInc$ tree ${{\mathcal T}}$ of order $N_0$. By it follows that $$\sigma(HD_{N_0-1})\geq \frac{A}{2} \,\sigma(HD_{N_0-2})\qquad \mbox{for $N_0>2$.}.$$ Taking also into account, identifying $HD_{N_0-2}$ with the root of ${{\mathcal T}}$ when $N_0=2$, we deduce that $$\label{eqdsa28} \sigma(HD_{N_0-1})\geq \frac{1}{2A} \,\sigma(HD_{N_0-2}) \qquad \mbox{for $N_0\geq2$.}$$ Using also and the stopping condition we get $$\label{eqdsa29} \sigma(HD_{N_0-1})\geq \frac12\,\sigma({\operatorname{Stop}}({{\widetilde{{{\mathcal T}}}}}_{N_0-1})) \geq \frac1{2A}\, \bigl (\sigma(HD_{N_0}) + \sigma(BR_{N_0}) + \sigma(LD_{N_0})\bigr).$$ Let ${{\mathcal G}}$ be the collection of those cubes $R\in HD_{N_0-2}$ such that $$\label{eqdefg22} \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P)\geq \max\Biggl(\frac{1}{20A}\,\sigma(R)\;,\;\; \frac1{10A} \sum_{\substack{L\in HD_{N_0}\cup BR_{N_0}\cup LD_{N_0}:\\L\subset R}} \sigma(L)\Biggr).$$ \[l.redtolast\] Let ${{\mathcal T}}$ be a $TInc$ tree and ${{\mathcal G}}$ as above. We have $$\sigma({{\mathcal G}})\geq C(B)\,\sigma(HD_{N_0-1}).$$ Let ${{\mathcal B}}= HD_{N_0-2}\setminus {{\mathcal G}}$. Let ${{\mathcal B}}_1$ the collection of cubes $R\in HD_{N_0-2}$ such that $$\frac{1}{20A}\,\sigma(R) > \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P)$$ and set ${{\mathcal B}}_2={{\mathcal B}}\setminus {{\mathcal B}}_1$. Thus the cubes from ${{\mathcal B}}_2$ belong to $HD_{N_0-2}$ and satisfy $$\sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P) < \frac1{10A} \sum_{\substack{L\in HD_{N_0}\cup BR_{N_0}\cup LD_{N_0}:\\L\subset R}} \sigma(L).$$ From the definition of ${{\mathcal B}}_1$ and we get $$\sum_{R\in{{\mathcal B}}_1} \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P) \leq \frac1{20A}\,\sum_{R\in{{\mathcal B}}_1}\sigma(R)\leq \frac 1{20A} \,\sigma(HD_{N_0-2}) \leq \frac 1{10}\, \sigma(HD_{N_0-1}).$$ On the other hand, from the definition of ${{\mathcal B}}_2$ and , $$\begin{aligned} \sum_{R\in{{\mathcal B}}_2} \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P) & \leq \frac{1}{10A} \sum_{R\in{{\mathcal B}}_2}\,\, \sum_{\substack{L\in HD_{N_0}\cup BR_{N_0}\cup LD_{N_0}:\\P\subset R}} \sigma(L) \\ & \leq \frac{1}{10A} \,\bigl (\sigma(HD_{N_0}) + \sigma(BR_{N_0})+ \sigma(LD_{N_0})\bigr) \leq \frac{1}{5}\,\sigma(HD_{N_0-1}).\end{aligned}$$ Therefore, $$\sum_{R\in{{\mathcal B}}} \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P) \leq \frac{3}{10}\,\sigma(HD_{N_0-1}),$$ and so $$\sum_{R\in{{\mathcal G}}} \sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P) \geq \frac{7}{10}\,\sigma(HD_{N_0-1}).$$ Observe now that each $R\in HD_{N_0-2}$ (in particular each $R\in{{\mathcal G}}$) satisfies $$\sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}} \sigma(P)\leq c\,B^2\,\sigma(R).$$ So we deduce $$\sigma({{\mathcal G}}) =\sum_{R\in{{\mathcal G}}} \sigma(R) \geq \frac1{B^2}\sum_{R\in{{\mathcal G}}} \,\sum_{\substack{P\in HD_{N_0-1}:\\P\subset R}}\! \sigma(P) \geq \frac{7c^{-1}}{10B^2}\,\sigma(HD_{N_0-1}).$$ For the proof of Lemma \[lemtame\], notice that by Lemmas \[l.sigmatinc\] and \[l.redtolast\] we have $$\begin{aligned} \sigma({{\mathcal T}}) &\leq C(A,B)\,\sigma(HD_{N_0-1})\leq {{\widetilde{C}}} (A,B)\,\sigma({{\mathcal G}}).\end{aligned}$$ We will show below that $$\label{e.redtolast} \sigma({{\mathcal G}})\leq C(A,B, M, \delta_{0}, \delta_{W}) \sum_{R\in {{\mathcal G}}} \sum_{P\in{{\mathcal T}}(R)} \|D_P({{\mathcal R}}\mu)\|_{L^2(\mu)}^2,$$ where ${{\mathcal T}}(R)$ is the tree formed by the cubes from ${{\mathcal T}}$ which are contained in $R$. Observe that the proof of Lemma \[lemtame\] is complete if we show . However the arguments involved are very delicate and so it will be addressed in Section 7. We introduce some necessary notation next. We are going to define a tractable tree. The reader may think of it as a $TInc$ tree of order $2$ that involves different constants in the stopping conditions. \[d.tract\] Let $R\in{\operatorname{Top}}$ be a cube. We say that ${{\mathcal T}}$ is a tractable tree with root $R$ if it is a collection of cubes of the form $${{\mathcal T}}:= {\operatorname{Tree}}(R)\cup \bigcup_{P\in HD(R)} {\operatorname{Tree}}(P)$$ satisfying the following conditions: $$\begin{aligned} \label{e.1stop} \frac{1}{20A}\sigma(R)\leq & \sigma(HD_{1}),\\ \label{e.2stop} \sigma(HD_{2}\cup BR_{2} \cup LD_{2})\leq &10A\sigma(HD_{1}),\\ \label{e.sbr} \sigma(BR_{1})\leq \delta_{W} \sigma(R)\; \text{ and } &\;\sigma(BR_{2})\leq \delta_{W} \sigma(HD_{1}),\end{aligned}$$ where the notation above is analogous to the one for $TInc$ trees, that is, $HD_{1}=HD(R)$, $LD_{1}=LD(R)$ and $BR_1=BR(R)$, $HD_{2}= \bigcup_{P\in HD(R)} HD(P)$, $LD_{2}=\bigcup_{P\in HD(R)} LD(P)$ and $BR_{2}=\bigcup_{P\in HD(R)} BR(P)$, ${\operatorname{Stop}}_{1}({{\mathcal T}})=HD_{1}\cup LD_{1}\cup BR_1$, and ${\operatorname{Stop}}_{2}({{\mathcal T}})=LD_{1}\cup BR_1\cup HD_{2}\cup LD_{2}\cup BR_2$. Tractable trees are essentially $TInc$ trees of order 2, in particular they satisfy $$\begin{aligned} \label{e.is} \sigma({\operatorname{Stop}}_{1}({{\mathcal T}})) & \gtrsim A^{-1}\sigma(R)\\ \label{e.ss} \sigma({\operatorname{Stop}}_{2}({{\mathcal T}})) & \lesssim A\sigma({\operatorname{Stop}}_{1}({{\mathcal T}})).\end{aligned}$$ Notice that given a $TInc$ tree ${{\widetilde{{{\mathcal T}}}}}$ and $Q\in {{\mathcal G}}$, then ${{\mathcal T}}(Q)$ is a tractable tree. We will study tractable trees in Section 7. Some Fourier calculus and a maximum principle ============================================= In this section we address some auxiliary questions which will be needed for the study of the tractable trees in the next section. The Fourier transform of the kernel $\dfrac{x}{|x|^{s+1}}$ is $c\,\dfrac\xi{|\xi|^{d-s+1}}$. Thus, if ${{\varphi}}$ is a ${{\mathcal C}}^\infty$ function which is compactly supported in ${{\mathbb R}}^d$, the function $\psi:{{\mathbb R}}^d\to{{\mathbb R}}^d$ whose Fourier transform is $${{\widehat{\psi}}}(\xi) = c\,|\xi|^{d-s-1}\,\xi\,{{\widehat{{{\varphi}}}}}(\xi)$$ satisfies ${{\mathcal R}}(\psi\,d{{\mathcal L}}^d)={{\varphi}}$, where $c$ is some appropriate constant. Notice that $${{\mathcal R}}(\psi\,{{\mathcal L}}^d)(x)=\int \frac{x-y}{|x-y|^{s+1}}\cdot \psi(y)\,d{{\mathcal L}}^d(y),$$ where the dot stands for the scalar product in ${{\mathbb R}}^d$. One can easily check that, although $\psi$ is not compactly supported, it is quite well localized, in the sense that it satisfies $$\label{equas1} |\psi(x)|\leq \frac{C_{{{\varphi}},\alpha}}{1+|x|^{d+\alpha}},$$ for $0\leq\alpha<d-s$. In the next section we will need to consider a $C^\infty$ function ${{\varphi}}_0$ which equals $1$ on the ball $B(0,1)$ and vanishes out of $B(0,1.1)$, say. Given a cube $Q\in{{\mathcal D}}$ and some point $z_Q\in Q$ that will be fixed below we denote ${{\varphi}}_Q = {{\varphi}}_0\circ T_Q$, where $T_Q$ is the affine map that maps $B(0,1)$ to $B(x_Q,\frac1{10}c_{sep}\ell(Q))$. We also denote by $\psi_Q$ a vectorial function such that ${{\mathcal R}}(\psi_Q\,{{\mathcal L}}^d)={{\varphi}}_Q$. That is, $${{\widehat{\psi}}}_Q(\xi) = c\,|\xi|^{d-s-1}\,\xi\,{{\widehat{{{\varphi}}}}}_Q(\xi).$$ It is straightforward to check that $\psi_Q = \ell(Q)^{s-d}\,\psi_0\circ T_Q$, where $\psi_0$ is given by ${{\mathcal R}}(\psi_0\,{{\mathcal L}}^d)={{\varphi}}_0$. Then it follows that $$\label{equas2} \|\psi_Q\|_1= \ell(Q)^{s-d}\,\|\psi_0\circ T_Q\|_1 = C_{1}\,\ell(Q)^s,$$ where $C_{1}=\|\psi_0\|_1$. Also from , $$\label{equas10} |\psi_Q(x)|\leq \frac{C_{{{\varphi}},\alpha}\,\ell(Q)^{s-d}}{1+\ell(Q)^{-d-\alpha}|x-z_Q|^{d+\alpha}} = \frac{C_{{{\varphi}},\alpha}\,\ell(Q)^{s+\alpha}}{\ell(Q)^{d+\alpha}+|x-z_Q|^{d+\alpha}}$$ for $0\leq\alpha<d-s$. \[lemmaxprin\] Let $\nu$ be a signed compactly supported Borel measure which is absolute continuous with respect to Lebesgue measure, and suppose that its density is an $L^\infty$ function. Let $h:{{\mathbb R}}^d\to{{\mathbb R}}^d$ be an $L^\infty$ vector field. Let $1<r<\infty$ and $a>0$. If $$|{{\mathcal R}}\nu(x)|^r + {{\mathcal R}}^*(h\,\nu)(x) \leq a\qquad\mbox{for all $x\in{\operatorname{supp}}\nu$},$$ then the same inequality hods in all ${{\mathbb R}}^d$. That is, $$\label{eqaaa16} |{{\mathcal R}}\nu(x)|^r + {{\mathcal R}}^*(h\,\nu)(x) \leq a\qquad\mbox{for all $x\in{{\mathbb R}}^d$},$$ where ${{\mathcal R}}^*$ is the operator dual to ${{\mathcal R}}$. Almost the same arguments as the ones in [@ENV12 Section 17] can be applied to prove this lemma. For the sake of completeness we give the detailed proof below. We will use the fact that for any $L^\infty$ vector field $f$, the maximum of ${{\mathcal R}}^*(f\,\nu)$ is attained in ${\operatorname{supp}}\nu$ if $\sup_{x\in{{\mathbb R}}^d}{{\mathcal R}}^*(f\,\nu)>0$. The fact that the maximum exists in this case is due to the fact that ${{\mathcal R}}^*(f\,\nu)$ is continuous in ${{\mathbb R}}^d$ and vanishes at $\infty$. On the other hand, in [@ENV12] it is shown that the maximum is attained on ${\operatorname{supp}}\nu$ if $f\,\nu$ has a ${{\mathcal C}}^\infty$ density with respect to the Lebesgue measure. However the same holds if the density is just $L^\infty$ and compactly supported. This can be deduced from a limiting argument that we omit. Let us turn our attention to $|{{\mathcal R}}\nu|^r + {{\mathcal R}}^*(h\,\nu)$ now. Again this is a continuous function vanishing at $\infty$ and thus the maximum is attained if its supremum in ${{\mathbb R}}^d$ is positive. To show that this is attained in ${\operatorname{supp}}\nu$ we linearize the problem in the following way. First notice that for any $y\in{{\mathbb R}}^d$ and $1<r<\infty$, $$\frac1r \,|y|^r = \sup_{\lambda\geq0,|e|=1} \lambda\,\langle e,y\rangle - \frac1{r'}\,\lambda^{r'},$$ where $r'=r/(r-1)$. This follows either from elementary methods or from the fact that $\frac1{r'}\,\lambda^{r'}$ is the Legendre transform of $F(t)=\frac1r \,t^r$. Then we have $$\begin{aligned} |{{\mathcal R}}\nu(x)|^r + {{\mathcal R}}^*(h\,\nu)(x)& = \sup_{\lambda\geq0,|e|=1} r\,\lambda\,\langle e,{{\mathcal R}}\nu(x)\rangle - \frac r{r'}\,\lambda^{r'} + {{\mathcal R}}^*(h\,\nu)(x)\\ & = \sup_{\lambda\geq0,|e|=1} -{{\mathcal R}}^*(r\,\lambda\,e\,\nu)(x) - \frac r{r'}\,\lambda^{r'} + {{\mathcal R}}^*(h\,\nu)(x)\\ & = \sup_{\lambda\geq0,|e|=1} {{\mathcal R}}^*(h\,\nu -r\,\lambda\,e\,\nu)(x) - \frac r{r'}\,\lambda^{r'}. \end{aligned}$$ As mentioned above, for each $\lambda>0$ and each $e$ with $|e|=1$, the maximum of ${{\mathcal R}}^*(h\,\nu -r\,\lambda\,e\,\nu)$ is attained in ${\operatorname{supp}}\nu$ if $\sup_{x\in{{\mathbb R}}^d}{{\mathcal R}}^*(h\,\nu -r\,\lambda\,e\,\nu)(x)>0$. Thus the maximum of $|{{\mathcal R}}\nu|^r + {{\mathcal R}}^*(h\,\nu)$ is attained in ${\operatorname{supp}}\nu$ whenever $$\label{eqaaa17} \sup_{x\in{{\mathbb R}}^d}|{{\mathcal R}}\nu(x)|^r + {{\mathcal R}}^*(h\,\nu)(x)>0.$$ The lemma is an immediate consequence of this statement. Indeed, notice that when proving one can assume that holds, since $a>0$. The tractable trees =================== This section is devoted to the proof of the key estimate . In fact, this is a consequence of the following lemma. \[l.tract\]\[lemtrac\] Let ${{\mathcal T}}$ be a tractable tree and $R$ its root. Then $$\sigma(\{R\})\leq C(A,B,M,\delta)\, \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.$$ Before turning to the arguments for the proof of the preceding result we show how Lemma \[lemtame\] follows from this. Recall that we have to prove that, if ${{\mathcal T}}$ be a $TInc$ tree, then $$\sigma({{\mathcal T}})\leq C(A,B,M,\delta_{0}, \delta_{W}') \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.$$ By Lemmas \[l.sigmatinc\] and \[l.redtolast\] we have $$\begin{aligned} \sigma({{\mathcal T}}) &\leq C(A,B)\,\sigma(HD_{N_0-1})\leq C(A,B)\,\sigma({{\mathcal G}}).\end{aligned}$$ where ${{\mathcal G}}$ is the family of cubes from $HD_{N_0-2}$ defined in . For $R\in{{\mathcal G}}$, it turns out that ${{\mathcal T}}(R)$, the family of cubes from ${{\mathcal T}}$ contained in $R$, is a tractable tree. Thus, by Lemma \[l.tract\] we have $$\begin{aligned} \sigma({{\mathcal G}})= \sum_{R\in{{\mathcal G}}}\sigma(R) &\leq C(A,B,M,\delta)\, \sum_{R\in{{\mathcal G}}}\sum_{Q\in{{\mathcal T}}(R)}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2\\ &\leq C(A,B,M,\delta)\, \sum_{Q\in{{\mathcal T}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2.\end{aligned}$$ We introduce now some additional notation. As in section 5, we use the notation $\Theta(HD_{k})$ for the dyadic density $\Theta_{d}(Q)$ constant for all cubes $Q\in HD_{k}$, $k\in \mathbb N$. We consider the following approximating measure for $\mu{{\lfloor}}R$: $$\label{e.approxm} \eta = \sum_{P\in{\operatorname{Stop}}_2({{\mathcal T}})} \frac{\mu(P)}{{{\mathcal L}}^d(P)}\, {{\mathcal L}}^d{{\lfloor}}P.$$ Recall that, by assumption, ${{\mathcal L}}^d(P)\approx\ell(P)^d$. In order to compare Riesz transforms with respect to $\mu$ and $\eta$, it is convenient to introduce some coefficients $q(\cdot)$: for $Q\in{{\mathcal T}}$, we set $$q(Q,{{\mathcal T}})=\sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})} \frac{\ell(P)\,\mu(P)}{D(P,Q)^{s+1}},$$ where $$D(P,Q) = \ell(P)+{\operatorname{dist}}(P,Q) + \ell(Q).$$ Notice that the coefficients $q(Q,{{\mathcal T}})$ depend on $Q$ and on the cubes from ${\operatorname{Stop}}_{2}({{\mathcal T}})$. \[lemdifer\] For every $Q\in{{\mathcal T}}$ and $x,x'\in Q$, $$|{{\mathcal R}}(\chi_{R\setminus Q} \,\mu)(x) - {{\mathcal R}}(\chi_{R\setminus Q} \,\eta)(x')|\lesssim p(Q,R)+ q(Q,{{\mathcal T}}).$$ For each $P\in{\operatorname{Stop}}_{2}({{\mathcal T}})$ let $z_P\in P$ be some fixed point. Taking into account that $\mu(P)=\eta(P)$ for each such cube, for $x\in Q$, we get $$\begin{aligned} {{\mathcal R}}(\chi_{R\setminus Q} \,\mu)(x) - {{\mathcal R}}&(\chi_{R\setminus Q} \,\eta)(x) \\ & = \sum_{\substack{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}):\,P\not\subset Q}}\left( \int_P K^s(x-y)\,d\mu(y) - \int_P K^s(x-y)\,d\eta(y)\right) \\ &=\sum_{\substack{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}):\, P\not\subset Q}} \int_P \bigl[K^s(x-y) - K^s(x-z_P)\bigr]\,d\mu(y)\\ &\quad - \sum_{\substack{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}):\, P\not\subset Q}}\int_P \bigl[K^s(x-y) - K^s(x-z_P)\bigr]\,d\eta(y).\end{aligned}$$ For $x$ and $y$ as in the preceding integrals we have $$\bigl|K^s(x-y) - K^s(x-z_P)\bigr|\lesssim \frac{\ell(P)}{D(P,Q)^{s+1}}.$$ Then we deduce that $$|{{\mathcal R}}(\chi_{R\setminus Q} \,\mu)(x) - {{\mathcal R}}(\chi_{R\setminus Q} \,\eta)(x)|\lesssim q(Q,{{\mathcal T}}).$$ The lemma follows by combining this estimate with the fact that for $x,x'\in P$, $$|{{\mathcal R}}(\chi_{R\setminus Q} \,\mu)(x) - {{\mathcal R}}(\chi_{R\setminus Q} \,\mu)(x')|\lesssim p(Q,R).$$ \[lemdifer22\] For $1\leq r<\infty$, we have $$\sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})} q(P,{{\mathcal T}})^r \mu(P)\lesssim c_{6}\,\sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})}p(P,R)^r\,\mu(P).$$ Let us consider the functions $p$ and $q$ defined on $R$ by $p(x):=p(P_{x})$ and $q(x):=q(P_{x}, {{\mathcal T}})$, where $P_{x}$ is the cube in ${\operatorname{Stop}}_{2}({{\mathcal T}})$ that contains $x$. We want to prove $\|q\|_{L^{r}}\leq c \|p\|_{L^{r}}$. Let us fix $\lambda>0$ and consider the set $\Omega_{\lambda}:=\{x\in {{\mathbb R}}^d: \, q(x)>\lambda \}$. Take a Whitney decomposition of $\Omega_{\lambda}$ into a family of “true cubes” $W(\Omega_{\lambda})$. In particular, we assume that the cubes $L\in W(\Omega_{\lambda})$ satisfy $${\operatorname{dist}}(3L,{{\mathbb R}}^d\setminus \Omega_\lambda)\approx\ell(L).$$ We claim that there exists an absolute constant $C>1$ such that $$\label{e.goodl} \mu(\{x\in {{\mathbb R}}^d: \,\, q(x)>C\lambda \}) \leq \sum_{L\in W(\Omega_{\lambda})} \mu\Bigl(\Bigl\{x\in L: \,\, \sum_{\substack{Q\in {\operatorname{Stop}}_{2}({{\mathcal T}})\\ Q\subset 3L}}\frac{\mu(Q)\ell(Q)}{D(P_{x},Q)^{s+1}}>\lambda \Bigl\}\Bigr).$$ We first show that the lemma follows from . By Chebychev, we deduce $$\begin{aligned} \mu(\{x\in {{\mathbb R}}^d: \,\, q(x)>C\lambda \}) & \leq \sum_{L\in W(\Omega_{\lambda})} \frac{1}{\lambda} \int_{L} \sum_{\substack{Q\in {\operatorname{Stop}}_{2}({{\mathcal T}})\\ Q\subset 3L}} \frac{\mu(Q)\ell(Q)}{D(P_{x},Q)^{s+1}} d\mu(x)\\ &= \sum_{L\in W(\Omega_{\lambda})} \frac{1}{\lambda} \int_{L}\int_{3L} \frac{\ell(P_{y})}{D(P_{x},P_{y})^{s+1}}d\mu(y) d\mu(x).\end{aligned}$$ By Fubini, the last double integral is bounded by $$\begin{aligned} \int_{3L} &\ell(P_{y})\int_{L} \frac{d\mu(x)}{\left( \ell(P_{y})+|x-y|\right)^{s+1}}\,d\mu(y)\\ &\lesssim \int_{3L} \ell(P_{y}) \sum_{\substack{Q\in \mathcal D\\ Q\supseteq P_{y}}}\left(\int_{ Q^1\setminus Q} \frac{d\mu(x)}{\left(\ell(P_{y})+|x-y| \right)^{s+1}}+ \int_{P_{y}} \frac{d\mu(x)}{\left(\ell(P_{y})+|x-y| \right)^{s+1}}\right)\,d\mu(y)\\ & \lesssim \int_{3L} \sum_{\substack{Q\in \mathcal D\\ Q\supseteq P_{y}}} \frac{\ell(P_{y})\mu(Q)}{\ell(Q)^{s+1}}\,d\mu(y) \lesssim\int_{3L} p(y)d\mu(y),\end{aligned}$$ where $Q^1$ stands for the parent cube of $Q$. Thus we get $$\mu(\{x\in {{\mathbb R}}^d: \,\, q(x)>C\lambda \}) \lesssim \frac{1}{\lambda} \int_{\Omega_{\lambda}}p(y)d\mu(y).$$ We use this estimate to get the desired result: $$\begin{aligned} \|q\|_{L_{r}}^{r} & \lesssim \int_{0}^{\infty} \lambda^{r-1} \mu(\{x\in {{\mathbb R}}^d: \,\, q(x)>C\lambda \}) d\lambda\\ & \lesssim \int_{0}^{\infty} \lambda^{r-2} \int_{\Omega_{\lambda}}p(x) d\mu(x) d\lambda\\ & = \int p(x) \int_{0}^{q(x)} \lambda^{r-2}d\lambda d \mu(x)\\ & \approx \int p(x) q(x)^{r-1}d\mu(x)\\ &\leq \left( \int p(x)^{r}d\mu(x)\right)^{1/r}\left( \int q(x)^{r'(r-1)} d\mu(x) \right)^{1/r'}= \|p\|_{L_{r}}\|q\|_{L_{r}}^{r/r'},\end{aligned}$$ where $r'$ is the dual exponent of $r$. So we obtain $\|q\|_{L_{r}}\lesssim \|p\|_{L_{r}}$, as wished. We are only left with the proof of . For $x\in \Omega_{\lambda}$, let $L \in W(\Omega_{\lambda})$ be such that $x\in L$. Take $x'\in \partial \Omega_{\lambda}$ such that $|x-x'|\approx \ell(L)$. As $x'\not\in\Omega_\lambda$, $$q(x')=\sum_{Q\in {\operatorname{Stop}}_{2}{{\mathcal T}}} \frac{ \mu(Q)\ell(Q)}{D(P_{x'}, Q)^{s+1}}\leq \lambda.$$ We split $q(x)$ in two summands, the local part and the non-local one: $$q(x)= \sum_{\substack{Q\in {\operatorname{Stop}}_{2}{{\mathcal T}}\\ Q\subset 3L}} \frac{ \mu(Q)\ell(Q)}{D(P_{x}, Q)^{s+1}} + \sum_{\substack{Q\in {\operatorname{Stop}}_{2}{{\mathcal T}}\\ Q\not\subset 3L}} \frac{ \mu(Q)\ell(Q)}{D(P_{x}, Q)^{s+1}}.$$ Observe that the local part is the one that appears on the right hand side of . Thus to prove the claim it is enough to show that $$\label{e.lclaim} \sum_{\substack{Q\in {\operatorname{Stop}}_{2}{{\mathcal T}}\\ Q\not\subset 3L}} \frac{ \mu(Q)\ell(Q)}{D(P_{x}, Q)^{s+1}}\leq C_{nl}\sum_{\substack{Q\in {\operatorname{Stop}}_{2}{{\mathcal T}}\\ Q\not\subset 3L}} \frac{ \mu(Q)\ell(Q)}{D(P_{x'}, Q)^{s+1}},$$ where $C_{nl}$ is some constant depending at most on $d$, $s$, and $c_{sep}$. We will then take $C:= C_{nl} +1$ in . To prove , observe that the numerators are the same in both sides. So it suffices to show that $$\label{eqdd392} D(P_{x'},Q)\leq C\, D(P_{x},Q) \quad \mbox{ when $Q\not\subset 3L$.}$$ First notice that, by the separation condition , if $P_{x'}\neq Q$, then $\ell(P_{x'})\leq c{\operatorname{dist}}(P_{x'},Q)$. If $P_{x'}= Q$, then obviously $\ell(P_{x'})= \ell(Q)$. So in any case $$\ell(P_{x'})\leq \ell(Q) + c{\operatorname{dist}}(P_{x'},Q).$$ Then to obtain , it is enough to prove $$\label{e.copyq} {\operatorname{dist}}(P_{x'},Q)\leq c(\ell(P_{x})+\ell(Q)+ {\operatorname{dist}}(P_{x},Q))$$ Since $${\operatorname{dist}}(P_{x'},Q)\leq {\operatorname{dist}}(P_{x'},P_{x}) + {\operatorname{dist}}(P_{x},Q)+ \ell(P_x),$$ it suffices to see that ${\operatorname{dist}}(P_{x'},P_x)\leq c(\ell(P_{x})+\ell(Q)+ {\operatorname{dist}}(P_{x},Q))$. This follows from the Whitney condition and the fact that $Q\not\subset 3L$: $${\operatorname{dist}}(P_{x'},P_{x})\leq |x-x'|\lesssim \ell(L) \lesssim {\operatorname{dist}}(P_{x},Q) + \ell(Q) + \ell(P_x).$$ So holds holds. This completes the proof of and the claim . Observe that the estimates obtained in Lemmas \[lemdifer\] and \[lemdifer22\] do not depend on the constants $A$, $B$, $M$, $\delta_0$ or $\delta_W$. In fact, they do not depend on the precise construction of the tree ${{\mathcal T}}$. \[lemboundeta\] For every $k\geq0$, let $HD_1^k$ be the collection of cubes $Q\in{{\mathcal T}}$ which satisfy $\Theta(Q)\geq 2^k\Theta(HD_1)$ and have maximal side length. Denote $$H_k=\bigcup_{Q\in HD_1^{k}} Q \qquad \text{and}\qquad {{\widetilde{H}}}_k=\bigcup_{Q\in HD_1^{k}} (1+\tfrac1{10}c_{sep})\,Q,$$ and let $\Phi(x)={\operatorname{dist}}(x, {{\widetilde{H}}}_k ^{c})$. Set $\Theta(HD_1^k) = 2^k\Theta(HD_1)$. Then, for $1<r<\infty$, ${{\mathcal R}}_{\Phi,\eta}$ is bounded in $L^r(\eta)$ with norm not exceeding $c(r,M)\,\Theta(HD_1^k)$. For $1<r<\infty$, the boundedness of ${{\mathcal R}}_{\Phi,\eta}$ in $L^r(\eta)$ with norm $\lesssim \Theta(HD_1^k)$ follows from the boundedness in $L^2(\eta)$. This is shown in [@Vo] (see also Lemma 5.27 of [@To]). So we only have to deal with the case $r=2$. Denote $$F = {{\widetilde{H}}}_k\cup \bigcup_{P\in{\operatorname{Stop}}_2({{\mathcal T}})}(1+\tfrac1{10}c_{sep})\,P,$$ and consider the auxiliary function $$\Psi(x)= {\operatorname{dist}}(x,F^c).$$ Notice that all the cubes $Q\in{{\mathcal D}}$ such that $\Theta(Q)\geq 2^k\Theta(HD_1)$ are contained in $F$. Then, arguing as in Lemma \[lemrieszmax\], it follows that ${{\mathcal R}}_{\Psi,\mu}$ is bounded in $L^2(\mu{{\lfloor}}R)$ with norm bounded by $C\Theta(HD_1^k)$, by an application of Theorem \[teontv\]. By approximation, we are going to show that ${{\mathcal R}}_{\Psi,\eta}$ is bounded in $L^2(\eta)$, with its norm not exceeding $C\Theta(HD_1^k)$. The first step consists in showing that ${{\mathcal R}}_{\Psi,\eta}$ is bounded from $L^2(\eta)$ to $L^2(\mu{{\lfloor}}R)$ with norm not exceeding $c\,\Theta(HD_1^k)$. To this end, we consider the auxiliary tree ${{\mathcal T}}''\subset{{\mathcal T}}$ whose stopping cubes are given by the family of maximal cubes from $HD_1^{k}\cup{\operatorname{Stop}}_{2}({{\mathcal T}})$, which we denote by ${\operatorname{Stop}}({{\mathcal T}}'')$. That is, ${{\mathcal T}}''$ is obtained from ${{\mathcal T}}$ after removing all the cubes from ${{\mathcal D}}$ contained in $HD_1^k$. Given $f\in L^2(\eta)$, we consider the function ${{\widetilde{f}}}\in L^2(\mu{{\lfloor}}R)$ which vanishes out of $R$ and is constantly equal to $\int_{P}f d\eta/\mu(P)$ on each $P\in {\operatorname{Stop}}({{\mathcal T}}'')$. So notice that $$\int_P f\,d\eta = \int_P {{\widetilde{f}}}\,d\mu\qquad\mbox{for all $P\in{\operatorname{Stop}}({{\mathcal T}}'').$}$$ Since $\|{{\widetilde{f}}}\|_{L^2(\mu)}\leq \|f\|_{L^2(\eta)}$, to prove the boundedness of ${{\mathcal R}}_{\Psi,\eta}$ from $L^2(\eta)$ to $L^2(\mu{{\lfloor}}R)$ with norm not exceeding $c\,\Theta(HD_1^k)$, it is enough to show that $$\label{eqdah3} \int_R |{{\mathcal R}}_\Psi(f\,\eta) - {{\mathcal R}}_\Psi({{\widetilde{f}}}\,\mu)|^2\,d\mu\leq C\,\Theta(HD_1^k)^2\,\|f\|_{L^2(\eta)}^2.$$ Now we operate as in Lemma \[lemdifer\], and then for each $x\in Q\in{\operatorname{Stop}}({{\mathcal T}}'')$, taking into account that $\int_P f\,d\eta= \int_P f\,d\mu$, we get $$\begin{aligned} {{\mathcal R}}_\Psi(f\,\eta)(x) - {{\mathcal R}}_\Psi({{\widetilde{f}}}\,\mu)(x) & = \sum_{P\in{\operatorname{Stop}}({{\mathcal T}}'')}\left( \int_P K_\Psi(x,y)\bigl[f(y)\,d\eta(y) - {{\widetilde{f}}}(y)\,d\mu(y)\bigr]\right) \\ &=\sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}'')} \int_P \bigl[K_\Psi(x,y) - K_\Psi(x,z_P)\bigr]f(y)\,d\eta(y)\\ &\quad - \sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}'')} \int_P \bigl[K_\Psi(x,y) - K_\Psi(x,z_P)\bigr]{{\widetilde{f}}}(y)\,d\mu(y).\end{aligned}$$ Since $\Psi(y)\gtrsim\ell(P)$ for every $y\in P$, we have $$\bigl|K_\Psi(x,y) - K_\Psi(x,z_P)\bigr|\lesssim \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}.$$ Then we deduce that $$\begin{aligned} |{{\mathcal R}}_\Psi(f\,\eta)(x) - & {{\mathcal R}}_\Psi({{\widetilde{f}}}\,\mu)(x)| \\ & \lesssim \sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}'')} \int_P \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}\,\bigl[|f(y)|\,d\eta(y) + |{{\widetilde{f}}}(y)|\,d\mu(y)\bigr]\\ & \lesssim \sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}}'')} \int_P \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}\, |f(y)|\,d\eta(y).\end{aligned}$$ To estimate the last integral we argue by duality. Given $g\in L^{p'}(\eta)$, $$\begin{aligned} \label{eqdafgq1} \int & \sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})} \int_P \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}\, |f(y)|\,d\eta(y)\,|g(x)|\,d\eta(x) \\ & = \sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})} \int_P |f(y)|\left(\int \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}\, |g(x)|d\eta(x)\right)d\eta(y). \nonumber\end{aligned}$$ Using the fact that $\eta(\lambda P)\leq \mu(\lambda P \cap R)\leq C\,\Theta(HD_1^k)\,\ell(\lambda P)^s$ for every $\lambda\geq 1$, by splitting the domain of integration into annuli we infer that $$\label{eqmumu2} \int \frac{\ell(P)}{|x-z_P|^{s+1} + \ell(P)^{s+1}}\, |g(x)|d\eta(x)\lesssim\Theta(HD_1^k)\, \inf_{z\in P} M_\eta g(z),$$ where $M_\eta$ stands for the centered Hardy-Littlewood operator with respect to $\eta$. So the left side of is bounded by $$C\,\Theta(HD_1^k)\sum_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})} \int_P | f(y)|\, M_\eta g(y)\,d\eta(y) \lesssim \Theta(HD_1^k)\, \|f\|_{L^2(\eta)}\,\|g\|_{L^2(\eta)}.$$ Thus is proved. By duality we deduce now that ${{\mathcal R}}_{\Psi,\mu}$ is bounded from $L^2(\mu)$ to $L^2(\eta)$, with norm not exceeding $C\,\Theta(HD_1^k)$. To prove the boundedness of ${{\mathcal R}}_{\Psi,\eta}$ in $L^2(\eta)$, we consider again $f$ and ${{\widetilde{f}}}$ as above. We claim that $$\int_R |{{\mathcal R}}_\Psi(f\,\eta) - {{\mathcal R}}_\Psi({{\widetilde{f}}}\,\mu)|^2\,d\eta\leq C\,\Theta(HD_1^k)^2\,\|f\|_{L^2(\eta)}^2.$$ This follows by almost the same arguments as the ones above for the proof of . The details are left for the reader. Now we turn our attention to the operator ${{\mathcal R}}_\Phi$. Given $x\in{\operatorname{supp}}\eta$, notice that $\Phi(x)\neq\Psi(x)$ only for the points $x\in{\operatorname{Stop}}_2({{\mathcal T}})\setminus H_k$. Indeed, for such points $\Phi(x)=0$ while $\Psi(x)=\ell(P_x)$, where $P_x$ is the cube from ${\operatorname{Stop}}_{2}({{\mathcal T}})$ which contains $x$. Now we just write$$|{{\mathcal R}}_{\eta}f(x)|\leq |{{\mathcal R}}_{c_{sep}\ell(P_x)/10,\eta}f(x)| + \int_{|x-y|\leq c_{sep}\ell(P_{x})/10}\frac{1}{|x-y|^{s}}|f(y)|d\eta(y).$$ The first term on the right is controlled by ${{\mathcal R}}_\Psi f(x)$ and an error in terms of $\Theta(HD_1^k)$, using . The last one also can be easily controlled since the separation condition gives that all $y$ such that $|x-y|\leq \frac1{10}\,c_{sep}\ell(P_{x})$ satisfies $y\in P_{x}$. On $P_{x}$ the measure $\eta$ is uniformly distributed with respect to Lebesgue and one can use standard analysis to controll it by $M_{\eta}f(x)\Theta(P_{x})\leq B^{2} \Theta(R)\, M_{\eta}f(x)$. \[l.t1eta\] Let ${{\mathcal T}}$ be a tractable tree with roor $R$ and $\eta$ be the measure defined in . Then for $1<r<\infty$ ${{\mathcal R}}_{\eta}$ is bounded in $L^{r}(\eta)$ with norm not exceeding $C_2(B, M)\Theta(R)$. This result can be considered just as a particular case of Lemma \[lemboundeta\]. Indeed, take $k\in{{\mathbb Z}}$ such that $2B< 2^k\leq4B$. Then it follows that the set $H_k$ in that lemma is empty, and so $\Phi(x)=0$. Thus ${{\mathcal R}}_\eta$ is bounded in $L^r(\eta)$ with norm not exceeding $c(M,r)\,\Theta(HD_1^k)\lesssim C_2(M,B)\,\Theta(R)$. \[lemvar1\] Let ${{\mathcal T}}$ be as above and set $$\label{eqdeff} f(x) = {{\mathcal R}}(\chi_{R^c}\mu)(x) - m_R({{\mathcal R}}\mu).$$ For every $1<r<\infty$ there exists some constant $c_r>0$ such that $$\|{{\mathcal R}}\eta +f\|_{L^r(\eta)}^r \geq c_r\,\Theta(HD_1)^r\,\eta(HD_1).$$ For a collection of cubes ${{\mathcal A}}\subset{{\mathcal D}}$ and $1<r<\infty$, we denote $$\sigma_r({{\mathcal A}}) = \sum_{P\in{{\mathcal A}}} \Theta(P)^r\,\mu(P),$$ so that $\sigma({{\mathcal A}})\equiv \sigma_2({{\mathcal A}})$. For every $k\geq0$, recall $HD_1^k$ is the collection of cubes $Q\in{{\mathcal T}}$ which satisfy $\Theta(Q)\geq 2^k\Theta(HD_1)$ and have maximal side length (this ensures that $\Theta(Q)\approx 2^k\Theta(HD_1)$) and $H_k=\bigcup_{Q\in HD_1^{k}} Q.$ Observe that $H_{k+1}\subset H_k$. Let $k_0\geq1$ be such that $\sigma_r(HD_1^{k_0})$ is maximal. For $l\geq1$, we have $$2^{(k_0+l)r} \Theta(HD_{1})^r \eta(H_{k_0+l}) \approx \sigma_r(HD_1^{k_0+l})\leq \sigma_r(HD_1^{k_0}) \approx 2^{k_0r} \Theta(HD_{1})^r \eta(H_{k_0}).$$ Thus there exists some fixed $l$ (absolute constant) such that $$\eta(H_{k_0+l})\leq \frac12\,\eta(H_{k_0}).$$ We consider the set $$H= H_{k_0}\setminus H_{k_0+l},$$ so that $\eta(H)\approx\eta(H_{k_0})$. We also denote $\Theta(H)=2^{k_0}\Theta(HD_1)$. Notice that $$\label{eqdif422} \sigma_r(HD_1)\leq \sigma_r(HD_1^{k_0}) \approx \Theta(H)^r\,\eta(H).$$ Let $z_Q\in Q$ be such that $$\eta\bigl(B(z_Q,\tfrac1{10}c_{sep}\ell(Q))\bigr)\approx \eta(Q),$$ with the comparability constant depending on $c_{sep}$. Consider the functions ${{\varphi}}_Q$ and $\psi_Q$ defined just above . Observe that ${\operatorname{supp}}{{\varphi}}_Q\subset \bar B(z_Q,0.11c_{sep}\ell(Q))$. In the arguments below we will need to work with the functions $${{\varphi}}= \sum_{Q\in HD_1^{k_0}}\Theta(Q)\,{{\varphi}}_Q$$ and $$\psi = \sum_{Q\in HD_1^{k_0}}\Theta(Q)\,\psi_Q,\qquad {{\widetilde{\psi}}} = \sum_{Q\in HD_1^{k_0}}\Theta(Q)\,|\psi_Q|,$$ where ${{\varphi}}_Q$ and $\psi_Q$ have been defined in the preceding section. ${{\mathcal R}}(\psi\,d{{\mathcal L}}^d)={{\varphi}}$ and, by , $$\label{eqnormpsi} \|\psi\|_1\leq\|{{\widetilde{\psi}}}\|_1\leq\!\sum_{Q\in HD_1^{k_0}}\Theta(Q)\,\|\psi_Q\|_1 = C_1\!\sum_{Q\in HD_1^{k_0}}\Theta(Q)\, \ell(Q)^s = C_1\,\eta(HD_1^{k_0})\approx \eta(H).$$ Recalling , to prove the lemma it suffices to show that $$\|{{\mathcal R}}\eta +f\|_{L^r(\eta)}^r \geq c_r\,\Theta(H)^r\,\eta(H).$$ We will argue by contradiction, following some ideas inspired by the techniques from [@ENV12] and [@NToV]. So suppose that $$\label{eqas12} \|{{\mathcal R}}\eta +f\|_{L^r(\eta)}^r \leq \lambda\,\Theta(H)^r\,\eta(H),$$ where $\lambda>0$ is some small constant to be fixed below. Consider the measures of the form $\nu=a\,\eta$, with $a\in L^\infty(\eta)$, $a\geq 0$, and let $F$ be the functional $$\label{e.functional} F(\nu) = \int |{{\mathcal R}}\nu + f|^r\,d\nu + \lambda\,\|a\|_{L^\infty(\eta)} \,\Theta(H)^r\,\eta(H).$$ Let $$m= \inf F(\nu),$$ where the infimum is taken over all the measures $\nu=a\,\eta$, with $a\in L^\infty(\eta)$, such that $\nu(H) = \eta(H)$. We call such measures admissible. Note that $\eta$ is admissible, and thus $$\label{e.admis} m \leq F(\eta) \leq 2\lambda\,\Theta(H)^r\,\eta(H),$$ by the assumption . This tells us that the infimum $m$ is attained over the collection of measures $\nu=a\eta$ with $\|a\|_{L^\infty(\eta)}\leq 3$. In particular, by taking weak $*$ limits in $L^\infty(\eta)$, this guaranties the existence of a minimizer. Let $\nu$ be an admissible measure such that $m=F(\nu)$. We claim that $$\label{eqclam43} |{{\mathcal R}}\nu(x) + f(x)|^r + r \,{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \nu\bigr)(x) \leq C\,\lambda^{1/r'}\,\Theta(H)^r \quad \text{ on ${\operatorname{supp}}\nu$.}$$ Let us assume this for the moment. Observe that, by Lemma \[lemcomp\], for all $x\in{\operatorname{supp}}\nu\subset R$, $$|f(x)| = |{{\mathcal R}}(\chi_{R^c}\mu)(x) - m_R({{\mathcal R}}\mu)|\lesssim p(R) \lesssim \Theta(R) = B^{-1}\,\Theta(HD_1) \leq B^{-1}\,\Theta(H).$$ Then, using the inequality $|\alpha + \beta|^r\leq 2^{r-1} \bigl (|\alpha|^r + |\beta|^r\bigr)$, we infer that $$\begin{aligned} |{{\mathcal R}}\nu(x)|^r & + 2^{r-1} r \,{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \nu\bigr)(x) \\ & \leq 2^{r-1}\bigl (|{{\mathcal R}}\nu(x) + f(x)|^r + |f(x)|^r\bigr) + 2^{r-1} r \,{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \nu\bigr)(x)\\ & \leq 2^{r-1}|f(x)|^r + C\,\lambda\,\Theta(H)^r\\&\leq C \bigl (B^{-r} + \lambda^{1/r'}\bigr)\,\Theta(H)^r\end{aligned}$$ for $x\in{\operatorname{supp}}\nu$. It is easily checked that the function on the left side of this inequality is continuous. Appealing then to the maximum principle in Lemma \[lemmaxprin\], with $h=2^{r-1} r ({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2}$, we infer the estimate above also holds out of ${\operatorname{supp}}\nu$. Therefore, integrating against ${{\widetilde{\psi}}}\,d{{\mathcal L}}^d$ (recall that ${{\widetilde{\psi}}}$ is non-negative), we get $$\label{eqpss3} \int \!|{{\mathcal R}}\nu|^r {{\widetilde{\psi}}}\,d{{\mathcal L}}^d + 2^{r-1} r\! \int \!{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \nu\bigr)\, {{\widetilde{\psi}}}\,d{{\mathcal L}}^d\leq C\bigl(B^{-r} + \lambda^{1/r'}\bigr)\,\Theta(H)^r\,\|{{\widetilde{\psi}}}\|_1.$$ Next we estimate the second integral on the left side of the preceding inequality, which we denote by $I$. For that purpose we will also need the estimate $$\label{eqacpsi} \int |{{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)|^r\,d\nu\lesssim \Theta(H)^r\,\eta(H),$$ that we will prove later. Using we have $$\begin{aligned} |I|& = \left|\int ({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \, {{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)\,d\nu\right| \\ &\leq \left(\int |{{\mathcal R}}\nu + f|^r \,d\nu\right)^{\frac{r-1}r} \left(\int |{{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)|^r\,d\nu\right)^{\frac{1}r}\\ & \lesssim \bigl (\lambda\,\Theta(H)^r\,\eta(H)\bigr)^{1/r'} \bigl (\Theta(H)^r\,\eta(H)\bigr)^{1/r} = \lambda^{1/r'}\,\Theta(H)^r\,\eta(H),\end{aligned}$$ where $r'=r/(r-1)$. From , the preceding estimate, and the fact that $\|{{\widetilde{\psi}}}\|_1\lesssim\eta(H)$, we deduce that $$\label{eqcont4} \int |{{\mathcal R}}\nu|^r {{\widetilde{\psi}}}\,d{{\mathcal L}}^d \lesssim \bigl (B^{-r} + \lambda^{1/r'}\bigr)\,\Theta(H)^r\, \eta(H).$$ To get a contradiction, note that by the construction of $\psi$ and ${{\varphi}}$, we have $$\left|\int {{\mathcal R}}\nu\,\psi\,d{{\mathcal L}}^d\right| = \left|\int {{\mathcal R}}^*(\psi\,d{{\mathcal L}}^d)\,d\nu \right|= \int {{\varphi}}\,d\nu = \sum_{Q\in HD_1^{k_0}} \Theta(Q)\,\nu(Q) \approx \Theta(H)\,\eta(H).$$ On other hand, since $|\psi|\leq{{\widetilde{\psi}}}$, by Hölder’s inequality, , and , $$\left|\int {{\mathcal R}}\nu\,\psi\,d{{\mathcal L}}^d\right| \leq \left(\int |{{\mathcal R}}\nu|^r {{\widetilde{\psi}}}\,d{{\mathcal L}}^d \right)^{1/r} \left(\int {{\widetilde{\psi}}}\,d{{\mathcal L}}^d\right)^{1/r'} \lesssim \bigl (B^{-r} + \lambda^{1/r'}\bigr)^{1/r}\,\Theta(H) \,\eta(H),$$ which contradicts the previous estimate if $B$ is big enough and $\lambda$ small enough. So to finish the proof of the lemma it only remains to prove the estimates and . This task is carried out below. We have to show that $$\int |{{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)|^r\,d\nu\lesssim \Theta(H)^r\,\eta(H).$$ Recall that $${{\widetilde{\psi}}} = \sum_{Q\in HD_1^{k_0}}\Theta(Q)\,|\psi_Q|.$$ We consider now the function $$g=\sum_{Q\in HD_1^{k_0}} g_Q,$$ where $g_Q = c_Q\,\chi_Q$, with $c_Q=\Theta(Q)\int|\psi_Q|\,d{{\mathcal L}}^d /\eta(Q)$, so that $$\int g_Q\,d\eta = \Theta(Q)\int |\psi_Q|\,d{{\mathcal L}}^d\qquad\mbox{for all $Q\in HD_1^{k_0}.$}$$ Recalling that $\|\Theta(Q)\psi_Q\|_1\lesssim \eta(Q)$, it follows that $|c_Q|\lesssim 1$ and $\|g_Q\|_{L^1(\eta)}\lesssim \eta(Q)$. The first step of our arguments consists in comparing ${{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)(x)$ to ${{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)$, with ${{\varepsilon}}(x)=\ell(Q)$ if $x\in Q\in HD_1^{k_0}$, and ${{\varepsilon}}(x)=0$ otherwise. We will prove that, for each $Q\in HD_1^{k_0}$, $$\label{eqtech4} |{{\mathcal R}}(\Theta(Q)|\psi_Q|\,d{{\mathcal L}}^d)(x) - {{\mathcal R}}_{{{\varepsilon}}(x)}(g_Q\,\eta)(x)|\lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}},$$ where $\beta$ is some positive absolute constant. The arguments to prove this estimate are similar to the ones in Lemma \[lemdifer\] or , although the fact that $\psi_Q$ may be not compactly supported introduces some additional difficulties. So we will show the details. Suppose first that ${\operatorname{dist}}(x,Q)\leq 2\,\ell(Q)=2\,{\operatorname{diam}}(Q)$. From , we obtain $$\begin{aligned} |{{\mathcal R}}(\Theta(Q)|\psi_Q|\,d{{\mathcal L}}^d)(x)| &\lesssim \Theta(Q)\int_{{\operatorname{dist}}(y,Q)|\leq 4\ell(Q)} \frac{\ell(Q)^{s-d}}{|x-y|^s}\,d{{\mathcal L}}^d(y)\\ &\quad + \Theta(Q)\int_{{\operatorname{dist}}(y,Q)|> 4\ell(Q)} \frac{\ell(Q)^{s+\alpha}}{|x-y|^s\,|y-z_Q|^{d+\alpha}}\,d{{\mathcal L}}^d(y)\\ & \lesssim \Theta(Q),\end{aligned}$$ taking into account that $|x-y|\approx |y-z_Q|>4\ell(Q)$ for the estimate of the last integral. On the other hand, we also have $$|{{\mathcal R}}_{\ell(Q)}(g_Q\,\eta)(x)| \leq \int_{|y-x|>2\ell(Q)} \frac{1}{|x-y|^s}\,|g_Q(y)|\,d\eta(y) \leq \frac{1}{\ell(Q)^s}\|g_Q\|_{L^1(\eta)}\lesssim \Theta(Q).$$ Thus holds when ${\operatorname{dist}}(x,Q)\leq 2\,\ell(Q)$. Suppose now that ${\operatorname{dist}}(x,Q)>2\,\ell(Q)$. For such points, ${{\mathcal R}}_{\ell(Q)}(g_Q\,\eta)(x) = {{\mathcal R}}(g_Q\,\eta)(x)$. Thus we can write $$\begin{aligned} |{{\mathcal R}}(\Theta(Q)|\psi_Q|\,&d{{\mathcal L}}^d)(x) - {{\mathcal R}}_{{{\varepsilon}}(x)}(g_Q\,\eta)(x)|\\ & \leq \int K^s(x-y) \bigl[\Theta(Q)|\psi_Q(y)|\,d{{\mathcal L}}^d(y) - g_Q(y)\,d\eta(y)\bigr] \\ & = \int_{|y-z_Q|\geq 2|x-z_Q|}\ldots + \int_{\frac12|x-z_Q|<|y-z_Q|<2|x-z_Q|}\ldots + \int_{|y-z_Q|\leq \frac12|x-z_Q|}\ldots\\& =:I_1 + I_2 + I_3.\end{aligned}$$ To estimate $I_1$, notice that $|y-x|\approx|y-z_Q|$ in the domain of integration. Also $g_Q(y)=0$ because $|y-z_Q|\geq 2|x-z_Q|> 4\ell(Q)$. So we have $$|K^s(x-y)|\leq \frac1{|x-y|^s}\lesssim \frac1{|y-z_Q|^s}$$ and, by the estimate , $$\begin{aligned} |I_1| &\lesssim \Theta(Q)\int_{|y-z_Q|\geq 2|x-z_Q|}\frac1{|y-z_Q|^s}\,|\psi_Q(y)|\,d{{\mathcal L}}^d(y)\\ &\lesssim \Theta(Q)\int_{|y-z_Q|\geq 2|x-z_Q|}\frac1{|y-z_Q|^s}\,\frac{\ell(Q)^{s+\alpha}}{|y-z_Q|^{d+\alpha}}\,d{{\mathcal L}}^d(y)\lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\alpha}}{|x-z_Q|^{s+\alpha}}.\end{aligned}$$ Let us turn our attention to $I_2$. Again in the domain of integration we have $g_Q(y)=0$ because $|y-z_Q|\geq \frac12|x-z_Q|> \ell(Q)$. By the estimate we get $$\begin{aligned} |I_2| &\lesssim \Theta(Q)\int_{\frac12|x-z_Q|<|y-z_Q|<2|x-z_Q|} \frac1{|x-y|^s}\,\frac{\ell(Q)^{s+\alpha}}{|y-z_Q|^{d+\alpha}}\,d{{\mathcal L}}^d(y)\\ & \lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\alpha}}{|x-z_Q|^{d+\alpha}}\int_{|y-z_Q|<2|x-z_Q|} \frac1{|x-y|^s}\,d{{\mathcal L}}^d(y)\lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\alpha}}{|x-z_Q|^{s+\alpha}}.\end{aligned}$$ Finally we deal with $I_3$. We write it as follows: $$\begin{aligned} I_3& = \int_{|y-z_Q|\leq \frac12|x-z_Q|} \bigl[K^s(x-y) - K^s(x-z_Q)\bigr]\, \bigl[\Theta(Q)|\psi_Q(y)|\,d{{\mathcal L}}^d(y) - g_Q(y)\,d\eta(y)\bigr]\\ &\quad + K^s(x-z_Q)\int_{|y-z_Q|\leq \frac12|x-z_Q|}\bigl[\Theta(Q)\, |\psi_Q(y)|\,d{{\mathcal L}}^d(y) - g_Q(y)\,d\eta(y)\bigr] \\ &=: I_3^a + I_3^b.\end{aligned}$$ To estimate $I_3^a$ we use the smoothness of the kernel $K^s$, and then we get $$\begin{aligned} \label{eqi3a} I_3^a&\lesssim \int_{|y-z_Q|\leq \frac12|x-z_Q|} \frac{|y-z_Q|^\beta}{|x-z_Q|^{s+\beta}}\, \bigl[\Theta(Q)|\psi_Q(y)|\,d{{\mathcal L}}^d(y)+ g_Q(y)\,d\eta(y)\bigr],\end{aligned}$$ where $0<\beta\leq1$. Using , we obtain $$\begin{aligned} \int_{|y-z_Q|\leq \frac12|x-z_Q|} &\frac{|y-z_Q|^\beta}{|x-z_Q|^{s+\beta}}\,\Theta(Q)|\psi_Q(y)|\,d{{\mathcal L}}^d(y)\\ &\lesssim \frac{\Theta(Q)}{|x-z_Q|^{s+\beta}} \int_{|y-z_Q|\leq \frac12|x-z_Q|} \frac{\ell(Q)^{s+\alpha}\,|y-z_Q|^{\beta}}{\ell(Q)^{d+\alpha} +|y-z_Q|^{d+\alpha}}\,d{{\mathcal L}}^d(y)\end{aligned}$$ The last integral is bounded by $$\begin{aligned} C\int_{|y-z_Q|\leq \ell(Q)} \frac{\ell(Q)^{s+\alpha+\beta}}{\ell(Q)^{d+\alpha}}\,d{{\mathcal L}}^d(y) + C\int_{|y-z_Q|> \ell(Q)} \frac{\ell(Q)^{s+\alpha}}{|y-z_Q|^{d+\alpha-\beta}}\,d{{\mathcal L}}^d(y).\end{aligned}$$ Both integrals are bounded by $C\,\ell(Q)^{s+\beta}$, assuming that $\beta<\alpha$ for the last one (for instance, we may choose $\beta=\alpha/2$). Thus, $$\int_{|y-z_Q|\leq \frac12|x-z_Q|} \frac{|y-z_Q|^\beta}{|x-z_Q|^{s+\beta}}\,\Theta(Q)|\psi_Q(y)|\,d{{\mathcal L}}^d(y) \lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\beta}}{|x-z_Q|^{s+\beta}}.$$ It remains now to estimate the integral involving the term $g_Q(y)\,d\eta(y)$ in . We have $$\begin{gathered} \int_{|y-z_Q|\leq \frac12|x-z_Q|} \frac{|y-z_Q|^\beta}{|x-z_Q|^{s+\beta}}\, g_Q(y)\,d\eta(y)\\ \lesssim \frac1{|x-z_Q|^{s+\beta}} \int_Q \ell(Q)^{\beta}\,g_Q(y)\,d\eta(y)\lesssim \frac{\ell(Q)^{\beta}\eta(Q)}{|x-z_Q|^{s+\beta}} =\frac{\Theta(Q)\,\ell(Q)^{s+\beta}}{|x-z_Q|^{s+\beta}}.\end{gathered}$$ Concerning $I_3^b$, since $\int g_Q\,d\eta = \Theta(Q)\int |\psi_Q|\,d{{\mathcal L}}^d$, we deduce that $$\begin{aligned} I_3^b&= -K^s(x-z_Q)\int_{|y-z_Q|>\frac12|x-z_Q|}\Theta(Q)\, |\psi_Q(y)|\,d{{\mathcal L}}^d(y) - g_Q(y)\,d\eta(y)\\ & = -K^s(x-z_Q)\int_{|y-z_Q|> \frac12|x-z_Q|}\Theta(Q)\, |\psi_Q(y)|\,d{{\mathcal L}}^d(y),\end{aligned}$$ taking into account that $g_Q(y)=0$ for $y$ in the integrals above. So by we get $$\begin{aligned} I_3^b &\lesssim \frac{\Theta(Q)}{|x-z_Q|^s} \int_{|y-z_Q|> \frac12|x-z_Q|} \frac{\ell(Q)^{s+\alpha}}{|y-z_Q|^{d+\alpha}} \,d{{\mathcal L}}^d(y)\lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\alpha}}{|x-z_Q|^{s+\alpha}}.\end{aligned}$$ Gathering the estimates obtained for $I_1$, $I_2$ and $I_3$, since $|x-z_Q|\approx {\operatorname{dist}}(x,Q)+\ell(Q)$ (recall that ${\operatorname{dist}}(x,Q)>2\,\ell(Q)$), we derive $$|{{\mathcal R}}(\Theta(Q)|\psi_Q|\,d{{\mathcal L}}^d)(x) - {{\mathcal R}}_{{{\varepsilon}}(x)}(g_Q\,\eta)(x)|\lesssim \frac{\Theta(Q)\,\ell(Q)^{s+\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}},$$ as claimed. So we infer that $$\begin{gathered} \int |{{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)(x) - {{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)|^r\,d\eta(x) \lesssim \int \biggl (\sum_{Q\in HD_1^{k_0}}\frac{\Theta(Q)\,\ell(Q)^{s+\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}}\biggr)^r\,d\eta(x).\end{gathered}$$ We estimate the last integral by duality. Consider a function $h\in L^{r'}(\eta)$. Then, $$\begin{gathered} \label{eqmul429} \int \sum_{Q\in HD_1^{k_0}}\frac{\eta(Q)\,\ell(Q)^{\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}}\,h(x)\,d\eta(x)\\= \sum_{Q\in HD_1^{k_0}} \eta(Q)\int \frac{\ell(Q)^{\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}}\,h(x)\,d\eta(x)\end{gathered}$$ For each $Q$, using the fact that $\eta(\lambda Q)\leq C\,\Theta(H)\,\ell(\lambda Q)^s$ for every $\lambda\geq 1$, as in we obtain $$\int \frac{\ell(Q)^{\beta}}{{\operatorname{dist}}(x,Q)^{s+\beta}+ \ell(Q)^{s+\beta}}\,h(x)\,d\eta(x)\lesssim C\,\Theta(H)\,\inf_{y\in Q} M_\eta h(y).$$ Therefore, the left side of does not exceed $$\begin{aligned} C\Theta(H)\sum_{Q\in HD_1^{k_0}} \eta(Q)\,\inf_{y\in Q} M_\eta h(y) & \lesssim \Theta(H) \int_{H_{k_0}}M_\eta h(y)\,d\eta(y) \\ & \lesssim \Theta(H)\,\eta(H_{k_0})^{1/r}\,\|h\|_{L^{r'}(\eta)} \lesssim \Theta(H)\,\eta(H)^{1/r}\,\|h\|_{L^{r'}(\eta)}.\end{aligned}$$ So we deduce that $$\label{eqmca2} \int |{{\mathcal R}}({{\widetilde{\psi}}}\,d{{\mathcal L}}^d)(x) - {{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)|^r\,d\eta(x)\\ \lesssim \Theta(H)^r\,\eta(H) = \Theta(H)^r\,\nu(H).$$ The estimate is a consequence of and the fact that $$\int |{{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)|^r\,d\eta(x)\\ \lesssim \Theta(H)^r\,\eta(H).$$ This inequality is easily derived from the $L^r(\eta)$ boundedness of the operator ${{\mathcal R}}_{\Phi,\eta}$, where $\Phi$ is the function defined in Lemma \[lemboundeta\] (with $k=k_0$). Indeed, notice that ${{\varepsilon}}(x)\approx\Phi(x)$ for all $x\in{\operatorname{supp}}\eta$. Then, from , it follows that $$\label{e.compsup} |{{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)|\lesssim |{{\mathcal R}}_{\Phi}(g\,\eta)(x)|+ \sup_{r\geq \Phi(x)}\frac{1}{r^s}\int_{B(x,r)} |g|\,d\eta\lesssim |{{\mathcal R}}_{\Phi}(g\,\eta)(x)| + \Theta(H).$$ So we have $$\begin{aligned} \int |{{\mathcal R}}_{{{\varepsilon}}(x)}(g\,\eta)(x)|^r\,d\eta(x)& \lesssim \|{{\mathcal R}}_{\Phi}(g\,\eta)\|_{L^r(\eta)}^r + \Theta(H)^r\,\|g\|_{L^r(\eta)}^r\\ & \lesssim \Theta(H)^r\,\|g\|_{L^r(\eta)}^r\lesssim \Theta(H)^r\,\eta(H_{k_0})\approx \Theta(H)^r\,\nu(H).\end{aligned}$$ Let $\nu=a\,\eta$ be a minimizing measure for $F(\nu)$ as in , so that, in particular, $$F(\nu)\leq 2\lambda\,\Theta(H)^r\,\eta(H).$$ We have to show that $$|{{\mathcal R}}\nu(x) + f(x)|^r + r \,{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \nu\bigr)(x) \leq C\,\lambda^{1/r'}\,\Theta(H)^r \quad \text{ on ${\operatorname{supp}}\nu$.}$$ Let $x_0\in{\operatorname{supp}}\nu$ and consider a ball $B=B(x_0,\rho)$, with $\rho>0$ small. For $0<t<1$ we construct a competing measure $\nu_t = a_t\,\eta$, where $a_t$ is defined as follows: $$a_t = (1-t\,\chi_B)a +t\,\frac{\nu(B\cap H)}{\nu(H)}\,\chi_{H}\,a.$$ Notice that, for each $0<t<1$, $a_t$ is a non-negative function such that $\nu_t(H)=\nu(H)$. Taking into account that $$\|a_t\|_{L^\infty(\eta)}\leq \|a\|_{L^\infty(\eta)} + t\,\frac{\nu(B\cap H)}{\nu(H)},$$ we deduce that $$\begin{aligned} F(\nu_t) & = \int |{{\mathcal R}}\nu_t + f|^r\,d\nu_{t} + \lambda\,\|a_t\|_{L^\infty(\eta)} \,\Theta(H)^r\,\eta(H)\\ & \leq \int |{{\mathcal R}}\nu_t + f|^r\,d\nu_{t} + \lambda\,\biggl (\|a\|_{L^\infty(\eta)} + t\,\frac{\nu(B\cap H)}{\nu(H)} \biggr) \,\Theta(H)^r\,\eta(H) =:{{\widetilde{F}}}(\nu_t).\end{aligned}$$ Since ${{\widetilde{F}}}(\nu_{0}) = F(\nu) \leq F(\nu_t)\leq {{\widetilde{F}}}(\nu_t)$ for $t\geq 0$, we infer that $$\label{eqvar49} \frac1{\nu(B)}\,\frac{d}{dt}\,{{\widetilde{F}}}(\nu_t)\biggr|_{t=0} \geq 0,$$ with the derivative taken from the right. An easy computation gives $$\begin{aligned} \frac{d}{dt}\,{{\widetilde{F}}}(\nu_t)\biggr|_{t=0} & = -\int_B |{{\mathcal R}}\nu + f|^r\,d\nu+\frac{\nu(B\cap H)}{\nu(H)}\int_{H}|{{\mathcal R}}\nu + f|^r\,d\nu \\ & \quad + r \int ({{\mathcal R}}\nu + f)\, |{{\mathcal R}}\nu + f|^{r-2} \,{{\mathcal R}}\biggl (-\chi_B \nu + \frac{\nu(B\cap H)}{\nu(H)} \,\chi_{H}\nu\biggr)\, d\nu \\ &\quad + \lambda\nu(B\cap H) \,\Theta(H)^r.\end{aligned}$$ So is equivalent to $$\begin{gathered} \label{eqmult82} \frac1{\nu(B)} \, \int_B |{{\mathcal R}}\nu + f|^r\,d\nu + \frac r{\nu(B)} \, \int ({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \,{{\mathcal R}}(\chi_B \nu)\, d\nu \\ \leq \frac{\nu(B\cap H)}{\nu(B)\,\nu(H)} \biggl[\int_{H}|{{\mathcal R}}\nu + f|^r\,d\nu+ r\int ({{\mathcal R}}\nu + f)\, |{{\mathcal R}}\nu + f|^{r-2} \,{{\mathcal R}}( \chi_{H}\nu) d\nu + \lambda \,\Theta(H)^r\,\nu(H)\biggr].\end{gathered}$$ To estimate the right side of the preceding inequality, first we use to notice that $$\int_{H}|{{\mathcal R}}\nu + f|^r\,d\nu\leq F(\nu) \lesssim \lambda \,\Theta(H)^r\,\nu(H).$$ We set $$\label{eqak4} \left|\int ({{\mathcal R}}\nu + f)\, |{{\mathcal R}}\nu + f|^{r-2} \,{{\mathcal R}}( \chi_{H}\nu)\,d\nu\right|\leq \left( \int |{{\mathcal R}}\nu + f|^r \,d\nu\right)^{1/r'} \!\! \left(\int |{{\mathcal R}}(\chi_{H}\nu)|^r \,d\nu\right)^{1/r}.$$ We claim now that $$\label{eqcla491} \int |{{\mathcal R}}(\chi_{H}\nu)|^r \,d\nu \leq C\,\Theta(H)^r\,\nu(H).$$ Assuming this to hold, tells us that $$\begin{aligned} \left|\int ({{\mathcal R}}\nu + f)\, |{{\mathcal R}}\nu + f|^{r-2} \,{{\mathcal R}}( \chi_{H}\nu)\,d\nu\right|& \lesssim F(\nu)^{1/r'} \bigl (\Theta(H)^r\,\nu(H)\bigr)^{1/r}\\& \lesssim \lambda^{1/r'}\,\Theta(H)^r\,\nu(H),\end{aligned}$$ recalling for the last inequality. Thus the right side of does not exceed $$C(\lambda^{1/r'}\,\Theta(H)^r + \lambda\,\Theta(H)^r)\lesssim \lambda^{1/r'}\,\Theta(H)^r,$$ assuming that $\lambda<1$ for the last estimate. Let us turn our attention to the left side of now. We rewrite it as $$\frac1{\nu(B)} \, \int_B |{{\mathcal R}}\nu + f|^r\,d\nu + \frac r{\nu(B)} \, \int_B {{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \,\nu\bigr)\, d\nu.$$ Taking into account that the functions in the integrands are continuous on ${\operatorname{supp}}(\nu)$, letting the radius $\rho$ of $B$ tend to $0$, it turns out that the above expression converges to $$|{{\mathcal R}}\nu(x_0) + f(x_0)|^r + r{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \,\nu\bigr)(x_0).$$ As a consequence, we derive $$|{{\mathcal R}}\nu(x_0) + f(x_0)|^r + r{{\mathcal R}}^*\bigl (({{\mathcal R}}\nu + f) |{{\mathcal R}}\nu + f|^{r-2} \,\nu\bigr)(x_0) \lesssim \lambda^{1/r'}\,\Theta(H)^r,$$ as wished. It remains to prove . To this end, recall that $H = H_{k_0}\setminus H_{k_0+l}$ for some $l\geq1$, and consider the enlarged set $${{\widetilde{H}}}_{k_0+l}=\bigcup_{Q\in HD_1^{k_0+l}} (1+\tfrac1{10}c_{sep})Q,$$ and the associated $1$-Lipschitz function $\Phi(x) = {\operatorname{dist}}(x,{{\mathbb R}}^d\setminus{{\widetilde{H}}}_{k_0+l})$. By Lemma \[lemboundeta\], ${{\mathcal R}}_{\Phi,\eta}$ is bounded in $L^r(\eta)$, with norm not greater than $C\,\Theta(HD_1^{k_0+l})\approx \Theta(H)$. Since $\nu=a\,\eta$ with $\|a\|_{L^\infty(\eta)}\leq3$, we infer that ${{\mathcal R}}_{\Phi,\nu}$ is bounded in $L^r(\nu)$, also with norm not greater than $C\,\Theta(H)$. As $\Phi$ vanishes in ${{\mathbb R}}^d\setminus{{\widetilde{H}}}_{k_0+n}$, it follows that $K_\Phi(x,y)=(x-y)/|x-y|^{s+1}$ for $y\in {\operatorname{supp}}\nu\cap H\subset{{\mathbb R}}^d\setminus {{\widetilde{H}}}_{k_0+n}$ and thus $${{\mathcal R}}(\chi_{H}\nu) = {{\mathcal R}}_\Phi(\chi_{H}\nu).$$ Therefore, $$\|{{\mathcal R}}(\chi_{H}\nu)\|_{L^r(\nu)}^r = \|{{\mathcal R}}_\Phi(\chi_{H}\nu)\|_{L^r(\nu)}^r\lesssim \Theta(H)^r \,\nu(H).$$ Let us remark that, from the conditions and , assuming also that the constant $\delta_W$ in is small enough and following the proof of we deduce that $$\label{equa22} \sigma(HD_1)\geq \frac12\,\sigma({\operatorname{Stop}}_{1}({{\mathcal T}})).$$ This estimate will be used below. \[lem22\] Let ${{\mathcal T}}$ be a tractable tree with root $R$ and $f$ as in . Then there exists some subset $G\subset \bigcup_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})}P$ satisfying $$\eta(G)\geq C_6(A,B,M)\,\eta(R)$$ such that, for every $x\in G$, $$\label{eqda59} |{{\mathcal R}}\eta(x) + f(x)|\geq C_7 A^{1/r}\,\bigl (p(P_x) + q(P_x,{{\mathcal T}})\bigr) + C_8(A,B)\,\Theta(R),$$ where $P_x$ stands for the cube from ${\operatorname{Stop}}_{2}({{\mathcal T}})$ that contains $x$. We distinguish two cases. In the first one we assume that $$\label{eqcase1} \sigma({\operatorname{Stop}}_{2}({{\mathcal T}}))\leq A^{-1}\,\sigma({\operatorname{Stop}}_{1}({{\mathcal T}})).$$ Then from Lemma \[lemvar1\] and we infer that $$\label{equari2} \|{{\mathcal R}}\eta + f\|_{L^2(\eta)}^2 \gtrsim\,\Theta(HD_1)^2\,\eta(HD_1)\gtrsim\sigma(HD_1) \geq \frac12\,\sigma({\operatorname{Stop}}_1({{\mathcal T}})) \geq \frac A2\,\sigma({\operatorname{Stop}}_{2}({{\mathcal T}})).$$ In the second case, does not hold. Then, taking also into account that ${{\mathcal T}}$ is a tractable tree, we have $$\label{eqcase2} A^{-1}\,\sigma({\operatorname{Stop}}_1({{\mathcal T}}))\leq \sigma({\operatorname{Stop}}_{2}({{\mathcal T}}))\lesssim A\,\sigma({\operatorname{Stop}}_1({{\mathcal T}})).$$ For a collection of cubes ${{\mathcal A}}\subset{{\mathcal D}}$ and $1<r<\infty$, we denote $$\sigma_r({{\mathcal A}}) = \sum_{P\in{{\mathcal A}}} \Theta(P)^r\,\mu(P),$$ so that $\sigma({{\mathcal A}})\equiv \sigma_2({{\mathcal A}})$. Our next objective consists in showing that for some $1<r<2$, the following holds: $$\label{equarir} \|{{\mathcal R}}\eta + f\|_{L^r(\eta)}^r \gtrsim\,A\,\sigma_r({\operatorname{Stop}}_{2}({{\mathcal T}})).$$ Notice that this estimate coincides with if we take $r=2$. To prove , notice that, by Lemma \[lemvar1\], choosing $1<r<2$ and using , we get $$\begin{aligned} \label{eqali2} \|{{\mathcal R}}\eta + f\|_{L^{r}(\eta)}^{r} & \gtrsim\Theta(HD_1)^{r}\,\mu(HD_1) \\ & \gtrsim \Theta(HD_1)^{r-2}\,\sigma(HD_1) \nonumber\\ & \gtrsim A^{-1}\,\Theta(HD_1)^{r-2}\,\sigma(HD_2)\nonumber\\ & \gtrsim A^{-1}\,\left(\frac{\Theta(HD_2)}{\Theta(HD_1)}\right)^{2-r}\,\Theta(HD_2)^r\,\mu(HD_2)\nonumber\\ & = A^{-1}\,B^{2-r}\,\Theta(HD_2)^{r}\,\mu(HD_2).\nonumber\end{aligned}$$ We claim now that $$\label{eqcla52} \sigma_r(HD_2)\approx \sigma_r({\operatorname{Stop}}_{2}({{\mathcal T}})).$$ Observe that follows from the two preceding estimates if we assume $B$ big enough so that $$A^{-1}\,B^{2-r} \geq A.$$ From now on we assume that $r=3/2$ if does not hold, say, and $r=2$ otherwise. To prove the claim (only in the case $r=3/2$) we will show that $$\sigma_r(LD_1) + \sigma_r(LD_2) + \sigma_r(BR_1)+ \sigma_r(BR_2)\leq \sigma_r(HD_2).$$ Using the left inequality in and we obtain $$\label{e.h2below} \sigma(HD_2) \gtrsim \sigma({\operatorname{Stop}}_{2}({{\mathcal T}}))\geq A^{-1}\sigma({\operatorname{Stop}}_1({{\mathcal T}}))\gtrsim A^{-2}\,\sigma(\{R\}).$$ Therefore, taking into account that $A\leq B$, $$\begin{aligned} \sigma_r(HD_2) & =\Theta(HD_2)^{r-2}\,\sigma(HD_2)\gtrsim A^{-2}\,\Theta(HD_2)^{r-2}\,\sigma(\{R\})\\ & \gtrsim A^{-2}\,B^{2(r-2)}\,\Theta(R)^{r-2}\,\sigma(\{R\}) \geq \frac1{B^{6-2r}}\,\sigma_r(\{R\}).\end{aligned}$$ Also, from and the preceding estimate $$\begin{aligned} \sigma_r(LD_1) + \sigma_r(LD_2) & \leq \delta_0^{\frac r{s+2}}\,\Theta(R)^r\,\mu(R) + \delta_0^{\frac r{s+2}}\,B^r\,\Theta(R)^r\,\mu(R) \\ & \lesssim \delta_0^{\frac r{s+2}}\,B^r\,\sigma_r(\{R\})\\ & \lesssim \delta_0^{\frac r{s+2}}\,B^{r+6-2r}\,\sigma_r(HD_2) .\end{aligned}$$ On the other hand, $$\begin{aligned} \sigma_r(BR_1)+ \sigma_r(BR_2) & \leq \Theta(HD_1)^{2-r} \,\sigma(BR_1) + \Theta(HD_2)^{2-r} \sigma(BR_2) \\ &\leq \Theta(HD_1)^{2-r}\delta_W B^{2}\sigma(R)+ \Theta(HD_2)^{2-r}\delta_W B^{4}\sigma(R)\\ &\lesssim 2 B^{2(2-r)}\delta_W B^{4}\Theta(R)^{2-r}\sigma(R)\\ &\lesssim \delta_W \,B^{4(2-r)+6}\,\sigma_r(HD_2).\end{aligned}$$ From the last estimates, assuming that $\delta_{0}$ and $\delta_W$ are small enough, the claim follows, and thus the proof of in the case $r=3/2$ is finished. Next we will use that either for $r=2$ or $r=3/2$ holds. First, note that from the first inequality in one obtains $$\begin{aligned} \label{eqali3} \|{{\mathcal R}}\eta + f\|_{L^{r}(\eta)}^{r} & \gtrsim \Theta(HD_1)^{r-2}\,\sigma(HD_1) \approx \Theta(HD_1)^{r-2}\,\sigma({\operatorname{Stop}}_1({{\mathcal T}}))\\ &\geq \Theta(HD_1)^{r-2}\,A^{-1} \,\sigma(\{R\}) \approx C(A,B,r)\,\sigma_r(\{R\}).\nonumber\end{aligned}$$ Denote $$\Theta(x) = \Theta(P_x),\qquad p(x)= p(P_x),\qquad q(x) = q(P_x,{{\mathcal T}}),$$ where $P_x$ stands for the cube from ${\operatorname{Stop}}_{2}({{\mathcal T}})$ that contains $x$. From Lemma \[lemdifer22\] and the fact that $p(x)\approx\Theta(x)$, we get $$\int q(x)^r\,d\eta(x) \leq c_6 \int p(x)^r\,d\eta(x) \leq c_7 \int \Theta(x)^r\,d\eta(x) = c_7\,\sigma_r({\operatorname{Stop}}_{2}({{\mathcal T}})).$$ Then, by and , we obtain $$\label{eqsaf3} \|{{\mathcal R}}\eta + f\|_{L^r(\eta)}^r \geq C_3 \,A\int \bigl (p(x) + q(x) + C_4(A,B,r)\,\Theta(R) \bigr)^r\,d\eta(x),$$ where $C_3$ is some absolute constant. We let $G$ be the subset of the points $x\in \bigcup_{P\in{\operatorname{Stop}}_{2}({{\mathcal T}})}P$ such that $$|{{\mathcal R}}\eta(x) + f(x)|\geq \left(\frac{C_3\,A}2\right)^{1/r}\bigl (p(P_x) + q(P_x,{{\mathcal T}})+ C_6(A,B,r) \Theta(R)\bigr).$$ Note that $$\int_{G^c}|{{\mathcal R}}\eta + f|^r\,d\eta \leq \frac{C_5\,A}2\, \int_R \bigl (p(P_x) + q(P_x,{{\mathcal T}}) + C_4(A,B) \Theta(R)\bigr)^r \,d\eta.$$ Thus, from we derive $$\begin{aligned} \int_{G}|{{\mathcal R}}\eta + f|^r\,d\eta & \geq \frac{C_3\,A}2\int_R \bigl (p(P_x) + q(P_x,{{\mathcal T}}) + C_4(A,B,r) \Theta(R)\bigr)^r \,d\eta\\ &\geq C_5(A,B,r)\,\Theta(R)^r\,\eta(R).\end{aligned}$$ On the other hand, by Hölder’s inequality, $$\int_{G}|{{\mathcal R}}\eta + f|^r\,d\eta\leq \left(\int|{{\mathcal R}}\eta + f|^{2r}\,d\eta\right)^{1/2} \eta(G)^{1/2}.$$ By Lemma \[l.t1eta\], ${{\mathcal R}}\eta$ is bounded in $L^{2r}(\eta)$ with norm not exceeding $C_2(B,M)\Theta(R)$. Recalling also that $|f(x)|\lesssim p(R)\lesssim \Theta(R)$ for all $x\in R$, it follows that $$\int|{{\mathcal R}}\eta + f|^{2r}\,d\eta\leq C_3(B,M)\Theta(R)^{2r}\,\eta(R).$$ By combining the last three estimates we obtain $$\Theta(R)^r\,\eta(R) \leq C(A,B,M,r) \bigl (\Theta(R)^{2r}\,\eta(R)\bigr)^{1/2}\eta(G)^{1/2},$$ which gives $\eta(G)\geq C_{6}(A,B,M,r)\eta(R)$. Let $G$ be the collection of the cubes $P\in{\operatorname{Stop}}_{2}({{\mathcal T}})$ such that there exists some $x\in P$ satisfying $$\label{eqffi2} |{{\mathcal R}}\eta(x)+f(x)|\geq C_7 A^{1/r}\,\bigl (p(P) + q(P,{{\mathcal T}})\bigr) + C_8(A,B)\,\Theta(R),$$ where $C_7,C_8$ are the constant implicit in . By the preceding lemma, $$\mu\biggl (\bigcup_{P\in G} P\biggr)= \eta\biggl (\bigcup_{P\in G} P\biggr)\geq C(A,B,M)\,\eta(R) = C(A,B,M)\,\mu(R).$$ It is easily checked that for $x\in P\in G$, $$|{{\mathcal R}}(\chi_{P} \,\eta)(x)|\lesssim \Theta(P)\leq p(P).$$ Further, by Lemma \[lemdifer\], for all $x,y\in P$, $$|{{\mathcal R}}(\chi_{R\setminus P} \,\mu)(x) - {{\mathcal R}}(\chi_{R\setminus P} \,\eta)(y)|\lesssim p(P)+ q(P,{{\mathcal T}}).$$ Then, for all $y\in P$, by and the preceding estimate, $$\begin{aligned} \bigl|{{\mathcal R}}(\chi_{P^c} \,\mu)(y) - m_R({{\mathcal R}}\mu&)\bigr| = |{{\mathcal R}}(\chi_{R\setminus P} \,\mu)(y)+f(y)|\\ & \geq C_7 A^{1/r}\,\bigl [p(P) + q(P,{{\mathcal T}})\bigr] + C_8(A,B)\,\Theta(R) - C\,\bigl (p(P) + q(P,{{\mathcal T}})\bigr)\\ & \geq \frac{C_7}2 A^{1/r}\,p(P) + C_8(A,B)\,\Theta(R),\end{aligned}$$ assuming $A$ big enough. Recall now that by Lemma \[lemcomp\], $$\bigl|{{\mathcal R}}(\chi_{ P^c}\mu)(y)- m_P({{\mathcal R}}\mu)\bigr|\leq C_{9} p(P).$$ Therefore, if $A$ is big enough again, $$\begin{aligned} \bigl|m_P({{\mathcal R}}\mu) - m_R({{\mathcal R}}\mu)\bigr| &\geq \frac{C_7}2 A^{1/r}\,p(P) + C_8(A,B)\,\Theta(R) - C_{9}p(P) \\ & \geq C_8(A,B)\,\Theta(R).\end{aligned}$$ Since, for $x\in P\in{\operatorname{Stop}}_{2}({{\mathcal T}})$, $$\sum_{Q\in{{\mathcal T}}} D_Q({{\mathcal R}}\mu)(x) = m_P({{\mathcal R}}\mu) - m_R({{\mathcal R}}\mu),$$ we infer that $$\begin{aligned} \sum_{Q\in{{\mathcal T}}} \|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 & \geq \int_{G} \Bigl|\sum_{Q\in{{\mathcal T}}} D_Q({{\mathcal R}}\mu)\bigr|^{2}\,d\mu \\ & \geq C_8(A,B)^2\,\Theta(R)^2\,\mu(G) \geq C(A,B,M)\,\Theta(R)^2\,\mu(R).\end{aligned}$$ Proof of Main Theorem ===================== We now finish the proof of Theorem \[teopri\]. We need to prove $$\displaystyle \sigma( \mathcal D)\lesssim \sum_{Q\in \mathcal D} \|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 .$$ First note that by stopping time arguments we can make a partition of ${{\mathcal D}}$ into disjoint maximal trees, i.e. trees of type $W$, $LW$, $MDec$ and $TInc$. So we can write $$\sigma( \mathcal D)= \sigma( W)+ \sigma( LW)+ \sigma( MDec)+ \sigma( TInc),$$ where, abusing notation, we have identified $W$, $LW$, $MDec$ and $TInc$ with the cubes contained in the trees of type $W$, $LW$, $MDec$ and $TInc$, respectively. We need the following series of lemmatas. \[l.sigmamdec\] Let ${{\mathcal T}}_{0}$ be the maximal tree with root $Q^0$. Then we have $$\label{e.sigmadecro} \sigma(MDec\setminus {{\mathcal T}}_{0}) \leq C(A,B,M,\delta_{0})( \sigma( W)+ \sigma( LW)+ \sigma( TInc)) \quad \text{if } {{\mathcal T}}_{0}\in MDec,$$ and $$\label{e.sigmadec} \sigma(MDec) \leq C(A,B,M,\delta_{0})( \sigma( W)+ \sigma( LW)+ \sigma( TInc)) \quad \text{if } {{\mathcal T}}_{0}\notin MDec.$$ Recall that $Q^0$ is the largest cube from ${{\mathcal D}}$. That is, it is the starting cube in the construction of $E$. We will prove , as follows by an almost identical argument. Let ${{\mathcal T}}\in MDec$ and let $R={\operatorname{Root}}({{\mathcal T}})$. We first note that $$\label{e.estimatedec} \sigma({{\mathcal T}})\leq C(B,A) \sigma(R)$$ To prove this we define $\mathcal S_{0}({{\mathcal T}}):=\{R\}$, and in general $\mathcal S_{k}({{\mathcal T}})$ is the collection of the cubes from ${\operatorname{Top}}\cap \,{{\mathcal T}}\setminus \bigcup_{j=0}^{k-1} \mathcal S_{j}$ which are maximal. Since ${{\mathcal T}}\in MDec$, then ${\operatorname{Tree}}(R')$ is $D\sigma$ for all $R'\in {\operatorname{Top}}\cap\, {{\mathcal T}}$. Therefore, $\sigma({\operatorname{Stop}}(R'))\leq A^{-1} \sigma(R)$ and also $\sigma(\mathcal S_{k}({{\mathcal T}}))\leq A^{-1} \sigma( \mathcal S_{k-1}({{\mathcal T}}))$. Iterating this estimate we get $$\sigma(\mathcal S_{k}({{\mathcal T}}))\leq A^{-k} \sigma(R).$$ And now the proof of follows easily using Lemma \[lemdif1\]: $$\begin{aligned} \sigma({{\mathcal T}})&\leq c'(B,\delta_0,M) \sum_{R'\in {\operatorname{Top}}\cap {{\mathcal T}}} \sigma(R') \leq c'(B,\delta_0,M) \sum_{k=0}^{\infty} \sigma(\mathcal S_{k}({{\mathcal T}}))\\ &\leq c'(B,\delta_0,M) \sum_{k=0}^{\infty}A^{-k} \sigma(R)\leq c'(B,\delta_0,M,A) \sigma(R).\end{aligned}$$ Next we observe that since $R={\operatorname{Root}}({{\mathcal T}})\neq Q_{0}$ and ${{\mathcal T}}\in MDec$, then there exists a tree ${{\mathcal T}}'\in W\cup LW\cup TInc$ such that $R\in {\operatorname{Stop}}({{\mathcal T}}')$. We use this fact together with to finish the proof of the lemma. $$\begin{aligned} \sigma(MDec\setminus {{\mathcal T}}_{0}) &= \sum_{{{\mathcal T}}\in MDec\setminus {{\mathcal T}}_{0}}\sigma({{\mathcal T}})\\ &\leq c'(B,\delta_0,M,A) \sum_{\substack{R={\operatorname{Root}}({{\mathcal T}})\\ {{\mathcal T}}\in MDec\setminus {{\mathcal T}}_{0}}}\sigma(R)\\ &\leq C(B,\delta_0,M,A) \sum_{{{\mathcal T}}'\in W\cup LW\cup TInc} \sigma({{\mathcal T}}'). \end{aligned}$$ The second inequality follows from the fact that any cube $Q\in{{\mathcal D}}$ satisfies $\Theta(Q)\lesssim\Theta(Q^1)$, since the $\ell(Q)\approx\ell( Q^1)$ where $Q^1$ is the parent cube of $Q$. \[l.rodec\] Let ${{\mathcal T}}_{0}$ be the maximal tree with ${\operatorname{Root}}({{\mathcal T}}_{0})= Q^{0}$. If ${{\mathcal T}}_{0}\in MDec$, then $$\sigma({{\mathcal T}}_{0})\lesssim \| {{\mathcal R}}\mu \|_{L^2(\mu)}^2.$$ The proof of the Lemma \[l.rodec\] will require the following result, whose proof we postpone. \[l.twosepcubes\] Let $Q^{0}$ be a cube that contains the support of measure $\mu$. Suppose there exist cubes $Q_{u}, Q_{d} \subset Q^0$ such that ${\operatorname{dist}}(Q_{u}, Q_{d} )\approx \ell(Q^0 )$ and $\mu(Q_{u})\approx \mu(Q_{d})\approx \mu(Q^0 )$. Then $$\|{{\mathcal R}}\mu\|_{L^2(\mu)}^2\geq C\,\sigma(Q^0).$$ Applying Lemma \[lemtouch\] to $R':=Q^0$, we deduce that there exist $\delta$ and $\eta$ such that $$\mu\biggl(\bigcup_{Q\in {{\mathcal S}}(Q^0 )} Q\biggr)\geq \eta \mu(Q^0 ),$$ where ${{\mathcal S}}(Q^0)= \{Q\in {\operatorname{Stop}}_{0}(Q^0): \,\, \ell(Q)\geq \delta \ell(Q^0 ) \}$. Let $Q_{u}\in {{\mathcal S}}(Q^0 )$ with maximal $\mu$ measure, then $$\label{e.mass} \mu(Q_{u})\gtrsim \delta^{d}\eta \, \mu(Q^0 ).$$ We now distinguish three possible cases. In the first one, let us suppose $Q_{u} \in BR(Q^0 )$. Then part (b) of Lemma \[lemt0\] and gives $$\begin{aligned} \| D_{Q^0 }{{\mathcal R}}\mu \|_{L^2(\mu)} \mu(Q^0 )^{1/2} & \geq \int_{Q^0 }| D_{Q^0 }{{\mathcal R}}\mu |d\mu \\ & \geq \int_{Q_{u}}|m_{Q_{u}}{{\mathcal R}}\mu- m_{Q^0 }{{\mathcal R}}\mu |d\mu \\ & \geq \frac{M}{2}\Theta(Q^0 ) \mu(Q_{u}) \geq \frac{M}{2} \delta^{d}\eta \, \Theta(Q^0 ) \mu(Q^0 ).\end{aligned}$$ We use this estimate to get the desired result: $$\| {{\mathcal R}}\mu \|_{L^2(\mu)}^{2} \geq \| D_{Q^0 }{{\mathcal R}}\mu \|_{L^2(\mu)}^{2}\geq \frac{M^{2}}{4} \delta^{2d}\eta^{2}\sigma(Q^0 ).$$ In the second case, let us assume $Q_{u}\in HD(Q^0 )$. We claim that $\mu(Q_{u})\leq \frac{1}{2}\mu(Q^0 )$. Suppose that is not the case, then $$\sigma({\operatorname{Stop}}(Q^0 ))\geq \sigma(Q_{u})> \frac{B^{2}}{2}\Theta(Q^0 )^{2}\mu(Q^0 ).$$ This contradicts the fact that the tree is $\sigma$-decreasing. Therefore $\mu(Q_{u})\leq \frac{1}{2}\mu(Q^0 )$ and obviously $\mu(Q^0 \setminus Q_{u})\geq \frac{1}{2}\mu(Q^0 )$. Consider the cube $Q_{d}$ such that $Q_{d}\subset Q^0 \setminus Q_{u}$, $\ell(Q_{d})=\ell(Q_{u})$ and $\mu(Q_{d})$ is maximal. Then $\mu(Q_{d})\approx \mu(Q^0 )$ with constant depending on $\delta$ and by the separation condition ${\operatorname{dist}}(Q_{u}, Q_{d}) \approx \ell(Q_{u}) \approx \ell(Q^0 )$ with constant also depending on $\delta$. We see that the cubes $Q_{u},Q_{d}$ verify the assumptions of Lemma \[l.twosepcubes\], therefore we conclude $$\| {{\mathcal R}}\mu \|_{L^2(\mu)}^2\geq C\sigma(Q^0 ).$$ The third case is very similar to the previous one. Let us now assume that $Q_{u}\in LD(Q^0 )$, then $\Theta(Q_{u})\leq \delta_{0} \Theta(Q^0 )$, which in particular gives that $\mu(Q_{u})\leq \delta_{0} \mu(Q^0 )$. We now proceed as in the second case to obtain the desired estimate. By replacing $Q_u$ and $Q_d$ by suitable descendants, we may assume that $$\label{eqdaqo39} \ell(Q^0)\approx{\operatorname{dist}}(Q_u,Q_d)>2\bigl(\ell(Q_u) + \ell(Q_d)\bigr).$$ Let $L_{1}$ be the shortest segment that joins the cubes $Q_{u}$ and $Q_{d}$ and let us call $u$ a unit vector parallel to the segment. Let $\mathcal H_{2}$ be the hyperplane that is perpendicular to $L_{1}$ and passes through the middle point of $L_{1}$. Then $\mathcal H_{2}$ divides $Q^0 $ in two regions, $D$ and $U$, so that by $D$ contains $Q_d$ and $U$ contains $Q_u$, say. Let us denote by ${{\mathcal R}}^u$ the singular integral operator associated with the kernel $K^s(x-y)\cdot u$. By the antisymmetry of the kernel, we have $$\begin{aligned} \| {{\mathcal R}}\mu \|_{L^2(\mu)}\, \mu(D)^{1/2} & \geq \left|\int_{D}{{\mathcal R}}^{u}\mu\,d\mu\right| = \left|\int_{D}{{\mathcal R}}^{u}(\chi_{U}\mu)d\mu\right|\\ &=\int_{Q_{u}}\int_{Q_{d}} \frac{|\langle x-y, u\rangle|}{|x-y|^{s+1}}d\mu(x)d\mu(y)\\ &\gtrsim \int_{Q_{u}}\int_{Q_{d}} \frac{{\operatorname{dist}}(Q_{u},Q_{d})}{\ell(Q^0 )^{s+1}}d\mu(x)d\mu(y)\\ &= \frac{{\operatorname{dist}}(Q_{u},Q_{d})}{\ell(Q^0 )^{s+1}}\mu(Q_{u})\mu(Q_{d}) \approx \frac{\mu(Q^0 )^{2}}{\ell(Q^0 )^{s}}.\end{aligned}$$ From the above estimates the lemma follows. Theorem \[teopri\] is an immediate consequence of the preceding results. Indeed, by Lemmas \[l.sigmamdec\] and \[l.rodec\] we have $$\sigma(MDec) \leq C(A,B,M,\delta_{0})\left( \sigma( W)+ \sigma( LW)+ \sigma( TInc) + \|{{\mathcal R}}\mu \|_{L^2(\mu)}^2 \right).$$ Thus, $$\sigma({{\mathcal D}}) \leq C(A,B,M,\delta_{0})\left( \sigma( W)+ \sigma( LW)+ \sigma( TInc) + \|{{\mathcal R}}\mu \|_{L^2(\mu)}^2 \right).$$ By Lemmas \[lemtw\] and \[lemtame\] we know that $$\begin{aligned} \sigma( W)+ \sigma( LW)+ \sigma( TInc)& \leq C(A,B,\delta_0,M,\delta_W)\, \sum_{Q\in{{\mathcal D}}}\|D_Q({{\mathcal R}}\mu)\|_{L^2(\mu)}^2 \\& =C(A,B,\delta_0,M,\delta_W)\,\|{{\mathcal R}}\mu \|_{L^2(\mu)}^2,\end{aligned}$$ and so we are done. [^1]: Both authors were partially supported by grants 2009SGR-000420 (Catalonia), and MTM-2010-16232 and MTM2013-44304-P (Spain). M.C.R. was also supported by the Juan de la Cierva programme 2011. X.T. was also funded by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC Grant agreement 320501 and by 2014 SGR 75 (Catalonia).

ArXiv

--- abstract: 'For any complete $n$-dim Riemannian manifold $M^n$ with nonnegative Ricci curvature, Kapovitch and Wilking proved that any finitely generated subgroup of the fundamental group $\pi_1(M^n)$ can be generated by $C(n)$ generators. Inspired by their work, we give a quantitative proof of the above theorem and show that $C(n)\leq n^{n^{20n}} $. Our main tools are quantitative Cheeger-Colding’s almost splitting theory, and the squeeze lemma for covering groups between two Riemannian manifolds with nonnegative Ricci curvature.' address: | Department of Mathematical Sciences,\ Tsinghua University, Beijing\ P. R. China, 100084 author: - Guoyi Xu title: Local estimate of fundamental groups --- [^1] \[0em\] \[1.5em\] Introduction {#introduction .unnumbered} ============ It is well-known that any compact Riemannian manifold $(M^n, g)$ has finitely generated fundamental group. For non-compact complete Riemannian manifolds, the conclusion is not always right. For example, the surface with infinite genus has infinitely generated fundamental group. A natural question is: [For which complete Riemannian manifold $(M^n, g)$, the fundamental group $\pi_1(M^n)$ is finitely generated? ]{} Note the above example has no non-negative sectional curvature metric, to obtain the finitely generatedness of fundamental group, we may consider adding some curvature assumption on complete Riemannian manifolds. For the group $\mathbf{G}$, we define that $\mathfrak{ng}(\mathbf{G})$ is the minimal number of generators needed. In $1911$ Bieberbach proved: For complete flat Riemannian manifold $M^n$, the fundamental group $\pi_1(M^n)$ is finitely generated and $\mathfrak{ng}\big(\pi_1(M^n)\big)\leq C(n)$. Bieberbach reduced the study of fundamental groups of flat manifolds to the study of the discrete subgroup of the isometry group of $\mathbb{R}^n$, which was later developed into a more general theory about the discrete subgroups of Lie groups (see [@Rag]). Later for hyperbolic manifolds, the following theorem was obtained (see [@BGS]): Any finite volume hyperbolic manifold has finitely generated fundamental group. [The above theorem is a corollary of the following general result in Lie group theory: Any lattice $\Gamma$ of a Lie group $\mathbf{G}$ is finitely generated, where lattice means that $\Gamma$ is a discrete subgroup of $\mathbf{G}$ and $\mathbf{G}/\Gamma$ has finite volume. ]{} The study of the above non-positive curvature case, has more algebraic flavor, which has strong contrast to the following non-negative curvature case. In $1972$, Cheeger and Gromoll [@Soul] studied non-negative sectional curvature Riemannian manifolds. Among other things, they obtained the following result: If $M^n$ has sectional curvature $sec\geq 0$, then $\pi_1(M^n)$ is finitely generated. In fact, in [@Soul] it was showed that $M^n$ is homotopic to a compact totally geodesic submanifold of $M^n$, through the deformation by the gradient flow of Busemann function. The above theorem follows as a corollary of this more general structure result. In $1978$, Gromov [@Gromov-almost-flat] used Toponogov Comparison Theorem to study the fundamental group directly, and proved: For $M^n$ with sectional curvature $sec\geq 0$, $\mathfrak{ng}\big(\pi_1(M^n)\big)\leq C(n)= \frac{\mathrm{Vol}\big(\mathbb{S}^{n- 1}\big)}{\mathrm{Vol}\big(\mathbb{D}^{n- 1}(\frac{\pi}{6})\big)}$, where $\mathbb{S}^{n- 1}$ is the unit sphere in $\mathbb{R}^n$, $\mathbb{D}^{n- 1}(r)$ is the geodesic ball with radius $r$ in $\mathbb{S}^{n- 1}$. For compact Riemannian manifolds with $Rc\geq 0$, under the additional assumption of conjugate radius, Guofang Wei [@Wei] gave a uniform estimate on the generators of the fundamental group similar as in Gromov’s result. Back in $1968$, Milnor [@Milnor] proved that for complete Riemannian manifold $M^n$ with $Rc\geq 0$, any finitely generated subgroup of $\pi_1(M^n)$, is polynomial growth of order $\leq n$. Furthermore, he posed the following conjecture: \[conj Milnor\] [For complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, $\pi_1(M^n)$ is finitely generated. ]{} In the $1980s$, Peter Li [@Li] used the heat kernel and Anderson [@Anderson] used the property of covering maps, to prove Milnor conjecture for Euclidean (maximal) volume growth case respectively (moreover, they proved the fundamental group is finite in fact). In $2000$, Sormani [@Sormani] used the excess estimate of Abresch and Gromoll [@AG] on the universal cover of manifolds, successfully proved Milnor conjecture for linear (minimal) volume growth case. Also in $2000$, B. Wilking [@Wilking] used Milnor’s Theorem and the theory of discrete subgroup in Lie group to prove: If for any complete Riemannian manifold $(M^n, g)$ with $Rc(g)\geq 0$ and $\pi_1(M^n)$ is an abelian group, we have $\pi_1(M^n)$ is finitely generated; then for any complete Riemannian manifold $(N^n, \tilde{g})$ with $Rc(\tilde{g})\geq 0$, $\pi_1(N^n)$ is finitely generated. Note the proof of the above Wilking’s Theorem do not need to use Bochner formula in Riemannian geometry, only relies on the Bishop-Gromov Volume Comparison Theorem. Recently, in $2011$, V. Kapovitch and B. Wilking [@KW] proved the following local estimate of fundamental groups among other things : \[thm KW\] [For complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, any finitely generated subgroup $\Gamma$ of $\pi_1(M^n)$ satisfies $\mathfrak{ng}(\Gamma)\leq C(n)$. ]{} Their proof was inspired by Fukaya and Yamaguchi’s work [@FY], started by contradiction, used the equivariant pointed Gromov-Hausdorff convergence and reduced the problem to the study of Ricci limit space with group actions, then do the induction on the dimension of Ricci limit space. The main technical tools are Cheeger-Colding’s theory of Ricci limit space and almost rigidity results. However this proof can not give the explicit estimate of the above $C(n)$. One main purpose of this paper is to give an explicit uniform estimate of $\mathfrak{ng}(\Gamma)$ for finitely generated subgroup $\Gamma\subseteq \pi_1(M^n)$. More precisely, we prove the following theorem, which can be thought as the quantitative version of Theorem \[thm KW\]. \[thm main of xu\] [For complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, any finitely generated subgroup $\Gamma$ of $\pi_1(M^n)$ satisfies $\mathfrak{ng}(\Gamma)\leq n^{n^{20n}} $. ]{} The above upper bound $n^{n^{20n}} $ is not sharp, and our method can not provide the sharp bound either. Also, comparing the concise proof of Theorem \[thm KW\] in [@KW], our proof is sort of lengthy. One advantage of our proof is self-contained. Basically, our argument only use Bishop-Gromov Volume Comparison Theorem and the Bochner formula for complete Riemannian manifolds with $Rc\geq 0$. We try to reveal the relation between discrete isometry group actions and $Rc\geq 0$ in an intrinsic way. Even the Compactness Theorem of Gromov-Hausdorff convergence is not used, let alone the theory of Ricci limit spaces, we only need the concept of $\epsilon$-Gromov-Hausdorff approximation. There are three key ingredients of our proof. The first one tells us how to transfer from Gromov-Hausdorff approximation (geometry assumption) to almost orthonormal linear harmonic functions (analysis result). The second one is doing reverse argument, i.e. transferring analysis to geometry. And the third one studies the generating set of the covering groups through analysis and geometry. We will describe those three key ingredients in the rest of this section, and conclude the section with a sketchy description of our proof. Firstly, we hope to reveal more explicit relation between group actions and the assumption $Rc\geq 0$, through the concrete analysis on distance function, in the similar spirit of Gromov’s proof about estimate of fundamental group’s generators. This hope starts from the following classical global result of Cheeger and Gromoll [@Cheeger-Gromoll] obtained in $1971$: \[thm Cheeger and Gromoll Ricci\] [If $(M^n, g)$ is a complete $n$-dim Riemannian manifold with $Rc\geq 0$, and $M^n$ contains a line, then $M^n$ is isometric to $N^{n- 1}\times \mathbb{R}$, where $N^{n- 1}$ is a complete $(n- 1)$-dim Riemannian manifold with $Rc\geq 0$. ]{} The proof of Theorem \[thm Cheeger and Gromoll Ricci\] used the harmonic function $\mathbf{b}$ globally defined on $M^n$, which is sort of the limit of the distance function of $M^n$ (more precisely, the Busemann function). In fact, the above manifold $N^{n- 1}$ is a level set of $\mathbf{b}$ and the splitting lines $\mathbb{R}$ are the gradient flow lines of $\mathbf{b}$. One crucial technical point is to get the gradient estimate of the harmonic function $\mathbf{b}$, which is $|\nabla \mathbf{b}|\equiv 1$. In $1975$, Cheng and Yau [@CY] established the well-known local gradient estimate of harmonic functions on complete Riemannian manifolds with $Rc\geq 0$, which enables us to obtain the gradient estimate of harmonic functions from the $C^0$-bound of harmonic functions locally. On such manifolds, in $1990$, Abresch and Gromoll [@AG] obtained the excess estimate, which gives the local estimate of the sum of two local Busemann functions with respect to a segment (such sum is $0$ for two Busemann functions with reverse directions in the proof of Theorem \[thm Cheeger and Gromoll Ricci\]). Then in the $1990$s, based on the gradient estimate and the excess estimate, Colding initiated the study of local properties of distance function, through harmonic function locally defined on manifolds with Ricci lower bound, in a series paper [@Colding-shape], [@Colding-large], [@Colding-volume]; while solving several important problems in the theory of Gromov-Hausdorff convergence, which was established by Gromov in the $1980$s (for more details see [@Gromov-book]). More precisely, among other things, Colding constructed the locally defined harmonic function based on the local Busemann function with respect to a segment, and proved such harmonic functions are ‘almost linear’ harmonic functions with bounded gradient, where the ‘almost linear’ is in average integral sense. The **first** ingredient of our proof is the above existence of almost orthonormal linear harmonic functions, which was originally established by Colding in [@Colding-shape], [@Colding-large], [@Colding-volume]. For our purpose, we need the explicit quantitative version, so we give all the details of the proof here. Although some calculation is sort of tedious, these explicit estimates possibly give an intrinsic expression of the transfer from geometry (Gromov-Hausdorff approximation) to analysis (existence of almost orthonormal linear harmonic functions). In $1996$, Cheeger and Colding [@CC-Ann] established the almost splitting theorem among other things, which transfers from analysis (existence of almost orthonormal linear harmonic functions) to geometry (Gromov-Hausdorff approximation). Very roughly, they proved that if there exist $k$ almost orthonormal linear harmonic functions on a geodesic ball $B_r(p)\subset M^n$ with $Rc\geq 0$, then a smaller concentric geodesic ball is close to a ball of $\mathbb{R}^k\times \mathbf{X}_k$ in the sense of Gromov-Hausdorff distance. This almost splitting theorem is the **second** ingredient of our proof. For the same reason as the above, we need the quantitative version. During the proof of the almost splitting theorem, one crucial thing is the existence of a suitable Gromov-Hausdorff approximation under the suitable Hessian integral bound assumption. Because the original proof of this existence result in [@CC-Ann] is concise, one of our contribution is a different detailed proof by modifying some argument in [@CN], for more details see Section \[SEC splitting on Ricci limit spaces\] of this paper. It is a natural question whether the topology of two metric spaces are the same when they are very close in Gromov-Hausdorff distance sense. Generally, the answer is no, although there is an intrinsic Reifenberg type theorem when one of the spaces is $n$-dim Euclidean space and the other space is $n$-dim Riemannian manifolds (for details, see [@CC1 Appendix]). However, for a family of Riemannian manifolds converging to a metric space in Gromov-Hausdorff distance sense, using the equivariant Gromov-Hausdorff convergence theory developed by Fukaya and Yamaguchi in the $1980$s (see [@Fukaya-Japan], [@FY]), Kapovitch and Wilking [@KW] obtained the results, which relate the fundamental groups of those converging Riemannian manifolds, to the limit group, which acts on the limit space of the universal covers of those Riemannian manifolds. The study of the fundamental groups can be put into a more general context, i.e. the covering group of a covering map between two Riemannian manifolds with $Rc\geq 0$. The **third** ingredient of our proof is to characterize the change of the covering groups by the Gromov-Hausdorff distance between two metric spaces in quantitative form, where one metric space is a geodesic ball in Riemannian manifolds with $Rc\geq 0$ and the other one is a ball in a product metric space $\mathbb{R}^k\times \mathbf{X}_k$. Our squeeze lemma provides a bridge linking analysis with group actions and geometry. Very roughly, if there are $k$ almost orthonormal linear harmonic functions on a geodesic ball $B_r(p)\subset M^n$ with $Rc\geq 0$ (analysis), then from the almost splitting theorem in the second ingredient, we know that $B_{cr}(p)$ is close to $B_{cr}(0, \hat{p})\subset \mathbb{R}^k\times \mathbf{X}_k$ in Gromov-Hausdorff sense (geometry). Then the group actions on $B_{cr}(p)$ are almost generated by the group actions on $B_{cr}(\hat{p})\subset \mathbf{X}_k$, for more details see Lemma \[lem first squeeze big Eucliean directions\]. Now we describe our proof in a rough way. We start with a geodesic ball $B_r(p)\subset M^n$ with $Rc\geq 0$, firstly we use the first ingredient tool to find one almost linear harmonic function, then apply Squeeze Lemma to shrink the group action to a group action on a ball $B_{cr}(0, \hat{p})\subset \mathbf{X}_{n- 1}$. Now we apply the first ingredient tool again to find two almost orthonormal linear harmonic functions on a smaller geodesic ball $B_{r_1}(p_1)\subset M^n$, and the group action on $B_r(p)$ can be ‘controlled’ or generated by the group action on $B_{r_1}(p_1)$. We repeat the above procedure by induction on the dimension of almost orthonormal linear harmonic functions. When the dimension is $n$, the group action is shown to be trivial, and we get our conclusion. Part I. G-H approximation yields A.O.L. harmonic functions {#part-i.-g-h-approximation-yields-a.o.l.-harmonic-functions .unnumbered} ========================================================== In Part I of this paper, we will prove the following version result about analytic characterization of Gromov-Hausdorff approximation, which was implied in the argument of Cheeger and Colding in a series of papers, [@Colding-shape], [@Colding-large], [@Colding-volume], [@CC-Ann] and [@Cheeger-GAFA]. Let us recall the definition of the pointed $\epsilon$-Gromov-Hausdorff approximation. \[def pted G-H appro\] [Let $(\mathbf{X}, d_\mathbf{X}, x_0)$ and $(\mathbf{Y}, d_\mathbf{Y}, y_0)$ be two pointed metric spaces. For $\epsilon> 0$, a map $f: \big(\mathbf{X}, x_0\big)\rightarrow (\mathbf{Y}, y_0)$ is called a **pointed $\epsilon$-Gromov-Hausdorff approximation** if $$\begin{aligned} &f(x_0)= y_0\ , \quad \quad \quad \quad \mathbf{Y}\subset \mathbf{U}_{\epsilon}\big(f(\mathbf{X})\big),\nonumber \\ &\Big|d_\mathbf{Y}\big(f(x_1), f(x_2)\big)- d_\mathbf{X}(x_1, x_2)\Big|< \epsilon \ , \quad \quad \quad \quad \forall x_1, x_2\in \mathbf{X}, \nonumber \end{aligned}$$ where $\mathbf{U}_{\epsilon}\big(f(\mathbf{X})\big)\vcentcolon = \Big\{z\in \mathbf{Y}: d\big(z, f(\mathbf{X})\big)\leq \epsilon\Big\}$. For simplicity, we also call that $f$ is an $\epsilon$-Gromov-Hausdorff approximation when the base points are fixed and clear. ]{} In the rest of the paper, unless otherwise mentioned, we use $\mathbf{X}, \mathbf{Y}, \mathbf{X}_k$ to denote metric spaces. Also, we are only interested in geometry and analysis on $n$-dim manifolds with $n\geq 3$, so we will always assume the dimension of any manifolds $\mathbf{n\geq 3}$ in the rest of this paper. For our application, we need the quantitative estimate, which relate the existence of the Gromov-Hausdorff approximation to the existence of almost orthonormal linear harmonic functions. Although many results in Part I are well-known to some experts in this field, we made the contribution to establish the suitable statement and the quantitative estimates. Also we elaborate the concise argument of Cheeger-Colding to present the proof of some known results in all the details for self-contained reason, and also hope to provide a backup reference for future study, besides the original works of Cheeger-Colding. \[thm AG-triple imply one more splitting-1\] [For $B_{10r}(q)\subset (M^n, g)$ with $Rc(g)\geq 0$, any $0< \epsilon< 1, \delta= n^{-3400n^3}\epsilon^{110n}$ and integer $0\leq k\leq n$, assume there is an $(\delta r)$-Gromov-Hausdorff approximation, $f: B_{10r}(q)\rightarrow B_{10r}(0, \hat{q})\subset \mathbb{R}^k\times \mathbf{X}_k$, and $\mathrm{diam}\big(B_r(\hat{q})\big)\geq \frac{1}{4}r$ where $B_r(\hat{q})\subset \mathbf{X}_k$. Then there are harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^{k+ 1}$ defined on $B_{s}(p)\subset B_{10r}(q)$, where $s= n^{-320n^3}\epsilon^{10n}r$, such that $$\begin{aligned} \sup_{B_{s}(p)\atop i= 1, \cdots, k+ 1} |\nabla \mathbf{b}_i|\leq 1+ \epsilon \quad \quad and \quad \quad \sup_{t\leq s}\fint_{B_{t}(p)}\sum_{i, j= 1}^{k+ 1} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \epsilon . \nonumber \end{aligned}$$ ]{} Almost orthonormal local Busemann functions =========================================== Cheng-Yau’s gradient estimates was originally proved in [@CY], the following form include an explicit form of constant, which is needed in the later proof. The proof is the same as in [@CY], so we omit it. \[thm Cheng-Yau’s lemma\] [Assume $Rc(M^n)\geq 0, p\in M^n$, $B_{2R}(p)\subset M^n$, $f: B_{2R}(p)\rightarrow \mathbb{R}$ is a harmonic function and $f\in C\big(\overline{B_{2R}(p)}\big)$, then $$\begin{aligned} \sup_{B_R(p)} |\nabla f|\leq \frac{60n}{R} \sup_{B_{2R}(p)} |f| .\nonumber \end{aligned}$$ If $\Delta f= c_0\geq 0$ for $f\in C\big(\overline{B_{2R}(p)}\big)$, then $\sup\limits_{B_R(p)} |\nabla f|\leq \frac{200n(R+ 1)}{R} \Big[\sup_{B_{2R}(p)} |f|+ c_0\Big]$. ]{} Set $b^{+}(\cdot)= d(q, \cdot)- d(q, p): M^n\rightarrow \mathbb{R}$, the function $b^+$ is called the **local Busemann function** with respect to the couple points $[p, q]$. And we define the positive part of a function $f$ as $$\begin{aligned} f_+(x)= \max\{f(x), 0\} .\nonumber \end{aligned}$$ \[lem estimate of Lap b\] [Given $R> 0$, $Rc(M^n)\geq 0$ and $p, q\in M$ with $d(p, q)> 2R$, then $$\begin{aligned} \frac{1}{V\big(B_R(p)\big)}\int_{B_R(p)} |\Delta b^{+}|\leq \frac{3n}{R} , \nonumber \end{aligned}$$ where $b^{+}(\cdot)= d(q, \cdot)- d(q, p): M^n\rightarrow \mathbb{R}$. ]{} \[lem harmonic b\] [Suppose $Rc(M^n)\geq 0$, $x\in M$, and the function $\mathbf{b}$ satisfies $$\nonumber \left\{ \begin{array}{rl} \Delta \mathbf{b}&= 0 \quad \quad \quad \quad \quad on\ B_{4R}(x) \\ \mathbf{b}&= b^+ \quad \quad \quad \quad \quad on\ \partial B_{4R}(x) , \end{array} \right.$$ where $b^{+}(\cdot)= d(q, \cdot)- d(q, p): M^n\rightarrow \mathbb{R}$. Then $$\begin{aligned} \fint_{B_{4R}(x)} |\nabla (\mathbf{b}- b^+)|^2 &\leq 8R\cdot \fint_{B_{4R}(x)} |\Delta b^+| , \label{eq 1.11} \\ \fint_{B_{2R}(x)} |\nabla^2 \mathbf{b}|^2 &\leq \frac{10^8\cdot n^3}{R^2}\Big[1+ \frac{d(p, x)}{R}\Big]^2 . \label{eq 1.13}\end{aligned}$$ ]{} \[lem good choice of balls\] Let $0< \delta< \frac{1}{2}$, $R> 0$, $D> 1$ be given, suppose $Rc(M^n)\geq 0$, $\tilde{p}\in M$, and $f, h\in L^1\big(B_R(\tilde{p})\big)$ with $\fint_{B_R(\tilde{p})} |f| \leq K$ , $\fint_{B_R(\tilde{p})} |h| \leq \hat{K}$. Then there exists finite many disjoint balls $\{B_{\frac{1}{2}\delta R}(x_j)\}_{j\in \mathbf{J}}$ such that we have the following: 1. For any $j\in \mathbf{J}$, $B_{\frac{1}{2}\delta R}(x_j)\subset B_{(1- \frac{3}{2}\delta)R}(\tilde{p})$ and $B_{2\delta R}(x_j)\subset B_R(\tilde{p})$. 2. $$\nonumber \left\{ \begin{array}{rl} V\big(\cup_{i\in \mathbf{J}}B_{\delta R}(x_i)\big)&\geq \big[(1- 2\delta)^n- \frac{2\cdot 10^n}{D}\big]V\big(B_R(\tilde{p})\big) \\ \sup\limits_{i\in \mathbf{J}}\fint_{B_{2\delta R}(x_i)} |f| &\leq DK\quad \quad and \quad \quad \sup\limits_{i\in \mathbf{J}}\fint_{B_{2\delta R}(x_i)} |h| \leq D\hat{K} . \end{array} \right.$$ \[lem appr by harm func\] [Suppose $Rc(M^n)\geq 0$, and $p, q\in M$ with $d(p, q)> 2r$. For all $\epsilon> 0$, there exists finitely many balls $B_{\delta r}(x_i)\subset B_r(p), \forall i\in \mathbf{A}$, where $\mathbf{A}$ is a finite set, and harmonic functions $\mathbf{b}_i$ with $\delta= 2^{-150n}\epsilon^{18}$ such that $$\begin{aligned} &V\big(\bigcup_{i\in \mathbf{A}} B_{\delta r}(x_i)\big)\geq (1- \epsilon)V\big(B_r(p)\big)\ , \quad \quad \quad \sum_{i\in \mathbf{A}}V\big(B_{\delta r}(x_i)\big)\leq 2^{2n} V\big(B_r(p)\big) , \nonumber \\ &\fint_{B_{2\delta r}(x_i)}|\nabla (\mathbf{b}_i- b^+)| \leq \epsilon \ , \quad \quad \quad \fint_{B_{2\delta r}(x_i)} |\nabla^2 \mathbf{b}_i|\leq \frac{\epsilon}{\delta r} ,\nonumber \end{aligned}$$ where $b^{+}(\cdot)= d(q, \cdot)- d(q, p): M^n\rightarrow \mathbb{R}$. ]{} Let $SM^n$ be the unit tangent bundle of $M^n$, if $\pi: SM^n\rightarrow M^n$ is the projection map, for any $\Omega\subset SM^n$, the **Liouville measure** of $\Omega$, denoted by $V(\Omega)$, is defined by $V(\Omega)= \mu\big(\pi(\Omega)\big)\cdot V(\mathbb{S}^{n- 1})$, where $V(\mathbb{S}^{n- 1})$ is the volume of the conical Euclidean $(n- 1)$-sphere, and $\mu$ is the volume measure of $(M^n, g)$ determined by the metric $g$. \[def geodesic flow\] [For $\nu\in S_xM^n$, let $\gamma_{\nu}(\cdot)$ be the geodesic starting from $x$ with $\gamma_{\nu}'(0)= \nu$, the geodesic flow $\mathfrak{g}^t(x, \nu): [0, \infty)\times SM^n\rightarrow SM^n$ is defined by $$\begin{aligned} \mathfrak{g}^t(x, \nu)= \big(\gamma_{\nu}(t), \gamma_{\nu}'(t)\big)\ , \quad \quad \quad \quad \forall t\geq 0 .\nonumber \end{aligned}$$ ]{} \[thm Liouville’s thm\] [For any region $D\subset SM^n$ we have $V(\mathfrak{g}^tD)= V(D)$, where $\mathfrak{g}^t: SM^n\rightarrow SM^n$ is the geodesic flow on $M^n$, and the measure on $D, \mathfrak{g}^tD$ is the Liouville measure. ]{} \[lem approximate by functions\] [Let $x\in M^n$, $l, r> 0$, suppose $f\in C^{\infty}\big(B_{r+ l}(x)\big)$, and $g$ is a Lipschitz function on $M^n$, then for any $0\leq t\leq l$, $$\begin{aligned} \fint_{SB_r(x)} \Big|\big(g\circ \gamma_v\big)'(t)- \frac{g\big(\gamma_v(l)\big)- g\big(\gamma_v(0)\big)}{l}\Big|\leq \frac{2}{V\big(B_r(x)\big)} \int_{B_{r+ l}(x)}\Big[ l|\nabla^2 f|+ \big|\nabla (g- f)\big|\Big] .\nonumber \end{aligned}$$ ]{} Now we prove Colding’s integral Toponogov theorem in quantitative form. \[thm Colding-1.28\] [Suppose $Rc(M^n)\geq 0$, $p, q\in M$ with $d(p, q)> 2r$. For all $1> \epsilon> 0$, there exists $\delta= 2^{-240n}\epsilon^{18}$ such that for all $0\leq t\leq \delta$, $$\begin{aligned} \fint_{SB_{r}(p)} \Big|(b^+\circ \gamma_v)'(tr)- \frac{(b^+\circ \gamma_v)(\delta r)- (b^+\circ \gamma_v)(0)}{\delta r}\Big|\leq \epsilon ,\label{eq 2.20.1}\end{aligned}$$ where $b^+(x)= d(x, q)- d(p, q)$. ]{} \[prop approximation component is almost orthogonal\] [Given $0< \epsilon< 1, 0\leq k\leq n, B_{10r}(q)\subset (M^n, g)$ with $Rc(g)\geq 0$, any $\delta\leq n^{-1250n}\cdot \epsilon^{100n}\big(\frac{r_1}{r}\big)$ and $\frac{r_1}{r}\leq n^{-110n}\epsilon^{10n}$. Assume there is an $(\delta r)$-Gromov-Hausdorff approximation $\Phi: B_{10r}(0, \hat{q})\rightarrow B_{10r}(q)$, where $B_{10r}(0, \hat{q})\subset \mathbb{R}^k\times \mathbf{X}_k$, and there are $\hat{q}_0, \hat{q}_1^+\in B_{3r}(\hat{q})\subset \mathbf{X}_k$ satisfying $d(\hat{q}_0, \hat{q}_1^+)= r_0\in \big[\frac{1}{16}r, 2r\big]$. Then we have $$\begin{aligned} \fint_{B_{r_1}(q_1)}\sum_{i, j= 1}^{k+ 1} \big|\langle \nabla b_i^+, \nabla b_j^+\rangle- \delta_{ij}\big|\leq \epsilon ,\label{almost o.n.}\end{aligned}$$ where $\big\{\mathbf{e}_i\big\}_{i= 1}^k$ is the standard basis for $\mathbb{R}^k, q_1= \Phi(0, \hat{q}_0)$ and $$\begin{aligned} b_i^+(\cdot)&= d(\cdot, p_i^+)- d(q_1, p_i^+)\ , \quad \quad \quad \quad p_i^+= \Phi(r_0\cdot \mathbf{e}_i, \hat{q}_0)\ , \quad \quad \quad \quad 1\leq i\leq k \nonumber \\ b_{k+ 1}^+(\cdot)&= d(\cdot, p_{k+ 1}^+)- d(q_1, p_{k+ 1}^+)\ , \quad \quad \quad \quad p_{k+ 1}^+= \Phi(0, \hat{q}_1^+) .\nonumber \end{aligned}$$ ]{} Existence of almost orthonormal linear (A.O.L.) harmonic functions ================================================================== \[def excess est\] [For $q^+, q^-, p\in \mathbf{X}$, where $\mathbf{X}$ is a metric space, we say that $[q^+, q^-, p]$ is an **AG-triple on $\mathbf{X}$ with the excess $s$ and the scale $t$** if $$\begin{aligned} \mathbf{E}(p)= s \quad\quad and \quad \quad \min\big\{d(p, q^+), d(p, q^-)\big\}= t ,\nonumber \end{aligned}$$ where $\mathbf{E}(\cdot)= d(\cdot, q^+)+ d(\cdot, q^-)- d(q^+, q^-)$. ]{} For $p\in M^n, r_2\geq r_1\geq 0$, we define $A_{r_1, r_2}(p)$ as the following: $$\begin{aligned} A_{r_1, r_2}(p)= \{x\in M^n|\ r_1< d(x, p)< r_2\} . \nonumber \end{aligned}$$ We have the following Abresch-Gromoll lemma. \[lem general Abresch-Gromoll\] [On complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, assume that $[q^+, q^-, p]$ is an AG-triple with the excess $\leq \frac{1}{n}\frac{r^2}{R}$ and the scale $\geq R$, furthermore assume $R\geq 2^{2n}r$, then $\sup_{B_{r}(p)} \mathbf{E}\leq 2^6\cdot \big(\frac{r}{R}\big)^{\frac{1}{n- 1}}r$, where $\mathbf{E}(\cdot)= d(\cdot, q^+)+ d(\cdot, q^-)- d(q^+, q^-)$. ]{} The following lemma provides the existence of good cut-off function on manifolds with $Rc\geq 0$, which will be used later. \[lem good cut-off function\] [If $Rc(M^n)\geq 0$ and $p\in M^n$, for any $\tau\in (0, 1)$, there is a nonnegative smooth function $\phi: M^n\rightarrow [0, 1]$ $$\nonumber \phi(x)= \left\{ \begin{array}{rl} &1 \quad \quad \quad \quad \quad \quad \quad \quad x\in B_{\tau r}(p) \\ &0 \quad \quad \quad \quad \quad \quad \quad \quad x\notin B_{r}(p) \nonumber \end{array} \right.$$ satisfying $\sup\limits_{x\in B_r(p)} |\Delta \phi(x)|\leq \frac{10^{15}n^{10}}{\tau^{2n}(1- \tau)^8}r^{-2}$. ]{} \[lem gradient est imply Hessian est\] [On complete Riemannian manifold $M^n$ with $Rc\geq 0$, for harmonic function $\mathbf{b}$ defined on $B_r(p)$ and any $\tau\in (0, 1)$, we have $$\begin{aligned} \fint_{B_{\tau r}(p)} \big|\nabla ^2 \mathbf{b}\big| \leq \frac{10^8n^5r^{-1}}{\tau^{\frac{3n}{2}}(1- \tau)^4}\cdot \sqrt{\sup_{B_r(p)}|\nabla \mathbf{b}|+ 1}\cdot \sqrt{\fint_{B_r(p)} \big||\nabla \mathbf{b}|- 1\big|} .\nonumber \end{aligned}$$ ]{} \[thm Li-Yau harnack\] [Let $(M^n, g)$ be a complete Riemannian manifold with $Rc\geq 0$, then the heat kernel $H(x, y, t)$ satisfies $$\begin{aligned} H(x, y, t)\leq (100n)^{2n+ 2}\frac{1}{V\big(B_{\sqrt{t}}(y)\big)}\exp\Big\{-\frac{d^2(x, y)}{100t}\Big\} .\nonumber \end{aligned}$$ ]{} Now we have the following existence result of almost linear harmonic function $\mathbf{b}$ with respect to the local Busemann function $b^+$. \[lem existence of harmonic function-r\] [On complete Riemannian manifold $M^n$ with $Rc\geq 0$, assume that $[q^+, q^-, p]$ is an AG-triple with the excess $\leq \frac{4}{n}\frac{r^2}{R}$ and the scale $\geq R$, also assume $R\geq 2^{2n+ 1}r$. Then there exists harmonic function $\mathbf{b}$ defined on $B_{2r}(p)$ such that $$\begin{aligned} \sup_{B_{r}(p)} |\nabla \mathbf{b}|\leq 1+ 2^{51n^2}\big(\frac{r}{R}\big)^{\frac{1}{4(n- 1)}} \quad \quad and \quad \quad \fint_{B_{r}(p)} \big|\nabla (\mathbf{b}- b^+)\big| \leq 2^{4n}\big(\frac{r}{R}\big)^{\frac{1}{2(n- 1)}} , \nonumber \end{aligned}$$ where $b^{+}(x)= d(x, q^{+})- d(p, q^+)$. ]{} [The bound of $|\nabla \mathbf{b}|$ was only a uniform bound $C(n)$ in [@CC-Ann], it was observed in [@ChN] that $|\nabla \mathbf{b}|$ can have the improved bound $1+ \epsilon$. The argument to get $|\nabla \mathbf{b}|\leq 1+ 2^{51n^2}\big(\frac{r}{R}\big)^{\frac{1}{4(n- 1)}}$, comes from $(3.42)- (3.45)$ of [@ChN], which was suggested to us by A. Naber. ]{} On Riemannian manifolds, if there is a segment $\gamma_{p, q}$ between two points $p, q$, we can choose the middle point of the segment $\gamma_{p, q}$, denoted as $z$. Then $[p, q, z]$ is an AG-triple with the excess $0$ and the scale $\frac{1}{2}d(p, q)$. For a metric space $\mathbf{X}_k$, generally we can not find the middle point as in Riemannian manifolds. However, if there exists a suitable Gromov-Hausdorff approximation from $\mathbb{R}^k\times \mathbf{X}_k$ to manifold $M^n$ locally, the following lemma provides the existence of almost middle point and AG-triple in metric space $\mathbf{X}_k$. \[lem almost mid pt\] [For $0\leq k\leq n, 0< \delta< \frac{1}{3}$, $B_{10r}(0, \hat{q})\subset \mathbb{R}^k\times \mathbf{X}_k$ and $B_{10r}(q)\subset (M^n, g)$, if there is an $(\delta r)$-Gromov-Hausdorff approximation $$\begin{aligned} \Phi: B_{10r}(0, \hat{q})\rightarrow B_{10r}(q) , \nonumber \end{aligned}$$ then for $\hat{q}_1^+, \hat{q}_1^-\in B_r(\hat{q})\subset \mathbf{X}_k$, there exists $\hat{q}_0 \in B_{3r}(\hat{q})$ such that $$\begin{aligned} \big|d(\hat{q}_0, \hat{q}_1^+)- \frac{1}{2}d(\hat{q}_1^+, \hat{q}_1^-)\big|\leq 8\sqrt{\delta}r \quad \quad and \quad \quad \big|d(\hat{q}_0, \hat{q}_1^-)- \frac{1}{2}d(\hat{q}_1^+, \hat{q}_1^-)\big|\leq 8\sqrt{\delta}r . \label{almost mid ineq conclusion} \end{aligned}$$ And if $\sqrt{\delta}r\leq \frac{d(\hat{q}_1^+, \hat{q}_1^-)}{100}$, then $\big[p_{k+1}^+, p_{k+ 1}^-, q_1\big]$ is an AG-triple with the excess $\leq 16\sqrt{\delta}r$ and the scale $\geq \frac{1}{4}d(\hat{q}_1^+, \hat{q}_1^-)$, where $$\begin{aligned} &q_1= \Phi(0, \hat{q}_0), \quad \quad \quad p_{k+ 1}^+= \Phi(0, \hat{q}_1^+), \quad \quad \quad p_{k+ 1}^-= \Phi(0, \hat{q}_1^-) . \nonumber \end{aligned}$$ ]{} \[thm AG-triple imply one more splitting-pre\] [For $B_{10r}(q)\subset (M^n, g)$ with $Rc(g)\geq 0$ and any $\epsilon\in (0, 1)$, any $\frac{r_1}{r}\leq n^{-300n^3}\epsilon^{10n}$ and $\delta\leq n^{-2000n^3}\epsilon^{70n}\big(\frac{r_1}{r}\big)^4$. If there is an $(\delta r)$-Gromov-Hausdorff approximation for $0\leq k\leq n$ and $B_{10r}(0, \hat{q})\subset \mathbb{R}^k\times \mathbf{X}_k$, $$\begin{aligned} \Phi: B_{10r}(0, \hat{q})\rightarrow B_{10r}(q) , \nonumber \end{aligned}$$ where $\mathrm{diam}\big(B_r(\hat{q})\big)= r_0\geq \frac{1}{4}r$. Then there are harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^{k+ 1}$ defined on some geodesic ball $B_{r_1}(q_1)\subset B_{10r}(q)$, such that $$\begin{aligned} \sup_{B_{r_1}(q_1)\atop i= 1, \cdots, k+ 1} |\nabla \mathbf{b}_i|\leq 1+ \epsilon \quad \quad and \quad \quad \fint_{B_{r_1}(q_1)}\sum_{i, j= 1}^{k+ 1} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \epsilon .\nonumber \end{aligned}$$ ]{} \[lem weak type 1-1 ineq-first\] [Suppose $(M^n, g)$ has $Rc\geq 0$, for $z\in M^n$, let $f: B_R(z)\rightarrow \mathbb{R}$ be a nonnegative function and $f\in L^1\big(B_R(z)\big)$, then there exists $p\in B_{\frac{R}{2}}(z)$ such that $$\begin{aligned} \sup_{t\leq \frac{R}{2}} \fint_{B_t(p)} f\leq 15^n\fint_{B_R(z)} f .\label{small ball integral average}\end{aligned}$$ ]{} The following lemma is well known so we omit its proof here. \[lem one side G-H map imply G-H dist\] [Let $(\mathbf{X}, d_\mathbf{X}, x_0)$ and $(\mathbf{Y}, d_\mathbf{Y}, y_0)$ be two pointed metric spaces, if there is a pointed $\epsilon$-Gromov-Hausdorff approximation $f: \big(\mathbf{X}, x_0\big)\rightarrow (\mathbf{Y}, y_0)$, then there exists a pointed $(3\epsilon)$-Gromov-Hausdorff approximation $h: \big(\mathbf{Y}, y_0\big)\rightarrow (\mathbf{X}, x_0)$. ]{} \[thm AG-triple imply one more splitting\] [For $B_{10r}(q)\subset (M^n, g)$ with $Rc(g)\geq 0$, any $0< \epsilon< 1, \delta= n^{-3400n^3}\epsilon^{110n}$ and integer $0\leq k\leq n$, assume there is an $(\delta r)$-Gromov-Hausdorff approximation , $$\begin{aligned} f: B_{10r}(q)\rightarrow B_{10r}(0, \hat{q})\subset \mathbb{R}^k\times \mathbf{X}_k ,\nonumber \end{aligned}$$ and $\mathrm{diam}\big(B_r(\hat{q})\big)= r_0\geq \frac{1}{4}r$. Then there are harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^{k+ 1}$ defined on some geodesic ball $B_{s}(p)\subset B_{10r}(q)$, where $s= n^{-320n^3}\epsilon^{10n}r$, such that $$\begin{aligned} \sup_{B_{s}(p)\atop i= 1, \cdots, k+ 1} |\nabla \mathbf{b}_i|\leq 1+ \epsilon \quad \quad and \quad \quad \sup_{t\leq s}\fint_{B_{t}(p)}\sum_{i, j= 1}^{k+ 1} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \epsilon .\nonumber \end{aligned}$$ ]{} Part II. A.O.L. harmonic functions produce G-H approximation {#part-ii.-a.o.l.-harmonic-functions-produce-g-h-approximation .unnumbered} ============================================================ In Part II of this paper, we will prove the following quantitative version of almost splitting theorem, which was implied in the argument of Cheeger and Colding in a series of papers, [@Colding-shape], [@Colding-large], [@Colding-volume], [@CC-Ann] and [@Cheeger-GAFA]. For our application, we need the quantitative estimate, which relate the Gromov-Hausdorff distance to the average integral bound of almost orthonormal linear harmonic functions. Although we believe that many results in Part II are well-known to some experts in this field, but we can not find the reference providing those quantitative estimates exactly. So we elaborate the concise argument of Cheeger-Colding to present the proof in all the details for self-contained reason. \[thm existence of harmonic map components imply GH-dist is small\] [For $\epsilon> 0$ and $1\leq k\leq n$, there is $\delta= n^{-700n^4}\epsilon^{18n^4}$ such that for complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, if there exist harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^k$, defined on $B_r(p)$, satisfying $\mathbf{b}_i(p)= 0$ and $$\begin{aligned} \sup_{B_r(p)}|\nabla \mathbf{b}_i|\leq 2 \ , \quad \quad \quad \quad \fint_{B_r(p)} \sum_{i, j= 1}^k \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \delta ,\label{assump on harmonic linear function}\end{aligned}$$ then we can find a metric space $\mathbf{X}_k$ and an $(\epsilon s)$-Gromov-Hausdorff approximation $f_k= (\mathbf{b}_1, \cdots, \mathbf{b}_k, \mathcal{P}_k): B_{s}(p)\rightarrow B_{s}(0, \hat{p})\subset \mathbb{R}^k\times \mathbf{X}_k$, where $s= \frac{1}{1280} r$. ]{} [Colding and Minicozzi [@CM] gave a characterization of Gromov-Hausdorff distance through the integral estimate of Hessian of harmonic functions among other things. ]{} Existence of almost linear function and Hessian estimate {#SEC existence of linear function} ======================================================== When there is a harmonic function $\mathbf{b}$ defined locally on manifold $M^n$, with bounded gradient and the average integral of $\big||\nabla \mathbf{b}|- 1\big|$ is small enough, we will show the existence of an almost linear function, which is a generalization of linear function in $\mathbb{R}^n$. The proof of Proposition \[prop existence of almost linear function\] has close relationship with the argument in [@Cheeger-GAFA]. [For $\mathbf{X}\subset M^n, t\in \mathbb{R}$, the function $\rho(x)= d(x, \mathbf{X})+ t: M^n\rightarrow \mathbb{R}$ is called **almost linear function**, which is the generalization of function $f(x)= d(x, \mathbb{R}^{n- 1})+ t$ defined on $\mathbb{R}^n$. ]{} \[def Lip constant function\] [For Lipschitz function $f$ defined on metric space $\mathbf{M}$, the **pointwise Lipschitz constant function** $\mathscr{L}(f)$ is defined by $$\begin{aligned} \mathscr{L}(f)(z)= \varlimsup_{d(z, z')\rightarrow 0} \frac{\big|f(z)- f(z')\big|}{d(z, z')}\ , \quad \quad \quad \quad z, z'\in \mathbf{M} ,\nonumber \end{aligned}$$ and the **Lipschtiz constant** $\mathbf{L}(f)$ is defined by $\mathbf{L}(f)= \sup_{z\in \mathbf{M}} \big\{\mathscr{L}(f)(z)\big\}$. ]{} \[rem Lip implies almost differentiable\] [From the classical Rademacher theorem, the Lipschitz function is almost differentiable on manifolds, for the general argument on metric measure spaces see [@Cheeger-GAFA]. Hence when the pointwise Lipschitz constant function $\mathscr{L}(f)(z)$ appears as the integrand function in an integral, we can replace it by $|\nabla f|(z)$, and we will use this fact freely in the following argument. ]{} \[prop existence of almost linear function\] [For any $\epsilon> 0$, there is $\delta= 2^{-100n^2}\epsilon^{2n^2}$ such that for any complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, if there exists one harmonic function $\mathbf{b}$ defined on $B_r(p)$ satisfying $\sup_{B_r(p)}|\nabla \mathbf{b}|\leq 2$ and $\fint_{B_r(p)} \big||\nabla \mathbf{b}|- 1\big|\leq \delta$. Then one can find $t_0, t_1\in \mathbb{R}$ and two functions $\rho, \tilde{\rho}$ defined on $B_{\frac{r}{10}}(p)$, such that $$\begin{aligned} &\rho(x)= d\big(x, \rho^{-1}(t_0)\big)+ t_0 \ , \quad \quad \tilde{\rho}(x)= t_1- d\big(x, \tilde{\rho}^{-1}(t_1)\big)\ , \quad \forall \ x\in B_{\frac{r}{160}}(p) \nonumber \\ &\frac{1}{320}r\leq d\big(x, \rho^{-1}(t_0)\big)\leq \frac{3}{320}r\ , \quad \quad \quad \quad \forall x\in B_{\frac{r}{320}}(p) \label{need 3.6.0} \\ &\sup_{B_{\frac{1}{20}r}(p)}\big|\mathbf{b}- \rho\big|\leq \epsilon\cdot r \ , \quad \quad \quad \quad \fint_{B_{\frac{1}{20}r}(p)} \big|\nabla (\mathbf{b}- \rho)\big|\leq \epsilon \nonumber \\ &\rho(x)\leq \tilde{\rho}(x)+ \frac{\epsilon}{2} r \ , \quad \quad \quad \quad \forall x\in B_{\frac{1}{10}r}(p) .\nonumber \end{aligned}$$ ]{} Segment inequality and measure of ‘good’ points =============================================== In the proofs of this section, when the context is clear, for simplicity, we use $B_r$ instead of $B_r(p)$, similar for $B_{4r}$ etc. \[lem segment ineq\] [Assume $(M^n, g)$ is a complete Riemannian manifold with $Rc\geq 0$, then for any nonnegative function $f$ defined on $B_{2r}(p)\subset M^n$, $$\begin{aligned} \int_{B_r(p)\times B_r(p)} \Big(\int_0^{d(y_1, y_2)} f\big(\gamma_{y_1, y_2}(s)\big) ds\Big) dy_1 dy_2\leq 2^{n+ 1}r\cdot V\big(B_r(p)\big)\cdot \int_{B_{2r}(p)} f . \nonumber \end{aligned}$$ ]{} For $x\in B_{2r}(p)\subset M^n$ and a closed subset $\mathbf{X}\subseteq M^n$, we define $$\begin{aligned} \rho(x)= d\big(x, \mathbf{X}\big)+ t_0\ , \quad \quad \quad \quad \hat{\rho}(x)= d\big(x, \mathbf{X}\big) ,\nonumber \end{aligned}$$ where $t_0\in \mathbb{R}$ is some constant, we define $\mathfrak{P}(x)\in \mathbf{X}$ by $d\big(x, \mathbf{X}\big)= d\big(x, \mathfrak{P}(x)\big)$ (if there are two points $y_1, y_2$ satisfying $d\big(x, \mathbf{X}\big)= d(x, y_1)= d(x, y_2)$, then define $\mathfrak{P}(x)= y_1$ or $y_2$ freely). We assume $$\begin{aligned} 0< r\leq \hat{\rho}(x)\leq 3r\ , \quad \quad \quad \quad \forall x\in B_r(p) .\label{hat of rho assume}\end{aligned}$$ For $0< \eta< \frac{1}{2}$, we have $\big|\frac{\hat{\rho}(y)- \hat{\rho}(x)}{\hat{\rho}(x)- \eta r}\big|\leq 4$. And we also define $$\begin{aligned} \mathfrak{G}_{\rho}(x)= \big(\hat{\rho}(x), \mathfrak{P}(x)\big): B_r(p)\rightarrow \mathbb{R}\times \mathbf{X} .\nonumber \end{aligned}$$ For $x, y\in B_r(p)$, define $$\begin{aligned} &\sigma_x(s)= \gamma_{\mathfrak{P}(x), x}(s+ \eta r)\ , \quad \quad \quad \quad \tilde{\sigma}_y(s)= \sigma_y\Big(\frac{\hat{\rho}(y)- \eta r}{\hat{\rho}(x)- \eta r}s\Big)= \gamma_{\mathfrak{P}(y), y}\Big(\frac{\hat{\rho}(y)- \eta r}{\hat{\rho}(x)- \eta r}s+ \eta r\Big)\nonumber \\ &\tau_s= \gamma_{\sigma_x(s), \tilde{\sigma}_y(s)}\ , \quad \quad \quad \quad l_s= d\big(\sigma_x(s), \tilde{\sigma}_y(s)\big) .\nonumber \end{aligned}$$ \[def local integral has good control\] [For $0< \eta< \frac{1}{2}$, we define $$\begin{aligned} Q_{\eta, \mathbf{b}}^{r, \rho}&= \Big\{x\in B_r(p): \int_ {0}^{\hat{\rho}(x)- \eta r} |\nabla \mathbf{b}- \nabla \rho|\big(\sigma_x(s)\big)ds\leq \eta r\Big\} \nonumber \\ T_{\eta, \mathbf{b}}^{r, \rho}&= \Big\{x\in B_r(p): \fint_{B_r(p)} dy\Big(\int_ 0^{\hat{\rho}(x)- \eta r} \big(\int_{\gamma_{\sigma_x(s), \tilde{\sigma}_y(s)}} |\nabla^2 \mathbf{b}|\big) ds\Big)\leq \eta r\Big\} \ , \nonumber \\ T_{\eta, \mathbf{b}}^{r, \rho}(x)&= \Big\{y\in B_r(p): \int_ 0^{\hat{\rho}(x)- \eta r} \big(\int_{\gamma_{\sigma_x(s), \tilde{\sigma}_y(s)}} |\nabla^2 \mathbf{b}|\big) ds\leq \sqrt{\eta} r\Big\}\ , \quad \quad \quad \quad \forall x\in T_{\eta, \mathbf{b}}^{r, \rho} .\nonumber \end{aligned}$$ ]{} For non-negative function $f$ defined on $B_r(p)$, we define $$\begin{aligned} \mathfrak{Q}_{\eta, f}^{r, \rho}\vcentcolon = \Big\{x\in B_r(p): \int_ {0}^{\hat{\rho}(x)- \eta r} f\big(\sigma_x(s)\big)ds\leq \eta r\Big\} . \nonumber \end{aligned}$$ \[lem lower bound of good local integra set-general\] [Assume (\[hat of rho assume\]), then for any non-negative function $f$ satisfying $\fint_{B_{4r}(p)} f\leq \delta$, we have $\frac{V(\mathfrak{Q}_{\eta, f}^{r, \rho})}{V\big(B_r(p)\big)}\geq 1- 3^n\eta^{-n}\delta$. ]{} \[lem lower bound of good local integra set\] [Assume (\[hat of rho assume\]), $\mathbf{b}$ is harmonic function on $B_{16r}(p)$ satisfying $\sup_{B_{16r}(p)}|\nabla \mathbf{b}|\leq 2$ and $\fint_{B_{16r}(p)} |\nabla \mathbf{b}- \nabla \rho|\leq \delta$. Then we have $$\begin{aligned} \frac{V(Q_{\eta, \mathbf{b}}^{r, \rho})}{V\big(B_r(p)\big)}&\geq 1- (12\eta^{-1})^n\delta\quad \quad and \quad \quad \frac{V(T_{\eta, \mathbf{b}}^{r, \rho})}{V\big(B_r(p)\big)}\geq 1- n^{20n}\eta^{-n}\sqrt{\delta}\ , \nonumber \\ \frac{V\big(T_{\eta, \mathbf{b}}^{r, \rho}(x)\big)}{V\big(B_r(p)\big)}&\geq 1- \sqrt{\eta}\ , \quad \quad \quad \quad \forall x\in T_{\eta, \mathbf{b}}^{r, \rho} .\nonumber \end{aligned}$$ ]{} For $x, y\in B_r(p)$, for each $\mathbf{b}_i$ from Theorem \[thm existence of harmonic map components imply GH-dist is small\] and the corresponding $\rho_i, \mathfrak{P}_i$, we define $$\begin{aligned} \sigma_{x, i}(s)= \gamma_{\mathfrak{P}_i(x), x}(s+ \eta r) \ , \quad \quad \quad \quad i=1, \cdots, k .\nonumber \end{aligned}$$ \[def local integral has good control-2\] [For $0< \eta< \frac{1}{2}c_1$, we define $$\begin{aligned} Q_{\eta, \mathbf{b}_i}^{r, \rho_i, \rho_j}&= \Big\{x\in B_r(p): \int_ {0}^{\hat{\rho}_j(x)- \eta r} \big|\nabla (\rho_i- \mathbf{b}_i)\big|\big(\sigma_{x, j}(s)\big)ds\leq \eta r\Big\} \nonumber \\ P_{\eta, \mathbf{b}_k, \mathbf{b}_l}^{r, \rho_j}&= \Big\{x\in B_r(p): \int_ {0}^{\hat{\rho}_j(x)- \eta r} \big|\langle\nabla \mathbf{b}_k, \nabla \mathbf{b}_l \rangle \big|\big(\sigma_{x, j}(s)\big)ds\leq \eta r\Big\} . \nonumber \end{aligned}$$ ]{} \[lem lower bound of good local integra set-2\] [Assume $0< r\leq \hat{\rho}_j(x)\leq 3r$ for any $x\in B_r(p)$, if $$\begin{aligned} \fint_{B_{4r}(p)} \big|\langle \nabla \mathbf{b}_k, \nabla \mathbf{b}_l \rangle\big|\leq \delta \quad \quad and \quad \quad \fint_{B_{4r}(p)}\big|\nabla (\mathbf{b}_i- \rho_i)\big| \leq \delta ,\nonumber \end{aligned}$$ then we have $$\begin{aligned} \frac{V(Q_{\eta, \mathbf{b}_i}^{r, \rho_i, \rho_j})}{V\big(B_r(p)\big)}\geq 1- (3\eta^{-1})^n\delta \quad \quad and \quad \quad \frac{V(P_{\eta, \mathbf{b}_k, \mathbf{b}_l}^{r, \rho_j})}{V\big(B_r(p)\big)}\geq 1- (3\eta^{-1})^n\delta .\nonumber \end{aligned}$$ ]{} Quantitative almost splitting theorem {#SEC splitting on Ricci limit spaces} ===================================== The main results of this section were sort of implied in Cheeger-Colding’s work (see [@CC-Ann] and [@CC1]), however we will not follow their argument there. Instead, we adapt the argument of Colding-Naber in [@CN] to prove the main result of this section. Although our argument has close relationship with [@CC-Ann] and [@CC1], the main difference is that the angle between two segments is not involved into our argument, and the first variation formula is applied to the case of both end points are moving. \[lem property 2 of G-H appr\] [For any $0< \eta< \frac{1}{2}$, assume (\[hat of rho assume\]), $\mathbf{b}$ is harmonic function on $B_{16r}(p)$ satisfying $\sup_{B_r(p)} |\nabla \mathbf{b}|\leq 2$ and $\sup_{B_r(p)}|\mathbf{b}- \rho|\leq \eta r $. Then for any $x\in T_{\eta, \mathbf{b}}^{r, \rho}\cap Q_{\eta, \mathbf{b}}^{r, \rho}$, $y\in T_{\eta, \mathbf{b}}^{r, \rho}(x)\cap Q_{\eta, \mathbf{b}}^{r, \rho}$, we have $\Big|d(x, y)- d\big(\mathfrak{G}_{\rho}(x), \mathfrak{G}_{\rho}(y)\big)\Big|\leq 5000\eta^{\frac{1}{8}}\cdot r$. ]{} \[cor dist of points in geodesic balls\] [For any $\eta\in (0, 1)$, assume (\[hat of rho assume\]) and $$\begin{aligned} &\sup_{B_{16r}(p)} |\nabla \mathbf{b}|\leq 2\ , \quad \quad \quad \quad \sup_{B_r(p)}|\mathbf{b}- \rho|\leq \eta r \nonumber \\ &\fint_{B_{16r}(p)} |\nabla \mathbf{b}- \nabla \rho| \leq 2^{-n}\eta^{3n} , \nonumber \end{aligned}$$ then there exists $\delta_1= n^{21}\eta^{\frac{1}{2n}}$ such that $\sup\limits_{x, y\in T_{\eta, \mathbf{b}}^{r, \rho}\cap Q_{\eta, \mathbf{b}}^{r, \rho}\cap B_{(1- \delta_1)r}(p)}\Big|d(x, y)- d\big(\mathfrak{G}_{\rho}(x), \mathfrak{G}_{\rho}(y)\big)\Big|\leq n^{22}\eta^{\frac{1}{3n}}r$. ]{} Part III. Covering groups of Riemannian manifolds with $Rc\geq 0$ {#part-iii.-covering-groups-of-riemannian-manifolds-with-rcgeq-0 .unnumbered} ================================================================= Squeeze Lemma and Dimension induction on harmonic functions =========================================================== In this section, unless otherwise mentioned, we assume $\varphi: \tilde{M}^n\rightarrow M^n$ is the covering map with covering group $\Gamma$ such that $M^n= \tilde{M}^n/\Gamma$, where $(\tilde{M}^n, \tilde{g})$ and $(M^n, g)$ are two complete Riemannian manifolds and the metric $g$ is the quotient metric of $\tilde{g}$ with respect to group action of $\Gamma$. We include the definition of quotient metric here for convenience. \[def quotient metric\] [Consider a subgroup $G\subset \mathrm{Isom}(\mathbf{X})$, where $\mathbf{X}$ is a metric space, for every $\bar{x}, \bar{y}\in \mathbf{X}/G$, set the **quotient metric** $\bar{d}$ on $\mathbf{X}/G$ by: $$\begin{aligned} \bar{d}(\bar{x}, \bar{y})= \inf\big\{d(x, y): x\in \bar{x}, y\in \bar{y}\big\}= \inf\big\{d(x, gy): g\in G\big\} .\nonumber \end{aligned}$$ ]{} For any function $f$ defined on a domain $\mathbf{D}\subset M^n$, we define $$\begin{aligned} \mathcal{L}(f)(\tilde{x})\vcentcolon = f\big(\varphi(\tilde{x})\big)\ , \quad \quad \quad \quad \forall\ \tilde{x}\in \varphi^{-1}(\mathbf{D}) . \nonumber \end{aligned}$$ For $\tilde{p}, \tilde{q}\in \tilde{M}^n$ and any $s> 0$, we define $$\begin{aligned} \Gamma_{\tilde{p}}(s)= \big\{\gamma\in \Gamma|\ d(\tilde{p}, \gamma\tilde{p})\leq s\big\} .\end{aligned}$$ Similarly we can define $\Gamma_{\tilde{q}}(s)$. When the point is fixed and clear in the context, we use $\Gamma(s)$ instead of $\Gamma_{\tilde{p}}(s)$ for simplicity. The following Lemma is motivated by the use of the canonical fundamental domain $\mathcal{F}$ in [@Anderson], and is needed for the proof of the squeeze lemma. \[lem average integral on covering spaces\] [For any function $f\geq 0$ defined on $B_r(p)\subset M^n$, $\tilde{p}$ is one lift of $p$ and $B_r(\tilde{p})$ is the geodesic ball centered at $\tilde{p}$ with radius $r$ in $\tilde{M}^n$. Then $$\begin{aligned} B_r(\tilde{p})\subset \varphi^{-1}\big(B_r(p)\big)\quad \quad and \quad \quad \fint_{B_r(\tilde{p})} \mathcal{L}(f)\leq 4^n\fint_{B_r(p)} f .\nonumber \end{aligned}$$ ]{} Before we state and prove our squeeze lemma, we would like to include the following well-known result and its proof here, because the squeeze lemma can be looked at the Gromov-Hausdorff perturbation version of the following Lemma. \[lem generator length for cocompact group action\] [For a locally compact, pointed length space $(\mathcal{Z}, q)$, assume $G$ is a closed subgroup of the isometry group $\mathrm{Isom}(\mathcal{Z})$ and $\mathcal{Z}/G= \mathbf{K}$, where $\mathbf{K}$ is a compact metric space with $\mathrm{diam}(\mathbf{K})= r_0$. Then for any $\epsilon> 0$, we have $$\begin{aligned} G= \langle G(2r_0+ \epsilon)\rangle .\nonumber \end{aligned}$$ ]{} \[lem first squeeze big Eucliean directions\] [For any $\epsilon> 0, \delta= n^{-900n^4}\epsilon^{40n^4}$, integer $0\leq k\leq n$, if there exist harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^k$ defined on $B_r(p)$, satisfying $\mathbf{b}_i(p)= 0$ and $$\begin{aligned} \sup_{B_{r}(p)\atop i= 1, \cdots ,k}|\nabla \mathbf{b}_i|\leq 2\ , \quad \quad \quad \quad \sup_{t\leq r}\fint_{B_{t}(p)}\sum_{i, j= 1}^{k} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \delta ,\label{average integral assumption}\end{aligned}$$ then there is a family of $(\epsilon s)$-Gromov-Hausdorff approximation for $s\in (0, 10r_c]$, $$\begin{aligned} f_s= (\mathbf{b}_1, \cdots, \mathbf{b}_k, \mathcal{P}_s): B_s(p)\rightarrow B_s(0, \hat{p}_s)\subset \mathbb{R}^k\times \mathbf{X}_{k, s} ,\nonumber \end{aligned}$$ where $r_c= \frac{r}{12800}$. And let $\mathrm{diam}\big(B_{r_c}(\hat{p}_{10r_c})\big)= r_0$, we have $$\begin{aligned} \Gamma(r_c)\subset \big\langle \Gamma(\epsilon r_c+ 2r_0)\big\rangle . \nonumber \end{aligned}$$ ]{} \[lem at most n almost o.n. vec\] [If $\epsilon< \frac{1}{10n^2}$, there do not exist $\{\nu_i\}_{i= 1}^{n+ 1}\subset \mathbb{R}^n$ such that $$\begin{aligned} \sup_{1\leq i< j\leq n+ 1 }\big|\langle \nu_i, \nu_j\rangle\big|\leq \epsilon \quad \quad and \quad \quad \sup_{1\leq i\leq n+ 1}\big||\nu_i|- 1\big|\leq \epsilon .\nonumber \end{aligned}$$ ]{} \[prop existence of n-dim harmonic map imply close to Euclidean ball\] [For complete Riemannian manifold $(M^n, g)$ with $Rc\geq 0$, if there are harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^n$, defined on $B_r(p)$, satisfying $\mathbf{b}_i(p)= 0$ and $$\begin{aligned} \sup_{B_{r}(p)\atop i= 1, \cdots ,n}|\nabla \mathbf{b}_i|\leq 2 \ , \quad \quad \quad \quad \sup_{s\leq r}\fint_{B_{s}(p)}\sum_{i, j= 1}^{n} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq n^{-80000n^7} ,\label{n-dim harmonic map assumption}\end{aligned}$$ then $\Gamma_{\tilde{p}}(\frac{r}{12800})= \{e\}$. ]{} We define $\mathfrak{ng}\big\langle \Gamma(r) \big\rangle$ as the minimal number of generators in $\Gamma(r)$ needed to generated $\big\langle \Gamma(r)\big\rangle$. \[prop local fund group counting at close different points\] [For $\tilde{p}, \tilde{q}\in (\tilde{M}^n,\tilde{g})$ with $Rc(\tilde{g})\geq 0$, assume $d(\tilde{p}, \tilde{q})\leq \delta< \frac{s}{2}$, then $$\begin{aligned} &\mathfrak{ng}\big\langle\Gamma_{\tilde{p}}(\lambda_2)\big\rangle \leq \Big(\frac{2\lambda_2+ \lambda_1}{\lambda_1}\Big)^n\mathfrak{ng}\big\langle\Gamma_{\tilde{p}}(\lambda_1)\big\rangle\ , \quad \quad \quad \quad if\ 0< \lambda_1\leq \lambda_2\nonumber \\ &\mathfrak{ng}\big\langle\Gamma_{\tilde{p}}(s)\big\rangle\leq \Big(\frac{3s+ 2\delta}{s- 2 \delta}\Big)^n \cdot \mathfrak{ng}\big\langle\Gamma_{\tilde{q}}(s- 2\delta)\big\rangle .\nonumber \end{aligned}$$ ]{} \[prop splitting one dim by group generator\] [For any $\epsilon\in (0, 1)$ and $0\leq k< n$, there exists $\delta= (n^{-30}\epsilon)^{5000n^7}$, such that if there exist harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^k$, satisfying $$\begin{aligned} \sup_{B_r(q)\atop i= 1, \cdots ,k}|\nabla \mathbf{b}_i|\leq 2 \quad \quad and \quad \quad \sup_{s\leq r}\fint_{B_s(q)}\sum_{i, j= 1}^k \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \delta .\label{assumption 4.2.6.1} \end{aligned}$$ Then there are harmonic functions $\big\{\mathbf{b}_i\big\}_{i= 1}^{k+ 1}$ defined on $B_{r_1}(q_1)\subset B_r(q)$, such that $$\begin{aligned} &\mathfrak{ng}\big\langle \Gamma_{\tilde{q}}(r)\big\rangle\leq n^{340n^4}\epsilon^{-10n^2}\cdot \mathfrak{ng}\big\langle \Gamma_{\tilde{q}_1}(r_1)\big\rangle \label{generator squeeze est} \\ &\sup_{B_{r_1}(q_1)\atop i= 1, \cdots ,k+ 1}|\nabla \mathbf{b}_i|\leq 2 \quad \quad and \quad \quad \sup_{t\leq r_1}\fint_{B_{t}(q_1)}\sum_{i, j= 1}^{k+ 1} \big|\langle \nabla \mathbf{b}_i, \nabla \mathbf{b}_j\rangle- \delta_{ij}\big|\leq \epsilon .\nonumber \end{aligned}$$ ]{} \[thm uniform est on generators of local funda group\] [Assume $\varphi: \tilde{M}^n\rightarrow M^n$ is the covering map with covering group $\Gamma$ such that $M^n= \tilde{M}^n/\Gamma$, where $(\tilde{M}^n, \tilde{g})$ and $(M^n, g)$ are two complete Riemannian manifolds and the metric $g$ is the quotient metric of $\tilde{g}$ with respect to group action of $\Gamma$, furthermore $Rc\geq 0$, then $\mathfrak{ng}\big\langle\Gamma_{\tilde{p}}(1)\big\rangle \leq n^{n^{20n}} $ for any $\tilde{p}\in \tilde{M}^n$. ]{} \[thm f.g. group has uniform generator bound\] [Suppose $(M^n, g)$ is a complete Riemannian manifold with $Rc\geq 0$, then for any finitely generated subgroup $\Gamma$ of $\pi_1(M^n)$, we have $\mathfrak{ng}(\Gamma)\leq n^{n^{20n}}$. ]{} Acknowledgments {#acknowledgments .unnumbered} =============== The author thank Aaron Naber for helpful suggestion, which greatly improves the quantitative estimates in the earlier version of this paper. We are grateful to Jiaping Wang for continuous encouragement, William P. Minicozzi II and Christina Sormani for comments. We are indebted to Vitali Kapovitch for several email reply about [@KW], which help us to understand the results there better. [^1]: The author was partially supported by NSFC-11771230

ArXiv

--- abstract: 'In the context of personalized medicine, text mining methods pose an interesting option for identifying disease-gene associations, as they can be used to generate novel links between diseases and genes which may complement knowledge from structured databases. The most straightforward approach to extract such links from text is to rely on a simple assumption postulating an association between all genes and diseases that co-occur within the same document. However, this approach (i) tends to yield a number of spurious associations, (ii) does not capture different relevant types of associations, and (iii) is incapable of aggregating knowledge that is spread across documents. Thus, we propose an approach in which disease-gene co-occurrences and gene-gene interactions are represented in an RDF graph. A machine learning-based classifier is trained that incorporates features extracted from the graph to separate disease-gene pairs into valid disease-gene associations and spurious ones. On the manually curated *Genetic Testing Registry*, our approach yields a 30 points increase in $\text{F}_1$ score over a plain co-occurrence baseline.' author: - 'Hendrik [ter Horst]{}' - Matthias Hartung - Roman Klinger - | \ Matthias Zwick - Philipp Cimiano bibliography: - 'bibliography.bib' title: 'Predicting Disease-Gene Associations using Cross-Document Graph-based Features' --- Introduction ============ Most current approaches in personalized medicine, irrespective of particular treatment modalities (e.g., small molecules, biologics, novel approaches like gene therapy), are centered around modulating a gene in order to modulate aspects of a disease [@Kisoretal2014]. Therefore, the detection of disease-gene links is an important starting point in drug discovery. Text mining methods pose an interesting option for identifying disease-gene links, as they can be used to generate new target (and therefore treatment) hypotheses and, in combination with experimental data, support the prioritization of research aimed at the discovery of new drug targets. Until now, text mining methods for disease-gene associations mostly rely on the degree of textual co-occurrence [@quan2014gene]. While those approaches are largely reliable in detecting well-known links for well-known diseases [@al2005new; @chun2006extraction; @pletscher2014diseases], such well-known links are of minor interest in drug research, as they do not support the discovery of new targets. Such novel targets are difficult to detect, as they often require the aggregation of evidence across individual documents; at the same time, they potentially shed light on yet unknown disease-gene links. In this paper, we propose a classification model for predicting novel disease-gene associations from biomedical text. The model combines *private* (intra-document) as well as *public* (cross-document) knowledge as defined by Swanson et al.[@swanson1996undiscovered] in terms of features based on local co-occurrences within documents and relations between diseases and genes that have been aggregated across individual documents. In an evaluation against an existing database, we address the following research questions: (i) Can such a combined model outperform a purely co-occurrence-based approach? (ii) What is the impact of features measuring the connectivity of diseases and genes? (iii) What is the impact of graph-based features capturing interactions between genes across documents? System Architecture {#sec:sysarc} =================== ![Overview of system architecture[]{data-label="fig:architecture"}](architecturesimplenocolor){width="90.00000%"} Our system architecture consists of the following components (cf. Fig. \[fig:architecture\]): 1. **Medline Corpus**: All analyses are done on Medline[^1], comprising a total of 21.5M abstracts. 2. **Information Extraction:** We rely on existing information extraction systems to identify disease names, genes/proteins and interactions between them. For **disease recognition**, we use a state-of-the-art CRF tagger [@Klingeretal2007] that has been trained on the NCBI Disease Corpus [@DoganLu2012]. Using the normalization procedure described in [@terHorst2015], disease names are reduced to a vocabulary of approx. 15K unique identifiers extracted from MeSH[^2] and OMIM[^3]. For the **identification of genes/proteins and their interactions**, we rely on TEES [@bjorne2009extracting], a state-of-the-art event extraction system that has been tailored to the detection of molecular interactions from biomedical text. All extracted genes are normalized using GeNo [@geno] and afterwards filtered by human genes using taxonomic information from EntrezGene [@maglott2011entrez]. ![TEES representation of an example sentence from [@Maetal2013].[]{data-label="fig:tees-example"}](teescomplexsentence2){width="\textwidth"} 3. **Postprocessing:** The complexity of the interaction graphs produced by TEES (see Fig. \[fig:tees-example\]) is incrementally reduced. Fig. \[fig:pp-example\] gives a running example of the individual steps described in the following: 1. **RDFication:** TEES graphs are broken down into binary relations by selecting all shortest paths that connect two proteins with respect to their semantic relation. These are represented as one RDF triple connecting two proteins. The path between the two proteins is serialized as a string and used as the name of an RDF property connecting both proteins. 2. **Simplification:** Semantic role information (i.e., <span style="font-variant:small-caps;">cause</span> and <span style="font-variant:small-caps;">theme</span>) is omitted. The direction of interactions is still captured in the directed edges of the graph. 3. **Generalization:** The TEES extraction scheme, originally consisting of 9 relations, was reduced to five relations after discussion with a domain expert: *Expression*, *Catabolism*, *Localization*, *Binding* and *Regulation*. 4. **Compression:** Consecutive occurrences of identical relations within a path signature are compressed by reducing them to one relation. 5. **Path joining:** To extract longer dependencies between genes/proteins as well, we join paths connecting two genes up to a distance of two edges. The join of the paths is serialized again, the above post-processing steps are applied and the results are stored as RDF triples. We refer to such serialized paths as *path signatures*. \[ref:pathsigdef\] - HSP27 – <span style="font-variant:small-caps;">theme</span>:Neg\_reg:<span style="font-variant:small-caps;">cause</span>:Pos\_reg:<span style="font-variant:small-caps;">theme</span>:Pos\_reg:<span style="font-variant:small-caps;">cause</span> – ActD\ HSP27 – <span style="font-variant:small-caps;">theme</span>:Neg\_reg:<span style="font-variant:small-caps;">cause</span>:Pos\_reg:<span style="font-variant:small-caps;">theme</span>:Pos\_reg:<span style="font-variant:small-caps;">theme</span> – caspase3\ ActD – <span style="font-variant:small-caps;">cause</span>:Pos\_reg:<span style="font-variant:small-caps;">theme</span> – caspase3\ - HSP27 – Neg\_reg:Pos\_reg:Pos\_reg – ActD\ HSP27 – Neg\_reg:Pos\_reg:Pos\_reg – caspase3\ ActD – Pos\_reg – caspase3 - HSP27 – Reg:Reg:Reg – ActD\ HSP27 – Reg:Reg:Reg – caspase3\ ActD – Reg – caspase3 - HSP27 – Reg3 – ActD\ HSP27 – Reg3 – caspase3\ ActD – Reg – caspase3 4. **Database**: The results of all the steps described above are stored in an RDF database, *Blazegraph*[^4]. 5. **Gene Classification System:** Given a disease as input, protein candidates are classified as to whether or not they interact with that disease. The classifier relies on features extracted for each pair of disease and protein. Being our main contribution, this component is described in the next section. Gene Classification System ========================== Our gene classification system takes a disease as input and predicts, for each gene in the database, whether it interacts with the given disease or not. The classifier is implemented as a Support Vector Machine relying on features that are extracted for each disease-gene pair. Seven feature groups are defined in total, which can be divided into *co-occurrence-based (CBF)* and *graph-based (GBF)* features, as described below. Our notation is as follows: Let $D$ be the set of all diseases and $G$ the set of all genes in the database[^5] and $P$ a vocabulary of predicates denoting semantic relations between them. Then, $T$ denotes the set of all triples in the database, such that: $T \subset (D \times P \times G) = \{\langle d,p,g\rangle | p=\mathit{coocc}\} \cup \{\langle g,p,g'\rangle | p=\mathit{interact}\}$. Co-occurrence-based Features {#sec:cooccurrencebased} ---------------------------- CBF features are based on the co-occurrence between diseases and genes. We consider $T_d \subset T = \{\langle d', \mathit{coocc}, g\rangle | d'=d\}$, the set of all genes co-occurring with a particular disease $d$, and analogously $T_g \subset T = \{\langle d, \mathit{coocc}, g'\rangle | g'=g\}$. Moreover, $T_{dg} \subset T= \{ \langle d', \mathit{coocc}, g'\rangle | d'=d, g'=g\}$ denotes the set of all co-occurrences of a particular disease $d$ and a particular gene $g$. #### Entropy. We compute the entropy $H(g)$ of a gene $g$ in order to measure the specificity of $g$ in terms of the diseases it co-occurs with. If $g$ co-occurs with only a few specific diseases, this results in low entropy and high specificity. Co-occurrence with many diseases yields high entropy and low specificity. We compute the entropy of $g$ as $H(g) = -\sum_d^D p(d|g) \cdot \log_2 p(d|g)$, where $p(d|g) = |T_{dg}|/|T_g|$. Analogously, we compute the entropy/specificity $H(d)$ of a disease $d$ in terms of the genes it co-occurs with. #### Co-occurrence Frequencies. This feature group combines relative co-occurrence frequencies of a disease-gene pair $(d,g)$: $$\label{equ:coocc1}\textit{Occ}(d,g) = \frac{|T_{dg}|}{\underset{d' \in D}\max{\ |T_{d'}|}} $$ Besides the normalization given in Equation , two other alternatives are used. In all variants, $\textnormal{Occ}(d,g)$ measures the strength of the connection of $d$ and $g$. Ranging from 0 to 1, small values indicate a weak connection, whereas larger values indicate a strong connection. #### Grades. This feature group consists of two features which capture a normalized frequency of triples that contain $d$ or $g$, respectively: $\textit{Grade}(d) = |T_{d}| / \underset{d' \in D}\max{\ |T_{d'}|}$ and $\textit{Grade}(g) = |T_{g}| / \underset{g' \in G}\max{\ |T_{g'}|}$. #### Odds Ratio is used to assess the degree of association between $d$ and $g$: $$\label{equ:odds} \textit{Odds}(d,g) = \frac{|T_{dg}|\cdot(|T|-|T_{dg}|)}{(|T_{dg}|-|T_{d}|)\cdot(|T_{dg}|-|T_{g}|)}$$ The higher $\mathit{Odds}(d,g)$, the stronger the association between $d$ and $g$. $\mathit{Odds}(d,g)=0$ can only be achieved if $|T_{dg}|=0$. #### TF-IDF. In order to assess the relevance of a gene $g$ for a disease $d$, we apply the $\mathit{tfidf}$ metric from information retrieval which takes term frequency ($\mathit{tf}$) and inverted document frequency ($\mathit{idf}$) into account [@Manningetal2008]: $\mathit{tfidf}(d,g) = \mathit{tf}(d,g)\cdot\mathit{idf}(g)$. Considering a disease as a “bag of genes”, $\mathit{tf}(d,g)$ is equivalent to $|T_{dg}|$, while $\mathit{idf}(g)$ can be computed in terms of Equation (\[equ:idf\]): $$\begin{aligned} \label{equ:idf} \textit{idf}(g) &= \log\left(\frac{|D|}{\sum_{d \in D} f(d,g)}\right),\ \textnormal{where}\ f(d,g) =\left\{\begin{array}{ll} 1 & \textnormal{if}\ |T_{dg}| > 0\\ 0 & \textnormal{else} \end{array}\right.\end{aligned}$$ High values of $\mathit{tfidf}(d,g)$ indicate that $g$ is mentioned frequently in the context of $d$, but still sufficiently specific to be informative for $d$, which we expect to be indicative of a relevant association between $d$ and $g$. Graph-based Features {#sec:genenetworkbased} -------------------- In contrast to the previously described feature groups which take a disease and a gene into account, GBF features are calculated *independently* of a particular disease in that they are entirely based on the gene interaction graph, i.e., the set of triples $I \subset T = \{\langle g,\mathit{interact},g'\rangle|g,g' \in G\}$. #### Path Signatures. Each gene $g$ is described in terms of a “bag of (outgoing) path signatures”, $S_{\textnormal{out}}(g) \subset I = \{\langle g', \mathit{interact}, g''\rangle|g=g'\}$, which have been constructed by joining individual edges in the gene interaction network (cf. Section \[sec:sysarc\], postprocessing step (3e)). We use *interact* as a placeholder for all predicates constructed in this process. The strength of interaction between a pair $\langle g, g'\rangle \in S_{\textnormal{out}}(g)$ is weighted by the $\mathit{tfidf}$ metric, expressing $\mathit{tf}(g,g')$ and $\mathit{idf}(g')$ analogously to the definitions in Section \[sec:cooccurrencebased\] (cf. Equation(\[equ:idf\])). Path signatures encoding an important relation between $g$ and $g'$ in terms of a high $\mathit{tfidf}$ value are considered useful for predicting novel disease-gene associations in cases where no direct evidence from co-occurrence relations is available yet. #### Gene Connectivity. {#sec:genepresence} This feature group describes the connectivity of a gene within the graph. Analogously to $S_{\textnormal{out}}(g)$ above, we define $S_{\textnormal{in}}(g,l)$ and $S_{\textnormal{out}}(g,l)$ as lists of all incoming and outgoing signatures of path length $l$, respectively. Further, $L$ denotes the maximum path length. Based on these definitions, we count the number of incoming and outgoing signatures for each path length $1 \leq l \leq L$ separately, as given in (\[equ:presenceout\]), and by accumulation over all path lengths. In these features, higher values indicate a higher connectivity of $g$ in the network. $$\begin{gathered} \label{equ:presenceout} \textit{Out}_l(g) = \frac{|S_{\textnormal{out}}(g,l)|}{\underset{g' \in G}\max{\ |S_{\textnormal{out}}(g',l)|}} \quad \textit{In}_l(g) = \frac{|S_{\textnormal{in}}(g,l)|}{\underset{g' \in G}\max{\ |S_{\textnormal{in}}(g',l)|}}\end{gathered}$$ We also measure the ratio of outgoing and incoming signatures per gene in terms of $\mathit{IORatio}(g)=|S_{out}(g)|/|S_{in}(g)|$. If $\mathit{IORatio}(g)>1$, $g$ has a manipulating role in the network; otherwise, $g$ tends to be manipulated by other genes. Experimental Evaluation ======================= Experimental Settings --------------------- #### Gold Standard. The *Genetic Testing Registry* (GTR; [@rubinstein2012nih]) is a manually curated database for results from biomedical experiments, mostly at the intersection of Mendelian disorders and human genes. Our GTR dump contains 5,800 disease-gene associations built from 4,200 diseases and 2,800 genes. #### Training and Testing Data are created from GTR as follows: All disease-gene associations in GTR are considered as positive examples. For each disease in GTR, we additionally generate the same amount of negative training examples by pairing the disease with genes that co-occur in Medline but are not attested in GTR as a valid disease-gene association. The resulting data set is split into 80% used for training and 20% for testing. The training set contains 3,665 diseases with 1,781 negative and 1,884 positive examples (i.e., associated genes). The test set comprises 910 diseases with 440 positive and 470 negative examples. #### Experimental Procedure. We train an SVM classifier using an RBF kernel [@19cortes95support] and apply grid search for meta-parameter optimization based on the LibSVM[^6] and WEKA[^7] toolkits. The trained model is applied to the task of predicting genes that are associated with a given disease. We evaluate the model on the GTR test set described above, reporting precision, recall and $\text{F}_1$ score. Results and Discussion ---------------------- **Feature Group** **Prec.** **Rec.** **$\text{F}_1$** ------------------- ----------- ---------- ------------------ Entropy 63.2 72.5 67.5 Co-occurrence 82.4 69.0 75.1 Grade 62.8 82.6 71.4 Odds Ratio 77.9 43.8 56.1 TF-IDF 85.3 62.8 72.3 Path Signatures 66.6 75.6 70.8 Connectivity 62.3 78.3 69.4 : Evaluation results of classification models based on indiviudal feature groups (left) and feature group combinations (right) on GTR testing data[]{data-label="tab:exp-results"} **Feat. Combination** **Prec.** **Rec.** **$\text{F}_1$** ----------------------- ----------- ---------- ------------------ CBF **89.6** 79.8 84.4 CBF+Connectivity 89.1 82.2 85.5 CBF+Conn+Best50Sig 89.4 **84.9** **87.1** CBF+Conn+Best100Sig 88.1 83.3 85.7 CBF+Conn+Best200Sig 88.2 84.1 86.1 Baseline 87.7 41.5 56.3 : Evaluation results of classification models based on indiviudal feature groups (left) and feature group combinations (right) on GTR testing data[]{data-label="tab:exp-results"} Evaluation results are reported in Table \[tab:exp-results\]. The left part displays classification performance of individudal feature groups; in the right part, testing performance of feature combinations (as selected by cross-validation on the training data) are shown. The baseline refers to the performance of a single-feature classifier relying on co-occurrence counts as described in Equation . The results clearly indicate a positive impact of both cooccurrence-based and graph-based features, as all feature combinations yield an increase over the baseline in both precision and recall. The CBF combination achieves highest overall precision, whereas connectivity and signature features improve recall (at the expense of slight losses in precision). As for path signatures, it is most effective to select only a small number of individual paths. We determine the best 50 path signature features based on information gain [@kullback1951information]. In the best-performing configuration (CBF+Conn+Best50Sig), our system outperforms the co-occurrence baseline by 1.7 points in precision and 43.4 points in recall. Increasing the recall relative to the co-occurrence baseline is a key prerequisite towards our goal of discovering novel disease-gene associations. Given that the GTR gold standard is relatively small and slightly biased towards Mendelian disorders, we subject the best-performing model to another evaluation in a practical use case, as described in the next section. Case Study: Pulmonary Fibrosis ------------------------------ Correct Plausible Candidate Incorrect -------------------- ------------ --------------------- ------------ -- -- Count (Percentage) 77 (38.5%) 79 (39.5%) 44 (22.0%) : Preliminary results of manual evaluation of 200 gene candidates predicted for pulmonary fibrosis.[]{data-label="tab:casestudyres"} In this experiment, our classification model was applied to the entire Medline corpus in order to predict genes related to *pulmonary fibrosis* (PF). The resulting hits were sorted by their corpus frequency and the 200 most frequent candidates were manually evaluated by a biomedical expert (who is not a PF researcher, though). Table \[tab:casestudyres\] shows the preliminary results of this analysis: 38.5% of the predictions are unanimously correct, whereas only 22% are clear errors. The missing mass is due to candidates for which hints were found that the gene may be associated with PF through relevant mechanisms or pathways. Thus, these candidates constitute plausible hypotheses which need further investigation by a PF expert. Main sources of erroneous predictions are false co-occurrences (e.g., due to negation contexts) or false positives as produced by the gene recognition component. Some errors of the latter type may be eliminated by incorporating the filtering approach proposed by [@Hartungetal2014]. In sum, this analysis clearly shows that our system is capable of generating promising candidates worth further investigation. Related Work ============ Three types of approaches have been proposed to tackle the problem of extracting explicitly mentioned disease-associated genes (DAGs) but also generating novel hypotheses from scientific publications. First, several authors extract DAGs from existing biomedical databases such as GeneSeeker [@van2003new] or PolySearch [@cheng2008polysearch]. Pi[ñ]{}ero et al. developed DisGeNET [@pinero2015disgenet], a database quantifying the degree of disease-gene associations by a combination of different sources of evidence, with textual co-occurrence being one of the main sources. Obviously, these approaches lack the ability to discover new target hypotheses. Second, text mining techniques have been considered as an alternative and are mostly based on textual co-occurrence (sentence or document-based). Such systems can be optimized on precision [@chun2006extraction] or recall [@pletscher2014diseases]. Al-Mubaid presents a technique using various information-theoretic concepts to support the co-occurrence-based extraction [@al2005new]. Third, a promising alternative to overcome mere textual co-occurrence is to aggregate knowledge across single publications (cf. [@van2013large]) into larger interaction graphs, as we also do in our approach. Nevertheless, the knowledge extraction to build those interaction graphs often relies on text mining techniques and natural language processing methods as in the BITOLA system [@baud2003improving] or in the approaches of Wren et al. [@wren2004knowledge] and Wilkinson et al. [@wilkinson2004method]. Closely related to our approach is the one by [Ö]{}zg[ü]{}r et al. [@ozgur2008identifying] who extract interaction paths from dependency networks and rely on graph centrality measures to rank proteins for a given disease. Contrary to our model, they do not use complex features extracted from the graph and do not combine different types of features. Conclusions and Outlook ======================= In this paper, we have presented a system and a model for predicting disease-gene associations from biomedical text, using a combination of features based on disease-gene co-occurrences and gene interactions that are represented in a graph database. In a classification experiment against a manually curated database used as gold standard, we were able to demonstrate the effectiveness of both types of features, outperforming a plain co-occurrence baseline by more than 30 points in $\text{F}_1$ score. Moreover, preliminary investigation of a practical use case from pharmaceutical industry suggests that almost 80% of the candidates predicted by our model are plausible and may support pharmaceutical researchers in hypothesis generation. In future work, we will carry out a more detailed evaluation of the case study and supplement our classification approach by a ranking model that not only separates positive and negative candidates but also reflects relative differences in these candidates’ plausibility. [^1]: <http://www.ncbi.nlm.nih.gov/pubmed> [^2]: <https://www.nlm.nih.gov/mesh/> [^3]: <http://omim.org> [^4]: <http://www.blazegraph.com> [^5]: $D$ and $G$ comprise all diseases and genes recognized during preprocessing the Medline corpus (cf. Section \[sec:sysarc\]). This amounts to 7.640 diseases and 11.201 genes, in total. [^6]: <http://www.csie.ntu.edu.tw/~cjlin/libsvm> [^7]: <http://www.cs.waikato.ac.nz/ml/weka/>

ArXiv

--- abstract: 'We study open-closed orbifold Gromov-Witten invariants of 3-dimensional Calabi-Yau smooth toric DM stacks with respect to Lagrangian branes of Aganagic-Vafa type. We prove an open mirror theorem for toric Calabi-Yau 3-orbifolds, which expresses generating functions of orbifold disk invariants in terms of Abel-Jacobi maps of the mirror curves. This generalizes a conjecture by Aganagic-Vafa and Aganagic-Klemm-Vafa, proved by the first and the second authors, on disk invariants of smooth toric Calabi-Yau 3-folds.' address: - 'Bohan Fang, Department of Mathematics, Columbia University, 2990 Broadway, New York, NY 10027' - 'Chiu-Chu Melissa Liu, Department of Mathematics, Columbia University, 2990 Broadway, New York, NY 10027' - 'Hsian-Hua Tseng, Department of Mathematics, Ohio State University, 100 Math Tower, 231 West 18th Avenue, Columbus, OH 43210' author: - Bohan Fang - 'Chiu-Chu Melissa Liu' - 'Hsian-Hua Tseng' date: 'December 25, 2012' title: 'Open-Closed Gromov-Witten Invariants of 3-dimensional Calabi-Yau Smooth Toric DM Stacks' --- Introduction ============ Open Gromov-Witten invariants of toric Calabi-Yau 3-folds have been studied extensively by both mathematicians and physicists. They correspond to “A-model topological open string amplitudes” in the physics literature and can be interpreted as intersection numbers of certain moduli spaces of holomorphic maps from bordered Riemann surfaces to the 3-fold with boundaries in a Lagrangian submanifold. The physics prediction of these open Gromov-Witten invariants comes from string dualities: [*mirror symmetry*]{} relates the A-model topological string theory of a Calabi-Yau 3-fold $X$ to the B-model topological string theory of the mirror Calabi-Yau 3-fold $X^\vee$; the [*large $N$ duality*]{} relates the A-model topological string theory of Calabi-Yau 3-folds to the Chern-Simons theory on 3-manifolds. Open GW invariants of smooth toric CY 3-folds --------------------------------------------- Aganagic-Vafa [@AV00] introduce a class of Lagrangian submanifolds in smooth toric Calabi-Yau 3-folds, which are diffeomorphic to $S^1\times { \mathbb{R} }^2$. By mirror symmetry, Aganagic-Vafa and Aganagic-Klemm-Vafa [@AV00; @AKV02] relate genus zero open GW invariants (disk invariants) of a smooth toric Calabi-Yau 3-fold $X$ relative to such a Lagrangian submanifold $L$ to the classical Abel-Jacobi map of the mirror Calabi-Yau 3-fold $X^\vee$, which can be further related to the Abel-Jacobi map to the mirror curve of $X$. This conjecture is proved in full generality in [@FL]. By the large $N$ duality, Aganagic-Klemm-Mariño-Vafa propose the topological vertex [@AKMV], an algorithm of computing all genera generating functions $F_{\beta',\mu_1, ..., \mu_h}(\lambda)$ of open Gromov-Witten invariants of $(X,L)$ obtained by fixing a topological type of the map (determined by the degree $\beta'\in H_2(X,L;{ \mathbb{Z} })$ and winding numbers $\mu_1,\ldots, \mu_h\in H_1(L;{ \mathbb{Z} })={ \mathbb{Z} }$) and summing over the genus of the domain. The algorithm of the topological vertex is proved in full generality in [@MOOP]. Bouchard-Klemm-Mariño-Pasquetti propose the remodeling conjecture [@BKMP1], an algorithm of constructing the $B$-model topological open string amplitudes in all genera of $X^\vee$ following [@Mar], using Eynard-Orantin’s recursive relation from the theory of matrix models [@EO07]. Combined with the mirror symmetry prediction, this gives an algorithm of computing generating functions $F_{g,h}(Q,X_1,\ldots,X_h)$ of open Gromov-Witten invariants of $(X,L)$ obtained by fixing a topological type of the domain (determined by the genus $g$ and number $h$ of boundary circles) and summing over the topological types of the map. The remodeling conjecture is proved in full generality very recently [@EO12]. Open GW invariants for toric CY 3-orbifolds ------------------------------------------- There have been attempts to generalize the above results to 3-dimensional Calabi-Yau smooth toric DM stacks. The closed GW theory of orbifolds has been studied for a long time. The physics literature dates back to early 1990s such as [@CV; @Za], which study the quantum cohomology ring of orbifolds. The mathematical definition is given by Chen-Ruan [@CR02] for symplectic orbifolds and by Abramovich-Graber-Vistoli [@AGV02; @AGV08] for Deligne-Mumford stacks. A toric Calabi-Yau 3-orbifold is a 3-dimensional Calabi-Yau smooth toric DM stack with trivial generic stabilizer. The concept of Aganagic-Vafa branes can be extended to the setting of 3-dimensional Calabi-Yau smooth toric DM stacks. These branes are diffeomorphic to $[(S^1\times { \mathbb{R} }^2)/G]$ where $G$ is a finite abelian group. The open Gromov-Witten invariants of 3-dimensional Calabi-Yau smooth toric DM stacks are defined via localization [@Ro11], generalizing the methods in [@KL01]. By localization, open and closed Gromov-Witten invariants of a smooth toric Calabi-Yau 3-fold can be obtained by gluing the Gromov-Witten vertex, a generating function of open Gromov-Witten invariants of ${ \mathbb{C} }^3$, which can be reduced to a generating function of certain cubic Hodge integrals [@DF]. Similarly, open and closed orbifold Gromov-Witten invariants of a 3-dimensional Calabi-Yau smooth toric DM stack can be obtained by gluing the orbifold Gromov-Witten vertex, a generating function of open Gromov-Witten invariants of $[{ \mathbb{C} }^3/G]$ (where $G$ is a finite abelian group acting trivially on $dz_1\wedge dz_2\wedge dz_3$), which can be reduced to a generating function of certain cubic abelian Hurwitz-Hodge integrals [@Ro11]. The Gromov-Witten vertex has been evaluated in the full 3-leg case [@LLLZ; @MOOP], but the orbifold Gromov-Witten vertex has only been evaluated in the 1-leg case for $[{ \mathbb{C} }^2/{ \mathbb{Z} }_n]\times { \mathbb{C} }$ in [@Zo11; @Zo12; @Ro12], where ${ \mathbb{C} }^2/{ \mathbb{Z} }_n$ is the $A_n$ surface singularity. As for mirror symmetry, a mirror theorem for disk invariants of $[\mathbb{C}^3/\mathbb{Z}_4]$ is proved in [@BC11]. The remodeling conjecture is also expected to predict higher genus open Gromov-Witten invariants of toric Calabi-Yau 3-orbifolds via mirror symmetry [@BKMP1; @BKMP2]. Summary of results ------------------ In this paper we study open-closed orbifold Gromov-Witten invariants of a 3-dimensional Calabi-Yau smooth toric DM stack ${\mathcal{X}}$ relative to a Aganagic-Vafa A-brane ${\mathcal{L}}$, and prove a mirror theorem for disk invariants of arbitrary toric Calabi-Yau 3-orbifolds. Open Gromov-Witten invariants of the pair $({\mathcal{X}}, {\mathcal{L}})$ count holomorphic maps from orbicurves to ${\mathcal{X}}$ with boundaries mapped to ${\mathcal{L}}$. Morally, the moduli space of such maps are characterized by the following data: - topological type $(g,h)$ of the domain orbicurve $(\Sigma,\partial \Sigma)$ , where $g$ is the genus and $h$ is the number of boundary holes; - number of interior marked points $n$; - topological type of the map $u:({\Sigma},{\partial{\Sigma}}=\coprod_{i=1}^h R_i )\to ({\mathcal{X}},{\mathcal{L}})$ given by the degree $\beta' =u_*[{\Sigma}]\in H_2({\mathcal{X}},{\mathcal{L}};{ \mathbb{Z} })$ and each $[u_*(R_i)]\in H_1({\mathcal{L}};{ \mathbb{Z} })$, collectively denoted by ${ {\vec{\mu}} }=([u_*(R_1)],\dots, [u_*(R_h)])$; - framing $f\in { \mathbb{Z} }$ of the Aganagic-Vafa A-brane ${\mathcal{L}}$. We denote this moduli space by ${\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}|\beta',{ {\vec{\mu}} })$. If we use the evaluation maps ${\mathrm{ev}}_i$, $i=1,\ldots, n$ at interior points to pull back classes in $H^*_{ {\mathrm{orb}} }({\mathcal{X}})$, we obtain open-closed Gromov-Witten invariants. More precisely, for any framed Aganagic-Vafa brane $({\mathcal{L}},f)$, where $f\in { \mathbb{Z} }$ there is a canonically associated orbidisk which passes through a torus fixed stacky point $p_{ {\sigma} }= BG_{ {\sigma} }$ in ${\mathcal{X}}$. Given $\gamma_1,\ldots, \gamma_n\in H_{ {\mathrm{orb}} }^*({\mathcal{X}};{ \mathbb{Q} })$, we define open-closed orbifold Gromov-Witten invariant $\langle \gamma_1,\ldots,\gamma_n\rangle^{{\mathcal{X}},({\mathcal{L}},f)}_{g,\beta',{ {\vec{\mu}} }}$ via localization using a circle action determined by the framing $f$; this is a rational number depending on $f$ and can be viewed as an equivariant invariant. For each topological type $(g,h)$ of the domain bordered Riemann surface, we define a generating function $F_{g,h}^{{\mathcal{X}},({\mathcal{L}},f)}$ of open-closed Gromov-Witten invariants which takes value in $H^*_{ {\mathrm{orb}} }(p_{ {\sigma} };{ \mathbb{C} })^{\otimes h}$, where $H^*_{ {\mathrm{orb}} }(p_{ {\sigma} };{ \mathbb{C} })=\oplus_{k\in G_{ {\sigma} }}{ \mathbb{C} }\mathbf 1_k$. In particular, the disk potential $F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}$ takes values in $H^*_{ {\mathrm{orb}} }(p_{ {\sigma} };{ \mathbb{C} })$. In case that ${\mathcal{L}}$ is an outer brane[^1], we denote ${ {\vec{\mu}} }=(\mu,k)\in H_1({\mathcal{L}};{ \mathbb{Z} })\subset { \mathbb{Z} }\times G_\sigma$. The disk potential $F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}$ is the following: $$F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}({\boldsymbol{\tau}}_2, X)=\sum_{\beta',n\geq 0} \sum_{(\mu, k)\in H_1({\mathcal{L}};{ \mathbb{Z} })} \frac{\langle ({\boldsymbol{\tau}}_2)^n\rangle^{{\mathcal{X}},({\mathcal{L}},f)}_{0,\beta',(\mu,k)}}{n!} \cdot X^\mu \mathbf 1_{k}\, .$$ In this paper, we prove a mirror theorem regarding $F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}$ when ${\mathcal{X}}$ is a semi-projective toric Calabi-Yau 3-orbifold. Mirror symmetry relates the topological A-model string theory to its mirror topological B-model string theory. When the A-model theory is a semi-projective toric Calabi-Yau 3-fold, its mirror is given by a Calabi-Yau hypersurface in ${ \mathbb{C} }^2\times ({ \mathbb{C} }^*)^2$ given by a equation $uv=H(x,y,q)$, where $(u,v,x,y)\in { \mathbb{C} }^2\times ({ \mathbb{C} }^*)^2$ and $q$ is the complex moduli parametrizing the B-model. The function $H(x,y,q)$ is determined by both the combinatorial toric data of ${\mathcal{X}}$ and the framed brane $({\mathcal{L}},f)$. The affine curve $H(x,y,q)=0$ is called the *mirror curve*. Let $y=y(x,q)$ be the solved function from the mirror curve. The B-model disk potential $W_{H,{\mathrm{inst}}}(x,q)$ is defined to be the power series part of the following Abel-Jacobi integral $$\int \log y(x,q) \frac{dx}{x}= \text{logarithm part}+W_{H,{\mathrm{inst}}}(x,q).$$ We define a ${ \mathbb{C} }$-linear function $L^{{\mathcal{L}},f}: H^*_{ {\mathrm{orb}} }(p_{ {\sigma} };{ \mathbb{C} })\to { \mathbb{C} }$, which sends each $1_k$ ($k\in G_{ {\sigma} }$) to a root of unity, and prove a [*mirror theorem*]{} for the disk potential: Under the open-closed mirror map ${\boldsymbol{\tau}}={\boldsymbol{\tau}}(q)$ and $X=X({\boldsymbol{\tau}},q)$, the complex valued function $$L^{{\mathcal{L}},f}(F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}({\boldsymbol{\tau}}(q),X(x,q)))=W_{H,{\mathrm{inst}}}(x,q).$$ There are other open Gromov-Witten invariants relative to different types of Lagrangian submanifolds. Jake Solomon defines open Gromov-Witten invariants of a compact symplectic manifold relative to a Lagrangian submanifold which is the fixed locus of an anti-symplectic involution [@So]. The mirror theorem for disk invariants for the quintic 3-fold relative to the real quintic is conjectured in [@Wa] and proved in [@PSW]. It has been generalized to compact Calabi-Yau 3-folds which are projective complete intersections [@PZ], where a mirror theorem for genus one open Gromov-Witten invariants (annulus invariants) is also proved. Open orbifold Gromov-Witten invariants of compact toric orbifolds with respect to Lagrangian torus fibers have recently been defined in [@CP], which generalizes the work of [@FOOO] on compact toric manifolds. There are mirror theorems on disk invariants in this context (see [@CLLT], [@CLT], and [@CCLT]). The rest of the paper is organized as follows. In Section 2 we review the necessary materials concerning toric DM stacks. In Section 3 we apply localizations to relate open-closed Gromov-Witten invariants and descendant Gromov-Witten invariants of 3-dimensional Calabi-Yau toric DM stacks. In Section 4 we prove a mirror theorem for orbifold disk invariants. Acknowledgments {#acknowledgments .unnumbered} --------------- The first author would like to thank Kwokwai Chan, Naichung Conan Leung and Yongbin Ruan for valuable discussions. The research of the first author is partially supported by NSF DMS-1206667. The research of the second author is partially supported by NSF DMS-1159416. Smooth Toric DM Stacks ====================== In this section, we follow the definitions in [@Ir09 Section 3.1], with slightly different notation. We work over ${ \mathbb{C} }$. Definition ---------- Let $N$ be a finitely generated abelian group, and let $N_{ \mathbb{R} }=N\otimes_{ \mathbb{Z} }{ \mathbb{R} }$. We have a short exact sequence of (additive) abelian groups: $$0 \to N_{ {\mathrm{tor}} }\to N \to \bar{N}=N/N_{ {\mathrm{tor}} }\to 0,$$ where $N_{ {\mathrm{tor}} }$ is the subgroup of torsion elements in $N$. Then $N_{ {\mathrm{tor}} }$ is a finite abelian group, and $\bar{N}={ \mathbb{Z} }^n$, where $n=\dim_{ \mathbb{R} }N_{ \mathbb{R} }$. The natural projection $N\to \bar{N}$ is denoted $b\mapsto \bar{b}$. A *smooth toric DM stack* is an extension of toric varieties [@Fu93; @BCS05]. A smooth toric DM stack is given by the following data: - vectors $b_1,\dots, b_{r'} \in N$. We require the subgroup $\oplus_{i=1}^{r'} { \mathbb{Z} }b_i$ is of finite index in $N$. - a simplicial fan ${\Sigma}$ in $N_{ \mathbb{R} }$ such that the set of $1$-cones is $$\{\rho_1,\dots,\rho_{r'}\},$$ where $\rho_i={ \mathbb{R} }_{\ge 0} b_i$, $i=1,\dots, r'$. The datum $\mathbf {\Sigma}=({\Sigma}, (b_1,\dots, b_{r'}))$ is the *stacky fan* in the sense of [@BCS05]. The vectors $b_1,\dots, b_{r'}$ may not generate $N$. We choose *additional* vectors $b_{r'+1},\dots, b_r$ such that $b_1,\dots, b_r$ generate $N$. There is a surjective group homomorphism $$\begin{aligned} \phi: & { {\widetilde{N}} }:=\oplus_{i=1}^r { \mathbb{Z} }{ {\tilde{b}} }_i & {\longrightarrow}N,\\ & { {\tilde{b}} }_i & \mapsto b_i.\end{aligned}$$ Define ${\mathbb{L}}:={\mathrm{Ker}}(\phi) \cong { \mathbb{Z} }^k$, where $k:=r-n$. Then we have the following short exact sequence of finitely generated abelian groups: $$\label{eqn:NtN} 0\to {\mathbb{L}}\stackrel{\psi }{{\longrightarrow}} { {\widetilde{N}} }\stackrel{\phi}{{\longrightarrow}} N\to 0.$$ Applying $ - \otimes_{ \mathbb{Z} }{ \mathbb{C} }^*$ to , we obtain an exact sequence of abelian groups: $$\label{eqn:bT} 1 \to K \to G \to { {\widetilde{{ \mathbb{T} }}} }\to { \mathbb{T} }\to 1,$$ where $$\begin{aligned} { \mathbb{T} }&:=& N\otimes_{ \mathbb{Z} }{ \mathbb{C} }^* =\bar{N}\otimes_{ \mathbb{Z} }{ \mathbb{C} }^* \cong ({ \mathbb{C} }^*)^n,\\ { {\widetilde{{ \mathbb{T} }}} }&:=& { {\widetilde{N}} }\otimes_{ \mathbb{Z} }{ \mathbb{C} }^* \cong ({ \mathbb{C} }^*)^r,\\ G &:=& {\mathbb{L}}\otimes_{ \mathbb{Z} }{ \mathbb{C} }^* \cong ({ \mathbb{C} }^*)^k,\\ K &:=& { {\mathrm{Tor}} }_1^{{ \mathbb{Z} }}(N,{ \mathbb{C} }^*)\cong N_{ {\mathrm{tor}} }.\end{aligned}$$ The action of ${ {\widetilde{{ \mathbb{T} }}} }$ on itself extends to a ${ {\widetilde{{ \mathbb{T} }}} }$-action on ${ \mathbb{C} }^r = {\mathrm{Spec}}{ \mathbb{C} }[Z_1,\dots, Z_r]$. $G$ acts on ${ \mathbb{C} }^r$ via the group homomorphism $G\to { {\widetilde{{ \mathbb{T} }}} }$ in , so $K\subset G$ acts on ${ \mathbb{C} }^r$ trivially. With the above preparation, we are now ready define the smooth toric DM stack ${\mathcal{X}}$. Let $${\mathcal{A}}=\{I\subset \{1,\dots, r\}: \text{$\sum_{i\notin I} { \mathbb{R} }_{\ge 0} b_i$ is a cone of ${\Sigma}$}\}$$ be the set of anti-cones. Given $I\in {\mathcal{A}}$, let ${ \mathbb{C} }^I$ be the subvariety of ${ \mathbb{C} }^r$ defined by the ideal in ${ \mathbb{C} }[Z_1,\ldots, Z_r]$ generated by $\{ Z_i \mid i\in I\}$. Define the smooth toric DM stack ${\mathcal{X}}$ as the stack quotient $${\mathcal{X}}:=[U_{\mathcal{A}}/ G],$$ where $$U_{\mathcal{A}}:={ \mathbb{C} }^r \backslash \bigcup_{I\notin \mathcal A} { \mathbb{C} }^I.$$ ${\mathcal{X}}$ contains the DM torus ${\mathcal{T}}:= [{ {\widetilde{{ \mathbb{T} }}} }/G]$ as a dense open subset, and the ${ {\widetilde{{ \mathbb{T} }}} }$-action on ${\mathcal{U}}_{\Sigma}$ descends to a ${\mathcal{T}}$-action on ${\mathcal{X}}$. The smooth toric DM stack ${\mathcal{X}}$ is a [*toric orbifold*]{} if the $G$-action on ${ {\widetilde{{ \mathbb{T} }}} }$ is free. The purpose of introducing additional vector $b_{r'+1},\dots, b_{r}$ is to ensure $G$ is a *connected* torus. The stacky fan $\mathbf {\Sigma}$ together with the extra vectors $b_{r'+1},\dots, b_r$ is an *extended stacky fan* in the sense of Jiang [@Ji08]. It follows from the definition that $\{r'+1,\dots, r\}\subset I$ for any $I\in \mathcal A$. An element of $\mathcal A$ is usually referred as an “anti-cone". We introduce the following character lattices: $$\begin{aligned} M &=& {\mathrm{Hom}}(N,{ \mathbb{Z} }) = {\mathrm{Hom}}({ \mathbb{T} },{ \mathbb{C} }^*), \\ { {\widetilde{M}} }&=& {\mathrm{Hom}}({ {\widetilde{N}} },{ \mathbb{Z} })= {\mathrm{Hom}}({ {\widetilde{{ \mathbb{T} }}} },{ \mathbb{C} }^*),\\ {\mathbb{L}}^\vee &=& {\mathrm{Hom}}({\mathbb{L}},{ \mathbb{Z} }) ={\mathrm{Hom}}(G,{ \mathbb{C} }^*).\end{aligned}$$ Applying ${\mathrm{Hom}}(-,{ \mathbb{Z} })$ to , we obtain the following exact sequence of (additive) abelian groups: $$0 \to M \stackrel{\phi^\vee}{\to} { {\widetilde{M}} }\stackrel{\psi^\vee}{\to} {\mathbb{L}}^\vee \to {\mathrm{Ext}}^1(N,{ \mathbb{Z} }) \to 0$$ Therefore, the group homomorphism $$\psi^\vee: { {\widetilde{M}} }\to {\mathbb{L}}^\vee$$ is surjective if and only if $N_{ {\mathrm{tor}} }=0$. We now consider a class of examples of 3-dimensional Calabi-Yau smooth toric DM stacks of the form $[{ \mathbb{C} }^3/{ \mathbb{Z} }_3]$. Let $\omega=e^{\frac{2\pi\sqrt{-1}}{3}}$ be the generator of ${ \mathbb{Z} }_3$. Given $i,j,k\in \{0,1,2\}$ such that $i+j+k\in 3{ \mathbb{Z} }$, we define ${\mathcal{X}}_{i,j,k}$ to be the quotient stack of the following ${ \mathbb{Z} }_3$-action on ${ \mathbb{C} }^3$: $$\omega \cdot (Z_1,Z_2, Z_3) = (\omega^i Z_1, \omega^j Z_2, \omega^k Z_3).$$ In the following example, we consider $${\mathcal{X}}_{1,1,1},\quad {\mathcal{X}}_{1,2,0}=[{ \mathbb{C} }^2/{ \mathbb{Z} }_3]\times { \mathbb{C} },\quad {\mathcal{X}}_{0,0,0}={ \mathbb{C} }^3\times B{ \mathbb{Z} }_3.$$ \[ex:definition\] 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$. The toric data are given as follows. $$\begin{gathered} N={ \mathbb{Z} }^3,\quad N_{ {\mathrm{tor}} }=0;\\ b_1=(1,0,1),b_2=(0,1,1),b_3=(-1,-1,1),b_4=(0,0,1);\\ r=4,r'=3,k=1;\\ {\Sigma}=\{\text{the $3$-cone spanned by $\{b_1,b_2,b_3\}$, and its faces, and faces of faces, etc.}\};\\ {\mathcal{A}}=\{I\subset\{1,2,3,4\}: 4\in I\};\\ {\mathbb{L}}\cong { \mathbb{Z} },\quad {\mathbb{L}}^\vee\cong { \mathbb{Z} };\\\end{gathered}$$  2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$, transversal $A_2$-singularity. The toric data are given as follows. $$\begin{gathered} N={ \mathbb{Z} }^3,\quad N_{ {\mathrm{tor}} }=0;\\ b_1=(1,0,1),b_2=(0,3,1),b_3=(0,0,1),b_4=(0,1,1), b_5=(0,2,1);\\ r=5,r'=3,k=2;\\ {\Sigma}=\{\text{the $3$-cone spanned by $\{b_1,b_2,b_3\}$, and its faces, and faces of faces, etc.}\};\\ {\mathcal{A}}=\{I\subset\{1,2,3,4,5\}: \{ 4, 5\} \subset I \};\\ {\mathbb{L}}\cong { \mathbb{Z} }^2,\quad {\mathbb{L}}^\vee\cong { \mathbb{Z} }^2.\end{gathered}$$  3. ${\mathcal{X}}={\mathcal{X}}_{0,0,0}$. The toric data is given as follows. $$\begin{gathered} N={ \mathbb{Z} }^3\oplus { \mathbb{Z} }_3 \quad N_{ {\mathrm{tor}} }={ \mathbb{Z} }_3;\\ b_1=(1,0,0,0),b_2=(0,1,0,0),b_3=(0,0,1,0),b_4=(1,0,0,1);\\ r=4,r'=3,k=1;\\ {\Sigma}=\{\text{the $3$-cone spanned by $\{b_1,b_2,b_3\}$, and its faces, and faces of faces, etc.}\};\\ {\mathcal{A}}=\{I\subset\{1,2,3,4\}: 4\in I \};\\ {\mathbb{L}}\cong { \mathbb{Z} },\quad {\mathbb{L}}^\vee\cong { \mathbb{Z} }.\end{gathered}$$ Equivariant line bundles and torus-invariant Cartier divisors ------------------------------------------------------------- A character $\chi\in { {\widetilde{M}} }$ gives a ${ {\widetilde{{ \mathbb{T} }}} }$-action on ${ \mathbb{C} }^r \times { \mathbb{C} }$ by $$(\tilde{t}_1,\ldots, \tilde{t}_r)\cdot (Z_1,\ldots, Z_r, u) =(\tilde{t}_1 Z_1,\ldots, \tilde{t}_r Z_r, \chi(\tilde{t}_1,\ldots, \tilde{t}_r) u),$$ where $$(\tilde{t}_1,\ldots, \tilde{t}_r)\in { {\widetilde{{ \mathbb{T} }}} }\cong ({ \mathbb{C} }^*)^r, \quad (Z_1,\ldots, Z_r)\in { \mathbb{C} }^r,\quad u\in { \mathbb{C} }.$$ Therefore ${ \mathbb{C} }^r\times { \mathbb{C} }$ can be viewed as the total space of a ${ {\widetilde{{ \mathbb{T} }}} }$-equivariant line bundle ${ {\widetilde{L}} }_\chi$ over ${ \mathbb{C} }^r$. If $$\chi(\tilde{t}_1,\ldots, \tilde{t}_r) =\prod_{i=1}^r \tilde{t}_i^{c_i},$$ where $c_1,\ldots, c_r\in { \mathbb{Z} }$, then $${ {\widetilde{L}} }_{\chi}= {\mathcal{O}}_{{ \mathbb{C} }^r}(\sum_{i=1}^r c_i { {\widetilde{D}} }_i),$$ where ${ {\widetilde{D}} }_i$ is the ${ {\widetilde{{ \mathbb{T} }}} }$-divisor in ${ \mathbb{C} }^r$ defined by $Z_i=0$. We have $${ {\widetilde{M}} }\cong { {\mathrm{Pic}} }_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r) \cong H^2_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r;{ \mathbb{Z} }),$$ where the first isomorphism is given by $\chi\mapsto { {\widetilde{L}} }_\chi$ and the second isomorphism is given by the ${ {\widetilde{{ \mathbb{T} }}} }$-equivariant first Chern class $(c_1)_{{ {\widetilde{{ \mathbb{T} }}} }}$. Define $$D_i^{\mathcal{T}}:= (c_1)_{{ {\widetilde{{ \mathbb{T} }}} }}({\mathcal{O}}_{{ \mathbb{C} }^r}({ {\widetilde{D}} }_i)) \in H^2_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r;{ \mathbb{Z} }) \cong H^2_{{\mathcal{T}}}([{ \mathbb{C} }^r/G];{ \mathbb{Z} }).$$ Then $\{ D_1^{\mathcal{T}},\ldots, D_r^{\mathcal{T}}\}$ is a ${ \mathbb{Z} }$-basis of $H^2_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r;{ \mathbb{Z} })\cong { {\widetilde{M}} }$ dual to the ${ \mathbb{Z} }$-basis $\{ { {\tilde{b}} }_1,\ldots, { {\tilde{b}} }_r\}$ of ${ {\widetilde{N}} }$. We have a commutative diagram $$\begin{CD} { {\mathrm{Pic}} }_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r) @>{\iota_{{\mathcal{T}}}^*}>> { {\mathrm{Pic}} }_{{ {\widetilde{{ \mathbb{T} }}} }}(U_{\mathcal{A}}) @>{\cong}>>{ {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}}) \\ @V{(c_1)_{{ {\widetilde{{ \mathbb{T} }}} }}}VV @V_{(c_1)_{{ {\widetilde{{ \mathbb{T} }}} }}}VV @V_{(c_1)_{{\mathcal{T}}}}VV \\ H^2_{{ {\widetilde{{ \mathbb{T} }}} }}({ \mathbb{C} }^r;{ \mathbb{Z} }) @>{\iota_{{\mathcal{T}}}^*}>> H^2_{{ {\widetilde{{ \mathbb{T} }}} }}(U_{\mathcal{A}};{ \mathbb{Z} }) @>{\cong}>> H^2_{{\mathcal{T}}}({\mathcal{X}};{ \mathbb{Z} }), \end{CD}$$ where $\iota_{{\mathcal{T}}}^*$ is a surjective group homomorphism induced by the inclusion $\iota: U_{\mathcal{A}}\hookrightarrow { \mathbb{C} }^r$, and $${\mathrm{Ker}}(\iota_{{\mathcal{T}}}^*)= \bigoplus_{i=r'+1}^r { \mathbb{Z} }D_i^{{\mathcal{T}}}$$ Therefore, $${ {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}})\cong H^2_{{\mathcal{T}}}({\mathcal{X}};{ \mathbb{Z} }) \cong { {\widetilde{M}} }/\oplus_{i=r'+1}^r { \mathbb{Z} }D^{\mathcal{T}}_i$$ Let $\bar{D}_i^{\mathcal{T}}:= \iota_{\mathcal{T}}^* D_i^{{\mathcal{T}}}$. Then $$\bar{D}_i^{\mathcal{T}}=0,\quad i=r'+1,\ldots, r,$$ and $$H^2_{{\mathcal{T}}}({\mathcal{X}};{ \mathbb{Z} }) = \bigoplus_{i=1}^{r'} { \mathbb{Z} }\bar{D}^{\mathcal{T}}_i \cong{ \mathbb{Z} }^{r'}.$$ For $i=1,\ldots,r'$, ${ {\widetilde{D}} }_i\cap U_{\mathcal{A}}$ is a ${ {\widetilde{{ \mathbb{T} }}} }$-divisor in $U_{\mathcal{A}}$, and it descends to a ${ \mathbb{T} }$-divisor ${\mathcal{D}}_i$ in ${\mathcal{X}}$. We have $$\bar{D}_i^{{\mathcal{T}}}=(c_1)_{{\mathcal{T}}}({\mathcal{O}}_{{\mathcal{X}}}({\mathcal{D}}_i)),\quad i=1,\ldots, r'.$$ For $i=r'+1,\ldots, r$, ${ {\widetilde{D}} }_i\cap U_{\mathcal{A}}$ is empty, so its the zero ${ {\widetilde{{ \mathbb{T} }}} }$-divisor. Line bundles and Cartier divisors --------------------------------- We have group isomorphisms $${\mathbb{L}}^\vee \cong { {\mathrm{Pic}} }_G({ \mathbb{C} }^r) \cong H^2_G({ \mathbb{C} }^r;{ \mathbb{Z} }),$$ where the first isomorphism is given by $\chi\in {\mathbb{L}}^\vee={\mathrm{Hom}}(G,{ \mathbb{C} }^*)\mapsto { {\widetilde{L}} }_\chi$, and the second isomorphism is given by the $G$-equivariant first Chern class $(c_1)_G$. We have a commutative diagram $$\begin{CD} { {\mathrm{Pic}} }_G({ \mathbb{C} }^r) @>{\iota^*}>> { {\mathrm{Pic}} }_G(U_{\mathcal{A}}) @>{\cong}>>{ {\mathrm{Pic}} }({\mathcal{X}}) \\ @V{(c_1)_G}VV @V_{(c_1)_G}VV @V_{c_1}VV \\ H^2_G({ \mathbb{C} }^r;{ \mathbb{Z} }) @>{\iota^*}>> H^2_G(U_{\mathcal{A}};{ \mathbb{Z} }) @>{\cong}>> H^2({\mathcal{X}};{ \mathbb{Z} }), \end{CD}$$ where $\iota^*$ is a surjective group homomorphism induced by the inclusion $\iota: U_{\mathcal{A}}\hookrightarrow { \mathbb{C} }^r$. The surjective map $H^2_G({ \mathbb{C} }^r;{ \mathbb{Z} })\to H^2({\mathcal{X}};{ \mathbb{Z} })$ is the restriction of the Kirwan map $$\kappa: H^*_G({ \mathbb{C} }^r;{ \mathbb{Z} }){\longrightarrow}H^*({\mathcal{X}};{ \mathbb{Z} }).$$ Define $$D_i:= (c_1)_G({\mathcal{O}}_{{ \mathbb{C} }^r}({ {\widetilde{D}} }_i)) \in H^2_G({ \mathbb{C} }^r;{ \mathbb{Z} }) \cong H^2([{ \mathbb{C} }^r/G];{ \mathbb{Z} }).$$ Then $${\mathrm{Ker}}(\iota^*)= \bigoplus_{i=r'+1}^r { \mathbb{Z} }D_i.$$ Therefore, $${ {\mathrm{Pic}} }({\mathcal{X}})\cong H^2({\mathcal{X}};{ \mathbb{Z} }) \cong {\mathbb{L}}^\vee/\oplus_{i=r'+1}^r { \mathbb{Z} }D_i.$$ Recall that $$\psi^\vee: { {\widetilde{M}} }\to {\mathbb{L}}^\vee$$ is surjective if and only if $N_{ {\mathrm{tor}} }=0$. Let $$\bar{D}_i = c_1({\mathcal{O}}_{{\mathcal{X}}}({\mathcal{D}}_i))\in H^2({\mathcal{X}};{ \mathbb{Z} }),\quad i=1,\ldots, r.$$ The map $$\bar{\psi}^\vee: { {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}})\cong H^2_{{\mathcal{T}}}({\mathcal{X}};{ \mathbb{Z} }) \to { {\mathrm{Pic}} }({\mathcal{X}})\cong H^2({\mathcal{X}};{ \mathbb{Z} }),$$ given by $$\bar{D}_i^{\mathcal{T}}\mapsto \bar{D}_i,\quad i=1,\ldots, r',$$ is surjective if and only if $N_{ {\mathrm{tor}} }=0$. In general, ${\mathrm{Coker}}(\psi^\vee)\cong {\mathrm{Coker}}(\bar{\psi}^\vee)$ is a finite abelian group. Pick a ${ \mathbb{Z} }$-basis $\{e_1,\ldots, e_k\}$ of ${\mathbb{L}}\cong { \mathbb{Z} }^k$, and let $\{e_1^\vee,\ldots, e_k^\vee\}$ be the dual ${ \mathbb{Z} }$-basis of ${\mathbb{L}}^\vee$. For each $a\in \{1,\ldots,k\}$, we define a [*charge vector*]{} $$l^{(a)}= (l^{(a)}_1,\ldots, l^{(a)}_r) \in { \mathbb{Z} }^r$$ by $$\psi(e_a)=\sum_{i=1}^r l^{(a)}_r { {\tilde{b}} }_r,$$ where $\psi: L \to { {\widetilde{M}} }$ is the inclusion map. Then $$\psi^\vee(D_i)=\sum_{a=1}^k l^{(a)}_r e_a^*,\quad i=1,\ldots,r,$$ and $$\sum_{i=1}^r l^{(a)}_r b_r = \phi\circ \psi(e_a)=0,\quad a=1,\ldots, k.$$ \[ex:picard\] We use the notation in Example \[ex:definition\]. 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$. $$\begin{gathered} D_1=D_2=D_3=1,\ D_4=-3;\\ l^{(1)}=(1,1,1,-3);\\ { {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}})\cong { \mathbb{Z} }^3,\quad { {\mathrm{Pic}} }({\mathcal{X}})\cong { \mathbb{Z} }/3{ \mathbb{Z} };\end{gathered}$$ 2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$. $$\begin{gathered} D_1=(0,0),\ D_2= (0,1),\ D_3=(1,0),\ D_4=(-2,1),\ D_5=(1,-2);\\ l^{(1)}=(0,0,1,-2,1),\ l^{(2)}= (0,1,0,1,-2);\\ { {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}})={ \mathbb{Z} }^3,\quad { {\mathrm{Pic}} }({\mathcal{X}}) ={ \mathbb{Z} }^2 \big/ \big({ \mathbb{Z} }(-2,1)\oplus { \mathbb{Z} }(1,-2)\big) \cong { \mathbb{Z} }/3{ \mathbb{Z} }.\end{gathered}$$ 3. ${\mathcal{X}}={\mathcal{X}}_{0,0,0}$. $$\begin{gathered} D_1=3,\ D_2= 0 ,\ D_3= 0,\ D_4=-3;\\ l^{(1)}=(3,0,0,-3);\\ { {\mathrm{Pic}} }_{{\mathcal{T}}}({\mathcal{X}})={ \mathbb{Z} }^3,\quad { {\mathrm{Pic}} }({\mathcal{X}}) = { \mathbb{Z} }/3{ \mathbb{Z} }.\end{gathered}$$ Torus invariant subvarieties and their generic stabilizers ---------------------------------------------------------- Let ${\Sigma}(d)$ be the set of $d$-dimensional cones. For each ${ {\sigma} }\in {\Sigma}(d)$, define $$I_{ {\sigma} }=\{ i\in \{1,\ldots,r\}\mid \rho_i\not\subset { {\sigma} }\} \in {\mathcal{A}},$$ and define $$I_{ {\sigma} }'= \{1,\ldots, r\}\setminus I_{ {\sigma} }.$$ Then $|I_{ {\sigma} }'|=d$ and $|I_{ {\sigma} }|=r-d$. Let $\tilde{V}({ {\sigma} })\subset U_{\mathcal{A}}$ be the closed subvariety defined by the ideal of ${ \mathbb{C} }[Z_1,\ldots, Z_r]$ generated by $$\{Z_i=0\mid \rho_i\subset { {\sigma} }\} = \{ Z_i=0\mid i \in I'_{ {\sigma} }\}.$$ Then ${\mathcal{V}}({ {\sigma} }) := [{ {\widetilde{V}} }({ {\sigma} })/G]$ is an $(n-d)$-dimensional ${\mathcal{T}}$-invariant closed subvariety of ${\mathcal{X}}= [U_{\mathcal{A}}/G]$. The group homomorphism $G\cong ({ \mathbb{C} }^*)^k \to { {\widetilde{{ \mathbb{T} }}} }\cong ({ \mathbb{C} }^*)^r$ is given by $$g\mapsto (\chi_1(g),\ldots, \chi_r(g)),$$ where $\chi_i\in {\mathrm{Hom}}(G,{ \mathbb{C} }^*)={\mathbb{L}}^\vee$ is given by $$\chi_i(u_1,\ldots, u_k)= \prod_{a=1}^k u_a^{ l_i^{(a)} }.$$ Let $$G_{ {\sigma} }:= \{ g\in G\mid g\cdot z = z \textup{ for all } z\in { {\widetilde{V}} }({ {\sigma} })\} =\bigcap_{i\in I_{ {\sigma} }} {\mathrm{Ker}}(\chi_i).$$ Then $G_{ {\sigma} }$ is the generic stabilizer of ${\mathcal{V}}({ {\sigma} })$. It is a finite subgroup of $G$. If $\tau\subset { {\sigma} }$ then $I_{ {\sigma} }\subset I_\tau$, so $G_\tau\subset G_{ {\sigma} }$. There are two special cases: - Let $\{0\}$ be the unique 0-dimensional cone. Then $G_{\{0\}}=K$ is the generic stabilizer of ${\mathcal{V}}(\{0\})={\mathcal{X}}$. - If ${ {\sigma} }\in {\Sigma}(r)$, then ${\mathfrak{p}}_{ {\sigma} }:= {\mathcal{V}}({ {\sigma} })$ is a ${\mathcal{T}}$ fixed point in ${\mathcal{X}}$, and ${\mathfrak{p}}_{ {\sigma} }=BG_{ {\sigma} }$. \[ex:stabilizer\] We use the notation in Example \[ex:definition\]. Let ${ {\sigma} }\subset N_{ \mathbb{R} }\cong { \mathbb{R} }^3$ denote the 3-dimensional cone spanned by $\bar{b}_1,\bar{b}_2,\bar{b}_3$. For $j=1,2,3$, let $\tau_j$ denote the 2-dimensional cone in $N_{ \mathbb{R} }$ spanned by $\{\bar{b}_i: i\in \{1,2,3\}-\{j\}\}$. 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$: $G_{ {\sigma} }={ \mathbb{Z} }_3,\quad G_{\tau_1}=G_{\tau_2}=G_{\tau_3}=\{1\}$. 2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$: $G_{ {\sigma} }={ \mathbb{Z} }_3 = G_{\tau_3}, \quad G_{\tau_1}=G_{\tau_2}=\{1\}$. 3. ${\mathcal{X}}={\mathcal{X}}_{0,0,0}$: $G_{ {\sigma} }={ \mathbb{Z} }_3 = G_{\tau_1}=G_{\tau_2}=G_{\tau_3}$. Define the set of flags in ${\Sigma}$ to be $$F({\Sigma})=\{(\tau,{ {\sigma} })\in {\Sigma}(r-1)\times {\Sigma}(r): \tau \subset { {\sigma} }\}.$$ Given $(\tau,{ {\sigma} })\in F({\Sigma})$, let ${\mathfrak{l}}_\tau:={ {\widetilde{V}} }(\tau)$ be the 1-dimensional ${ {\widetilde{T}} }$-invariant subvariety of ${\mathcal{X}}$. Then ${\mathfrak{p}}_{ {\sigma} }$ is contained in ${\mathfrak{l}}_\tau$. There is a unique $i\in \{1,\ldots, r'\}$ such that $i\in I'_{ {\sigma} }\setminus I'_\tau$. The representation of $G_{ {\sigma} }$ on the tangent line $T_{{\mathfrak{p}}_{ {\sigma} }}{\mathfrak{l}}_\tau$ to ${\mathfrak{l}}_\tau$ at the stacky point ${\mathfrak{p}}_{ {\sigma} }$ is given by $\chi_i|_{G_{ {\sigma} }}: G_{ {\sigma} }\to { \mathbb{C} }^*$. The image $\chi_i(G_{ {\sigma} }) \subset { \mathbb{C} }^*$ is a cyclic subgroup of ${ \mathbb{C} }^*$; we define the order of this group to be $r(\tau,{ {\sigma} })$. Then there is a short exact sequence of finite abelian groups: $$1\to G_\tau\to G_{ {\sigma} }\to {\boldsymbol{\mu}}_{r(\tau,{ {\sigma} })}\to 1,$$ where $\mu_a$ is the group of $a$-th roots of unity. The extended nef cone and the extended Mori cone {#sec:nef-NE} ------------------------------------------------ In this paragraph, ${ \mathbb{F} }={ \mathbb{Q} }$, ${ \mathbb{R} }$, or ${ \mathbb{C} }$. Given a finitely generated abelian group $\Lambda$ with $\Lambda/\Lambda_{ {\mathrm{tor}} }\cong { \mathbb{Z} }^m$, define $\Lambda_{ \mathbb{F} }= \Lambda\otimes_{ \mathbb{Z} }{ \mathbb{F} }\cong { \mathbb{F} }^m$. We have the following short exact sequences of vector spaces: $$\begin{aligned} && 0\to {\mathbb{L}}_{ \mathbb{F} }\to { {\widetilde{N}} }_{ \mathbb{F} }\to N_{ \mathbb{F} }\to 0,\\ && 0\to M_{ \mathbb{F} }\to { {\widetilde{M}} }_{ \mathbb{F} }\to {\mathbb{L}}^\vee_{ \mathbb{F} }\to 0. \end{aligned}$$ We also have the following isomorphisms of vector spaces over ${ \mathbb{F} }$: $$\begin{aligned} && H^2({\mathcal{X}};{ \mathbb{F} })\cong H^2(X;{ \mathbb{F} }) \cong {\mathbb{L}}^\vee_{ \mathbb{F} }/\oplus_{i=r'+1}^r { \mathbb{F} }D_i,\\ && H^2_{{\mathcal{T}}}({\mathcal{X}};{ \mathbb{F} }) \cong H^2_{ \mathbb{T} }(X;{ \mathbb{F} }) \cong { {\widetilde{M}} }_{ \mathbb{F} }/\oplus_{i=r'+1}^r { \mathbb{F} }D_i^{\mathcal{T}},\end{aligned}$$ where $X$ is the coarse moduli space of ${\mathcal{X}}$. From now on, we assume all the maximal cones in ${\Sigma}$ are $n$-dimensional, where $n=\dim_{ \mathbb{C} }{\mathcal{X}}$. Given a maximal cone ${ {\sigma} }\in {\Sigma}(n)$, we define $${\mathbb{K}}_{ {\sigma} }^\vee := \bigoplus_{i\in I_{ {\sigma} }}{ \mathbb{Z} }D_i.$$ Then ${\mathbb{K}}_{ {\sigma} }^\vee$ is a sublattice of ${\mathbb{L}}^\vee$ of finite index. We define the [*extended ${ {\sigma} }$-nef cone*]{} to be $${\widetilde{{ {\mathrm{Nef}} }}}_{ {\sigma} }= \sum_{i\in I_{ {\sigma} }}{ \mathbb{R} }_{\geq 0} D_i,$$ which is a $k$-dimensional cone in ${\mathbb{L}}^\vee_{ \mathbb{R} }\cong { \mathbb{R} }^k$. The [*extended nef cone*]{} of the extended stacky fan $({\Sigma},b_1,\ldots, b_r)$ is $${\widetilde{{ {\mathrm{Nef}} }}}_{{\mathcal{X}}}:=\bigcap_{{ {\sigma} }\in {\Sigma}(n)} {\widetilde{{ {\mathrm{Nef}} }}}_{ {\sigma} }.$$ The [*extended ${ {\sigma} }$-Kähler cone*]{} ${ {\widetilde{C}} }_{ {\sigma} }$ is defined to be the interior of ${\widetilde{{ {\mathrm{Nef}} }}}_{ {\sigma} }$; the [*extended Kähler cone*]{} of ${\mathcal{X}}$, ${ {\widetilde{C}} }_{{\mathcal{X}}}$, is defined to be the interior of the extended nef cone ${\widetilde{{ {\mathrm{Nef}} }}}_{{\mathcal{X}}}$. Let ${\mathbb{K}}_{ {\sigma} }$ be the dual lattice of ${\mathbb{K}}_{ {\sigma} }^\vee$; it can be viewed as an additive subgroup of ${\mathbb{L}}_{ \mathbb{Q} }$: $${\mathbb{K}}_{ {\sigma} }=\{ \beta\in {\mathbb{L}}_{ \mathbb{Q} }\mid \langle D, \beta\rangle \in { \mathbb{Z} }\ \forall D\in {\mathbb{K}}_{ {\sigma} }^\vee \},$$ where $\langle-, -\rangle$ is the natural pairing between ${\mathbb{L}}^\vee_{ \mathbb{Q} }$ and ${\mathbb{L}}_{ \mathbb{Q} }$. Define $${\mathbb{K}}:= \bigcup_{{ {\sigma} }\in {\Sigma}(n)} {\mathbb{K}}_{ {\sigma} }.$$ Then ${\mathbb{K}}$ is a subset (which is not necessarily a subgroup) of ${\mathbb{L}}_{ \mathbb{Q} }$, and ${\mathbb{L}}\subset {\mathbb{K}}$. We define the [*extended ${ {\sigma} }$-Mori cone*]{} ${\widetilde{{ {\mathrm{NE}} }}}_{ {\sigma} }\subset {\mathbb{L}}_{ \mathbb{R} }$ to be the dual cone of ${\widetilde{{ {\mathrm{Nef}} }}}_{ {\sigma} }\subset {\mathbb{L}}_{ \mathbb{R} }^\vee$: $${\widetilde{{ {\mathrm{NE}} }}}_{ {\sigma} }=\{ \beta \in {\mathbb{L}}_{ \mathbb{R} }\mid \langle D,\beta\rangle \geq 0 \ \forall D\in {\widetilde{{ {\mathrm{Nef}} }}}_{ {\sigma} }\}.$$ It is a $k$-dimensional cone in ${\mathbb{L}}_{ \mathbb{R} }$. The [*extended Mori cone*]{} of the extended stacky fan $({\Sigma},b_1,\ldots, b_r)$ is $${\widetilde{{ {\mathrm{NE}} }}}_{{\mathcal{X}}}:= \bigcup_{{ {\sigma} }\in {\Sigma}(n)} {\widetilde{{ {\mathrm{NE}} }}}_{ {\sigma} }.$$ Finally, we define $${\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}:= {\mathbb{K}}_{ {\sigma} }\cap {\widetilde{{ {\mathrm{NE}} }}}_{ {\sigma} },\quad {\mathbb{K}}_{{\mathrm{eff}}}:= {\mathbb{K}}\cap {\widetilde{{ {\mathrm{NE}} }}}({\mathcal{X}})= \bigcup_{{ {\sigma} }\in {\Sigma}(n)} {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}.$$ \[ex:effective\] 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$. $$\begin{gathered} {\mathbb{K}}^\vee \cong 3{ \mathbb{Z} }, \quad {\widetilde{{ {\mathrm{Nef}} }}}_{{\mathcal{X}}}={ \mathbb{R} }_{\leq 0};\\ {\mathbb{K}}\cong \frac{1}{3}{ \mathbb{Z} }, \quad {\widetilde{{ {\mathrm{NE}} }}}_{{\mathcal{X}}}={ \mathbb{R} }_{\leq 0},\quad {\mathbb{K}}_{\mathrm{eff}}=\frac{1}{3}{ \mathbb{Z} }_{\leq 0}.\end{gathered}$$  2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$. $$\begin{gathered} {\mathbb{K}}^\vee\cong { \mathbb{Z} }(-2,1)\oplus { \mathbb{Z} }(1,-2),\quad {\widetilde{{ {\mathrm{Nef}} }}}_{{\mathcal{X}}}= { \mathbb{R} }_{\geq 0}(-2,1) +{ \mathbb{R} }_{\geq 0}(1,-2);\\ {\mathbb{K}}\cong { \mathbb{Z} }(-\frac{2}{3}, -\frac{1}{3})\oplus { \mathbb{Z} }(-\frac{1}{3}, -\frac{2}{3}), \quad {\widetilde{{ {\mathrm{NE}} }}}_{{\mathcal{X}}} = { \mathbb{R} }_{\geq 0}(-\frac{2}{3}, -\frac{1}{3}) + { \mathbb{R} }_{\geq 0}(-\frac{1}{3}, -\frac{2}{3}),\\ {\mathbb{K}}_{\mathrm{eff}}={ \mathbb{Z} }_{\geq 0}(-\frac{2}{3},-\frac{1}{3}) + { \mathbb{Z} }_{\geq 0}(-\frac{1}{3},-\frac{2}{3}).\end{gathered}$$   3. ${\mathcal{X}}={\mathcal{X}}_{0,0,0}$. $$\begin{gathered} {\mathbb{K}}^\vee\cong 3{ \mathbb{Z} },\quad {\widetilde{{ {\mathrm{Nef}} }}}_{{\mathcal{X}}}= { \mathbb{R} }_{\leq 0};\\ {\mathbb{K}}\cong \frac{1}{3}{ \mathbb{Z} }, \quad {\widetilde{{ {\mathrm{NE}} }}}_{{\mathcal{X}}} = { \mathbb{R} }_{\leq 0},\quad {\mathbb{K}}_{\mathrm{eff}}=\frac{1}{3}{ \mathbb{Z} }_{\leq 0}.\end{gathered}$$ \[semi-proj\] From now on, we make the following assumptions on the toric orbifold ${\mathcal{X}}$. 1. The coarse moduli space $X_{\Sigma}$ of ${\mathcal{X}}$ is semi-projective. 2. We may choose $b_{r'+1},\ldots, b_r$ such that $\hat \rho:=D_1+\dots + D_r$ is contained in the closure of the extended Kähler cone ${ {\widetilde{C}} }_{\mathcal{X}}$. <!-- --> 1. We make the above assumptions (a) and (b) so that the equivariant mirror theorem in [@CCIT] is applicable to ${\mathcal{X}}$. 2. By [@CLS Proposition 14.4.1], $X_{\Sigma}$ is semi-projective if and only if $|{\Sigma}|$ is equal to the cone spanned by $b_1,\ldots, b_r$. For example, ${\mathcal{O}}_{{ \mathbb{P} }^1}(-3)\oplus {\mathcal{O}}_{{ \mathbb{P} }^1}(1)$ is a smooth toric Calabi-Yau 3-fold which is not semi-projective:  3. When ${\mathcal{X}}$ is a Calabi-Yau smooth toric DM stack, Assumption (b) holds if its coarse moduli space $X_{\Sigma}$ has a toric crepant resolution of singularities; see [@Ir09 Remark 3.4]. By [@CLS Proposition 11.4.19], any 3-dimensional Gorenstein toric variety $X_{\Sigma}$ has a resolution of singularities $\phi:X_{{\Sigma}'}\to X_{{\Sigma}}$ such that $\phi$ is projective and crepant. So Assumption \[semi-proj\] (b) holds for any 3-dimensional Calabi-Yau smooth toric DM stacks. Smooth toric DM stacks as symplectic quotients ---------------------------------------------- Let $G_{ \mathbb{R} }\cong U(1)^k$ be the maximal compact subgroup of $G\cong ({ \mathbb{C} }^*)^k$. Then the Lie algebra of $G_{ \mathbb{R} }$ is ${\mathbb{L}}_{ \mathbb{R} }$. Let $${ {\widetilde{\mu}} }: { \mathbb{C} }^r\to {\mathbb{L}}^\vee_{ \mathbb{R} }=\bigoplus_{a=1}^k { \mathbb{R} }e_a ^\vee$$ be the moment map of the Hamiltonian $G_{ \mathbb{R} }$-action on ${ \mathbb{C} }^r$, equipped with the Kähler form $$\sqrt{-1} \sum_{i=1}^r dZ_i\wedge d\bar{Z}_i.$$ Then $${ {\widetilde{\mu}} }(Z_1,\ldots, Z_r) = \sum_{i=1}^r \sum_{a=1}^k l_i^{(a)}|Z_i|^2 e_a.$$ If ${\mathbf{r}}=\sum_{a=1}^k r_a e_a^\vee$ is in the extended Kähler cone of ${\mathcal{X}}$, then $${\mathcal{X}}=[{ {\widetilde{\mu}} }^{-1}({\mathbf{r}})/G_{ \mathbb{R} }].$$ The real numbers $r_1,\ldots, r_k$ are extended Käler parameters. Let $T_a=-r_a+\sqrt{-1}\theta_a$ be complexified extended Kähler parameters of ${\mathcal{X}}$. The inertia stack and the Chen-Ruan orbifold cohomology {#sec:CR} ------------------------------------------------------- Given ${ {\sigma} }\in {\Sigma}$, define $${\mathrm{Box}({ {\sigma} })}:=\{ v\in N: \bar{v}=\sum_{i\in I'_{ {\sigma} }} c_i \bar{b}_i, \quad 0\leq c_i <1\}.$$ Then $N_{ {\mathrm{tor}} }\subset {\mathrm{Box}({ {\sigma} })}\subset N$. If $\tau\subset \sigma$ then $I'_\tau\subset I'_{ {\sigma} }$, so $\mathrm{Box}(\tau)\subset {\mathrm{Box}({ {\sigma} })}$. Let ${ {\sigma} }\in {\Sigma}(n)$ be a maximal cone in ${\Sigma}$. We have a short exact sequence of abelian groups $$0\to {\mathbb{K}}_{ {\sigma} }/{\mathbb{L}}\to {\mathbb{L}}_{ \mathbb{R} }/{\mathbb{L}}\to {\mathbb{L}}_{ \mathbb{R} }/{\mathbb{K}}_{ {\sigma} }\to 0,$$ which can be identified with the following short exact sequence of multiplicative abelian groups $$1\to G_{ {\sigma} }\to G_{ \mathbb{R} }\to (G/G_{ {\sigma} })_{ \mathbb{R} }\to 0$$ where $G_{ \mathbb{R} }\cong U(1)^k$ is the maximal compact subgroup of $G\cong ({ \mathbb{C} }^*)^k$, and $(G/G_{ {\sigma} })_{ \mathbb{R} }\cong U(1)^k$ is the maximal compact subgroup of $G_{ {\sigma} }\cong({ \mathbb{C} }^*)^k$. Given a real number $x$, we recall some standard notation: $\lfloor x \rfloor$ is the greatest integer less than or equal to $x$, $\lceil x \rceil$ is the least integer greater or equal to $x$, and $\{ x\} = x-\lfloor x \rfloor$ is the fractional part of $x$. Define $v: {\mathbb{K}}_{ {\sigma} }\to N$ by $$v(\beta)= \sum_{i=1}^r \lceil \langle D_i,\beta\rangle\rceil b_i.$$ Then $$\overline{v(\beta)} = \sum_{i\in I'_{ {\sigma} }} \{ -\langle D_i,\beta\rangle \}\bar{b}_i,$$ so $v(\beta)\in {\mathrm{Box}({ {\sigma} })}$. Indeed, $v$ induces a bijection $K_{ {\sigma} }/{\mathbb{L}}\cong {\mathrm{Box}({ {\sigma} })}$. For any $\tau\in {\Sigma}$ there exists ${ {\sigma} }\in {\Sigma}(n)$ such that $\tau\subset { {\sigma} }$. The bijection $G_{ {\sigma} }\to {\mathrm{Box}({ {\sigma} })}$ restricts to a bijection $G_\tau\to \mathrm{Box}(\tau)$. Define $${\mathrm{Box}(\mathbf {\Sigma})}:=\bigcup_{{ {\sigma} }\in {\Sigma}}{\mathrm{Box}({ {\sigma} })}=\bigcup_{{ {\sigma} }\in{\Sigma}(n)}{\mathrm{Box}({ {\sigma} })}.$$ Then $N_{ {\mathrm{tor}} }\subset {\mathrm{Box}(\mathbf {\Sigma})}\subset N$. There is a bijection ${\mathbb{K}}/{\mathbb{L}}\to {\mathrm{Box}(\mathbf {\Sigma})}$. Given $v\in {\mathrm{Box}({ {\sigma} })}$, where ${ {\sigma} }\in {\Sigma}(d)$, define $c_i(v)\in [0,1)\cap { \mathbb{Q} }$ by $$\bar{v}= \sum_{i\in I'_{ {\sigma} }} c_i(v) \bar{b}_i.$$ Suppose that $k \in G_{ {\sigma} }$ corresponds to $v\in {\mathrm{Box}({ {\sigma} })}$ under the bijection $G_{ {\sigma} }\cong{\mathrm{Box}({ {\sigma} })}$, then $$\chi_i(k) = \begin{cases} 1, & i\in I_{ {\sigma} },\\ e^{2\pi\sqrt{-1} c_i(v)},& i \in I'_{ {\sigma} }. \end{cases}$$ Define $${\mathrm{age}}(k)={\mathrm{age}}(v)= \sum_{i\notin I_{ {\sigma} }} c_i(v).$$ Let $IU=\{(z,k)\in U_{\mathcal{A}}\times G\mid k\cdot z = z\}$, and let $G$ acts on $IU$ by $h\cdot(z,k)= (h\cdot z,k)$. The inertia stack ${\mathcal{I}}{\mathcal{X}}$ of ${\mathcal{X}}$ is defined to be the quotient stack $${\mathcal{I}}{\mathcal{X}}:= [IU/G].$$ Note that $(z=(Z_1,\ldots,Z_r), k)\in U$ if and only if $$k\in \bigcup_{{ {\sigma} }\in {\Sigma}}G_{ {\sigma} }\textup{ and } Z_i=0 \textup{ whenever } \chi_i(k) \neq 1.$$ So $$IU=\bigcup_{v\in {\mathrm{Box}(\mathbf {\Sigma})}} U_v,$$ where $$U_v:= \{(Z_1,\ldots, Z_m)\in U_{\mathcal{A}}: Z_i=0 \textup{ if } c_i(v) \neq 0\}.$$ The connected components of ${\mathcal{I}}{\mathcal{X}}$ are $$\{ {\mathcal{X}}_v:= [U_v/G] : v\in {\mathrm{Box}(\mathbf {\Sigma})}\}.$$ The involution $IU\to IU$, $(z,k)\mapsto (z,k^{-1})$ induces involutions ${\mathrm{inv}}:{\mathcal{I}}{\mathcal{X}}\to {\mathcal{I}}{\mathcal{X}}$ and ${\mathrm{inv}}:{\mathrm{Box}(\mathbf {\Sigma})}\to {\mathrm{Box}(\mathbf {\Sigma})}$ such that ${\mathrm{inv}}({\mathcal{X}}_v)={\mathcal{X}}_{{\mathrm{inv}}(v)}$. In the remainder of this subsection, we consider rational cohomology, and write $H^*(-)$ instead of $H^*(-;{ \mathbb{Q} })$. The Chen-Ruan orbifold cohomology [@CR04] is defined to be $$H^*_{ {\mathrm{orb}} }({\mathcal{X}})=\bigoplus_{v\in {\mathrm{Box}(\mathbf {\Sigma})}} H^*({\mathcal{X}}_v)[2{\mathrm{age}}(v)].$$ Denote $\mathbf 1_v$ to be the unit in $H^*({\mathcal{X}}_v)$. Then $\mathbf 1_v\in H^{2{\mathrm{age}}(v)}_{ {\mathrm{orb}} }({\mathcal{X}})$. In particular, $$H^0_{ {\mathrm{orb}} }({\mathcal{X}}) =\bigoplus_{v\in N_{ {\mathrm{tor}} }} { \mathbb{Q} }\mathbf 1_v.$$ Suppose that ${\mathcal{X}}$ is a [*proper*]{} toric DM stack. Then the orbifold Poincaré pairing on $H^*_{ {\mathrm{orb}} }({\mathcal{X}})$ is defined as $$\label{eqn:Poincare} (\alpha,\beta):=\int_{{\mathcal{I}}{\mathcal{X}}} \alpha\cup {\mathrm{inv}}^*(\beta),$$ We also have an equivariant pairing on $H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }}({\mathcal{X}})$: $$\label{eqn:T-Poincare} (\alpha,\beta)_{{ \mathbb{T} }} := \int_{{\mathcal{I}}{\mathcal{X}}_{{ \mathbb{T} }}} \alpha\cup {\mathrm{inv}}^*(\beta),$$ where $$\int_{ {\mathcal{I}}{\mathcal{X}}_{{ \mathbb{T} }}}: H_{{ {\mathrm{orb}} },{ \mathbb{T} }}^*({\mathcal{X}}) \to H_{{ \mathbb{T} }}^*({\mathrm{point}}) = H^*(B{ \mathbb{T} })$$ is the equivariant pushforward to a point. When ${\mathcal{X}}$ is not proper, is not defined, but we can still define via an equivariant pairing $H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }}({\mathcal{X}})\otimes H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }}({\mathcal{X}}) \to {\mathcal{Q}}_{ \mathbb{T} }$, where ${\mathcal{Q}}_{ \mathbb{T} }$ is the fractional field of the ring $H^*(B{ \mathbb{T} })$. \[ex:Horb\] 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$. $$\begin{gathered} N={ \mathbb{Z} }^3,\quad {\mathrm{Box}(\mathbf {\Sigma})}=\{(0,0,0),(0,0,1),(0,0,2)\};\\ H^0_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_{(0,0,0)},\ H^2_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_{(0,0,1)},\ H^4_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_{(0,0,2)}.\end{gathered}$$ 2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$. $$\begin{gathered} N={ \mathbb{Z} }^3,\quad {\mathrm{Box}(\mathbf {\Sigma})}=\{ (0,0,0), (0,2,1), (0,1,1)\};\\ H^0_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_{(0,0,0)},\quad H^2_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_{(0,2,1)} \oplus { \mathbb{Q} }\mathbf 1_{(0,1,1)}.\end{gathered}$$ 3. ${\mathcal{X}}={\mathcal{X}}_{0,0,0}$. $$\begin{gathered} N={ \mathbb{Z} }^3\oplus { \mathbb{Z} }_3,\quad {\mathrm{Box}(\mathbf {\Sigma})}=N_{ {\mathrm{tor}} }={ \mathbb{Z} }_3=\{0,1,2\};\\ H^0_{ {\mathrm{orb}} }({\mathcal{X}})={ \mathbb{Q} }{\mathbf{1}}_0\oplus { \mathbb{Q} }{\mathbf{1}}_1 \oplus { \mathbb{Q} }{\mathbf{1}}_2.\end{gathered}$$ Open-Closed GW invariants {#sec:GW-inv} ========================= From now on, we consider 3-dimensional Calabi-Yau smooth toric DM stacks. Then $n=3$, ${ \mathbb{T} }\cong({ \mathbb{C} }^*)^3$ and ${ \mathbb{T} }'\cong \{ (t_1,t_2,t_3)\in ({ \mathbb{C} }^*)^3\mid t_1t_2t_3=1\}$. Agananic-Vafa A-branes {#sec:AV-brane} ---------------------- In [@AV00], Aganagic-Vafa introduced a class of Lagrangian submanifolds of semi-projective smooth toric Calabi-Yau 3-folds. It is straightforward to generalize this construction to 3-dimensional Calabi-Yau smooth toric DM stacks with semi-projective coarse moduli spaces. Let ${\mathcal{X}}=[{ {\widetilde{\mu}} }^{-1}({\mathbf{r}})/G_{ \mathbb{R} }]$ be a 3-dimensional Calabi-Yau smooth toric DM stack, where ${\mathbf{r}}\in { {\widetilde{C}} }({\mathcal{X}})\subset {\mathbb{L}}^\vee_{ \mathbb{R} }$. ${ {\widetilde{\mu}} }^{-1}(\mathbf{r})$ is defined by $$\sum_{i=1}^{k+3} l_i^{(a)} |X_i|^2 =r_a,\quad a=1,\ldots, k.$$ Write $X_i=\rho_i e^{\sqrt{-1}\phi_i}$, where $\rho_i=|X_i|$. An Aganagic-Vafa brane is a Lagrangian sub-orbifold of ${\mathcal{X}}$ of the form $${\mathcal{L}}=[{ {\widetilde{L}} }/G_{ \mathbb{R} }]$$ where $${ {\widetilde{L}} }=\{ (X_1,\ldots, X_r)\in { {\widetilde{\mu}} }^{-1}(\mathbf{r}): \sum_{i=1}^{k+3}\hat{l}^1_i |X_i|^2=c_1, \sum_{i=1}^{k+3} \hat{l}_i^2 |X_i|^2 = c_2, \sum_{i=1}^{k+3}\phi_i =\textup{const} \}$$ for some $\hat{l}^\alpha_i \in { \mathbb{Z} }, \sum_{i=1}^{k+3}\hat{l}^\alpha_i =0$, $\alpha=1,2$. An Aganagic-Vafa brane intersects a unique 1-dimensional orbit closure ${\mathfrak{l}}_\tau := {\mathcal{V}}(\tau)$, $\tau\in {\Sigma}(2)$. Let $\ell_\tau$ be the coarse moduli of ${\mathfrak{l}}_\tau$. We say ${\mathcal{L}}$ is an inner (resp. outer) brane if $\ell_\tau\cong { \mathbb{P} }^1$ (resp. $\ell_\tau\cong { \mathbb{C} }$). We next describe the first homology group of an Aganagic-Vafa brane. [^2] We pick a fixed point ${\mathfrak{p}}_{ {\sigma} }$ in $\ell_\tau$, so that $(\tau,{ {\sigma} })\in F({\Sigma})$. Suppose that $I'_{ {\sigma} }=\{i_1, i_2, i_3\}$. Then ${\mathcal{X}}_{ {\sigma} }=[{ \mathbb{C} }^3/G_{ {\sigma} }]$. Let $(\tau_1,\sigma)= (\tau,\sigma)$, $(\tau_2,\sigma)$ and $(\tau_3,\sigma)$ be three flags in the toric graph in the counter-clockwise direction, such that $$I'_{\tau_1}= \{i_2, i_3\},\quad I'_{\tau_2}=\{i_3, i_1\},\quad I'_{\tau_3}=\{i_1,i_2\}.$$ For $j=1,2,3$, let $G_j= G_{\tau_j}$ and let $s_j=r(\tau_j,{ {\sigma} })$. Then there are exact short sequences of finite abelian groups: $$1\to G_j\to G_{ {\sigma} }\stackrel{\chi_{i_j}}{\to} {\boldsymbol{\mu}}_{s_j}\to 1,\quad j=1,2,3.$$ By the Calabi-Yau condition, for any $k\in G_{ {\sigma} }$, $$\chi_{i_1}(k)\chi_{i_2}(k)\chi_{i_3}(k)=1,\quad {\mathrm{age}}(k)\in { \mathbb{Z} }.$$ There are canonical identifications $$\begin{aligned} \label{eqn:g-sigma-identification} &G_{ {\sigma} }\cong \{v\in N: \bar{v}= c_1 \bar{b}_{i_1}+ c_2 \bar{b}_{i_2}+ c_3 \bar{b}_{i_3}\in N_{ \mathbb{Q} },\ 0\le c_i<1\},\\ &G_1\cong \{v\in N: \bar{v}= c_2 \bar{b}_{i_2}+ c_3 \bar{b}_{i_3}\in N_{ \mathbb{Q} },\ 0\le c_i< 1\},\nonumber \\ &G_2\cong \{v\in N: \bar{v}= c_1 \bar{b}_{i_1}+ c_3 \bar{b}_{i_3}\in N_{ \mathbb{Q} },\ 0\le c_i< 1\},\nonumber \\ &G_3\cong \{v\in N: \bar{v}= c_1 \bar{b}_{i_1}+ c_2 \bar{b}_{i_2}\in N_{ \mathbb{Q} },\ 0\le c_i< 1\}.\nonumber\end{aligned}$$ Then $${\mathcal{L}}= [{ {\widetilde{L}} }_{ {\sigma} }/G_{ {\sigma} }]\subset {\mathcal{X}}_{ {\sigma} }=[{ \mathbb{C} }^3/G_{ {\sigma} }]\subset {\mathcal{X}}.$$ There is a $G_{ {\sigma} }$-equivariant diffeomorphism ${ {\widetilde{L}} }_{ {\sigma} }\cong S^1\times { \mathbb{C} }$, where $G_{ {\sigma} }$ acts on $S^1\times { \mathbb{C} }$ by $$k\cdot (e^{\sqrt{-1}\theta}, u)= (\chi_{i_1}(k) e^{\sqrt{-1} \theta}, \chi_{i_2}(k) u).$$ We have a commutative diagram $$\begin{CD} G_{ {\sigma} }@>>> { {\widetilde{L}} }_{ {\sigma} }@>>> {\mathcal{L}}\\ @VVV @VVV @VVV\\ \mu_{s_1} @>>> S^1 @>>> S^1/\mu_{s_1} \end{CD}$$ where the columns are fibrations. So we have a commutative diagram $$\begin{CD} 0 @>>> \pi_1({ {\widetilde{L}} }_{\sigma}) @>>> \pi_1({\mathcal{L}}) @>>> \pi_0(G_{ {\sigma} }) @>>> 0 \\ && @VVV @VVV @VVV\\ 0 @>>> \pi_1(S^1) @>>> \pi_1(S^1/{\boldsymbol{\mu}}_s) @>>> \pi_0({\boldsymbol{\mu}}_{s_1}) @>>> 0 \end{CD}$$ where the rows are short exact sequences of abelian groups. The above diagram can be rewritten as: $$\begin{CD} 0 @>>> { \mathbb{Z} }@>>> H_1({\mathcal{L}};{ \mathbb{Z} }) @>{\pi_{ {\sigma} }}>> G_{ {\sigma} }@>>> 0 \\ && @V{\mathrm{id}}VV @V{\pi}VV @V{\chi_{i_1}}VV\\ 0 @>>> { \mathbb{Z} }@>{\times s_1}>> { \mathbb{Z} }@>>> { \mathbb{Z} }_{s_1} @>>> 0 \end{CD}$$ The group homomorphism $\pi\times \pi_{ {\sigma} }:H_1({\mathcal{L}};{ \mathbb{Z} })\to { \mathbb{Z} }\times G_{ {\sigma} }$ is injective, and the image is $$\{(d_0,k)\in { \mathbb{Z} }\times G_{ {\sigma} }: \exp(2\pi\sqrt{-1}\frac{d_0}{s_1}) = \chi_{i_1}(k)\}.$$ Let $V_\tau=\{ { {\sigma} }\in {\Sigma}(3): (\tau,{ {\sigma} })\in F({\Sigma})\}$. If $V_\tau=\{{ {\sigma} }\}$ then define $$H_{\tau,{ {\sigma} }} =\{(\pi(\gamma), \pi_{ {\sigma} }(\gamma)): \gamma\in H_1({\mathcal{L}};{ \mathbb{Z} })\}\subset { \mathbb{Z} }\times G_{ {\sigma} }$$ If $V_\tau=\{{ {\sigma} }_+,{ {\sigma} }_-\}$ then define $$H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-} =\{(\pi(\gamma), \pi_{{ {\sigma} }_+}(\gamma), \pi_{{ {\sigma} }_-}(\gamma)): \gamma\in H_1({\mathcal{L}};{ \mathbb{Z} })\} \subset { \mathbb{Z} }\times G_{{ {\sigma} }_+}\times G_{{ {\sigma} }_-}.$$ Let ${\Sigma}_c(2)=\{\tau\in {\Sigma}(2): |V_\tau|=2\}$. Then ${\mathcal{L}}$ is an inner brane iff $\tau\in {\Sigma}_c(2)$. We may identify $H_1({\mathcal{L}};{ \mathbb{Z} })$ with $H_{\tau,{ {\sigma} }}$ (resp. $H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}$) if ${\mathcal{L}}$ is an outer (resp. inner) brane. The identification gives a bijection $H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}\to H_{\tau,{ {\sigma} }_-,{ {\sigma} }_+}$, $(d_0,k^+,k^-)\mapsto (-d_0, k^-,k^+)$. Moduli spaces of stable maps to $({\mathcal{X}},{\mathcal{L}})$ --------------------------------------------------------------- Stable maps to orbifolds with Lagrangian boundary conditions and their moduli spaces have been introduced in [@CP Section 2]. Let $({\Sigma},x_1,\ldots, x_n)$ be a prestable bordered orbifold Riemann surface with $n$ interior marked points in the sense of [@CP Section 2]. Then the coarse moduli space $(\bar{{\Sigma}},\bar{x}_1,\ldots, \bar{x}_n)$ is a prestable bordered Riemann surface with $n$ interior marked points, defined in [@KL01 Section 3.6] and [@Liu02 Section 3.2]. We define the topological type $(g,h)$ of ${\Sigma}$ to be the topological type of $\bar{{\Sigma}}$ (see [@Liu02 Section 3.2]). Let $({\Sigma},{\partial{\Sigma}})$ be a prestable bordered orbifold Riemann surface of type $(g,h)$, and let ${\partial{\Sigma}}= R_1\cup \cdots \cup R_h$ be union of connected component. Each connected component is a circle which contains no orbifold points. Let $u:({\Sigma},{\partial{\Sigma}})\to ({\mathcal{X}},{\mathcal{L}})$ be a (bordered) stable map in the sense of [@CP Section 2]. The topological type of $u$ is given by the degree $\beta'= f_*[{\Sigma}]\in H_2({\mathcal{X}},{\mathcal{L}};{ \mathbb{Z} })$, and $$\begin{cases} (\mu_i, k_i)=f_*[R_i]\in H_1({\mathcal{L}};{ \mathbb{Z} }) =H_{\tau,{ {\sigma} }}\subset { \mathbb{Z} }\times G_{ {\sigma} }, & \textup{if } \ell_\tau={ \mathbb{C} },\\ (\mu_i,k_i^+, k_i^-) = f_*[R_i]\in H_1({\mathcal{L}};{ \mathbb{Z} })= H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-} \subset { \mathbb{Z} }\times G_{{ {\sigma} }_+}\times G_{{ {\sigma} }_-}, & \textup{if } \ell_\tau={ \mathbb{P} }^1. \end{cases}$$ We call $\mu_i$ winding numbers and $k_i, k_i^\pm$ twistings. Given $\beta'\in H_2({\mathcal{X}},{\mathcal{L}};{ \mathbb{Z} })$ and $${ {\vec{\mu}} }= \begin{cases} ((\mu_1,k_1),\ldots, (\mu_h,k_h)) \in H_1({\mathcal{L}};{ \mathbb{Z} })^h &\textup{if } \ell_\tau={ \mathbb{C} },\\ ((\mu_1,k_1^+, k_1^-),\ldots, (\mu_h,k_h^+, k_h^-)) \in H_1({\mathcal{L}};{ \mathbb{Z} })^h & \textup{if } \ell_\tau={ \mathbb{P} }^1. \end{cases}$$ let ${\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} })$ be the moduli space of stable maps of type $(g,h)$, degree $\beta'$, winding numbers and twisting ${ {\vec{\mu}} }$, with $n$ interior marked points. The tangent space ${\mathcal{T}}_\xi ^1$ and the obstruction space ${\mathcal{T}}_\xi^2$ at $$\xi=[u:( ({\Sigma}, x_1,\ldots,x_n), {\partial{\Sigma}})\to ({\mathcal{X}},{\mathcal{L}})] \in {\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} })$$ fit into the following exact sequence: $$\begin{aligned} 0 &\to & {\mathrm{Aut}}(({\Sigma},x_1,\ldots,x_n),{\partial{\Sigma}})\to H^0({\Sigma},{\partial{\Sigma}},u^*T{\mathcal{X}}, (u|_{{\partial{\Sigma}}})^*T{\mathcal{L}})\to {\mathcal{T}}^1_\xi \\ &\to & {\mathrm{Def}}(({\Sigma},x_1,\ldots, x_n),{\partial{\Sigma}})\to H^1({\Sigma},{\partial{\Sigma}}, u^*T{\mathcal{X}}, (u|_{{\partial{\Sigma}}})^* T{\mathcal{L}})\to {\mathcal{T}}^2_\xi,\end{aligned}$$ where ${\mathrm{Aut}}(({\Sigma},x_1,\ldots, x_n),{\partial{\Sigma}})$ (resp. ${\mathrm{Def}}(({\Sigma},x_1,\ldots, x_n),{\partial{\Sigma}})$) is the space of infinitesimal automorphisms (resp. deformations) of the domain. When ${\Sigma}$ is smooth, $$\begin{aligned} {\mathrm{Aut}}(({\Sigma},x_1,\ldots, x_n),{\partial{\Sigma}}) &=& H^0({\Sigma},{\partial{\Sigma}}, T{\Sigma}(-\sum_{j=1}^nx_j), T{\partial{\Sigma}}), \\ {\mathrm{Def}}(({\Sigma},x_1,\ldots, x_n),{\partial{\Sigma}}) &=& H^1({\Sigma},{\partial{\Sigma}}, T{\Sigma}(-\sum_{j=1}^nx_j), T{\partial{\Sigma}}). \end{aligned}$$ There are evaluation maps (at interior marked points) $${\mathrm{ev}}_j:{\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} }) \to {\mathcal{I}}{\mathcal{X}},\quad j=1,\ldots,n.$$ Given $\vec{v}=(v_1,\ldots, v_n)$, where $v_1,\ldots, v_n\in {\mathrm{Box}(\mathbf {\Sigma})}$, define $${\overline{{\mathcal{M}}}}_{(g,h),\vec{v}}({\mathcal{X}},{\mathcal{L}}\mid\beta',{ {\vec{\mu}} }) := \bigcap_{j=1}^n {\mathrm{ev}}_j^{-1}({\mathcal{X}}_{v_j}).$$ Then the virtual (real) dimension of ${\overline{{\mathcal{M}}}}_{g,\vec{v}}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} })$ is $$2 \sum_{j=1}^n(1-{\mathrm{age}}(v_j)).$$ where ${\mathrm{age}}(v_j)\in \{0,1,2\}$. Torus action and equivariant invariants --------------------------------------- Let ${ \mathbb{T} }'_{ \mathbb{R} }\cong U(1)^2$ be the maximal compact subgroup of ${ \mathbb{T} }'\cong ({ \mathbb{C} }^*)^2$. Then the ${ \mathbb{T} }'_{ \mathbb{R} }$-action on ${\mathcal{X}}$ is holomorphic and preserves ${\mathcal{L}}$, so it acts on the moduli spaces ${\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid\beta',{ {\vec{\mu}} })$. Given $\gamma_1,\ldots, \gamma_n \in H^*_{{ \mathbb{T} }',{ {\mathrm{orb}} }}({\mathcal{X}};{ \mathbb{Q} })=H^*_{{ \mathbb{T} }'_{ \mathbb{R} },{ {\mathrm{orb}} }}({\mathcal{X}},{ \mathbb{Q} })$, we define $$\langle \gamma_1,\ldots, \gamma_n\rangle^{{\mathcal{X}},{\mathcal{L}},{ \mathbb{T} }_{ \mathbb{R} }'}_{g,\beta',{ {\vec{\mu}} }}:= \int_{[F]^{{ {\mathrm{vir}} }}} \frac{\prod_{j=1}^n ({\mathrm{ev}}_j^*\gamma_i)|_F}{e_{{ \mathbb{T} }_{ \mathbb{R} }'}(N^{ {\mathrm{vir}} }_F)} \in {\mathcal{Q}}_{{ \mathbb{T} }'}$$ where $F\subset {\overline{{\mathcal{M}}}}_{(g,h), n}({\mathcal{X}},{\mathcal{L}}\mid\beta',{ {\vec{\mu}} })$ is the ${ \mathbb{T} }_{ \mathbb{R} }'$-fixed points set of the ${ \mathbb{T} }_{ \mathbb{R} }'$-action on ${\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid\beta',{ {\vec{\mu}} })$ and ${\mathcal{Q}}_{{ \mathbb{T} }_{ \mathbb{R} }'}\cong { \mathbb{Q} }({\mathsf{w}}_1,{\mathsf{w}}_2)$ is the fractional field of $H^*_{{ \mathbb{T} }_{ \mathbb{R} }'}({\mathrm{point}};{ \mathbb{Q} })\cong { \mathbb{Q} }[{\mathsf{w}}_1,{\mathsf{w}}_2]$. Disk factor as equivariant open GW invariants --------------------------------------------- Suppose that ${\mathcal{L}}$ is an inner brane. Let $(d,k_+, k_-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-} \subset { \mathbb{Z} }\times G_{{ {\sigma} }_+}\times G_{{ {\sigma} }_-}$. Let $$s_1^\pm= r(\tau,{ {\sigma} }_\pm).$$ Suppose that $$I'_{{ {\sigma} }_+}=\{i_1, i_2, i_3\},\quad I'_{{ {\sigma} }_-}=\{ i_4, i_3, i_2\}.$$ Let $$L_2 = {\mathcal{O}}_{{\mathcal{X}}}({\mathcal{D}}_{i_2})\big|_{{\mathfrak{l}}_\tau},\quad L_3={\mathcal{O}}_{{\mathcal{X}}}({\mathcal{D}}_{i_3})|_{{\mathfrak{l}}_\tau}.$$ Then $$N_{{\mathfrak{l}}_\tau/{\mathcal{X}}} = L_2\oplus L_3$$ Let $${\mathsf{w}}_j =(c_1)_{{ \mathbb{T} }'}\big({\mathcal{O}}_{\mathcal{X}}({\mathcal{D}}_{i_j})\big)|_{{\mathfrak{p}}_+},\quad j=1,2,3.$$ Then ${\mathsf{w}}_1+{\mathsf{w}}_2+{\mathsf{w}}_3=0$. $$\begin{aligned} && c_1(T_{{\mathfrak{p}}_+}{\mathfrak{l}}_\tau) ={\mathsf{w}}_1= \frac{{\mathsf{u}}}{s_1^+},\quad c_1(L_2)|_{{\mathfrak{p}}_+}= {\mathsf{w}}_2,\quad c_1(L_3)|_{{\mathfrak{p}}_+}= {\mathsf{w}}_3,\\ && c_1(T_{{\mathfrak{p}}_-}{\mathfrak{l}}_\tau) = -\frac{{\mathsf{u}}}{s_1^-},\quad c_1(L_2)|_{{\mathfrak{p}}_+} ={\mathsf{w}}_2-a_2 {\mathsf{u}},\quad c_1(L_3)|_{{\mathfrak{p}}_-}= {\mathsf{w}}_3-a_3{\mathsf{u}},\end{aligned}$$ where $\deg L_i = a_i\in { \mathbb{Q} }$. The Lagrangian ${\mathcal{L}}$ intersects ${\mathfrak{l}}_\tau$ along a circle, which divides ${\mathfrak{l}}_\tau$ into two orbi-disks $D_+ \cong [D/{ \mathbb{Z} }_{s_1^+}]$ and $D_-\cong[D/{ \mathbb{Z} }_{s_1^-}]$, where $D=\{z\in { \mathbb{C} }\mid |z|\leq 1\}$, and ${\mathfrak{p}}_\pm$ is the unique ${\mathcal{T}}$ fixed point in $D_\pm$. Let $$b =[D_+]\in H_2({\mathcal{X}},{\mathcal{L}}), \quad \alpha\in [{\mathfrak{l}}_\tau]\in H_2({\mathcal{X}}), \quad \alpha-b\in H_2({\mathcal{X}},{\mathcal{L}}).$$ Let $(d_0,k^+,k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}$, where $d_0\neq 0$. Define $${\overline{{\mathcal{M}}}}(d_0,k^+,k^-) := \begin{cases} {\overline{{\mathcal{M}}}}_{(0,1),1}({\mathcal{X}},{\mathcal{L}}\mid d_0 b, (d_0,k^+, k^-)), & d_0>0,\\ {\overline{{\mathcal{M}}}}_{(0,1),1}({\mathcal{X}},{\mathcal{L}}\mid -d_0(\alpha-b), (d_0, k^+, k^-)), & d_0<0. \end{cases}$$ $$\textup{virtual dimension of }{\overline{{\mathcal{M}}}}(d_0,k^+,k^-)= \begin{cases} 1 -{\mathrm{age}}(k^+), & d_0 >0,\\ 1-{\mathrm{age}}(k^-), & d_0<0. \end{cases}$$ Define $$D(d_0,k^+, k^-):= \begin{cases} \langle 1_{k^+}\rangle_{0,d_0 b, (d_0 ,k^+,k^-)}^{{\mathcal{X}},{\mathcal{L}}}, & d_0 >0,\\ & \\ \langle 1_{k^-}\rangle_{0,-d_0 (\alpha-b), (d_0,k^+, k^-)}^{{\mathcal{X}}, {\mathcal{L}}}, & d_0 <0. \end{cases}.$$ Then $D(d_0,k^+, k^-)$ is a rational function in ${\mathsf{w}}_1,{\mathsf{w}}_2$, homogeneous of degree ${\mathrm{age}}(k_+)-1$ (resp. ${\mathrm{age}}(k_-)-1$) if $d_0>0$ (resp. $d_0<0$). Similarly, when ${\mathcal{L}}$ is an outer brane, we define $D(d_0,k)\in { \mathbb{Q} }({\mathsf{w}}_1,{\mathsf{w}}_2)$, where $d_0>0$ and $k\in G_{ {\sigma} }$. The disk factor is computed in [@BC11] when $G_{ {\sigma} }$ is cyclic, and in [@Ro11 Section 3.3] for general $G_{ {\sigma} }$. In our notation, the formula in [@Ro11 Section 3.3] says[^3] $$\label{eqn:disk} \begin{aligned} D(d_0,k) =& (\frac{s_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k)-1} \frac{s_1}{d_0|G_{ {\sigma} }|} \cdot \frac{ \prod_{a=1}^{ \frac{d_0}{s_1} +c_{i_2}(k) +c_{i_3}(k) -1} (\frac{d_0 {\mathsf{w}}_2}{{\mathsf{w}}_1} + a-c_{i_2}(k)) }{\lfloor \frac{d_0}{s_1}\rfloor!} \\ =& (\frac{s_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k)-1} \frac{1}{d_0 |G_1|} \cdot \frac{ \prod_{a=1}^{ \lfloor \frac{d_0}{s_1}\rfloor+ {\mathrm{age}}(k) -1}(\frac{d_0 {\mathsf{w}}_2}{{\mathsf{w}}_1} + a-c_{i_2}(k) ) }{\lfloor \frac{d_0}{s_1}\rfloor!} \end{aligned}$$ where $c_i(k)\in { \mathbb{Q} }\cap [0,1)$ is defined as in Section \[sec:CR\]. Disk factor as equivariant relative GW invariants ------------------------------------------------- The disk factors $D(d_0, k^+, k^-)$ and $D(d_0,k)$ are rational functions in ${\mathsf{w}}_1, {\mathsf{w}}_2$. In order to obtain a rational number, we specialize to a 1-dimensional subtorus of ${ \mathbb{T} }'_{ \mathbb{R} }$ determined by the framing of the Aganagic-Vafa A-brane ${\mathcal{L}}$. When ${\mathcal{X}}$ is a smooth toric Calabi-Yau 3-fold, the framing is an integer. In this section, we clarify the framing of an Aganagic-Vafa A-brane ${\mathcal{L}}$ in a 3-dimensional Calabi-Yau smooth toric DM stack. We then reinterpret the disk factors $D(d_0,k^+,k^-)$ and $D(d_0,k)$ as equivariant relative Gromov-Witten invariants, which gives a canonical choice of the sign of the disk factor. Let ${\mathcal{L}}$ be an inner brane. Then there exists $f^+,f^-\in { \mathbb{Z} }$ such that $$a_2=\deg L_2= \frac{f^+}{s_1^+} - \frac{f^-+1}{s_1^-},\quad a_3=\deg L_3 =\frac{f^-}{s_1^-}-\frac{f^++1}{s_1^+}.$$ We call such a choice $(f^+,f^-)$ a [*framing*]{} of the Aganagic-Vafa brane ${\mathcal{L}}$. Given a framing $(f^+,f^-)$ of ${\mathcal{L}}$, we degenerate ${\mathfrak{l}}_\tau$ to a nodal curve with two irreducible components ${\mathfrak{l}}_+$ and ${\mathfrak{l}}_-$, such that the stabilizer of the node ${\mathfrak{p}}_0$ is $G_\tau$. Over the coarse curves $\ell_\tau$ and $\ell_\pm$ of ${\mathfrak{l}}_\tau$ and ${\mathfrak{l}}_\pm$: - we degenerate $L_2$ on $\ell_\tau$ to $L_2^+ ={\mathcal{O}}(\frac{f^+}{s^+_1})$ on $\ell_+$ and $L_2^- ={\mathcal{O}}(-\frac{f^-+1}{s_1^-})$ on $\ell_-$; - we degenerate $L_3$ on $\ell_\tau$ to $L_3^+ ={\mathcal{O}}(-\frac{f^++1}{s^+_1})$ on $\ell_+$ and $L_3^-= {\mathcal{O}}(\frac{f^-}{s_1^-})$ on $\ell_-$. $$(c_1)_{{ \mathbb{T} }'}(L_2^\pm)_{{\mathfrak{p}}_0}={\mathsf{w}}_2-f^+{\mathsf{w}}_1, = -(c_1)_{{ \mathbb{T} }'}(L_3^\pm)_{{\mathfrak{p}}_0}$$ We compute the disk factor $D(d_0 ,k^+,k^-)$ by computing genus zero relative stable maps to of $({\mathfrak{l}}_+,{\mathfrak{p}}_+)$ or $({\mathfrak{l}}_-,{\mathfrak{p}}_-)$. More precisely, let ${\mathcal{M}}={\mathcal{M}}_{0,1}({\mathfrak{l}}_+ ,{\mathfrak{p}}_0, (d_0))$ be the moduli space of relative stable maps $u:({\mathcal{C}},x,y)\to {\mathfrak{l}}_+$ such that $u^* {\mathfrak{p}}_0 = d_0 y$, where $({\mathcal{C}},x,y)$ is a genus 0 orbicurve with 2 marked points. Let $\pi: {\mathcal{U}}\to {\mathcal{M}}$ be the universal curve and let $u:{\mathcal{U}}\to {\mathcal{T}}$ be the universal map to the universal target ${\mathcal{T}}$. Let ${\mathrm{ev}}:{\mathcal{M}}\to {\mathcal{I}}{\mathfrak{l}}_+$ be the evaluation map at the (stacky) point $x$. We define $$\langle 1_{k^+} \rangle^{{\mathcal{X}},{\mathcal{L}}}_{0,d_0 b, k^+,k^-} = \int_{[{\mathcal{M}}_{0,1}({\mathfrak{l}}_+,{\mathfrak{p}}_0,(d_0))]^{ {\mathrm{vir}} }} {\mathrm{ev}}^*(1_{k^+})e_{{ \mathbb{T} }'}(-R^\bullet\pi_* (u^* L_2^+ \oplus u^*L_3^+ \ominus u^*L_2^+ \otimes {\mathcal{O}}_R)).$$ Let $u:({\mathcal{C}},x,y)\to {\mathfrak{l}}_+ $ be relative stable map which represents a point in ${\mathcal{M}}$. Suppose that $u$ is fixed by the torus action. Recall that $c_i:G_{ {\sigma} }\to [0,1)\cap { \mathbb{Q} }$ is defined by $\chi_i(k)=\exp(2\pi\sqrt{-1}c_i(k))$. For $j=1, 2,3$, let ${\epsilon}_j= c_{i_j}(k^+)$. Then ${\epsilon}_1=\langle\frac{d_0}{s_1^+}\rangle$. $$\begin{aligned} { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^0({\mathcal{C}},u^*L_1^+) \bigr) &=& \sum_{a=0}^{\lfloor \frac{d_0}{s^+_1}\rfloor} e^{a \frac{s_1^+{\mathsf{w}}_1}{d_0}}\\ { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^1({\mathcal{C}},u^*L_1^+) \bigr) &=& 0 \\ { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^0({\mathcal{C}},u^*L_2^+\otimes {\mathcal{O}}_y) \bigr) &=& \delta_{\langle\frac{fd_0}{s_1^+}-{\epsilon}_2\rangle, 0} e^{{\mathsf{w}}_2-f^+{\mathsf{w}}_1} \\ { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^1({\mathcal{C}},u^*L_2^+\otimes {\mathcal{O}}_y) \bigr) &=& 0\\ \end{aligned}$$ $$\begin{aligned} { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^0({\mathcal{C}}, u^*L^+_2) \bigr) &=& \begin{cases} \displaystyle{ \sum_{a= -\lfloor \frac{f^+ d_0 }{s^+_1}-{\epsilon}_2\rfloor}^0 e^{{\mathsf{w}}_2+(a-{\epsilon}_2) \frac{s_1^+{\mathsf{w}}_1}{d_0 } } }, & f^+ \geq 0,\\ 0, & f^+<0,\\ \end{cases} \\ { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^1({\mathcal{C}}, u^*L^+_2) \bigr) &=& \begin{cases} 0, & f^+\geq 0,\\ \displaystyle{ \sum_{a=1}^{-\lfloor \frac{f^+ d_0 }{s^+_1}-{\epsilon}_2\rfloor-1} e^{{\mathsf{w}}_2+(a-{\epsilon}_2)\frac{s^+_1{\mathsf{w}}_1}{d_0 } } }, & f^+<0 \end{cases}\end{aligned}$$ $$\begin{aligned} { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^0({\mathcal{C}}, u^*L^+_3) \bigr) &=& \begin{cases} \displaystyle{ \sum_{a=-\lfloor-\frac{(f^+ +1)d_0 }{s^+_1}-{\epsilon}_3\rfloor}^0 e^{-{\mathsf{w}}_1-{\mathsf{w}}_2+(a-{\epsilon}_3) \frac{s^+_1{\mathsf{w}}_1}{d_0 } } }, & f^+<0,\\ 0, & f^+\geq 0,\\ \end{cases} \\ { {\mathrm{ch}} }_{{ \mathbb{T} }'}\bigl( H^1({\mathcal{C}}, u^*L^+_3) \bigr) &=& \begin{cases} 0, & f^+ <0 ,\\ \displaystyle{ \sum_{a=1}^{-\lfloor \frac{-(f^+ +1)d_0 }{s^+_1} +1 - {\epsilon}_3\rfloor} e^{-{\mathsf{w}}_1-{\mathsf{w}}_2+(a-{\epsilon}_3)\frac{s^+_1{\mathsf{w}}_1}{d_0 } } }, & f^+\geq 0. \end{cases}\end{aligned}$$ where $$\begin{aligned} \sum_{a=-\lfloor-\frac{(f^+ +1)d_0 }{s^+_1}-{\epsilon}_3 \rfloor}^0 e^{-{\mathsf{w}}_1-{\mathsf{w}}_2+(a-{\epsilon}_3) \frac{s^+_1{\mathsf{w}}_1}{d_0 } } &=& \sum_{a=\frac{d_0 }{s^+_1}+{\epsilon}_2+{\epsilon}_3}^{- \lfloor\frac{f^+ d_0 }{s^+_1}-{\epsilon}_2\rfloor -1 + \delta_{\langle \frac{f^+ d_0 }{s^+_1}-{\epsilon}_2 \rangle , 0}} e^{-w_2 +({\epsilon}_2-a)\frac{s^+_1{\mathsf{w}}_1}{d_0 }} \\ \sum_{a=1}^{-\lfloor \frac{-(f^+ +1)d_0 }{s_1} +1 - {\epsilon}_3 \rfloor} e^{-{\mathsf{w}}_1-{\mathsf{w}}_2+(a-{\epsilon}_3)\frac{s^+_1{\mathsf{w}}_1}{d_0 } } &=& \sum_{a=-\lfloor \frac{f^+ d_0 }{s^+_1}-{\epsilon}_2\rfloor +\delta_{\langle \frac{f^+ d_0 }{s^+_1} \rangle, {\epsilon}_2}}^{\frac{d_0 }{s^+_1} +{\epsilon}_2+{\epsilon}_3-1} e^{-w_2 +({\epsilon}_2-a)\frac{s^+_1{\mathsf{w}}_1}{d_0 }} \end{aligned}$$ $$\frac{d_0}{s^+_1} + {\epsilon}_2 +{\epsilon}_3 = \lfloor \frac{d_0 }{s^+_1}\rfloor + {\mathrm{age}}(k^+)$$ We have $$\begin{aligned} e_{{ \mathbb{T} }'}(B_1^m) &=& 1\\ e_{{ \mathbb{T} }'}(B_2^m) &=& \lfloor \frac{d_0}{s_1^+} \rfloor! (\frac{s_1^+{\mathsf{w}}_1}{d_0})^{\lfloor \frac{d_0}{s_1^+} \rfloor} \\ \frac{e_{{ \mathbb{T} }'}(B_5^m)}{e_{{ \mathbb{T} }'}(B_4^m)} &=& (-1)^{\lceil \frac{f^+ d_0}{s^+_1}-{\epsilon}_2\rceil +\lfloor \frac{d_0}{s^+_1}\rfloor +{\mathrm{age}}(k^+)-1} \prod_{a =1}^{\lfloor \frac{d_0 }{s^+_1}\rfloor + {\mathrm{age}}(k^+) -1}({\mathsf{w}}_2 + (a-{\epsilon}_2)\frac{s^+_1{\mathsf{w}}_1}{d_0}). \end{aligned}$$ $$|{\mathrm{Aut}}(f)| = d_0|G_1|$$ $$\begin{aligned} && D(d_0,k^+,k^-) = \frac{1}{|{\mathrm{Aut}}(f)|} \frac{e_{{ \mathbb{T} }'}(B_1^m) e_{{ \mathbb{T} }'}(B_5^m)}{e_{{ \mathbb{T} }'}(B_2^m) e_{{ \mathbb{T} }'}(B_4^m)} \\ &=& \frac{1}{d_0|G_1|}\frac{\prod_{a=1}^{\lfloor \frac{d_0}{s^+_1}\rfloor + {\mathrm{age}}(k^+) -1}({\mathsf{w}}_2 + (a-{\epsilon}_2)\frac{s^+_1{\mathsf{w}}_1}{d_0})} {\lfloor\frac{d_0}{s^+_1}\rfloor! (\frac{s^+_1{\mathsf{w}}_1}{d_0})^{\lfloor \frac{d_0}{s^+_1}\rfloor}} \cdot (-1)^{\lceil \frac{f^+ d_0}{s^+_1}-{\epsilon}_2\rceil +\lfloor \frac{d_0}{s^+_1}\rfloor +{\mathrm{age}}(k^+)-1 }. \\ &=&(-1)^{\lceil \frac{f^+ d_0}{s^+_1}-{\epsilon}_2\rceil +\lfloor \frac{d_0}{s^+ _1}\rfloor +{\mathrm{age}}(k^+)-1 } (\frac{s^+_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k^+)-1} \cdot \frac{1}{d_0|G_1|}\frac{\prod_{a=1}^{\lfloor \frac{d_0}{s^+_1}\rfloor +{\mathrm{age}}(k^+)-1} (\frac{d_0 {\mathsf{w}}_2}{s^+_1{\mathsf{w}}_1} +a-{\epsilon}_2)} {\lfloor\frac{d_0 }{s^+_1}\rfloor !} \\ &=& - (-1)^{\frac{d_0}{s_1^+}(-f^+-1) -{\epsilon}_3 -\{ {\epsilon}_2 -f^+ {\epsilon}_1 \} } (\frac{s^+_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k^+)-1} \cdot \frac{1}{d_0|G_1|}\frac{\prod_{a=1}^{\lfloor \frac{d_0}{s^+_1}\rfloor +{\mathrm{age}}(k^+)-1} (\frac{d_0 {\mathsf{w}}_2}{s^+_1{\mathsf{w}}_1} +a-{\epsilon}_2)} {\lfloor\frac{d_0}{s^+_1}\rfloor !} \\\end{aligned}$$ If $d_0 >0$, define $$\begin{aligned} && D(d_0 ,k^+,k^-;f^+,f^-) \\ &=& -(-1)^{\frac{d_0 }{s_1^+}(-f^+-1)-{\epsilon}_3 -\{ {\epsilon}_2- f^+ {\epsilon}_1\} } (\frac{s^+_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k^+)-1} \cdot \frac{1}{d_0|G_1|}\frac{\prod_{a=1}^{\lfloor \frac{d_0 }{s_1}\rfloor +{\mathrm{age}}(k^+)-1} (\frac{f^+ d_0 }{s^+_1}-c_{i_2}(k^+) +a)} {\lfloor\frac{d_0 }{s^+_1}\rfloor !}.\end{aligned}$$ Let $${\epsilon}_1'=c_{i_1}(k^-)=\langle -\frac{d_0}{s_1^-}\rangle,\quad {\epsilon}_2'=c_{i_2}(k^-),\quad {\epsilon}_3'=c_{i_3}(k^-).$$ If $d_0 <0$, define $$\begin{aligned} && D(d_0 ,k^+,k^-;f^+,f^-)\\ &=& -(-1)^{\frac{d_0 }{s_1^-}(f^-+1)-{\epsilon}'_2 -\{ {\epsilon}_3'- f^-{\epsilon}_1' \} } (\frac{s^+_1{\mathsf{w}}_1}{d_0 })^{{\mathrm{age}}(k^-)-1} \cdot \frac{1}{-d_0 |G_1|}\frac{\prod_{a=1}^{\lfloor \frac{-d_0 }{s^-_1}\rfloor +{\mathrm{age}}(k^-)-1} (\frac{-f^- d_0 }{s^-_1}-c_{i_3}(k^-) +a)} {\lfloor\frac{-d_0 }{s^-_1}\rfloor !} \\ &=& -(-1)^{\frac{d_0 }{s_1^-}(-f^--1)+ {\epsilon}'_2 -\{ {\epsilon}_2 - f^+ {\epsilon}_1 \} } (\frac{s^+_1{\mathsf{w}}_1}{d_0 })^{{\mathrm{age}}(k^-)-1} \cdot \frac{1}{-d_0 |G_1|}\frac{\prod_{a=1}^{\lfloor \frac{-d_0 }{s^-_1}\rfloor +{\mathrm{age}}(k^-)-1} (\frac{-f^- d_0 }{s^-_1}-c_{i_3}(k^-) +a)} {\lfloor\frac{-d_0 }{s^-_1}\rfloor !}\end{aligned}$$ where we use the identity $$\{ {\epsilon}_3'-f^-{\epsilon}_1' \} = \{ - ({\epsilon}_2-f^+{\epsilon}_1) \}.$$ If ${\mathcal{L}}$ is an outer brane, then the framing of ${\mathcal{L}}$ is an integer $f$. Define $$D(d_0 ,k;f) = -(-1)^{\frac{d_0 }{s_1}(-f-1) -{\epsilon}_3 -\{ {\epsilon}_2- f{\epsilon}_1 \} } (\frac{s_1{\mathsf{w}}_1}{d_0 })^{{\mathrm{age}}(k)-1} \cdot \frac{1}{d_0 |G_1|}\frac{\prod_{a=1}^{\lfloor \frac{d_0 }{s_1}\rfloor +{\mathrm{age}}(k)-1} (\frac{fd_0 }{s_1}-c_{i_2}(k) +a)} {\lfloor\frac{d_0 }{s_1}\rfloor !}.$$ Open-closed GW invariants and descendant GW invariants ------------------------------------------------------ For any torus fixed point ${\mathfrak{p}}_{ {\sigma} }$ of ${\mathcal{X}}$, where ${ {\sigma} }\in {\Sigma}(3)$, we have $$H^*_{ {\mathrm{orb}} }({\mathfrak{p}}_{ {\sigma} }) =\bigoplus_{k \in G_{ {\sigma} }} { \mathbb{Q} }{\mathbf{1}}_k,\quad H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathfrak{p}}_{ {\sigma} }) =\bigoplus_{k \in G_{ {\sigma} }}{ \mathbb{Q} }[u_1,u_2] {\mathbf{1}}_k.$$ The inclusion $\iota_{ {\sigma} }: {\mathfrak{p}}_{ {\sigma} }\hookrightarrow {\mathcal{X}}$ induces $$\iota_{{ {\sigma} }*} : H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathfrak{p}}_{ {\sigma} }) = H^*_{{ \mathbb{T} }'}({\mathcal{I}}{\mathfrak{p}}_{ {\sigma} }) \to H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathcal{X}}) = H^*_{{ \mathbb{T} }'}({\mathcal{I}}{\mathcal{X}}).$$ Define $$\phi_{{ {\sigma} },k} = \iota_{{ {\sigma} }*} {\mathbf{1}}_k\in H^*_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathcal{X}}).$$ We first assume that ${\mathcal{L}}$ is an inner brane, let ${\mathfrak{l}}_\tau$ be the unique 1-dimensional orbit closure which intersects ${\mathcal{L}}$, and let $V_\tau=\{{ {\sigma} }_+,{ {\sigma} }_-\}$. Let $${ {\vec{\mu}} }=((\mu_1,k_1^+, k_1^-), \ldots, (\mu_h,k_h^+, k_h^-)),$$ where $(\mu_j, k_j^+, k_j^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}$. Let $J_\pm=\{ j\in \{1,\ldots, h\}: \pm \mu_j>0\}$. Then there exists $\beta\in H_2({\mathcal{X}})$ such that $$\beta'=\beta +(\sum_{j\in J_+} \mu_j)b + \sum_{j\in J_-}(-\mu_j)(\alpha-b).$$ Let $\langle k_j^+\rangle$ be the cyclic subgroup generated by $k_j^+$, and let $r_j^+$ be the cardinality of $\langle k_j^+\rangle$. Similarly, let $r_j^-$ be the cardinality of $\langle k_j^-\rangle$. We have $$\begin{aligned} {\overline{{\mathcal{M}}}}_{(g,h),n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} })^{{ \mathbb{T} }'} =\{ F_{ {\Gamma} }\mid { {\Gamma} }\in G_{g,n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} }) \}\\ {\overline{{\mathcal{M}}}}_{g,n+h}({\mathcal{X}}, \beta)^{{ \mathbb{T} }'} =\{ F_{ {\hat{{ {\Gamma} }}} }\mid { {\hat{{ {\Gamma} }}} }\in G_{g,n+h}({\mathcal{X}},\beta) \}\end{aligned}$$ In the remaining part of this subsection, we use the following abbreviations: $$\begin{aligned} && {\mathcal{M}}= {\overline{{\mathcal{M}}}}_{g,n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} }),\quad {\hat{{\mathcal{M}}}}= {\overline{{\mathcal{M}}}}_{g,n+h}({\mathcal{X}}, \beta), \\ && {\mathcal{M}}_j = \begin{cases} {\overline{{\mathcal{M}}}}_{(0,1),1}({\mathcal{X}},{\mathcal{L}}\mid \mu_j b, (\mu_j,k_j^+,k_j^-)), & j\in J_+\\ {\overline{{\mathcal{M}}}}_{(0,1),1}({\mathcal{X}}, {\mathcal{L}}\mid -\mu_j(\alpha-b), (\mu_j, k_j^+, k_j^-)), & j\in J_- \end{cases} \\ && {\mathcal{G}}= G_{g,n}({\mathcal{X}},{\mathcal{L}}\mid \beta',{ {\vec{\mu}} }),\quad {\hat{{\mathcal{G}}}}= G_{g,n+h}({\mathcal{X}},\beta)\\ && {\mathbf{x}}=(x_1,\ldots, x_n),\quad {\mathbf{y}}=(y_1,\ldots, y_h). \end{aligned}$$ Given $u:({\Sigma},{\mathbf{x}},{\partial{\Sigma}})\to ({\mathcal{X}}, {\mathcal{L}})$ which represents a point $\xi\in {\mathcal{M}}^{{ \mathbb{T} }'}$, we have $${\Sigma}= {\mathcal{C}}\cup \bigcup_{j=1}^h D_j,$$ where ${\mathcal{C}}$ is an orbicurve of genus $g$, $x_1,\ldots, x_n \in {\mathcal{C}}$, $D_j = [\{ z\in { \mathbb{C} }\mid |z|\leq 1\}/{ \mathbb{Z} }_{r_j^\pm}]$, ${\Sigma}$ and $D_j$ intersect at $y_j = B{ \mathbb{Z} }_{r_j^\pm}$ if $j\in J_\pm$. Let $u_j=u|_{D_j}$ and ${\hat{u}}= u|_{\mathcal{C}}$. Then 1. For $j=1,\ldots, h$, $u_j: (D_j,\partial D_j)\to ({\mathcal{X}}, {\mathcal{L}})$ represents a point in ${\mathcal{M}}_j^{{ \mathbb{T} }'}$. 2. ${\hat{u}}:({\mathcal{C}},{\mathbf{x}},{\mathbf{y}})\to {\mathcal{X}}$ represents a point $\hat{\xi}\in {\hat{{\mathcal{M}}}}^{{ \mathbb{T} }'}$, and ${\hat{u}}({\mathbf{y}}_j) =[{\mathfrak{p}}_\pm, (k_j^\pm)^{-1}] \in {\mathcal{I}}{\mathfrak{p}}_\pm \subset {\mathcal{I}}{\mathcal{X}}$ if $j\in J_\pm$. Let $x_{n+j}=y_j$. Let $F_{ {\Gamma} }$ be the connected component of ${\mathcal{M}}^{{ \mathbb{T} }'}$ associated to the decorated graph ${ {\Gamma} }\in {\mathcal{G}}$, and let $F_{{ {\hat{{ {\Gamma} }}} }}$ be the connected component of ${\hat{{\mathcal{M}}}}^{{ \mathbb{T} }'}$ associated to the decorated graph ${ {\Gamma} }'\in {\mathcal{G}}$. Then for any ${ {\Gamma} }\in {\mathcal{G}}$ there exists ${ {\hat{{ {\Gamma} }}} }\in {\hat{{\mathcal{G}}}}$ such that $${\mathrm{ev}}_{n+j}(F_{{ {\hat{{ {\Gamma} }}} }}) = ({\mathfrak{p}}_\pm,(k_j^\pm)^{-1}) \in {\mathcal{I}}{\mathfrak{p}}_\pm \subset {\mathcal{I}}{\mathcal{X}}$$ if $j\in J_\pm$, and $F_{{ {\Gamma} }}$ can be identified with $F_{{ {\hat{{ {\Gamma} }}} }}$ up to a finite morphism. More precisely, $$\begin{aligned} [F_{ {\Gamma} }]^{ {\mathrm{vir}} }&=& \prod_{j\in J_+} \frac{|G_{{ {\sigma} }_+}|}{r^+_j|{\mathrm{Aut}}(u_j)|} \prod_{j\in J_-} \frac{|G_{{ {\sigma} }_-}|}{r^-_j|{\mathrm{Aut}}(u_j)|} [F_{{ {\hat{{ {\Gamma} }}} }}]^{ {\mathrm{vir}} }\\ &=&\prod_{j\in J_+} \frac{s^+_1}{r^+_j\mu_j} \prod_{j\in J_-} \frac{s^-_1}{-r^-_j\mu_j} [F_{{ {\hat{{ {\Gamma} }}} }}]^{ {\mathrm{vir}} }. \end{aligned}$$ We have $$\frac{1}{e_{{ \mathbb{T} }'}(N_{ {\Gamma} }^{ {\mathrm{vir}} })} =\frac{e_{{ \mathbb{T} }'}(B_1^m) e_{{ \mathbb{T} }'}(B_5^m)}{e_{{ \mathbb{T} }'}(B_2^m) e_{{ \mathbb{T} }'}(B_4^m)},\quad \frac{1}{e_{{ \mathbb{T} }'}(N_{ {\hat{{ {\Gamma} }}} }^{ {\mathrm{vir}} })} =\frac{e_{{ \mathbb{T} }'}({\hat{B}}_1^m) e_{{ \mathbb{T} }'}({\hat{B}}_5^m) }{e_{{ \mathbb{T} }'}({\hat{B}}_2^m) e_{{ \mathbb{T} }'}({\hat{B}}_4^m)},$$ where $$\begin{aligned} e_{{ \mathbb{T} }'}(B_1^m)&=& e_{{ \mathbb{T} }'}({\hat{B}}_1^m), \\ e_{{ \mathbb{T} }'}(B_4^m) &=& e_{{ \mathbb{T} }'}({\hat{B}}_4^m) \prod_{j\in J_+}(\frac{s^+_1 {\mathsf{w}}_1}{r^+_j \mu_j} -\frac{\bar{\psi}_j}{r^+_j}) \prod_{j\in J_-}(\frac{s^+_1 {\mathsf{w}}_1}{r^-_j \mu_j}-\frac{\bar{\psi}_j}{r^-_j})\end{aligned}$$ For $k=0,1$ and $j=1,\ldots, h$, let $$H^k(D_j) = H^k\bigl(D_j, \partial D_j, u_j^*T{\mathcal{X}}, (u_j|_{\partial D_j})^*T{\mathcal{L}}\bigr).$$ Then there is a long exact sequence $$\begin{aligned} && 0 \to B_2 \to {\hat{B}}_2 \oplus \bigoplus_{j=1}^h H^0(D_j) \to \bigoplus_{j\in J_+} (T_{{\mathfrak{p}}_+}{\mathcal{X}})^{k_j^+}\oplus \bigoplus_{j\in J_-} (T_{{\mathfrak{p}}_-}{\mathcal{X}})^{k_j^-} \\ && \to B_5 \to {\hat{B}}_5 \oplus \bigoplus_{j=1}^h H^1(D_j)\to 0,\end{aligned}$$ where $(T_{{\mathfrak{p}}_\pm}{\mathcal{X}})^{k_j^\pm}$ denote the $k_j^\pm$-invariant part of $T_{{\mathfrak{p}}_\pm}{\mathcal{X}}$. Note that $$(T_{{\mathfrak{p}}_\pm}{\mathcal{X}})^{k_j^\pm} = T_{({\mathfrak{p}}_\pm, k_j^\pm)} {\mathcal{I}}{\mathcal{X}}= T_{({\mathfrak{p}}_\pm, k_j^{-1})} {\mathcal{I}}{\mathcal{X}}.$$ $$\begin{aligned} \frac{e_{{ \mathbb{T} }'}(H^1(D_j)^m)}{e_{{ \mathbb{T} }'}(H^0(D_j)^m)} = |\mu_j ||G_1| D(\mu_j ,k^+,k^-;f^+,f^-)\end{aligned}$$ Then $$\begin{aligned} && \int_{[F_{ {\Gamma} }]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i\big)|_{F_\Gamma}}{e_{{ \mathbb{T} }'}(N_{ {\Gamma} }^{ {\mathrm{vir}} })} \\ &=& \prod_{j\in J_+ }\frac{s_1^+}{r^+_j\mu_j} \prod_{j\in J_-}\frac{s_1^-}{- r^-_j\mu_j} \prod_{j=1}^h (|\mu_j||G_1|D(\mu_j,k_j^+,k_j^-;f^+,f^-)) \\ &&\cdot \int_{[F_{{ {\hat{{ {\Gamma} }}} }}]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i \prod_{j\in J_+}{\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_+, (k_j^+)^{-1}} \prod_{j\in J_-} {\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_-, (k_j^-)^{-1}}\big)|_{F_\Gamma}} {\prod_{j\in J_+}(\frac{s^+_1 {\mathsf{w}}_1}{r^+_j \mu_j} -\frac{\bar{\psi}_{n+j}}{r^+_j}) \prod_{j\in J_-}(\frac{s^+_1 {\mathsf{w}}_1}{r^-_j \mu_j} -\frac{\bar{\psi}_{n+j}}{r^-_j})e_{{ \mathbb{T} }'}(N_{ {\hat{{ {\Gamma} }}} }^{ {\mathrm{vir}} })} \\ &=& \prod_{j=1}^hD'(\mu_j,k^+_j,k_j^-;f^+,f^-) \\ && \cdot \int_{[F_{{ {\hat{{ {\Gamma} }}} }}]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i \prod_{j\in J_+}{\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_+, (k_j^+)^{-1}} \prod_{j\in J_-}{\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_-, (k_j^-)^{-1}}\big)|_{F_\Gamma}}{ \prod_{j=1}^h\frac{s_1^+{\mathsf{w}}_1}{\mu_j}(\frac{s^+_1{\mathsf{w}}_1}{\mu_j} -\bar{\psi}_{n+j})e_{{ \mathbb{T} }'}(N_{ {\hat{{ {\Gamma} }}} }^{ {\mathrm{vir}} })} \end{aligned}$$ where $$D'(d_0,k^+,k^-;f^+,f^-) = \begin{cases} \displaystyle{ -(-1)^{\frac{d_0}{s_1^+}(-f^+-1)-c_{i_3}(k_+) -\{ c_{i_2}(k_+)-f^+ c_{i_1}(k_+) \} } \frac{s_1^+}{d_0}(\frac{s^+_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k^+)}} & \\ \quad \quad \displaystyle{ \cdot \frac{\prod_{a=1}^{\lfloor \frac{d_0 }{s_1}\rfloor +{\mathrm{age}}(k^+)-1} (\frac{f^+ d_0 }{s^+_1}-c_{i_2}(k^+) +a)} {\lfloor\frac{d_0}{s^+_1}\rfloor !} }, \quad \quad d_0>0, & \\ & \\ \displaystyle{ -(-1)^{\frac{d_0}{s_1^-}(-f^--1) +c_{i_2}(k_-)-\{ c_{i_2}(k_+)-f^+ c_{i_1}(k_+) \} }\frac{s_1^-}{-d_0}(\frac{s^+_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k^-)} } & \\ \quad \quad \displaystyle{ \cdot \frac{\prod_{a=1}^{\lfloor \frac{-d_0}{s^-_1}\rfloor +{\mathrm{age}}(k^-)-1} (\frac{-f^- d_0}{s^-_1}-c_{i_3}(k^-) +a)} {\lfloor\frac{-d_0}{s^-_1}\rfloor !} } , \quad\quad d_0<0. & \\ \end{cases}$$ \[inner-psi\] Suppose that $({\mathcal{L}},f^+,f^-)$ is a framed inner brane, and $${ {\vec{\mu}} }=((\mu_1,k_1^+,k_1^-),\ldots, (\mu_h,k_h^+, k_h^-)),$$ where $(\mu_j,k_j^+, k_j^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}$. Let $J_\pm=\{ j\in \{1,\ldots,h\}: \pm \mu_j>0\}$. Then $$\begin{aligned} && \langle \gamma_1,\ldots, \gamma_n\rangle_{g,\beta',{ {\vec{\mu}} }}^{{\mathcal{X}}, ({\mathcal{L}},f^+,f^-)} =\prod_{j=1}^h D'(\mu_j,k_j^+,k_j^-;f^+,f^-) \\ && \quad \cdot \int_{[{\overline{{\mathcal{M}}}}_{g,n+h}({\mathcal{X}},\beta)]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i \prod_{j\in J_+}{\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_+, (k_j^+)^{-1}} \prod_{j\in J_-} {\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }_-, (k_j^-)^{-1}} \big)} {\prod_{j=1}^h\frac{s_1^+{\mathsf{w}}_1}{\mu_j}(\frac{s^+_1{\mathsf{w}}_1}{\mu_j} -\bar{\psi}_{n+j})} \end{aligned}$$ where $$\beta\in H_2({\mathcal{X}}),\quad \beta'= \beta + \big(\sum_{j\in J_+} \mu_j\big)b -\big(\sum_{j\in J_-} \mu_j\big)(\alpha-b) \in H_2({\mathcal{X}},{\mathcal{L}}).$$ Suppose that $({\mathcal{L}},f)$ is a framed outer brane, and $(d_0,k)\in H_{{ {\sigma} },\tau}\subset { \mathbb{Z} }_{>0}\times G_{ {\sigma} }$. Define $$D'(d_0,k;f)= -(-1)^{\frac{d_0}{s_1}(-f-1)-c_{i_3}(k)-\{ c_{i_2}(k) -fc_{i_1}(k) \} } \frac{s_1}{d_0}(\frac{s_1{\mathsf{w}}_1}{d_0})^{{\mathrm{age}}(k)} \cdot \frac{\prod_{a=1}^{\lfloor \frac{d_0}{s_1}\rfloor +{\mathrm{age}}(k)-1} (\frac{f d_0}{s_1}-c_{i_2}(k) +a)}{\lfloor \frac{d_0}{s_1}\rfloor!} .$$ \[outer-psi\] Suppose that $({\mathcal{L}},f)$ is a framed inner brane, and ${ {\vec{\mu}} }=((\mu_1,k_1),\ldots, (\mu_h,k_h))$, where $(\mu_j,k_j)\in H_{\tau,{ {\sigma} }}$. Then $$\langle \gamma_1,\ldots, \gamma_n\rangle_{g,\beta',{ {\vec{\mu}} }}^{{\mathcal{X}}, ({\mathcal{L}},f)} =\prod_{j=1}^h D'(\mu_j,k_j;f) \cdot \int_{[{\overline{{\mathcal{M}}}}_{g,n+h}({\mathcal{X}},\beta)]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i \prod_{j=1}^h{\mathrm{ev}}_{n+j}^*\phi_{{ {\sigma} }, (k_j)^{-1}}\big)} {\prod_{j=1}^h\frac{s_1 {\mathsf{w}}_1}{\mu_j}(\frac{s_1{\mathsf{w}}_1}{\mu_j} -\bar{\psi}_{n+j})}$$ where $$\beta\in H_2({\mathcal{X}}),\quad \beta'= \beta + \big(\sum_{j=1}^h \mu_j\big)b.$$ Generating functions of open-closed GW invariants ------------------------------------------------- Introduce variables $\{ X_j \mid j=1,\ldots,h\}$ and let $${\boldsymbol{\tau}}_2 = \sum_{i=1}^m \tau_i u_i$$ where $u_1,\ldots, u_m$ form a basis of $H^2_{ {\mathrm{orb}} }({\mathcal{X}};{ \mathbb{Q} })$. We choose ${ \mathbb{T} }'$-equivariant lifting of ${\boldsymbol{\tau}}_2$ as follows: for each $u_i\in H^2_{ {\mathrm{orb}} }({\mathcal{X}};{ \mathbb{Q} })$, we choose the unique ${ \mathbb{T} }'$-equivariant lifting $u_i^{{ \mathbb{T} }'} \in H^2_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathcal{X}};{ \mathbb{Q} })$ such that $\iota_{ {\sigma} }^* u_i^{{ \mathbb{T} }'} = 0\in H^2_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathfrak{p}}_{ {\sigma} };{ \mathbb{Q} })$, where $\iota_{ {\sigma} }^*: H^2_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathcal{X}};{ \mathbb{Q} })\to H^2_{{ {\mathrm{orb}} },{ \mathbb{T} }'}({\mathfrak{p}}_{ {\sigma} };{ \mathbb{Q} })$ is induced by the inclusion map $\iota_{ {\sigma} }:{\mathfrak{p}}_{ {\sigma} }\to {\mathcal{X}}$. If ${\mathcal{L}}$ is an outer brane, define $$\begin{aligned} F_{g,h}^{{\mathcal{X}},{\mathcal{L}}}({\boldsymbol{\tau}}_2, Q^b, X_1,\ldots,X_h) =\sum_{\beta', n\geq 0}\sum_{(\mu_j, k_j)\in H_{\tau,{ {\sigma} }}} \frac{\langle ({\boldsymbol{\tau}}_2) ^n\rangle_{g,\beta,(\mu_1,k_1),\ldots, (\mu_h,k_h)}^{{\mathcal{X}},{\mathcal{L}}}}{n!} \prod_{j=1}^h (Q^bX_j)^{\mu_j} ({\mathbf 1}_{k_1}\otimes \cdots \otimes{\mathbf 1}_{k_h})\end{aligned}$$ which is a function which takes values in $H^*_{ {\mathrm{orb}} }({\mathfrak{p}}_{ {\sigma} };{ \mathbb{C} })^{\otimes h}$, where $$H^*_{ {\mathrm{orb}} }({\mathfrak{p}}_{ {\sigma} };{ \mathbb{C} })=\bigoplus_{k\in G_{ {\sigma} }} { \mathbb{C} }1_k.$$ If ${\mathcal{L}}$ is an inner brane, define $$\begin{aligned} F_{g,h}^{{\mathcal{X}},{\mathcal{L}}}({\boldsymbol{\tau}}_2, Q^b, X_1,\ldots, X_h) = & \sum_{\beta', n\geq 0}\sum_{(\mu_j, k_j^+,k_j^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}} \frac{\langle ({\boldsymbol{\tau}}_2)^n\rangle_{g,\beta,(\mu_1,k_1^+, k_1^-),\ldots, (\mu_h,k_h^+, k_h^-)}^{{\mathcal{X}},{\mathcal{L}}}}{n!} \\ & \cdot \prod_{\substack{ j\in \{1,\ldots, h\}\\ \mu_j>0} } (Q^b X_j)^{\mu_j} \prod_{\substack{ j\in \{1,\ldots, h\}\\ \mu_j<0} } (Q^{b-\alpha}X_j)^{\mu_j} ({\mathbf 1}_{k_1^+}\otimes \cdots \otimes {\mathbf 1}_{k_h^+}) \end{aligned}$$ which is a function which takes values in $H^*_{ {\mathrm{orb}} }({\mathfrak{p}}_{{ {\sigma} }_+}{ \mathbb{C} })^{\otimes h}$, where $$H^*_{ {\mathrm{orb}} }({\mathfrak{p}}_{{ {\sigma} }_+}{ \mathbb{C} })=\bigoplus_{k\in G_{{ {\sigma} }_+}}{ \mathbb{C} }1_k.$$ The equivariant $J$-function and the disk potential --------------------------------------------------- Let $\{u_i\}_{i=1}^N$ be a homogeneous basis of $H^*_{{ \mathbb{T} },{ {\mathrm{orb}} }}({\mathcal{X}};{ \mathbb{Q} })$, and $\{u^i\}_{i=1}^N$ be its dual basis. Define $${\boldsymbol{\tau}}=\sum_{i=1}^N \tau_i u_i ={\boldsymbol{\tau}}_0+ {\boldsymbol{\tau}}_2 + {\boldsymbol{\tau}}_{>2}$$ where $${\boldsymbol{\tau}}_0\in H^0_{{ \mathbb{T} },{ {\mathrm{orb}} }}({\mathcal{X}};{ \mathbb{C} }),\quad {\boldsymbol{\tau}}_2\in H^2_{{ \mathbb{T} },{ {\mathrm{orb}} }}({\mathcal{X}};{ \mathbb{C} }),\quad {\boldsymbol{\tau}}_{>2} \in H^{>2}_{{ \mathbb{T} },{ {\mathrm{orb}} }}({\mathcal{X}};{ \mathbb{C} }).$$ The $J$-function [@T; @CG07; @Gi96] is a $H^*_{{ \mathbb{T} },{ {\mathrm{orb}} }}({\mathcal{X}})$-valued function: $$J({\boldsymbol{\tau}},z):= 1+\sum_{\beta\ge 0, n\ge 0} \frac{1}{n!}\sum_{i=1}^N \langle 1,{\boldsymbol{\tau}}^n, \frac{u_i}{z-\bar \psi}\rangle^{\mathcal{X}}_{0,\beta} u^i.$$ Then $$\iota_{ {\sigma} }^* J({\boldsymbol{\tau}},z)|_{{\mathsf{w}}_1=f{\mathsf{w}}_2} =\sum_{k\in G_{ {\sigma} }} J^f_{{ {\sigma} },k}({\boldsymbol{\tau}},z) {\mathbf{1}}_k,$$ where $$J_{{ {\sigma} },k}^f({\boldsymbol{\tau}},z)= 1+\sum_{\beta\ge 0, n\ge 0} \frac{1}{n!} \sum_{i=1}^N \langle 1, {\boldsymbol{\tau}}^n, \frac{\phi_{{ {\sigma} },k^{-1}}}{z-\bar \psi}\rangle^{\mathcal{X}}_{0,\beta}.$$ As a special case of Proposition \[outer-psi\], $$\begin{aligned} \langle \gamma_1,\ldots, \gamma_n\rangle_{0,\beta+d_0 b,(d_0 ,k)}^{{\mathcal{X}},({\mathcal{L}},f)} &=& D'(d_0 ,k;f) \int_{[{\overline{{\mathcal{M}}}}_{0,n+1}({\mathcal{X}},\beta)]^{ {\mathrm{vir}} }} \frac{ \big(\prod_{i=1}^n{\mathrm{ev}}_i^*\gamma_i\cup {\mathrm{ev}}_{n+1}^*\phi_{{ {\sigma} },k^{-1}}\big)}{\frac{s_1{\mathsf{w}}_1}{d_0}(\frac{s_1{\mathsf{w}}_1}{d_0} -\bar{\psi}_{n+1})}\\ &=& D'(d_0 ,k;f) \langle 1,\gamma_1,\dots,\gamma_n,\frac{\phi_{{ {\sigma} },k^{-1}}}{\frac{s_1{\mathsf{w}}_1}{d_0} -\bar{\psi}} \rangle^{\mathcal{X}}_{0,\beta};\end{aligned}$$ $$\begin{aligned} F^{{\mathcal{X}},({\mathcal{L}},f)}_{0,1}({\boldsymbol{\tau}}_2,Q^b, X_1) &=& \sum_{\beta,n\geq 0} \sum_{(d_0 ,k)\in H_{\tau,{ {\sigma} }}} \frac{1}{n!}\langle ({\boldsymbol{\tau}}_2)^n\rangle^{{\mathcal{X}},({\mathcal{L}},f)}_{0,\beta+d_0 b,(d_0 ,k)} (Q^bX)^{d_0} \\ &=& \sum_{(d_0,k)\in H_{\tau,{ {\sigma} }}} (Q^b X_1)^{d_0} D'(d_0 ,k;f)J^f_{{ {\sigma} },k}({\boldsymbol{\tau}}_2 ,\frac{s_1{\mathsf{w}}_1}{d_0}){\mathbf{1}}_k .\end{aligned}$$ \[pro:F-J\] Let $X=Q^b X_1$. If $({\mathcal{L}},f)$ is a framed outer brane, then $$F^{{\mathcal{X}},({\mathcal{L}},f)}_{0,1}({\boldsymbol{\tau}}_2,Q^b, X_1) = \sum_{(d_0 ,k)\in H_{\tau,{ {\sigma} }}} X^{d_0} D'(d_0 ,k;f)J^f_{{ {\sigma} },k}({\boldsymbol{\tau}}_2,\frac{s_1{\mathsf{w}}_1}{d_0}){\mathbf{1}}_k$$ If $({\mathcal{L}},f^+,f^-)$ is a framed inner brane, then $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}_{0,1}({\boldsymbol{\tau}}_2, Q^b, X_1 ) \\ &=& \sum_{(d_0 ,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 >0} X^{d_0} D'(d_0 ,k^+,k^-;f^+,f^-)J^{f^+}_{{ {\sigma} }^+,k^+}({\boldsymbol{\tau}}_2,\frac{s_1^+{\mathsf{w}}_1}{d_0}){\mathbf{1}}_{k^+} \\ && + \sum_{(d_0 ,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 <0} X^{d_0} Q^{-d_0 \alpha} D'(d_0 ,k^+,k^-;f^+,f^-)J^{f^-}_{{ {\sigma} }^-,k^-}({\boldsymbol{\tau}}_2,\frac{s_1^+{\mathsf{w}}_1}{d_0}) \cdot \frac{s_1^+}{s_1^-} {\mathbf{1}}_{k^+} \end{aligned}$$ Mirror symmetry for the disk amplitudes ======================================= The equivariant $I$-function and the equivariant mirror theorem {#sec:I} --------------------------------------------------------------- We choose a $p_1,\ldots, p_k \in {\mathbb{L}}^\vee\cap { {\widetilde{C}} }_{\mathcal{X}}$ such that - $\{p_1,\ldots,p_k\}$ is a ${ \mathbb{Q} }$-basis of ${\mathbb{L}}^\vee_{ \mathbb{Q} }$. - $\{\bar{p}_1,\ldots, \bar{p}_{k'}\}$ is a ${ \mathbb{Q} }$-basis of $H^2({\mathcal{X}};{ \mathbb{Q} })$, and $\bar{p}_a=0$ for $k'+1\leq q\leq k$. Let $q'_0,q_1,\ldots, q_k$ be $k$ formal variables, and define $q^\beta = q_1^{\langle p_1,\beta\rangle} \cdots q_k^{\langle p_k,\beta\rangle}$ for $\beta\in {\mathbb{K}}$. The equivariant $I$-function is an $H^*_{{ {\mathrm{orb}} },{\mathcal{T}}} ({\mathcal{X}})$-valued power series defined as follows [@Ir09]: $$\begin{aligned} I(q_0',q,z) &=& e^{\frac{\log q'_0+\sum_{a=1}^{k'} \bar p_a^{\mathcal{T}}\log q_a}{z}} \sum_{\beta\in {\mathbb{K}}_{\mathrm{eff}}} q^\beta \prod_{i=1}^{r'} \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty (\bar D^{\mathcal{T}}_i +(\langle D_i,\beta\rangle -m)z)}{\prod_{m=0}^\infty(\bar D^{\mathcal{T}}_i+(\langle D_i,\beta\rangle -m)z)} \\ && \quad \cdot \prod_{i=r'+1}^r \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty (\langle D_i,\beta\rangle -m)z} {\prod_{m=0}^\infty(\langle D_i,\beta\rangle -m)z} {\mathbf{1}}_{v(\beta)} \end{aligned}$$ where $q^\beta=\prod_{a=1}^k q_a^{\langle p_a,\beta\rangle}$. Note that $\langle p_a,\beta\rangle\geq 0$ for $\beta\in {\mathbb{K}}_{{\mathrm{eff}}}$. The equivariant $I$-function can be rewritten as $$\begin{aligned} I(q_0',q,z) &=& e^{\frac{\log q'_0+\sum_{a=1}^{k'} \bar p_a^{\mathcal{T}}\log q_a}{z}} \sum_{\beta\in {\mathbb{K}}_{\mathrm{eff}}} \frac{q^\beta}{z^{\langle \hat \rho ,\beta\rangle + {\mathrm{age}}(v(\beta))} } \prod_{i=1}^{r'} \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty ( \frac{\bar D^{\mathcal{T}}_i}{z} +\langle D_i,\beta\rangle -m)} {\prod_{m=0}^\infty(\frac{\bar D^{\mathcal{T}}_i}{z}+ \langle D_i,\beta\rangle -m)} \\ && \quad \cdot \prod_{i=r'+1}^r \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty (\langle D_i,\beta\rangle -m)}{\prod_{m=0}^\infty(\langle D_i,\beta\rangle -m)} {\mathbf{1}}_{v(\beta)} \end{aligned}$$ where $\hat \rho = D_1+\cdots +D_r \in { {\widetilde{C}} }_{\mathcal{X}}$. Suppose that ${\mathcal{X}}$ is Calabi-Yau, so that ${\mathrm{age}}(v)$ is an integer for any $v\in {\mathrm{Box}(\mathbf {\Sigma})}$. Then $$H^{\leq 2}_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}}) = H^0_{{ {\mathrm{orb}} }.{\mathcal{T}}}({\mathcal{X}}) \oplus H^2_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}}).$$ Let ${\mathcal{Q}}={ \mathbb{Q} }({\mathsf{u}}_1,{\mathsf{u}}_2,{\mathsf{u}}_3)$ be the fractional field of $H^*_{{\mathcal{T}}}(\mathrm{point};{ \mathbb{Q} }) =H^*_{{ \mathbb{T} }}(\mathrm{point};{ \mathbb{Q} })$. $$\begin{aligned} H^0_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}};{\mathcal{Q}}) &=& \bigoplus_{v\in N_{ {\mathrm{tor}} }} {\mathcal{Q}}{\mathbf{1}}_v , \\ H^2_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}};{\mathcal{Q}}) &=& \bigoplus_{i=1}^{k'}\bigoplus_{v\in N_{ {\mathrm{tor}} }} {\mathcal{Q}}\bar{p}_a {\mathbf{1}}_v \oplus \bigoplus_{\substack{ v\in {\mathrm{Box}(\mathbf {\Sigma})}\\ {\mathrm{age}}(v)=1} } {\mathcal{Q}}{\mathbf{1}}_v \end{aligned}$$ From now on, we further assume that $N_{ {\mathrm{tor}} }=0$, or equivalently, ${\mathcal{X}}$ has trivial generic stabilizer. In this case, we may assume that $\hat{\rho}=0$, and $$\{ v\in {\mathrm{Box}(\mathbf {\Sigma})}: {\mathrm{age}}(v)=1\} = \{ b_{r'+1},\ldots, b_r\}.$$ Then $$H^0_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}};{\mathcal{Q}}) = {\mathcal{Q}}{\mathbf{1}},\quad H^2_{{ {\mathrm{orb}} },{\mathcal{T}}}({\mathcal{X}};{\mathcal{Q}}) = \bigoplus_{a=1}^{k'} {\mathcal{Q}}\bar{p}_a \oplus \bigoplus_{a=r'+1}^r {\mathcal{Q}}{\mathbf{1}}_{b_i}.$$ We choose $\bar{p}_a$ such that $\iota^*\bar{p}_a=0$ $$\begin{aligned} I(q_0',q,z) &=& e^{\frac{\log q'_0+\sum_{a=1}^{k'} \bar p_a^{\mathcal{T}}\log q_a}{z}} \sum_{\beta\in {\mathbb{K}}_{\mathrm{eff}}} \frac{q^\beta}{z^{{\mathrm{age}}(v(\beta))} } \prod_{i=1}^{r'} \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty (\frac{\bar D^{\mathcal{T}}_i}{z} +\langle D_i,\beta\rangle -m)} {\prod_{m=0}^\infty(\frac{\bar D^{\mathcal{T}}_i}{z}+ \langle D_i,\beta\rangle -m)} \\ && \quad \cdot \prod_{i=r'+1}^r \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty (\langle D_i,\beta\rangle -m)}{\prod_{m=0}^\infty(\langle D_i,\beta\rangle -m)} {\mathbf{1}}_{v(\beta)} \end{aligned}$$ For $i=1,\ldots, r$, we will define $\Omega_i \subset {\mathbb{K}}_{\mathrm{eff}}-\{0\}$. and $A_i(q)$ supported on $\Omega_i$. We observe that, if $\beta\in {\mathbb{K}}_{\mathrm{eff}}$ and $v(\beta)=0$ then $\langle D_i,\beta\rangle\in { \mathbb{Z} }$ for $i=1,\ldots, r$. - For $i=1,\ldots,r'$, let $$\Omega_i =\left \{ \beta\in {\mathbb{K}}_{\mathrm{eff}}: v(\beta)=0, \langle D_i,\beta\rangle <0 \textup{ and } \langle D_j, \beta \rangle \geq 0 \textup{ for } j\in \{1,\ldots, r\}-\{i\} \right \}.$$ Then $\Omega_i\subset \{ \beta\in {\mathbb{K}}_{\mathrm{eff}}: v(\beta)=0, \beta\neq 0\}$. We define $$A_i(q):=\sum_{\beta \in \Omega_i} q^\beta \frac{(-1)^{-\langle D_i,\beta\rangle-1}(-\langle D_i, \beta\rangle -1)! }{ \prod_{j\in\{1,\ldots, r\}-\{i\}}\langle D_j,\beta\rangle!}.$$ - For $i=r'+1,\ldots, r$, let $$\Omega_i := \{ \beta\in {\mathbb{K}}_{\mathrm{eff}}: v(\beta)= b_i, \langle D_j, \beta\rangle \notin { \mathbb{Z} }_{<0} \textup{ for }j=1,\ldots, r\},$$ and define $$A_i(q) = \sum_{\beta\in \Omega_i} q^\beta \prod_{j=1}^r \frac{\prod_{m=\lceil \langle D_j, \beta \rangle \rceil}^\infty (\langle D_j,\beta\rangle -m)}{\prod_{m=0}^\infty(\langle D_j,\beta\rangle -m)}.$$ Let ${ {\sigma} }$ be the smallest cone containing $b_i$. Then $$b_i=\sum_{j\in I_{ {\sigma} }'}c_j(b_i)b_j,$$ where $c_j(b_i)\in (0,1)$ and $\sum_{j\in I_{ {\sigma} }'}c_j(b_i)=1$. There exists a unique $D_i^\vee \in {\mathbb{L}}_{ \mathbb{Q} }$ such that $$\langle D_j, D_i^\vee\rangle =\begin{cases} 1, & j=i,\\ -c_j(b_i), & j\in I'_{ {\sigma} },\\ 0, & j\in I_{ {\sigma} }-\{i\}. \end{cases}$$ Then $$A_i(q)=q^{D_i^\vee} + \textup{ higher order terms }$$ $$I(q_0',q,z)= 1 +\frac{1}{z}(\log q_0' {\mathbf{1}}+\sum_{a=1}^{k'} \log(q_a)\bar{p}^{{\mathcal{T}}}_a + \sum_{i=1}^{r'} A_i(q) \bar{D}_i^{{\mathcal{T}}}+ \sum_{i=r'+1}^r A_i(q) {\mathbf{1}}_{b_i}) + o(z^{-1}).$$ For $i=1,\ldots, r'$, $$\bar{{\mathcal{D}}}^{{\mathcal{T}}}_i =\sum_{a=1}^{k'}\bar{l}_i^{(a)} \bar{p}_a^{{\mathcal{T}}} +\lambda_i$$ where $\lambda_i\in H^2(B{ \mathbb{T} };{ \mathbb{Q} })$. Let $S_a(q):=\sum_{i=1}^{r'} \bar{l}_i^{(a)} A_i(q)$. Then $$I(q_0',q,z)= 1 +\frac{1}{z}( (\log q_0' +\sum_{i=1}^{r'}\lambda_i A_i(q)) {\mathbf{1}}+\sum_{a=1}^{k'} (\log(q_a) + S_a(q)) \bar{p}^{{\mathcal{T}}}_a + \sum_{i=r'+1}^r A_i(q) {\mathbf{1}}_{b_i}) + o(z^{-1}).$$ Recall that the equivariant small $J$-function for ${\mathcal{X}}$ is $$J({\boldsymbol{\tau}},z)=1+\sum_{\beta\ge 0, n\ge 0} \sum_{i=1}^N \frac{1}{n!} \langle 1,{\boldsymbol{\tau}}^n, \frac{u_i}{z-\bar \psi}\rangle^{\mathcal{X}}_{0,\beta} u^i,$$ where $\{u_i\}_{i=1}^N$ is a $H^*(B{ \mathbb{T} })$-basis of $H^*_{ \mathbb{T} }({\mathcal{X}};{ \mathbb{Q} })$ and $\{u^i\}_{i=1}^N$ is the dual basis. Mirror theorem [@Gi96; @Gi98; @LLY97; @LLY99] relates the small $J$-function $$J({\boldsymbol{\tau}}_0+{\boldsymbol{\tau}}_2,z ) = J(\tau_0 1 +{\boldsymbol{\tau}}_2,z)= e^{\frac{\tau_0}{z}} J({\boldsymbol{\tau}}_2,z)$$ to the $I$-function up to a mirror transform. A mirror theorem for toric orbifolds will be proved in [@CCIT]. One version of that orbifold mirror theorem is stated in [@Ir09 Conjecture 4.3]. We cite its equivariant version as the following mirror theorem. \[con:mirror\] If the toric orbifold ${\mathcal{X}}$ satisfies Assumption \[semi-proj\], then $$e^{\frac{\tau_0(q_0',q)}{z} } J({\boldsymbol{\tau}}_2(q),z) = I(q_0',q, z),$$ where the equivariant closed mirror map $(q_0',q)\mapsto \tau_0(q_0',q)1 + {\boldsymbol{\tau}}_2(q)$ is determined by the first-order term in the asymptotic expansion of the $I$-function $$I(q_0',q,z)=1+\frac{\tau_0 (q_0',q) 1 +{\boldsymbol{\tau}}_2(q)}{z}+o(z^{-1}).$$ More explicitly, the equivariant closed mirror map is given by $$\begin{aligned} \tau_0 &=& \log(q_0') +\sum_{i=1}^{r'} \lambda_i A_i(q),\\ \tau_a &=& \begin{cases} \log(q_a)+ S_a(q), & 1\leq a\leq k',\\ A_{a-k'+r'}(q), & k'+1\leq a\leq k. \end{cases}.\end{aligned}$$ ${\mathcal{X}}={\mathcal{X}}_{i,j,k}$, $k'=0$. $$\bar{D}^{\mathcal{T}}_i =\begin{cases}{\mathsf{w}}_i, & i\in \{1,2,3\},\\ 0, & i>3. \end{cases}$$ 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$, $k=1$, $p_1=-3$, ${\mathbb{K}}_{\mathrm{eff}}=\frac{1}{3}{ \mathbb{Z} }_\leq 0$. $$A_1(q)=A_2(q)=A_3(q)=0, \quad A_4(q) =\sum_{\alpha=0}^\infty \frac{(-1)^{3\alpha} q_1^{3\alpha+1}}{(3\alpha+1)!} \Big(\frac{\Gamma(\alpha+\frac{1}{3})}{\Gamma(\frac{1}{3})}\Big)^3.$$ $$\begin{aligned} I(q'_0, q_1,z)&=&e^{\frac{\log q'_0}{z}}\bigl( \sum_{d\in { \mathbb{Z} }_{\geq 0}} \frac{q^d}{d! z^d} \prod_{i=1}^3\prod_{m=\lceil -\frac{d}{3}\rceil}^{-1}({\mathsf{w}}_i-(\frac{d}{3}+m)z) {\mathbf{1}}_{\omega^d} \Bigr) \\ &=& 1+\frac{1}{z}\Bigl( (\log q'_0) {\mathbf{1}}+ A_4(q){\mathbf{1}}_\omega\Bigr) +o(z^{-1})\end{aligned}$$ Let $r=|q_1|$. The closed mirror map ${\boldsymbol{\tau}}=\tau_0 {\mathbf{1}}+ \tau_1 {\mathbf{1}}_\omega$ is given by $$\begin{aligned} \tau_0 &=& \log q'_0,\\ \tau_1 &=& A_4(q) = q_1 + O(r^4)\end{aligned}$$ 2. ${\mathcal{X}}={\mathcal{X}}_{1,2,0}$, $k=2$, $p_1=(-2,1)$, $p_2=(1,-2)$, ${\mathbb{K}}_{\mathrm{eff}}={ \mathbb{Z} }_{\geq 0} (-\frac{2}{3}, -\frac{1}{3}) \oplus { \mathbb{Z} }_{\geq 0} (-\frac{1}{3}, -\frac{2}{3})$, $$\begin{aligned} && A_1(q)=A_2(q)=A_3(q)=0\\ && A_4(q)=\sum_{ \substack{d_1,d_2\geq 0\\ 2d_1+d_2\in 2+3{ \mathbb{Z} }} } q_1^{d_1} q_2^{d_2}\frac{(-1)^{d_1+d_2-1}}{d_1! d_2!} \frac{\Gamma(\frac{2d_1+d_2}{3}) \Gamma(\frac{d_1+2d_2}{3})}{\Gamma(\frac{2}{3})\Gamma(\frac{1}{3})} \\ && A_5(q) =\sum_{ \substack{d_1,d_2\geq 0\\ 2d_1+d_2 =1+3{ \mathbb{Z} }} } q_1^{d_1} q_2^{d_2}\frac{(-1)^{d_1+d_2-1}}{d_1! d_2!} \frac{\Gamma(\frac{2d_1+d_2}{3}) \Gamma(\frac{d_1+2d_2}{3})}{\Gamma(\frac{1}{3})\Gamma(\frac{2}{3})}\end{aligned}$$ $$\begin{aligned} && I(q'_0,q_1,q_2,z)\\ &=& e^{\frac{\log q'_0}{z}} \sum_{\substack{d_1,d_2\in { \mathbb{Z} }_{\geq 0} } } \frac{q_1^{d_1}q_2^{d_2}}{d_1! d_2! z^{d_1+d_2} } \prod_{m=\lceil -\frac{2d_1+d_2}{3}\rceil}^{-1}({\mathsf{w}}_3 - (\frac{2d_1+d_2}{3}+m)z)\\ &&\cdot \prod_{m=\lceil -\frac{d_1+2d_2}{3}\rceil}^{-1}({\mathsf{w}}_2 - (\frac{d_1+2 d_2}{3}+m)z) {\mathbf{1}}_{\omega^{\frac{d_1+2d_2}{3}} } \\ &=& 1+\frac{1}{z}\Bigg( \log q'_0 {\mathbf{1}}+ A_4(q) {\mathbf{1}}_\omega +A_5(q) {\mathbf{1}}_{\omega^2} \Bigg) + o(z^{-1})\end{aligned}$$ Let $r=\sqrt{|q_1|^2+ |q_2|^2}$. The closed mirror map ${\boldsymbol{\tau}}=\tau_0 {\mathbf{1}}+ \tau_1 {\mathbf{1}}_\omega+ \tau_2 {\mathbf{1}}_{\omega^2}$ is given by $$\begin{aligned} \tau_0 &=& \log q'_0,\\ \tau_1 &=& A_4(q) = q_1 + O(r^2),\\ \tau_2 &=& A_5(q)= q_2 + O(r^2).\end{aligned}$$ The pullback of the disk potential under the mirror map ------------------------------------------------------- By Proposition \[pro:F-J\], if $({\mathcal{L}},f)$ is a framed outer brane, then $$F^{{\mathcal{X}},({\mathcal{L}},f)}_{0,1}({\boldsymbol{\tau}}_2,Q^b,X_1) = \sum_{(d_0 ,k)\in H_{\tau,{ {\sigma} }}} X^{d_0} D'(d_0,k;f)J^f_{{ {\sigma} },k}({\boldsymbol{\tau}}_2,\frac{s_1{\mathsf{w}}_1}{d_0}){\mathbf{1}}_k.$$ Let $F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X)$ be the pullback of $F^{{\mathcal{X}},({\mathcal{L}},f)}_{0,1}({\boldsymbol{\tau}}_2,Q^b, X_1)$ under the closed mirror map. By Proposition \[pro:F-J\], if $({\mathcal{L}},f^+,f^-)$ is a framed inner brane, then $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}_{0,1}({\boldsymbol{\tau}}_2 ,Q^b, X_1) \\ &=& \sum_{(d_0,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 >0} X^{d_0} D'(d_0,k^+,k^-;f^+,f^-)J^{f^+}_{{ {\sigma} }^+,k^+}({\boldsymbol{\tau}}_2,\frac{s_1^+{\mathsf{w}}_1}{d_0}){\mathbf{1}}_{k^+} \\ && + \sum_{(d_0,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 <0} X^{d_0} Q^{-d_0 \alpha} D'(d,k^+,k^-;f^+,f^-)J^{f^-}_{{ {\sigma} }^-,k^-}({\boldsymbol{\tau}}_2,\frac{s_1^+{\mathsf{w}}_1}{d_0 })\cdot \frac{s_1^+}{s_1^-} {\mathbf{1}}_{k^+}. \end{aligned}$$ Let $F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}(q,X)$ be the pullback of $F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}_{0,1}({\boldsymbol{\tau}}_2,Q^b, X_1)$ under the closed mirror map. Given ${ {\sigma} }\in {\Sigma}(3)$, $k\in G_{ {\sigma} }$, and $f\in { \mathbb{Z} }$, define $I_{{ {\sigma} },k}^f(q,z)$ by $$\iota_{ {\sigma} }^* I(q,z)|_{{\mathsf{w}}_2=f{\mathsf{w}}_1} =\sum_{k\in G_{ {\sigma} }}I_{{ {\sigma} },k}^f(q,z){\mathbf{1}}_k.$$ Since a toric Calabi-Yau orbifold satisfies the weak Fano condition, by the equivariant mirror theorem (Conjecture \[con:mirror\]), we may write $F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X)$ in terms of $I_{{ {\sigma} },k}^f(q,z)$, and write $F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}(q,X)$ in terms of $I_{{ {\sigma} }_+,k^+}^{f^+}(q,z)$ and $I_{{ {\sigma} }_-,k^-}^{f^-}(q,z)$. If $({\mathcal{L}},f)$ is a framed outer brane, then $$F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X) = \sum_{(d_0,k)\in H_{\tau,{ {\sigma} }}} X^{d_0} D'(d_0,k;f) e^{\frac{-d_0 \tau_0(q) }{s_1{\mathsf{w}}_1}} I^f_{{ {\sigma} },k}(q,\frac{s_1{\mathsf{w}}_1}{d_0}){\mathbf{1}}_k$$ If $({\mathcal{L}},f^+,f^-)$ is a framed inner brane, then $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f^+,f^-)}(q,X) \\ &=& \sum_{(d_0,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 >0} X^{d_0} D'(d_0,k^+,k^-;f^+,f^-) e^{\frac{-d_0 \tau_0(q) }{s_1^+{\mathsf{w}}_1}}I^{f^+}_{{ {\sigma} }^+,k^+}(q,\frac{s_1^+{\mathsf{w}}_1}{d_0}){\mathbf{1}}_{k^+} \\ && + \sum_{(d_0,k^+, k^-)\in H_{\tau,{ {\sigma} }_+,{ {\sigma} }_-}, d_0 <0} X^{d_0} Q^{-d_0 \alpha} D'(d_0 ,k^+,k^-;f^+,f^-) e^{\frac{-d_0 \tau_0(q)}{s_1^+{\mathsf{w}}_1}} I^{f^-}_{{ {\sigma} }^-,k^-}(q,\frac{s_1^+{\mathsf{w}}_1}{d_0}) \cdot \frac{s_1^+}{s_1^-}{\mathbf{1}}_{k^+} \end{aligned}$$ ### A framed outer brane $({\mathcal{L}},f)$ Let $({\mathcal{L}},f)$ be a framed outer brane. Let ${\mathfrak{l}}_\tau$ be the unique 1-dimensional orbit closure intersecting ${\mathcal{L}}$, and let ${\mathfrak{p}}_{ {\sigma} }$ be the unique torus fixed (stacky) point in ${\mathfrak{l}}_\tau$. Recall that $$\begin{aligned} && I'_{ {\sigma} }= \{ i\in \{1,\ldots,r'\}: \rho_i\subset { {\sigma} }\} =\{i_1,i_2, i_3 \},\quad I_{ {\sigma} }= \{1,\ldots, r\} \setminus I'_{ {\sigma} },\\ && I'_\tau = \{i\in \{1,\ldots, r'\}:\rho_i\subset \tau\}= \{i_2, i_3\},\quad I_\tau =\{1,\ldots, r\} \setminus I'_\tau,\\ && {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }} = \{ \beta\in {\mathbb{L}}_{ \mathbb{Q} }: \langle D_i, \beta\rangle \in { \mathbb{Z} }_{\geq 0} \textup{ for } i\in I_{ {\sigma} }\}\\ && H_{\tau,{ {\sigma} }} = \{ (d_0,k)\in { \mathbb{Z} }\times G_{ {\sigma} }: \exp(2\pi\sqrt{-1}\frac{d_0}{s_1}) = \chi_{i_1}(k)\}.\end{aligned}$$ Let ${\mathsf{b}}_{{ {\sigma} },i}=\iota_{ {\sigma} }^* \bar{D}^{\mathcal{T}}_i \in H^2_{{\mathcal{T}}}({\mathfrak{p}}_{ {\sigma} };{ \mathbb{Q} })=H^2(B{ \mathbb{T} };{ \mathbb{Q} })$ for $1\leq i\leq r'$, and let ${\mathsf{b}}_i=0$ for $r'+1\leq i\leq r$. For $\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}$, define $$I({ {\sigma} },\beta): = \prod_{i=1}^r \frac{\prod_{m=\lceil \langle D_i, \beta \rangle \rceil}^\infty ({\mathsf{b}}_{{ {\sigma} },i} +(\langle D_i,\beta\rangle -m)\frac{s_1{\mathsf{w}}_1}{d_0})} {\prod_{m=0}^\infty({\mathsf{b}}_{{ {\sigma} },i} +(\langle D_i,\beta\rangle -m)\frac{s_1{\mathsf{w}}_1}{d_0})}$$ Recall that $\iota_{ {\sigma} }^*\bar{p}_a^{{\mathcal{T}}}=0$, so $$\iota_{ {\sigma} }^* I(q,z)|_{z=\frac{s_1{\mathsf{w}}_1}{d_0}} = \sum_{\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }} } e^{\frac{d_0}{s_1{\mathsf{w}}_1}\log q_0'} q^\beta I^f({ {\sigma} },\beta) {\mathbf{1}}_{v(\beta)}$$ With the above notation, we can rewrite $F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X)$ as $$F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X)=\sum_{(d_0,k)\in H_{\tau,{ {\sigma} }}}\sum_{\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}, v(\beta)=k} x^{d_0} q^\beta D'(d_0,k,f)I^f({ {\sigma} },\beta) {\mathbf{1}}_k$$ where $I^f({ {\sigma} },\beta)=I({ {\sigma} },\beta)|_{{\mathsf{w}}_2=f{\mathsf{w}}_1, {\mathsf{w}}_3=-(f+1){\mathsf{w}}_1}$, and $$x=X \exp\big(\frac{\log q_0'-\tau_0(q)}{s_1 {\mathsf{w}}_1}\big)$$ is the B-brane moduli parameter. Recall that $$\tau_0 + \sum_{a=1}^{k'} \tau_a \bar{p}_a^{{\mathcal{T}}} + \sum_{a=k'+1}^k \tau_a {\mathbf{1}}_{b_{a-k'+r'}} = \log q_0' + \sum_{a=1}^{k'} \log q_a \bar{p}_a^{{\mathcal{T}}} +\sum_{i=1}^{r'} A_i(q)\bar{{\mathcal{D}}}^{{\mathcal{T}}}_i +\sum_{i=r'+1}^r A_i(q){\mathbf{1}}_{b_i}.$$ We pull back the above identity under $\iota_{ {\sigma} }^*$, and recall that $$\iota_{ {\sigma} }^*\bar{p}_a^{{\mathcal{T}}}=0,\quad \iota_{ {\sigma} }^*{\mathcal{D}}^{{\mathcal{T}}'}_i \Bigr|_{{\mathsf{w}}_2=f{\mathsf{w}}_1, {\mathsf{w}}_3 = (-f-1){\mathsf{w}}_1} = l^{(0)}_i {\mathsf{w}}_1,$$ we get $$\tau_0(q_0',q)=\log q'_0+ \sum_{i=1}^{r'} l^{(0)}_i A_i(q) {\mathsf{w}}_1$$ So the open mirror map is given by $$\log X = \log x + \frac{1}{s_1}\sum_{i=1}^{r'} l^{(0)}_i A_i(q).$$ We define $\{ W^{{\mathcal{X}},({\mathcal{L}},f)}_v (q,x)\}_{v \in {\mathrm{Box}({ {\sigma} })}}$ by $$F^{{\mathcal{X}},({\mathcal{L}},f)}(q,X) = \sum_{v \in {\mathrm{Box}({ {\sigma} })}} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(q,x) {\mathbf{1}}_v.$$ Then $$\label{eqn:W} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(q,x)= \sum_{\substack{ d_0 \in { \mathbb{Z} }_{>0} \\ \{ \frac{d_0}{s_1} \} = c_{i_1}(v)} } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}\\ v(\beta)=v} } x^{d_0} q^\beta D'(d_0 ,v,f)I^f({ {\sigma} },\beta)$$ Following [@LM01; @Ma01], we define [*extended charge vectors*]{} $$\{\tilde l^{(a)}_i\}=\begin{pmatrix} & & & \{ l^{(a)}_i\} & & & & 0 & 0\\ \dots & 1 & \dots & f & \dots & -f-1 & \dots & 1 & -1 \end{pmatrix},$$ where $a=0,\dots, k$, and the last row is an additional vector $\tilde l^{(0)}$ containing the phase and the framing of the A-brane ${\mathcal{L}}$. Given ${{\widetilde{\beta}}}=(d_0,\beta)\in { \mathbb{Z} }\times {\mathbb{K}}_{ {\sigma} }$, define the extended or open sector pairing to be $$\langle D_i, {{\widetilde{\beta}}}\rangle=\frac{d_0}{s_1} \tilde l_i^{(0)}+\langle D_i, \beta \rangle.$$ Notice in this particular notation we allow $i=r+1,r+2$ since the corresponding $\tilde l_i^{(a)}$ exist, although $D_i$ are not actual divisors. Recall that $\{D_i: i\in I_{ {\sigma} }\}$ is a ${ \mathbb{Q} }$-basis of ${\mathbb{L}}^\vee_{ \mathbb{Q} }\cong { \mathbb{Q} }^k$ and a ${ \mathbb{Z} }$-basis of ${\mathbb{K}}_{ {\sigma} }\cong { \mathbb{Z} }^k$. Let $\{ p_a\}_{a=1,\ldots,k} = \{D_i\}_{i\in I_{ {\sigma} }}$, and let $\{ e_a\}_{a=1,\ldots,k}$ be the dual ${ \mathbb{Q} }$-basis of ${\mathbb{L}}_{ \mathbb{Q} }$, so that $\langle p_a,e_b\rangle =\delta_{ab}$. Then $\{e_a\}_{a=1,\ldots,k}$ is a ${ \mathbb{Z} }$-basis of ${\mathbb{K}}_{ {\sigma} }\cong { \mathbb{Z} }^k$, and $${\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }} =\sum_{a=0}^k{ \mathbb{Z} }_{\geq 0} e_a.$$ Given any $(d_0,\beta)\in { \mathbb{Z} }\times {\mathbb{K}}_{ {\sigma} }$, define $$q^{{{\widetilde{\beta}}}} = x^{d_0} q^\beta = x^{d_0} \prod_{a=1}^k q_a^{\langle p_a,\beta\rangle}.$$ With the above notation, we have: \[thm:amplitude-term-outer\] $$\begin{aligned} \label{eqn:amplitude-term-outer} W_v^{{\mathcal{X}},({\mathcal{L}},f)}(q,x)=\sum_{\substack{{{\widetilde{\beta}}}= (d_0,\beta) \in {\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})\\ v(\beta)=v } } q^{{\widetilde{\beta}}}A^{{\mathcal{X}},({\mathcal{L}},f)}_{{{\widetilde{\beta}}}}. \end{aligned}$$ where $$\begin{aligned} {\mathbb{K}}_{{\mathrm{eff}}}({\mathcal{X}},{\mathcal{L}}) &=&\{{{\widetilde{\beta}}}=(d_0,\beta)\in { \mathbb{Z} }_{>0} \times {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}: \langle D_{i_1},{{\widetilde{\beta}}}\rangle\in { \mathbb{Z} }_{\geq 0} \} \\ A^{{\mathcal{X}},({\mathcal{L}},f)}_{{{\widetilde{\beta}}}=(d_0,\beta)} &=& \frac{ -(-1)^{\frac{d_0}{s_1}(-f-1)- \{ c_{i_2}(v) - f c_{i_1}(v)\} + \langle D_{i_3}, \beta\rangle} }{ \frac{d_0}{s_1} \prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle!} \cdot \frac{\Gamma(-\langle D_{i_3},{{\widetilde{\beta}}}\rangle) }{\Gamma(\langle D_{i_2},{{\widetilde{\beta}}}\rangle+1)} \end{aligned}$$ To simplify the calculations, we introduce the symbol for any two numbers $a$ and $b$ with $a-b\in { \mathbb{Z} }$. $$\begin{bmatrix} a\\b \end{bmatrix}=\frac{\prod_{i=0}^\infty (a-i) }{\prod_{i=0}^\infty (b-i)}.$$ It has the following properties $${\begin{bmatrix}a\\#2\end{bmatrix}}={\begin{bmatrix}b\\#2\end{bmatrix}}^{-1},\quad {\begin{bmatrix}a\\#2\end{bmatrix}}\cdot{\begin{bmatrix}b\\#2\end{bmatrix}}={\begin{bmatrix}a\\#2\end{bmatrix}},\quad {\begin{bmatrix}a\\#2\end{bmatrix}}=(-1)^{a-b}{\begin{bmatrix}-b-1\\#2\end{bmatrix}}.$$ Let ${{\widetilde{\beta}}}=(d_0,\beta)$, and let ${\epsilon}_j =c_{i_j}(v)$ for $j=1,2,3$. By , $$W^{{\mathcal{X}},({\mathcal{L}},f)}_v(q,x)= \sum_{\substack{ d_0 \in { \mathbb{Z} }_{>0} \\ \{ \frac{d_0}{s_1} \} = {\epsilon}_1 } } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}\\ v(\beta)=v} } x^{d_0} q^\beta D'(d_0,v;f)I^f({ {\sigma} },\beta),$$ where $$\begin{aligned} D'(d_0 ,v;f) &=& -(-1)^{\frac{d_0}{s}(-f-1)-{\epsilon}_3 - \{ {\epsilon}_2-f{\epsilon}_1\} } \frac{s_1}{d_0} \left ( \frac{s_1{\mathsf{w}}_1}{d_0} \right)^{{\epsilon}_1+{\epsilon}_2+{\epsilon}_3} \frac{\prod_{a=1}^{\lfloor \frac{d_0}{s_1} \rfloor + {\epsilon}_1 + {\epsilon}_2 + {\epsilon}_3 -1} (\frac{d_0}{s_1}f + (a-{\epsilon}_2))}{\lfloor \frac{d_0}{s_1} \rfloor!} \\ &=& (-1)^{\frac{d_0}{s}(-f-1)-{\epsilon}_3- \{ {\epsilon}_2-f{\epsilon}_1\} } \frac{s_1}{d_0} \left ( \frac{s_1{\mathsf{w}}_1}{d_0} \right)^{{\epsilon}_1+{\epsilon}_2+{\epsilon}_3}{\begin{bmatrix}0\\#2\end{bmatrix}} {\begin{bmatrix}\frac{(f+1)d_0}{s_1}+{\epsilon}_3-1\\#2\end{bmatrix}}\end{aligned}$$ Given any $\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}$, we have $\lceil \langle D_{i_j},\beta\rangle \rceil -{\epsilon}_j =\langle D_{i_j},\beta\rangle$ for $j=1, 2,3$, and $\lceil\langle D_i, \beta\rangle \rceil =\langle D_i,\beta\rangle$ for $i\in I_{ {\sigma} }$. By straightforward calculation, $$\begin{aligned} && I^f({ {\sigma} },\beta) \\ & =& (\frac{s_1{\mathsf{w}}_1}{d})^{-{\epsilon}_1 -{\epsilon}_2 -{\epsilon}_3 } \prod_{i\in I_{ {\sigma} }} \frac{\prod_{a=-\infty}^0 a}{\prod_{a=-\infty}^{ \lceil \langle D_i,\beta \rangle \rceil }a } \cdot \frac{\prod_{a=-\infty}^0(\frac{d_0}{s_1}-{\epsilon}_1 +a)} {\prod_{a=-\infty}^{\lceil \langle D_{i_1},\beta \rangle \rceil}(\frac{d_0}{s_1}-{\epsilon}_1 +a)} \\ && \cdot \frac{\prod_{a=-\infty}^0 (f \frac{d_0}{s_1}-{\epsilon}_2 +a)}{\prod_{a=-\infty}^{\lceil \langle D_{i_2},\beta \rangle\rceil}(f \frac{d_0}{s_1}-{\epsilon}_2+a)} \cdot \frac{\prod_{a=-\infty}^0((-1-f)\frac{d_0}{s_1}-{\epsilon}_3 +a)}{\prod_{a=-\infty}^{\lceil \langle D_{i_3},\beta \rangle\rceil }((-1-f)\frac{d_0}{s_1}-{\epsilon}_3 +a)} \\ &=& (\frac{s_1{\mathsf{w}}_1}{d_0})^{-{\epsilon}_1 -{\epsilon}_2 -{\epsilon}_3 } {\begin{bmatrix}\frac{d_0}{s_1} -{\epsilon}_1\\#2\end{bmatrix}} {\begin{bmatrix}\frac{fd_0}{s_1}-{\epsilon}_2\\#2\end{bmatrix}} {\begin{bmatrix} \frac{(-f-1)d_0}{s_1}-{\epsilon}_3\\#2\end{bmatrix}} \prod_{i\in I_{ {\sigma} }} {\begin{bmatrix}0\\#2\end{bmatrix}}\end{aligned}$$ $$W^{{\mathcal{X}},({\mathcal{L}},f)}_v (q,x)=\sum_{\substack{d_0\in { \mathbb{Z} }_{>0}\\ \{ \frac{d_0}{s_1} \} ={\epsilon}_1 } } \sum_{\substack{\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }}\\ v(\beta)=v}} x^{\frac{d_0}{s_1}} q^\beta A^{{\mathcal{X}},({\mathcal{L}},f)}_{(d,\beta)}$$ where $$A^{{\mathcal{X}}, ({\mathcal{L}},f)}_{{{\widetilde{\beta}}}} = -(-1)^{ \frac{d_0}{s_1}(-f-1) - \{ {\epsilon}_2-f{\epsilon}_1\} + \langle D_{i_3},\beta\rangle} \frac{s_1}{d_0} {\begin{bmatrix}-\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1 \\#2\end{bmatrix}} \prod_{i\in I_\tau} {\begin{bmatrix}0\\#2\end{bmatrix}}.$$ Note that $A^{{\mathcal{X}}, ({\mathcal{L}},f)}_{{{\widetilde{\beta}}}} =0$ if $\langle D_i,{{\widetilde{\beta}}}\rangle <0$ for some $i\in I_\tau$, so $$W^{{\mathcal{X}},({\mathcal{L}},f)}_v (q,x)=\sum_{\substack{{{\widetilde{\beta}}}=(d_0,\beta)\in {\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})\\ v(\beta)=v} } q^{{\widetilde{\beta}}}A^{{\mathcal{X}},({\mathcal{L}},f)}_{{{\widetilde{\beta}}}},$$ where $$A^{{\mathcal{X}}, ({\mathcal{L}},f)}_{{{\widetilde{\beta}}}} = \frac{ -(-1)^{ \frac{d_0}{s}(-f-1) - \{ c_{i_2}(v)-fc_{i_1}(v) \} + \langle D_{i_3}, \beta\rangle } }{ \frac{d_0}{s_1} \prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle!} \cdot \frac{\Gamma(-\langle D_{i_3},{{\widetilde{\beta}}}\rangle) }{\Gamma(\langle D_{i_2},{{\widetilde{\beta}}}\rangle+1)}$$ ${\mathcal{X}}={\mathcal{X}}_{i,j,k}$, the open mirror map is given by $x=X$. \[exp:prepotential-A\] 1. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}$, $(i_1, i_2, i_3)=(1,2,3)$, $s_1=3$. The extended charge vectors are $$\left( \begin{array} {cccccc} 1 & 1 & 1 & -3 & 0 & 0 \\ 1 & f & -f-1 & 0 & 1 & -1 \end{array} \right)$$ $$F^{{\mathcal{X}},({\mathcal{L}},f)}(q_1,X) = \sum_{ \substack{d_0 \in { \mathbb{Z} }_{>0} \\ d_1 \in { \mathbb{Z} }_{\geq 0} \\ d_0-d_1\in 3{ \mathbb{Z} }} } x^{d_0} q_1^{d_1} \frac{-(-1)^{ \lfloor \frac{d_1-f d_0 }{3}\rfloor -\frac{d_0+2d_1}{3}} }{\frac{d_0}{3}\cdot d_1 !(\frac{d_0-d_1}{3})! } \cdot \frac{ \Gamma(\frac{(f+1)d_0+d_1}{3})}{ \Gamma(\frac{fd_0-d_1}{3}+1) } {\mathbf{1}}_{\omega^{d_1} }$$ 2. ${\mathcal{X}}={\mathcal{X}}_{0,1,2}$, $(i_1, i_2, i_3) =(1, 2,3)$, $s_1=1$. The extended charge vectors are $$\left( \begin{array} {ccccccc} 0 & 0 & 1 & -2 & 1 & 0 & 0 \\ 0 & 1 & 0 & 1 & -2 & 0 & 0 \\ 1 & f & -f-1 & 0 & 0 & 1 & -1 \end{array} \right)$$ $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f)}(q_1,q_2,X) \\ &=& \sum_{ \substack{ d_0 \in{ \mathbb{Z} }_{> 0} \\ d_1, d_2\in { \mathbb{Z} }_{\geq 0}} } x^{d_0} q_1^{d_1}q_2^{d_2} \frac{-(-1)^{(-f-1)d_0 + \lfloor \frac{-2d_1-d_2}{3} \rfloor} }{d_0 \cdot d_0!d_1!d_2!} \cdot \frac{\Gamma( (f+1)d_0 +\frac{2d_1+d_2}{3}) }{\Gamma(fd_0-\frac{d_1+2 d_2}{3}+1) } {\mathbf{1}}_{\omega^{d_1-d_2} }\end{aligned}$$ 3. ${\mathcal{X}}={\mathcal{X}}_{0,1,2}$, $(i_1, i_2, i_3)=(2,3,1)$. $s_1=3$. The extended charge vectors are $$\left( \begin{array} {ccccccc} 0 & 0 & 1 & -2 & 1 & 0 & 0 \\ 0 & 1 & 0 & 1 & -2 & 0 & 0 \\ -f-1 & 1 & f & 0 & 0 & 1 & -1 \end{array} \right)$$ $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f)}(q_1,q_2,X) \\ &=& \sum_{ \substack{ d_0 \in{ \mathbb{Z} }_{> 0} \\ d_1, d_2\in { \mathbb{Z} }_{\geq 0} \\ d_0 - d_1 - 2d_2\in 3{ \mathbb{Z} }} } x^{d_0} q_1^{d_1}q_2^{d_2} \frac{-(-1)^{ \lfloor \frac{(-f-1)d_0}{3}\rfloor} }{\frac{d_0}{3} \cdot (\frac{d_0-d_1-2d_2}{3})! \cdot d_1!d_2!} \cdot \frac{\Gamma(\frac{(f+1)d_0}{3})}{ \Gamma(\frac{fd_0-2d_1-d_2}{3}+1)} {\mathbf{1}}_{\omega^{d_1-d_2 } }\end{aligned}$$ 4. ${\mathcal{X}}={\mathcal{X}}_{0,1,2}$, $(i_1, i_2, i_3)=(3,1,2)$. $s_1=3$. The extended charge vectors are $$\left( \begin{array} {ccccccc} 0 & 0 & 1 & -2 & 1 & 0 & 0 \\ 0 & 1 & 0 & 1 & -2 & 0 & 0 \\ f & -f-1 & 1 & 0 & 0 & 1 & -1 \end{array} \right)$$ $$\begin{aligned} && F^{{\mathcal{X}},({\mathcal{L}},f)}(q_1,q_2,X) \\ &=& \sum_{ \substack{ d_0 \in{ \mathbb{Z} }_{> 0} \\ d_1, d_2\in { \mathbb{Z} }_{\geq 0} \\ d_0 - 2d_1 - d_2\in 3{ \mathbb{Z} }} } x^{\frac{d_0}{3}} q_1^{d_1}q_2^{d_2} \frac{-(-1)^{ \frac{-d_0-d_1-2d_2}{3} + \lfloor \frac{-fd_0}{3}\rfloor} }{\frac{d_0}{3} \cdot (\frac{d_0-2d_1-d_2}{3})! \cdot d_1!d_2!} \cdot \frac{\Gamma(\frac{(f+1)d_0+ d_1+2d_2}{3})}{ \Gamma(\frac{fd_0}{3}+1)} {\mathbf{1}}_{\omega^{d_1-d_2}}\end{aligned}$$ ### A framed inner brane \[thm:amplitude-term-inner\] $$\begin{aligned} \label{eqn:amplitude-term-inner} W_v^{{\mathcal{X}},({\mathcal{L}},f)}(q,x)=\sum_{\substack{{{\widetilde{\beta}}}= (d_0,\beta) \in {\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})\\ v(\beta)=v } } q^{{\widetilde{\beta}}}A^{{\mathcal{X}},({\mathcal{L}},f)}_{{{\widetilde{\beta}}}}. \end{aligned}$$ where $$\begin{aligned} {\mathbb{K}}_{{\mathrm{eff}}}({\mathcal{X}},{\mathcal{L}}) &=&\{{{\widetilde{\beta}}}=(d_0,\beta)\in ({ \mathbb{Z} }-\{0\}) \times {\mathbb{K}}_{{\mathrm{eff}}}: \langle D_i,{{\widetilde{\beta}}}\rangle\in { \mathbb{Z} }_{\geq 0} \textup{ for }i\in I_\tau \} \\ A^{{\mathcal{X}},({\mathcal{L}},f)}_{{{\widetilde{\beta}}}=(d,\beta)} &=& -(-1)^{\frac{d_0}{s_1^+}(-f_+-1)-\{ c_{i_2}(v)-f c_{i_1}(v)\}+ \langle D_{i_3},\beta\rangle }\frac{s_1}{d_0} {\begin{bmatrix}-\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1 \\#2\end{bmatrix}} \prod_{i\in I_\tau}{\begin{bmatrix}0\\#2\end{bmatrix}} .\end{aligned}$$ $$W_v^{{\mathcal{X}},({\mathcal{L}},f)}(q,x) = I_+ + I_-,$$ where $$\begin{aligned} I_+ &=& \sum_{\substack{ d_0\in { \mathbb{Z} }_{>0}\\ \langle \frac{d_0}{s^+_1}\rangle ={\epsilon}_1} } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }_+}\\ v(\beta)=v} } -x^{d_0} q^\beta (-1)^{\frac{d_0}{s_1^+}(-f^+-1)-\{ {\epsilon}_1-f{\epsilon}_2\} + \langle D_{i_3},\beta\rangle} \frac{s_1^+}{d_0} {\begin{bmatrix} \frac{(f^+ +1)d_0}{s_1^+} -\langle D_{i_3}, \beta\rangle -1\\#2\end{bmatrix}}\\ &&\cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot \prod_{I-\{i_1,i_2, i_3, i_4\} } {\begin{bmatrix}0\\#2\end{bmatrix}} \\ I_-&= & \sum_{\substack{ d_0\in { \mathbb{Z} }_{<0}\\ \langle \frac{-d_0}{s^+_1}\rangle ={\epsilon}_1} } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }_-}\\ v(\beta-d_0 \alpha)=v} } -x^{d_0} q^{\beta-d_0\alpha} (-1)^{ \frac{d_0}{s_1^-}(-f^-1) -\{ {\epsilon}_1-f{\epsilon}_2\} - \langle D_{i_2},\beta\rangle } \frac{s_1^-}{d_0} {\begin{bmatrix} \frac{-(f^- +1)d_0}{s_1^-} -\langle D_{i_2}, \beta\rangle -1\\#2\end{bmatrix}}\\ &&\cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot \prod_{I-\{i_1,i_2, i_3, i_4\} } {\begin{bmatrix}0\\#2\end{bmatrix}} \frac{s_1^+}{s_1^-}\end{aligned}$$ We have $$\langle D_{i_1}, \alpha \rangle = \frac{1}{s_1^+},\quad \langle D_{i_2}, \alpha \rangle =\frac{f^+}{s_1^+}-\frac{f^-+1}{s_1^-},\quad \langle D_{i_3}, \alpha \rangle = \frac{f^-}{s_1^-}-\frac{f^+ +1}{s_1^+},\quad \langle D_{i_4}, \alpha \rangle = \frac{1}{s_1^-},$$ and $\langle D_i,\alpha\rangle =0$ for $i\in I-\{i_0, i_1, i_3, i_4\}$. So for $\beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }_-}$, $$\begin{aligned} \langle D_{i_1},\beta\rangle &=& \langle D_{i_1}, \beta-d_0\alpha\rangle + \frac{d_0}{s_1^+}\\ -\frac{(f^-+1)d_0}{s_1^-} -\langle D_{i_2}, \beta \rangle &=& -\frac{f^+d_0}{s_1^+} -\langle D_{i_2}, \beta-d_0\alpha\rangle \\ -\frac{f^-d_0}{s_1^-} +\langle D_{i_3}, \beta \rangle &=& -\frac{(f^+ +1)d_0}{s_1^+} +\langle D_{i_3}, \beta-d_0\alpha\rangle \\ -\frac{d_0}{s_1^-} +\langle D_{i_4},\beta\rangle &=& \langle D_{i_4}, \beta-d_0\alpha\rangle\end{aligned}$$ So $$\begin{aligned} I_- &= & \sum_{\substack{ d_0\in { \mathbb{Z} }_{<0}\\ \langle \frac{-d_0}{s^+_1}\rangle ={\epsilon}_1} } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }_-}\\ v(\beta-d_0\alpha)=v} } - x^{d_0} q^{\beta-{d_0}\alpha} (-1)^{ \frac{d_0}{s_1^-}(-f^- - 1) -\{ {\epsilon}_2-f{\epsilon}_1\} - \langle D_{i_2},\beta\rangle } \frac{s_1^+}{d_0} {\begin{bmatrix} \frac{-f^+d_0}{s_1^+} -\langle D_{i_2}, \beta-d_0\alpha\rangle -1\\#2\end{bmatrix}}\\ &&\cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot \prod_{I-\{i_1,i_2, i_3, i_4\} } {\begin{bmatrix}0\\#2\end{bmatrix}} \\ &=& \sum_{\substack{ d_0\in { \mathbb{Z} }_{<0}\\ \langle \frac{-d_0}{s^+_1}\rangle ={\epsilon}_1} } \sum_{\substack{ \beta\in {\mathbb{K}}_{{\mathrm{eff}},{ {\sigma} }_-}\\ v(\beta-d_0\alpha)=v} } - x^{d_0} q^{\beta-d_0\alpha} (-1)^{ \frac{d_0}{s_1^+}(-f^+-1)-\{ {\epsilon}_2-f{\epsilon}_1\} + \langle D_{i_3},\beta-d_0\alpha \rangle } \frac{s_1^+}{d_0} {\begin{bmatrix}\frac{(f^++1) d_0}{s_1^+} - \langle D_{i_3},\beta-d_0\alpha \rangle -1\\#2\end{bmatrix}} \\ &&\cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot {\begin{bmatrix}0\\#2\end{bmatrix}} \cdot \prod_{I-\{i_1,i_2, i_3, i_4\} } {\begin{bmatrix}0\\#2\end{bmatrix}} \end{aligned}$$ The B-model and the mirror curve -------------------------------- The mirror B-model to the toric Calabi-Yau threefold ${\mathcal{X}}$ is another non-compact Calabi-Yau hypersurface $Y\subset { \mathbb{C} }^2\times ({ \mathbb{C} }^*)^2$, constructed as the Hori-Vafa mirror [@HV00]. It contains a distinguished mirror curve $C\subset\{ (0,0) \}\times ({ \mathbb{C} }^*)^2$. We simply state the relevant results in the most elementary way, and refer to [@AV00; @AKV02] for the mirror prediction of the disk amplitudes from $Y$. ### Mirror curve and the prepotential The mirror curve $C$ is given by the following equations $$\begin{aligned} x_1+\dots+x_r&=0,\\ \prod_{i=1}^r x_i^{l^{(a)}_i}&={\hat{q}}_a,\ a=1,\dots,k.\end{aligned}$$ After a change of variable $$x_{i_1}={\hat{q}}_0 y^{-f}, x_{i_2}=y, x_{i_3}=1,$$ and writing other $x_i$ in terms of $x,y$, we arrive at an equation $$\begin{aligned} \label{eqn:curve_F} F(y,{\hat{q}}_0,{\hat{q}}_1,\dots, {\hat{q}}_k)=0,\end{aligned}$$ which prescribes an curve in $({ \mathbb{C} }^*)^2$. This is called the *mirror curve* $C$ to the Calabi-Yau three orbifold ${\mathcal{X}}$ and the Aganagic-Vafa brane $({\mathcal{L}},f)$. We denote ${\hat{q}}_0=x^{s_1}$. The definition ${\mathbb{K}}_{{\mathrm{eff}},\sigma}=\{\beta\in {\mathbb{L}}_{ \mathbb{Q} }: \langle D_i, \text{$\beta \rangle \in { \mathbb{Z} }_{\geq0}$ for $i\in I_\sigma$}\}$ prompts us to identify ${\mathbb{K}}_{{\mathrm{eff}},\sigma}=\bigoplus_{i=1}^k { \mathbb{Z} }_{\ge 0} e^k$. We choose $\{p_a\}_{a=1}^k =\{D_i\}_{i\in I_{ {\sigma} }}$ such that $\langle p_a,e^b \rangle = \delta_{ab}$. In particular, we denote $i(a)\in I_\sigma$ with $p_a=D_{i(a)}$. We set $p_0=D_{i_1}$ and $i(0)=i_1$. Let $(\tilde{p}_{ab})_{0\leq a,b \leq k} = (\tilde{l}_{i(a)}^{(b)})$, and $(p_{ab})_{1\leq a,b \leq k}=(\tilde{l}_{i(a)}^{(b)})$. Then we have $$\tilde{p}_{00}=1,\quad \tilde{p}_{a0}=0,\quad \tilde{p}_{0b}= l_{i_1}^{(b)}, \quad \tilde{p}_{ab}= p_{ab},\quad \text{when $a,b\in \{1,\ldots,k\}$.}$$ Let $(\tilde{p}^{ab})_{0\leq a,b\leq k}$ be the inverse matrix of $(\tilde{p}_{ab})_{0\leq a,b\leq k}$, and $(p^{ab})_{1\leq a,b \leq k}$ be the inverse matrix of $(p_{ab})_{1\leq a,b\leq k}$. Then $$\tilde{p}^{00}=1 ,\quad \tilde{p}^{a0}=0 ,\quad \tilde{p}^{0b}= -\sum_{a=1}^k l_{i_1}^{(a)} p^{ab}=-\langle D_{i_1},e^b\rangle ,\quad \tilde{p}^{ab}= p^{ab},\quad \text{when $a,b\in \{1,\ldots,k\}$.}$$ Define ${ {\tilde{e}} }^b=(s_1 \tilde{p}^{0b}, e^b)$. Then $\langle p_a, { {\tilde{e}} }^b \rangle =\delta_{ab}$. Let $${ \mathbb{D} }_{\mathrm{eff}}=\bigoplus_{a=0}^k { \mathbb{Z} }_{\geq 0}{ {\tilde{e}} }^a.$$ It is obvious that ${\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})=\{(d,\beta)\in { \mathbb{D} }_{\mathrm{eff}}: d\neq 0\}$. Define $$\begin{aligned} \label{eqn:q-change} \tilde{q}_a := \prod_{b=0}^k \hat{q}_b^{\tilde{p}^{ba}},&\ a=0,\dots, k;\quad q_a:=\prod_{b=1}^k {\hat{q}}_b^{p^{ba}},\ a=1,\dots, k;\\ &\tilde{q}_0 = \hat{q}_0=q_0=x^{s_1}. \notag\end{aligned}$$ Then $$\tilde{q}_a = \hat{q}_0^{\tilde{p}^{0a}} \prod_{b=1}^k \hat{q}_b^{p^{ba}} = q_a \hat{q}_0^{-\langle D_{i_1},e^a\rangle}.$$ We have $$x_{i_2}^{\tilde{l}^{(a)}_{i_2}} x_{i_3}^{\tilde{l}^{(a)}_{i_3}}\prod_{i\in I_\tau} x_i^{\tilde{l}^{(a)}_i} = \hat{q}_a,\quad a=1,\ldots,k.$$ Let $\{ y_0, y_1,\ldots, y_k \} = \{x_i\}_{i\in I_\tau}$, $x_{i_2}=y$, $x_{i_3}=1$. Then $$\prod_{b=0}^k y_b^{\tilde{p}_{ba}} =\hat{q}_a y^{-\tilde{l}^{(a)}_{i_2}}.$$ So $$\begin{aligned} y_a = \prod_{b=0}^k (\hat{q}_b y^{\tilde{l}^{(b)}_{i_2}})^{\tilde{p}^{ba}}= \tilde{q}_a y^{-\sum_{b=0}^k \tilde{l}^{(b)}_{i_2} \tilde{p}^{ba}}. \end{aligned}$$ Let $$\epsilon_a = -\sum_{b=0}^k \tilde{l}^{(b)}_{i_2} \tilde{p}^{ba},\quad a=0,\ldots,k.$$ Then $\epsilon_0= -f$, $$\begin{aligned} \epsilon_a &= \sum_{b=1}^k (fl_{i_1}^{(b)}- l_{i_2}^{(b)})p^{ba}\\ &=f\langle D_{i_1},e^a\rangle - \langle D_{i_2},e^a\rangle ,\quad a=1,\ldots,k. \end{aligned}$$ The framed mirror curve equation, in terms of $y$ and new variables ${{\widetilde{q}}}_0,\dots, {{\widetilde{q}}}_k$ is $$\begin{aligned} \label{eqn:curve_H} H(y,\tilde{q}_0,\ldots, \tilde{q}_k)=0,\end{aligned}$$ where $$H(y,\tilde{q}_0,\ldots, \tilde{q}_k) = 1+y + \sum_{a=0}^k y_a = 1 +y + \sum_{a=0}^k \tilde{q}_a y^{\epsilon_a} = 1 + y +\sum_{a=0}^k \tilde{q}_a y^{{\epsilon}_a}.$$ We call this $H(y,{{\widetilde{q}}}_0,\dots,{{\widetilde{q}}}_k)$ *normalized curve equation*. So $$\frac{\partial H}{\partial y}(y,\tilde{q}_0,\ldots,\tilde{q}_k )= 1 + \sum_{a=0}^k \tilde{q}_a {\epsilon}_a y^{{\epsilon}_a-1}.$$ Write $\tilde{q}=(\tilde{q}_0,\ldots,\tilde{q}_k)$. Choose a branch of $\log y$ near $y=-1$ such that $\log(-1) = i\pi$. Then $H(y,\tilde{q})$ is holomorphic near $y=-1$, $\tilde{q}=0$, and $$H(-1,0,\dots,0)=0,\quad \frac{\partial H}{\partial y}(-1,0,\dots,0)=1 \neq 0.$$ By implicit function theorem, there exists a holomorphic function $h(\tilde{q}_0,\ldots, \tilde{q}_k)$ defined in an open neighborhood of $\tilde{q}=0$, such that $y(0,\dots,0)=-1$ and $H(y(\tilde{q}), \tilde{q})=0$. We conclude: \[lemm:y\] The implicit function $y({{\widetilde{q}}}_0,\dots, {{\widetilde{q}}}_k)$ is a power series in $\tilde{q}_0,\ldots, \tilde{q}_k$ with constant term $-1$, and $\log y$ is a power series in $\tilde{q}_0, \ldots, \tilde{q}_k$ with constant term $i\pi$. If we write $y$ as a Laurent series in $q_0, q_1,\dots,q_k$, [^4] the B-model prepotential $W_H$ is simply given as an anti-derivative: $$W_H=\int \frac{\log y(x,q_1,\dots,q_k)}{q_0} d q_0.$$ The Laurent series $W_H$ is decomposed as $$W_H= f(q_1,\dots,q_k)\log q_0 + W_{H,{\mathrm{inst}}}(x,q_1,\dots,q_k).$$ The *instanton* part $W_{H,{\mathrm{inst}}}$ is a Laurent series in $x,q_1,\dots,q_k$, with no degree $0$ term in $x$. Let $W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,\dots, q_k)$ be the pullback of the disk potential $F^{{\mathcal{X}},({\mathcal{L}},f)}_{(0,1),v}(Q,X)$ in sector $v\in \mathrm{Box}(\sigma)$ under the open-closed mirror map. Define a character $$\chi^{{\mathcal{L}},f}: {\mathrm{Box}({ {\sigma} })}= G_{ {\sigma} }\to U(1),\quad v\mapsto e^{2\pi\sqrt{-1}(c_{i_2}(v)-fc_{i_1}(v))}.$$ Define $$W^{{\mathcal{X}},({\mathcal{L}},f)}(x,q_1,\dots,q_k):=\sum_{v\in \mathrm{Box(\sigma)}} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,\dots, q_k) \sqrt{\chi^{{\mathcal{L}},f}(v)},$$ where $\sqrt{\chi^{{\mathcal{L}},f}(v)} =e^{\pi\sqrt{-1}(c_{i_2}(v)-fc_{i_1}(v))} $ The mirror conjecture for disk amplitudes, proposed in [@AV00; @AKV02], is the following. \[conj:akv\] The pullback of the B-model prepotential $W^{{\mathcal{X}},({\mathcal{L}},f)}(x,q_1,\dots, q_k)$ of the disk potential $F_{0,1}^{{\mathcal{X}},({\mathcal{L}},f)}(Q,X)$ under the open-closed mirror map is equal to the instanton part of the B-model super potential: $$W^{{\mathcal{X}},({\mathcal{L}},f)}(x,q_1,\dots,q_k)=W_{H,{\mathrm{inst}}}(x,q_1,\dots,q_k).$$ If ${\mathcal{X}}$ is a smooth variety, then $\chi_0=1$, where $v=0$ is the only element in ${\mathrm{Box}({ {\sigma} })}$. We get back to the original form of the conjecture in [@AV00; @AKV02]. Open mirror theorem for disk amplitudes {#sec:open-mirror} --------------------------------------- The solution $v$ to the exponential polynomial equation $$s_1 e^{r_1 v}+s_2 e^{r_2 v}+\dots+ s_k e^{r_k v} - e^v +1=0,$$ around $s_1=\dots=s_k=0, v=0$ is in the following power series form (see Appendix \[app:exppoly\] for a proof) $$\begin{aligned} v=\sum_{\substack{n_1,\dots,n_k=0\\(n_1,\dots,n_k)\neq 0}}^\infty \frac{(r_1n_1 + \dots r_k n_k -1)_{(n_1+\dots+n_k-1)}}{n_1!\dots n_k!} s_1^{n_1}\dots s_k^{n_k}. \label{eqn:exppoly}\end{aligned}$$ Here we adopt the Pochhammer symbol $$(a)_n=\frac{\Gamma(a+1)}{\Gamma(a-n+1)}=\begin{cases} a(a-1)\cdots(a-n+1),\quad n>0;\\ 1,\quad n=0;\\\frac{1}{(a+1)\dots(a-n)},\quad n<0;\end{cases}$$ where $a\in { \mathbb{C} }$ and $n\in { \mathbb{Z} }$. Starting from this simple observation, we prove Conjecture \[conj:akv\] in this section. The mirror curve $H(y, {{\widetilde{q}}}_0,\dots,{{\widetilde{q}}}_k)$ is $$1+y+\sum_{a=0}^k {{\widetilde{q}}}_a y^{\epsilon_a}=0.$$ After a change of variable $y=-y'$ and ${{\widetilde{q}}}_a=(-1)^{-\epsilon_a}{{\widetilde{q}}}'_a$, it becomes $$1-y'+\sum_{a=0}^k {{\widetilde{q}}}'_a y'^{\epsilon_a}=0.$$ For $\beta \in {\mathbb{K}}_\sigma$ and ${{\widetilde{\beta}}}=(d_0,\beta)\in { \mathbb{Z} }\times {\mathbb{K}}_\sigma$, recall the definition of power notations $$\begin{aligned} q^\beta=\prod_{a=1}^k q_a^{\langle p_a,\beta\rangle},&\quad q^{{\widetilde{\beta}}}=x^{d_0} \prod_{a=1}^k q_a^{\langle p_a,\beta\rangle},\\ {{\widetilde{q}}}^{\prime {{\widetilde{\beta}}}}=\prod_{a=0}^k {{\widetilde{q}}}_a^{\prime \langle \tilde p_a,{{\widetilde{\beta}}}\rangle},&\quad {{\widetilde{q}}}^{{\widetilde{\beta}}}=\prod_{a=0}^k {{\widetilde{q}}}_a^{\langle \tilde p_a,{{\widetilde{\beta}}}\rangle}.\end{aligned}$$ Hence $$\begin{aligned} \log y'&=\sum_{\substack{{{\tilde{d}}}_0,\dots,{{\tilde{d}}}_k=0\\({{\tilde{d}}}_0,\dots,{{\tilde{d}}}_k)\neq 0}}^\infty \frac{(\epsilon_0 {{\tilde{d}}}_0 + \dots \epsilon_k {{\tilde{d}}}_k -1)_{({{\tilde{d}}}_0+\dots+{{\tilde{d}}}_k-1)}}{{{\tilde{d}}}_0!\dots {{\tilde{d}}}_k!} {{\widetilde{q}}}_0^{\prime {{\tilde{d}}}_0}\dots {{\widetilde{q}}}_k^{\prime {{\tilde{d}}}_k}\\ &=\sum_{{{\widetilde{\beta}}}=(d_0,\beta)\in { \mathbb{D} }_{\mathrm{eff}}, {{\widetilde{\beta}}}\neq 0} \frac{(-f \langle D_{i_1},{{\widetilde{\beta}}}\rangle +f\langle D_{i_1}, \beta\rangle -\langle D_{i_2},\beta \rangle-1)_{(\sum_{i\in I_\tau}\langle D_i,{{\widetilde{\beta}}}\rangle-1)}}{\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} {{\widetilde{q}}}'^{{\widetilde{\beta}}}\\ &=\sum_{{{\widetilde{\beta}}}\in { \mathbb{D} }_{\mathrm{eff}}, {{\widetilde{\beta}}}\neq 0} \frac{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} {{\widetilde{q}}}'^{{\widetilde{\beta}}}\\ &=\sum_{{{\widetilde{\beta}}}\in { \mathbb{D} }_{\mathrm{eff}}, {{\widetilde{\beta}}}\neq 0} (-1)^{-\langle D_{i_2},{{\widetilde{\beta}}}\rangle}\frac{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} {{\widetilde{q}}}^{{\widetilde{\beta}}}\\ &=\sum_{{{\widetilde{\beta}}}\in { \mathbb{D} }_{\mathrm{eff}}, {{\widetilde{\beta}}}\neq 0} (-1)^{\langle D_{i_3},{{\widetilde{\beta}}}\rangle+1}\frac{(-\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} x^{d_0} q^\beta.\end{aligned}$$ It follows that $$\begin{aligned} \int \frac{\log y}{q_0} dq_0&=f(q_1,\dots, q_n) \log q_0\\&+\sum_{{{\widetilde{\beta}}}=(d_0,\beta)\in { \mathbb{D} }_{\mathrm{eff}}, d_0\neq 0} (-1)^{\langle D_{i_3},{{\widetilde{\beta}}}\rangle+1}\frac{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\frac{d_0}{s_1}\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} x^{d} q^\beta.\end{aligned}$$ We conclude that $$\begin{aligned} \label{eqn:W_H} W_{H,{\mathrm{inst}}}(x,q_1,\dots,q_k)=\sum_{{{\widetilde{\beta}}}=(d_0,\beta)\in {\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})} (-1)^{\langle D_{i_3},\beta \rangle-\frac{(f+1)d_0}{s_1}+1}\frac{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\frac{d_0}{s_1}\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !} x^{d_0} q^\beta.\end{aligned}$$ In Theorems \[thm:amplitude-term-outer\] and \[thm:amplitude-term-inner\] the pulled-back disk amplitude $$W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,\dots,q_k)=\sum_{\substack{{{\widetilde{\beta}}}=(d_0,\beta)\in {\mathbb{K}}_{\mathrm{eff}}({\mathcal{X}},{\mathcal{L}})\\v=v(\beta)}} x^{d_0} q^\beta A^{{\mathcal{X}},({\mathcal{L}},f)}_{{\widetilde{\beta}}},$$ where $$\begin{aligned} \label{eqn:A_beta} A^{{\mathcal{X}},({\mathcal{L}},f)}_{{\widetilde{\beta}}}=(-1)^{ \langle D_{i_3},\beta\rangle-\frac{d_0}{s_1}(f+1) +1 -\{ c_{i_2}(v)-fc_{i_1}(v)\} } \frac{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -1)_{(-\langle D_{i_2},{{\widetilde{\beta}}}\rangle -\langle D_{i_3},{{\widetilde{\beta}}}\rangle -1)}} {\frac{d_0}{s_1}\prod_{i\in I_\tau} \langle D_i,{{\widetilde{\beta}}}\rangle !}.\end{aligned}$$ It follows that $$W_{H,{\mathrm{inst}}}(x,q_1,\dots,q_k)= \sum_{v\in {\mathrm{Box}({ {\sigma} })}} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,\dots,q_k) \sqrt{\chi^{{\mathcal{L}},f}(v)}.$$ This gives Conjecture \[conj:akv\]. ${\mathcal{X}}={\mathcal{X}}_{1,1,1}=[{ \mathbb{C} }^3/{ \mathbb{Z} }_3]$, $s_1=3$. The framed mirror curve is $$x^3 y^{-f} + y +1 + {\hat{q}}_1^{-\frac{1}{3}} x y^{\frac{1-f}{3}}=0.$$ After a change of variables $${{\widetilde{q}}}_0={\hat{q}}_0=x^3,\quad {{\widetilde{q}}}_1={\hat{q}}_1^{-\frac{1}{3}}{\hat{q}}_0^{\frac{1}{3}},$$ the curve equation becomes $${{\widetilde{q}}}_0 y^{-f}+{{\widetilde{q}}}_1 y^{\frac{1-f}{3}} +y +1 =0.$$ The B-model pre-potential $$\log y=\sum_{{{\tilde{d}}}_0,{{\tilde{d}}}_1\geq 0} \frac{(-f {{\tilde{d}}}_0 + \frac{1-f}{3} {{\tilde{d}}}_1 -1)_{({{\tilde{d}}}_0+{{\tilde{d}}}_1-1)}}{{{\tilde{d}}}_0!{{\tilde{d}}}_1!}{{\widetilde{q}}}_0^{{{\tilde{d}}}_0}{{\widetilde{q}}}_1^{{{\tilde{d}}}_1}.$$ Let $d_0=3{{\tilde{d}}}_0+{{\tilde{d}}}_1$ and $d_1={{\tilde{d}}}_1$. $$W_{H,{\mathrm{inst}}}=\sum_{\substack{d_0\in { \mathbb{Z} }_{>0}, d_1\in { \mathbb{Z} }_{\geq 0}\\ d_0-d_1\in 3{ \mathbb{Z} }_{\geq 0}}} (-1)^{-\frac{(f+1)d_0}{3}-\frac{d_1}{3}}\frac{1}{\frac{d_0}{3}( \frac{d_0-d_1}{3} )!d_1!}\cdot \frac{\Gamma(\frac{f+1}{3}d_0+\frac{d_1}{3})}{\Gamma(\frac{fd_0-d_1}{3}+1)} x^d q_1^{d_1}.$$ Setting $\chi_v=-e^{-\sqrt{-1}\pi(\lfloor \frac{(f-1)d}{3}\rfloor+(f+4)\{\frac{d}{3}\})}$, $$W_{H,{\mathrm{inst}}}(x,q_1)=\sum_{v\in \mathrm{Box}(\sigma)} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1)\sqrt{\chi^{{\mathcal{L}},f}(v)},$$ where the pullback prepotential $W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1)$ in the twisted-sector $v$ is given in Example \[exp:prepotential-A\], and $$\sqrt{\chi^{{\mathcal{L}},f}(1)}=1,\quad \sqrt{\chi^{{\mathcal{L}},f}(e^{2\pi\sqrt{-1}/3})}=e^{\pi\sqrt{-1}(1-f)/3}, \quad \sqrt{\chi^{{\mathcal{L}},f}(e^{4\pi\sqrt{-1}/3})}=e^{2\pi\sqrt{-1}(1-f)/3},$$ The framed mirror curve of an outer brane in $K_{{ \mathbb{P} }^2}$, the crepant resolution of the coarse moduli space of ${\mathcal{X}}_{1,1,1}$, is $$\label{eqn:localPtwo} \hat{q}\hat{x}^3 \hat{y}^{-3\hat{f}-1} +y +1 +\hat{x} y ^{-\hat{f}} =0,$$ or equivalently, $$\label{eqn:BKMP} y^{3\hat{f}+2} + y^{3\hat{f}+1} + \hat{q} \hat{x}^3 + \hat{x}y^{2\hat{f}+1} =0.$$ Equation coincides with Equation (3.19) in [@BKMP2]. The framed mirror curve of ${\mathcal{X}}_{1,1,1}$ is $$\label{eqn:mirror-curve} x^3y^{-f} + y + 1 + q_{ {\mathrm{orb}} }x y^{\frac{1-f}{3}} =0.$$ and are equivalent under the following change of B-model coordinates and framing. $$q_{ {\mathrm{orb}} }= \hat{q}^\frac{-1}{3},\quad x= \hat{x}\hat{q}^{\frac{1}{3}}, \quad f= 3\hat{f}+1.$$ The change of coordinates $$(q_{ {\mathrm{orb}} }, x)\mapsto (q_{ {\mathrm{orb}} }^{-3}, xq_{ {\mathrm{orb}} }^{-1})=(\hat{q},\hat{x})$$ is a 3-to-1 map, and the inverse map $$(\hat{q},\hat{x})\mapsto (\hat{q}^{\frac{-1}{3}}, \hat{x}\hat{q}^{\frac{1}{3}}) =(q_{ {\mathrm{orb}} }, x)$$ is multi-valued. ${\mathcal{X}}={\mathcal{X}}_{0,1,2}=[{ \mathbb{C} }^2/{ \mathbb{Z} }_3]\times { \mathbb{C} }$, $s_1=1$. The framed mirror curve is $$x y^{-f} +y + 1 + {\hat{q}}_1^{-\frac{2}{3}} {\hat{q}}_2^{-\frac{1}{3}} y^{\frac{1}{3}} + {\hat{q}}_1^{-\frac{1}{3}} {\hat{q}}_2^{-\frac{2}{3}} y^{\frac{2}{3}}=0.$$ After a change of variables $${{\widetilde{q}}}_0={\hat{q}}_0=x,\quad {{\widetilde{q}}}_1={\hat{q}}_1^{-\frac{2}{3}} {\hat{q}}_2^{-\frac{1}{3}},\quad {{\widetilde{q}}}_1={\hat{q}}_1^{-\frac{2}{3}} {\hat{q}}_2^{-\frac{1}{3}},$$ the curve equation becomes $${{\widetilde{q}}}_0y^{-f}+{{\widetilde{q}}}_1y^\frac{1}{3}+{{\widetilde{q}}}_2 y^\frac{2}{3}+y+1=0.$$ The B-model superpotential $$\log y=\sum_{{{\tilde{d}}}_0,{{\tilde{d}}}_1,{{\tilde{d}}}_2\geq 0} \frac{(-f{{\tilde{d}}}_0+\frac{1}{3}{{\tilde{d}}}_1+\frac{2}{3}{{\tilde{d}}}_2-1)_{({{\tilde{d}}}_0+{{\tilde{d}}}_1+{{\tilde{d}}}_2-1)}}{{{\tilde{d}}}_0!{{\tilde{d}}}_1!{{\tilde{d}}}_2!} {{\widetilde{q}}}_0^{{{\tilde{d}}}_0}{{\widetilde{q}}}_1^{{{\tilde{d}}}_1} {{\widetilde{q}}}_2^{{{\tilde{d}}}_2}.$$ Let $d_0={{\tilde{d}}}_0$, $d_1={{\tilde{d}}}_1$ and $d_2={{\tilde{d}}}_2$. $$\begin{aligned} &W_{H,{\mathrm{inst}}}\\ =& \sum_{\substack{d\in { \mathbb{Z} }_{>0}\\d_1,d_2\in { \mathbb{Z} }_{\geq 0}}} \frac{(-1)^{-fd_0-\frac{2d_1+d_2}{3}}}{d_0 d_0! d_1! d_2!}\cdot \frac{\Gamma((f+1)d_0+\frac{2d_1+d_2}{3})}{\Gamma(fd_0-\frac{d_1+2d_2}{3}+1)}x^d_0 q_1^{d_1} q_2^{d_2}.\end{aligned}$$ $$W_{H,{\mathrm{inst}}}=\sum_{v\in {\mathrm{Box}({ {\sigma} })}} W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,q_2)\sqrt{\chi^{{\mathcal{L}},f}(v)}$$ where the pullback prepotential $W^{{\mathcal{X}},({\mathcal{L}},f)}_v(x,q_1,q_2)$ in the twisted-sector $v$ is given in Example \[exp:prepotential-A\], and $$\sqrt{\chi^{{\mathcal{L}},f}(1)}=1,\quad \sqrt{\chi^{{\mathcal{L}},f}(e^{2\pi\sqrt{-1}/3})}=e^{\pi\sqrt{-1}/3}, \quad \sqrt{\chi^{{\mathcal{L}},f}(e^{4\pi\sqrt{-1}/3})}=e^{2\pi\sqrt{-1}/3}.$$ Proof of Equation {#app:exppoly} ================== The solution to the following exponential polynomial $$\label{eqn:eqnexppoly} 1-e^{v'}+{{\widetilde{q}}}_0 e^{r_0 v'}+ {{\widetilde{q}}}_1 e^{r_1 v'}+\dots +{{\widetilde{q}}}_k e^{r_k v'}=0$$ at ${{\widetilde{q}}}_0=\dots={{\widetilde{q}}}_k=0$ is indeed obtained by solving a system of GKZ-type Picard-Fuchs equations. The issue here is that since we are dealing with orbifolds, the parameters $r_0,\dots,r_k$ might be rational numbers. We minimize our argument by quoting known results for integer parameters. Consider a system of *extended charge vectors* $$({{\tilde{l}}}^{(a)}_i)_{i=1,\dots,k+2}^{a=0,\dots,k}= \begin{pmatrix} 1 & -r_0 & r_0-1 & 0 & 0 & 0 & \cdots & 0 & 1 & -1\\ 0 & -r_1 & r_1-1 & 1 & 0 & 0 & \cdots & 0 & 0 & 0 \\ 0 & -r_2 & r_2-1 & 0 & 1 & 0 & \cdots & 0 & 0 & 0 \\ 0 & -r_3 & r_3-1 & 0 & 0 & 1 & \cdots & 0 & 0 & 0 \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots &\vdots &\vdots \\ 0 & -r_k & r_k-1 & 0 & 0 & 1 & \cdots & 1 & 0 & 0 \end{pmatrix}$$ Assume these charge vectors consist of integers, i.e. $r_a\in { \mathbb{Z} },\ a=0,\dots, k.$ As a set of extended charge vectors, these charge vectors prescribe a B-model as a threefold $$Y=\{(z,w,x,y)\in { \mathbb{C} }^2\times ({ \mathbb{C} }^*)^2, zw=H(x,y,q_1,\dots,q_k)\}$$ where $H(x,y)=x_1+\dots+x_k$ is given by the following equations $$\begin{aligned} & \prod_{i=1}^k x_i^{{{\tilde{l}}}_i^{(a)}}=q_a,\quad a=1,\dots, k,\\ & x_1=-x^{r_1},\ x_2=y=e^v, \ x_3=1.\end{aligned}$$ Define two $2$-cycles ${\mathcal{C}}_*$ given by $$\begin{aligned} w=0=H(x,y), \quad x=x_*, \quad y=y_*.\end{aligned}$$ This definition involves a choice of $x_*$ and $y_*$ with $H(x_*,y_*)=0$. Define another $2$-cycle ${\mathcal{C}}_x$ given by $$\begin{aligned} w=0=F(x(|z|),y(|z|)),&\quad x(0)=x,\\ x(|z|)=x_*,\ y(|z|)=y_*,&\quad \text{when $|z|>\Lambda$ for some $\Lambda>0$},\end{aligned}$$ The $2$-cycle ${\mathcal{C}}_x$ together with $x(|z|),y(|z|)$ as functions in $|z|$ are regarded as $2$-branes wrapping over ${\mathcal{C}}_x$ in physics literature [@AV00; @AKV02]. These two $2$-cycles ${\mathcal{C}}_*$ and ${\mathcal{C}}_*$ bound a compact $3$-cycle $\Gamma(x)$, i.e. $\partial \Gamma={\mathcal{C}}_x-{\mathcal{C}}_*$. We denote that the chain integral, which does not depend on the choice of $\Gamma(x)$, to be $$P=\frac{1}{2\pi i}\int_{\Gamma(x)} \Omega.$$ Lerche and Mayr [@LM01] show that the $3$-chain integral $P$ prescribed by this system of charge vectors are annihilated by the following Picard-Fuchs operators ($q_0=-x$) $${\mathcal{D}}_a=\prod_{{{\tilde{l}}}_i^{(a)}>0}\prod_{j=0}^{{{\tilde{l}}}^{(a)}_i-1} \left (\sum_{b=0}^k q_b \frac{\partial}{\partial q_b}-j\right) - q_a \prod_{{{\tilde{l}}}_i^{(a)}<0} \prod_{j=0}^{-{{\tilde{l}}}^{(a)}_i-1} \left (\sum_{\beta=0}^k q_b \frac{\partial}{\partial q_b}-j\right),\quad a=0,\dots,k.$$ Solving these Picard-Fuchs equations ${\mathcal{D}}_a P=0,\ a=0,\dots,k$ $$\begin{aligned} P&=\text{double logarithm terms}\\ &+\sum_{(d_0,\dots, d_k)\in { \mathbb{Z} }_{> 0}\times { \mathbb{Z} }_{\geq 0}^k} \frac{(-1)^{d_0+r_0 d_0+\dots + r_k d_k} (r_0 d_0+\dots + r_k d_k-1)_{(d_1+\dots+d_k-1)} } {d_0\cdot d_0! d_1!\dots d_k!} x^{d_0} q_1^{d_1}\dots q_k^{d_k}.\end{aligned}$$ Notice that Lerche and Mayr do not assume the charge vectors prescribe a toric variety. Their argument is completely on the B-model side. On the other hand, the power series part of this three chain integral is obtained by solving the spectral curve $H(x,y)=0$, as shown by Aganagic and Vafa [@AV00], i.e. $$\int v \frac{dx}{x}= \text{logarithm terms} + \text{power series part of $P$}.$$ Given these charge vectors, the spectral curve equation $H(x,y)=0$ is $$1 + e^v - x e^{r_0 v}+q_1e^{r_1 v}+q_k e^{r_k v}+\dots+q_k e^{r_k v}=0.$$ Hence the explicit form of the power series part of $P$ implies $$\begin{aligned} v=&\sqrt{-1} \pi + f_k(q_1,\dots,q_k)\\ & + \sum_{(d_0,\dots, d_k)\in { \mathbb{Z} }_{> 0}\times { \mathbb{Z} }_{\geq 0}^k} \frac{(-1)^{d_0+r_0 d_0+\dots + r_k d_k} (r_0 d_0+\dots + r_k d_k-1)_{(d_1+\dots+d_k-1)} } {d_0! d_1!\dots d_k!} x^{d_0} q_1^{d_1}\dots q_k^{d_k}.\end{aligned}$$ where $f(q_1,\dots,q_k)$ is an analytic function with $f(0)=0$. After a change of variables $$\begin{aligned} &y=-y'=-e^{v'},\ v'=v-\sqrt{-1} \pi, x=-(-1)^{-r_0} {{\widetilde{q}}}'_0,\\ & q_a=(-1)^{-r_a} {{\widetilde{q}}}'_a, \ a=1,\dots, k,\end{aligned}$$ the curve equation becomes $$1-e^{v'}+{{\widetilde{q}}}_0 e^{r_0 v'}+ {{\widetilde{q}}}_1 e^{r_1 v'}+\dots +{{\widetilde{q}}}_k e^{r_k v'}=0,$$ The solution becomes $$v'=f'_k({{\widetilde{q}}}_1,\dots,{{\widetilde{q}}}_k)+\sum_{(d_0,\dots, d_k)\in { \mathbb{Z} }_{> 0}\times { \mathbb{Z} }_{\geq 0}^k} \frac{(r_0 n_0+\dots + r_k n_k-1)_{(d_1+\dots+d_k-1)} }{d_0! d_1!\dots d_k!} {{\widetilde{q}}}_0^{\prime d_0} {{\widetilde{q}}}_1^{\prime d_1}\dots {{\widetilde{q}}}_k^{\prime d_k}.$$ Setting ${{\widetilde{q}}}_0=0$, $v'(0,{{\widetilde{q}}}_1,\dots,{{\widetilde{q}}}_k)=f'_k({{\widetilde{q}}}_1,\dots, {{\widetilde{q}}}_k)$ is again the solution to the Equation in the same form with one less variable. Repeating the same argument, we have $$f'_{k}=f'_{k-1}({{\widetilde{q}}}_2,\dots,{{\widetilde{q}}}_{k-1})+\sum_{(d_1,\dots, d_k)\in { \mathbb{Z} }_{>0}\times { \mathbb{Z} }_{\geq 0}^{k-1}} \frac{(r_1 d_1+\dots + r_k d_k-1)_{(d_1+\dots+d_k-1)} }{d_1! d_2!\dots d_k!} {{\widetilde{q}}}_1^{\prime d_1} {{\widetilde{q}}}_2^{\prime d_2}\dots {{\widetilde{q}}}_k^{\prime d_k}.$$ By induction and the observation that $v'(0,\dots,0)=0$, we have $$v'=\sum_{\substack{d_1,\dots,d_k=0\\(d_1,\dots,d_k)\neq 0}}^\infty \frac{(r_0 d_0+\dots + r_k d_k-1)_{(d_1+\dots+d_k-1)} }{d_0! d_1!\dots d_k!} {{\widetilde{q}}}_0^{\prime d_0} {{\widetilde{q}}}_1^{\prime d_1}\dots {{\widetilde{q}}}_k^{\prime d_k}.$$ Let $$v'=\sum_{\substack{d_1,\dots,d_k=0\\(d_1,\dots,d_k)\neq 0}}^\infty c_{d_0,\dots, d_k} {{\widetilde{q}}}_0^{\prime d_0}\dots {{\widetilde{q}}}_k^{\prime d_k}$$ be the power series solution to the Equation . We conclude that $$c_{d_0,\dots, d_k}=\frac{(r_0 d_0+\dots + r_k d_k-1)_{(d_1+\dots+d_k-1)}}{d_0! d_1!\dots d_k!}$$ when $r_0,\dots, r_k \in { \mathbb{Z} }$. Since $c_{d_0,\dots,d_k}$ is a rational function of $r_0,\dots, r_k$, this expression has to be true for all complex-valued $r_a$. [ABC]{} D. Abramovich, T. Graber, A. Vistoli, “Angelo Algebraic orbifold quantum products," [*Orbifolds in mathematics and physics (Madison, WI, 2001),*]{} 1–24, Contemp. Math., 310, Amer. Math. Soc., Providence, RI, 2002. D. Abramovich, T. Graber, A. Vistoli, “Gromov-Witten theory of Deligne-Mumford stacks," Amer. J. Math. 130 (2008), no. 5, 1337–1398. M. Aganagic, A. Klemm, M. Mariño, C. Vafa, “The topological vertex," Comm. Math. Phys. 254 (2005), no. 2, 425–478. M. Aganagic, A. Klemm, C. Vafa, “Disk instantons, mirror symmetry and the duality web, " Z. Naturforsch. A [**57**]{} (2002), no. 1-2, 1–28. M. Aganagic, C. Vafa, “Mirror Symmetry, D-Branes and Counting Holomorphic Discs", [hep-th/0012041]{}. L. Borisov, L. Chen, G. Smith, “The orbifold Chow ring of toric Deligne-Mumford stacks,” J. Amer. Math. Soc. [**18**]{} (2005), no. 1, 193–215. V. Bouchard, A. Klemm, M. Mariño, and S. Pasquetti, “Remodeling the B-model," Comm. Math. Phys. [**287**]{} (2009), no. 1, 117–178. V. Bouchard, A. Klemm, M. Mariño, and S. Pasquetti, “Topological open strings on orbifolds,” Comm. Math. Phys. [**296**]{} (2010), 589–623. A. Brini, “Open topological strings and integrable hierarchies: remodeling the A-model,” Comm. Math. Phys. [**312**]{} (2012), no. 3, 735–780. A. Brini, R. Cavalieri, “Open orbifold Gromov-Witten invariants of $[{ \mathbb{C} }^3/{ \mathbb{Z} }_n]$: localization and mirror symmetry,” Selecta Math. (N.S.) 17 (2011), no. 4, 879–933. J. Bryan, C. Cadman, B. Young, “The orbifold topological vertex,”, Adv. Math. [**229**]{} (2012), no. 1, 531–595. D.E. Diaconescu, B. Florea, “Localization and Gluing of Topological Amplitudes,” Comm. Math. Phys. [**257**]{} (2005) 119-149. R. Cavalieri, D. Ross, “Open Gromov-Witten theory and the crepant resolution conjecture,” [arXiv:1102.0717]{}, to appear in Michigan Mathematical Journal. S. Cecotti, C. Vafa, “Massive orbifolds", Modern Phys. Lett. A [**7**]{} (1992), no. 19, 1715–1723. K. Chan, C.-H. Cho, S.-C. Lau, H.-H. Tseng, “Lagrangian Floer superpotentials and crepant resolutions for toric orbifolds,” [arXiv:1208.5282]{}. K. Chan, S.-C. Lau, N. C. Leung, H.-H. Tseng, “Open Gromov-Witten invariants and mirror maps for semi-Fano toric manifolds”, [arXiv:1112.0388]{}. K. Chan, S.-C. Lau, H.-H. Tseng, “Enumerative meaning of mirror maps for toric Calabi-Yau manifolds”, [arXiv:1110.4439]{}. W. Chen, Y. Ruan, “Orbifold Gromov-Witten theory,” [*Orbifolds in mathematics and physics (Madison, WI, 2001),*]{} 25–85, Contemp. Math., [**310**]{}, Amer. Math. Soc., Providence, RI, 2002. W. Chen, Y. Ruan, “A new cohomology theory of orbifold,” Comm. Math. Phys. [**248**]{} (2004), no. 1, 1–31. C.-H. Cho, M. Poddar, “Holomorphic orbidiscs and Lagrangian Floer cohomology of symplectic toric orbifolds”, [arXiv:1206.3994]{}. T. Coates, “On the crepant resolution conjecture in the local case,” Comm. Math. Phys. [**287**]{} (2009), no. 3, 1071–1108. T. Coates, A. Givental, “Quantum Riemann-Roch, Lefschetz and Serre,” Ann. of Math. (2) [**165**]{} (2007), no. 1, 15–53. T. Coates, A. Corti, H. Iritani, H.-H. Tseng, “Quantum cohomology of toric stacks”, in preparation. D. Cox, J. Little, H. Schenck, [*Toric varieties*]{}, Graduate Studies in Mathematics [**124**]{}, American Mathematical Society, 2011. E. Eynard, N. Orantin, “Invariants of algebraic curves and topological expansion," Commun. Number Theory Phys. [**1**]{} (2007), no. 2, 347–45 E. Eynard, N. Orantin, “Computation of open Gromov-Witten invariants for toric Calabi-Yau 3-folds by topological recursion, a proof of the BKMP conjecture,” [arXiv:1205.1103]{}. B. Fang, C.-C. Liu, “Open Gromov-Witten invariants of toric Calabi-Yau 3-folds,” [arXiv:1103.0693]{}. K. Fukaya, Y.-G. Oh, H. Ohta. K. Ono. “Lagrangian Floer theory on compact toric manifolds, I”, Duke Math. J. [**151**]{} (2010), no. 1, 23–174. W. Fulton, *Introduction to toric varieties*, Annals of Mathematics Studies, [**131**]{}. The William H. Roever Lectures in Geometry. Princeton University Press, Princeton, NJ, 1993. A. Givental, “A mirror theorem for toric complete intersections," Topological field theory, primitive forms and related topics (Kyoto, 1996), 141–175, Progr. Math., [**160**]{}, *Birkhäuser Boston*, Boston, MA, 1998. A. Givental, “Elliptic Gromov-Witten invariants and the generalized mirror conjecture," Integrable systems and algebraic geometry (Kobe/Kyoto, 1997), 107155, World Sci. Publ., River Edge, NJ, 1998. E. Gonzalez, C. Woodward, “Quantum cohomology and toric minimal model programs,” [arXiv:1207.3253]{}. T. Graber, E. Zaslow, “Open-string Gromov-Witten invariants: calculations and a mirror “theorem”,” Orbifolds in mathematics and physics (Madison, WI, 2001), 107–121, Contemp. Math., [**310**]{}, Amer. Math. Soc., Providence, RI, 2002. K. Hori, A. Iqbal, C. Vafa, “D-Branes And Mirror Symmetry", [hep-th/0005247]{}. K. Hori, C. Vafa, “Mirror Symmetry", [arXiv:hep-th/0002222]{}. S. Hosono, A. Klemm, S. Theisen, S.-T. Yau, “Mirror symmetry, mirror map and applications to complete intersection Calabi-Yau spaces," Nuclear Phys. B [**433**]{} (1995), no. 3, 501–552. H. Iritani, “An integral structure in quantum cohomology and mirror symmetry for toric orbifolds,” Adv. Math. [**222**]{} (2009), no. 3, 1016–1079. Y. Jiang, “The orbifold cohomology ring of simplicial toric stack bundles," Illinois J. Math. [**52**]{} (2008), no. 2, 493–514. S. Katz, C.C. Liu, “Enumerative geometry of stable maps with Lagrangian boundary conditions and multiple covers of the disc,” Adv. Theor. Math. Phys. 5 (2001), no. 1, 1–49. W. Lerche, P. Mayr, “On N=1 Mirror Symmetry for Open Type II Strings," [arXiv:hep-th/0111113.]{} B. H. Lian, K. Liu, S.-T. Yau, “Mirror principle. I," Asian J. Math. [**1**]{} (1997), no. 4, 729–763. B. H. Lian, K. Liu, S.-T. Yau, “Mirror principle. II," Asian J. Math. [**3**]{} (1997), no. 1, 109–146. C. Lin, “Bouchard-Klemm-Marino-Pasquetti Conjecture for $\mathbb{C}^3$," [arXiv:0910.3739]{}. C.-C. M. Liu, “Moduli of $J$-holomorphic curves with Lagrangian boundary conditions and open Gromov-Witten invariants for an $S^1$-equivariant pair,” [arXiv:math/0211388]{}. C.-C. M. Liu, “Localization in Gromov-Witten theory and orbifold Gromov-Witten theory", [arXiv:1107.4712]{}. J. Li, C.-C. M. Liu, K. Liu, J. Zhou, “A mathematical theory of the topological vertex," Geom. Topol. [**13**]{} (2009), no. 1, 527–621. M. Mari\~ no, “Open string amplitudes and large order behavior in topological string theory", J. High Energy Phys. (2008), no. 3, 060, 34 pp. P. Mayr, “$\mathcal N=1$ mirror symmetry and open/closed string duality," Adv. Theor. Math. Phys. [**5**]{} (2001), no. 2, 213–242. D. Maulik, A. Oblomkov, A. Okounkov, “Gromov-Witten/Donaldson-Thomas correspondence for toric 3-folds," Invent. Math. [**186**]{} (2011), no. 2, 435-479. A. Okounkov, R. Pandharipande, “Hodge integrals and invariants of the unknot," Geom. Topol. [**8**]{} (2004), 675–699. R. Pandharipande, J. Solomon, J. Walcher, “Disk enumeration on the quintic 3-fold,” J. Amer. Math. Soc. [**21**]{} (2008), no. 4, 1169–1209. A. Popa, A. Zinger, “Mirror symmetry for closed, open, and unoriented Gromov-Witten Invariants.” D. Ross, “Localization and gluing of orbifold amplitudes: the Gromov-Witten orbifold vertex", [arXiv:1109.5995]{}, to appear in Trans. Amer. Math. Soc. D. Ross, “The Gerby Gopakumar-Mariño-Vafa Formula,” [arXiv:1208.4342]{}. J. Solomon, “Intersection theory on the moduli space of holomorphic curves with Lagrangian boundary conditions,” [arXiv:math/0606429]{}. H.-H. Tseng, “Orbifold quantum Riemann-Roch, Lefschetz and Serre,” Geom. Topol. [**14**]{} (2010), no. 1, 1–81. J. Walcher, “Opening mirror symmetry on the quintic,” Comm. Math. Phys. [**276**]{} (2007), no. 3, 671–689. E. Zaslow, “Topological orbifold models and quantum cohomology rings", Comm. Math. Phys. [**156**]{} (1993), no. 2, 301–331. J. Zhou, “Local mirror symmetry for one-legged topological vertex," [arXiv:0910.4320]{}. J. Zhou, “Open string invariants and mirror curve of the resolved conifold,” [arXiv:1001.0447]{}. Z. Zong, “Generalized Mariño-Vafa formula and local Gromov-Witten theory of orbi-curves," [arXiv:1109.4992]{}. Z. Zong, “A Formula of the one-leg orbifold Gromov-Witten Vertex and Gromov-Witten invariants of the local $\mathcal{B}{ \mathbb{Z} }_m$ gerbe,” [arXiv:1204.1753]{}. [^1]: We work with both inner and outer branes. See Section \[sec:AV-brane\] for the definition. [^2]: If the $G_{ \mathbb{R} }$-action on ${ {\widetilde{L}} }$ is free, then ${\mathcal{L}}$ is a smooth manifold diffeomorphic to $S^1\times { \mathbb{R} }^2$, so $H_1({\mathcal{L}};{ \mathbb{Z} })={ \mathbb{Z} }$ [^3]: The disk function in [@Ro11 Section 3.3] and our disk factor are the same when $k\neq 0$. When $k=0$, the disk function is $\langle \ \rangle^{{\mathcal{X}},{\mathcal{L}}}_{\ldots}$ (no insertion), while the disk factor is $\langle 1 \rangle^{{\mathcal{X}},{\mathcal{L}}}_{\ldots}$ (one insertion of $1$), so there is an additional factor of $(\frac{s_1}{{\mathsf{w}}_1})^{\delta_{0,k}}$ in the disk function in [@Ro11 Section 3.3]. [^4]: After a change of variables , negative powers may appear.

ArXiv

--- abstract: | In this paper we propose to use elements of the mathematical formalism of Quantum Mechanics to capture the idea that agents’ preferences, in addition to being typically uncertain, can also be *indeterminate*. They are determined (i.e., realized, and not merely revealed) only when the action takes place. An agent is described by a *state*** **that is a superposition of potential types (or preferences or behaviors). This superposed state is projected (or collapses) onto one of the possible behaviors at the time of the interaction. In addition to the main goal of modelling uncertainty of preferences that is not due to lack of information, this formalism seems to be adequate to describe widely observed phenomena of noncommutativity in patterns of behavior. We explore some implications of our approach in a comparison between classical and type indeterminate rational choice behavior. The potential of the approach is illustrated in two examples. JEL: D80, C65, B41 Keywords: indeterminacy, non-commutativity, quantum state. author: - 'Ariane Lambert Mogiliansky,[^1] Shmuel Zamir,[^2] Hervé Zwirn[^3]' title: | Type Indeterminacy:\ A Model of the KT(Kahneman–Tversky)-Man[^4] --- Introduction ============ It has recently been proposed that models of quantum games can be used to study how the extension of classical moves to quantum ones (i.e., complex linear combinations of classical moves) can affect the analysis of a game. For example Eisert et al. (1999) show that allowing the players to use quantum strategies in the Prisoners’ Dilemma is a way of escaping the well-known ‘bad feature’ of this game.[^5] From a game-theoretical point of view the approach consists in changing the strategy spaces, and thus the interest of the results lies in the appeal of these changes.[^6] This paper also proposes to use elements of the mathematical formalism of Quantum Mechanics but with a different intention: to model uncertain preferences.[^7] The basic idea is that the Hilbert space model of Quantum Mechanics** **can be thought of as a very general contextual predictive tool particularly well-suited to describing experiments in psychology or in revealing preferences. The well-established Bayesian approach suggested by Harsanyi to model incomplete information consists of a chance move that selects the types of the players and informs each player of his own type. For the purposes of this paper, we underline the following essential implication of this approach: all uncertainty about a player’s type exclusively reflects the others player’s incomplete knowledge of it. This follows from the fact that a Harsanyi type is fully determined. It is a complete well-defined description of the characteristics of a player that is known to him. Consequently, from the point of view of the other players, uncertainty as to the type can only be due to lack of *information*. Each player has a probability distribution over the type of the other players, but her own type is fully determined and is known to her. This brings us to the first important point at which we depart from the classical approach: we propose that in addition to informational reasons, the uncertainty about preferences is due to *indeterminacy*:* * prior to the moment a player acts, her (behavior) type is indeterminate. The *state* representing the player is a *superposition* of potential types. It is only at the moment when the player selects an action that a specific type is actualized.[^8] It is not merely revealed but rather determined in the sense that prior to the choice, there is an irreducible multiplicity of potential types. Thus we suggest that in modelling a decision situation, we do not assume that the preference characteristics can always be fully known with certainty (neither to the decision-maker nor even to the analyst). Instead, what can be known is the state of the agent: a vector in a Hilbert space which encapsulates all existing information to predict how the agent is expected to behave in different decision situations. This idea, daringly imported from Quantum Mechanics to the context of decision and game theory, is very much in line with Tversky and Simonson (Kahneman and Tversky 2000) according to whom *There is a growing body of evidence that supports an alternative conception according to which preferences are often constructed – not merely revealed – in the elicitation process. These constructions are contingent on the framing of the problem, the method of elicitation, and the context of the choice.* In Ariely, Prelec and Loewenstein (2003), the authors show in a series of experiments that *valuations are initially malleable but become imprinted after the agent is called upon to make an initial decision* (p. 74). This view is also consistent with that of cognitive psychology, which teaches one to distinguish between objective reality and the proximal stimulus to which the observer is exposed, and to further distinguish between those and the mental representation of the situation that the observer eventually constructs. More generally, this view fits in with the observation that players (even highly rational ones) may act differently in game theoretically equivalent situations that differ only in seemingly irrelevant aspects (framing, prior unrelated events, etc.). Our theory as to why agents act differently in game theoretically equivalent situations is that they are not in the same state;(revealed) preferences are contextual because of (intrinsic) indeterminacy. The basic analogy with Physics, which makes it appealing to adopt the mathematical formalism of Quantum Mechanics to the social sciences, is the following: we view decisions and choices as something similar to the result of a *measurement* (of the player’s type). A situation** **of** **decision is then similar to an experimental setup to measure the player’s type. It is modelled by** **an operator (called *observable*), and the resulting behavior is an eigenvalue of that operator. The non-commutativity of observables and its consequences (very central features of Quantum Mechanics) is reminiscent of many empirical phenomena like the following one exhibited in a well-known experiment conducted by Leon Festinger.[^9] In this experiment people were asked to sort a batch of spools into lots of twelve and give a square pegs a quarter turn to the left. They all agreed that the task was very boring. Then, they were told that one subject was missing for the experiment and asked to convince a potential female subject in the waiting room to participate. They were offered \$1 for expressing their enthusiasm for the task.[^10] Some refused, but others accepted. Those who accepted maintained afterwards that the task was enjoyable. This experiment aimed at showing that attitudes change in response to cognitive dissonance. The dissonance faced by those who accepted to fake enthusiasm for \$1 was due to the contradiction between the self-image of being “a good guy” and that of“being ready to lie for a dollar”. Changing one’s attitude and persuading oneself that the task really was interesting was a way to resolve the dissonance. Similar phenomena have been documented in hazardous industries, with employees showing very little caution in the face of a danger. Here too, experimental and empirical studies (e.g., Daniel Ben–Horing 1979) showed how employees change their attitude after they have decided to work in a hazardous industry. More generally, suppose that an agent is presented with the same situation of decision in two different contexts. The contexts may vary with respect to the situation of decision that precedes the investigated one. In Festinger’s experiment the two measurements of the attitude toward the task differ in that the second one was preceded by a question about willingness to lie about the task for a dollar. Two contexts may also vary with respect to the framing of the situation of decision, (cf. Selten 1998)). If we do not observe the same decision in the two contexts, then the classical approach considers that the two situations are not identical and hence that** **they should be modeled so as to incorporate the context. In many cases, however, such an assumption, i.e., that the situations are different, is difficult to justify. In contrast, we propose that the difference between** **the two** **decisions comes from the fact that the agent is not in the same state. The context, e.g., a past situation of decision to which the agent has been exposed, is represented** **by an operator that does not commute with the operator associated with the situation of decision** **currently considered. The consequence is that** **the initial agent’s state has changed and that the agent is therefore expected to behave differently from what she would have done if confronted directly with the situation. As in Quantum Mechanics, the non-commutativity of certain** **situations of** **decision (measurements) leads us to conjecture that the preferences of** **an agent are represented by a state that is indeterminate and gets determined (with respect to any particular type characteristics) in the course of interaction with the environment. In section 3, we show how this approach can explain the reversal of preferences in** **a model of rational choice and that it provides a framework for explaining cognitive dissonance and framing effects. The objective of this paper is to propose a theoretical framework for modelling the KT(Kahneman–Tversky)–man, i.e., for the constructive preference perspective. Our approach amounts to extending the classical representation of uncertainty in Harsanyi’s style to non-classical indeterminacy. This work is a contribution to Behavioral Economics.** **A distinguishing feature of behavioral theories is that they often focus on rather specific anomalies (e.g., ‘trade-off contrast’ or ‘extremeness aversion’ (Kahneman and Tversky 2000). Important insights have been obtained by systematically investigating the consequences on utility maximization of ‘fairness concerns’ (Rabin 1993), ‘temptation and costly self-control’ (Gul and Pesendorfer 2001) or ‘concerns for self-image’ (Benabou and Tirole 2002). Yet, other explanations appeal to bounded rationality, e.g., ‘superficial reasoning’ or ‘choice of beliefs’ (Selten 1998, Akerlof and Dickens 1982). In contrast, the type indeterminacy model is a framework model that addresses structural properties of preferences, i.e., their intrinsic indeterminacy. A value of our approach is in providing a unified explanation for a wide variety of behavioral phenomena. In section 2, we present the framework and some basic notions of quantum theory. In Section 3, we develop applications of the theory to social sciences. In** **Section 4, we discuss some basic assumptions of the model. The appendix provides a brief exposition of some basic concepts of Quantum Mechanics. The basic framework =================== In this section we present the basic notions of our framework. They are heavily inspired by the mathematical formalism of Quantum Mechanics ( see e.g., Cohen-Tannoudji, Diu, Laloë 1973 and Cohen 1989) from which we also borrow the notation.  The notions of state and superposition -------------------------------------- The object of our investigation is the** **individual choice behavior, which we interpret as the revelation of an agent’s preferences in a situation of decision that we shall call a** ***DS*** (**Decision Situation**).** In this paper, we focus on decision situations that do not involve any strategic thinking. Examples of such *DS* include the choice between buying a Toshiba or a Compaq laptop, the choice between investing in a project or not, the choice between a sure gain of \$100 or a bet with probability 0.5 to win \$250 and 0.5 to win \$0, etc. When considering games, we view them as decision situations from the perspective of a single player who plays once.[^11] An agent is represented by a state which encapsulates all the** **information on the agent’s expected choice behavior. The formalism that we present* *below allows for a variety of underlying models. In particular, some states may correspond to a choice. For instance if the choice set is $\left\{ a,b,c\right\} ,$ we may identify three possible states corresponding to the respective choices *a*, *b* and *c*. A state may also** **correspond to an ordering of the 3 items in which case we have six possible states (e.g. a state could be (a,b,c) in this order). We may also consider other models. In section 3.1 we examine consequences of the second case. For the ease of exposition, the basic framework is presented in terms of the first case where states are identified with choices (but the reader is invited to keep in mind that other interpretations of the model are possible**)**. However what we just wrote must be considered in the light of the principle of superposition as explained below.** ** Mathematically, a state $|\psi \rangle $ is a vector in a Hilbert space $% \mathcal{H}$ of finite or countably infinite dimensions, over the field of the real numbers $\mathbb{R}$.[^12] The relationship between $\mathcal{H}$ and a decision situation will be specified later. For technical reasons related to the probabilistic content of the state, each state vector has to be of length one, that is, $\langle \psi \left\vert \psi \right\rangle ^{2}=1\ $(where $% \langle \cdot \left\vert \cdot \right\rangle $ denotes the inner product in $% \mathcal{H)}$. So all vectors of the form $\lambda \left\vert \psi \right\rangle $ where $\lambda \in \mathbb{R},$ correspond to the same state, which we represent by a vector of length one. A key ingredient in the formalism of indeterminacy is *the principle of superposition.* This principle states that the linear combination of any two states$\;$is itself a possible state.[^13] Consider two states $% \left\vert \varphi _{1}\right\rangle ,\left\vert \varphi _{2}\right\rangle \in \mathcal{H}$. If** **$\left\vert \psi \right\rangle =\lambda _{1}\left\vert \varphi _{1}\right\rangle +\lambda _{2}\left\vert \varphi _{2}\right\rangle $ with $\lambda _{1},\lambda _{2}\in \mathbb{R}\ $then** **$\left\vert \psi \right\rangle \in \mathcal{H}.$ The principle of superposition implies that, unlike the Harsanyi type space, the state space is non-Boolean.[^14] The notions of measurement* *and of observable ---------------------------------------------- Measurement is a central notion in our framework**.** A measurement is an operation (or an** **experiment) performed on a system. It yields a result, the outcome of the measurement. A defining feature of a measurement is the so-called first-kindness property.[^15] It refers to the fact that if one performs a measurement on a system and obtain**s** a result, then one will get the same result if one** **performs again that measurement on the same system immediately after. Thus, the outcome of a first-kind measurement is reproducible but only in a next subsequent measurement. First-kindness does not entail that the first outcome is obtained when repeating a measurement if other measurements are performed on the system in between. Whether an operation is a measurement, i.e., is endowed with the property of first-kindness is an empirical issue. When it comes to decision theory, it means that we do not, a priory, assume that any choice set can be used to measure preferences. In particular, the product set of two sets each of which is associated with a first-kind measurement, is not in general associated with a first-kind measurement.** **The set of decision problems that we consider consists exclusively of decision problems that can be associated with first-kind measurements. We call those decision problems Decision Situations (*DS*).[^16] A Decision Situation $A$ can be thought of as an experimental setup where the agent is invited to choose a particular action among all the** **possible actions** **allowed by this Decision Situation. In this paper, we will consider only the case of finitely many possible outcomes. They will be labelled from 1 to $n$ by convention. When an agent selects an action**,** we say that she ‘plays’ the *DS*. To every Decision Situation $A$, we will** **associate an *observable,* namely, a specific symmetric operator on $\mathcal{H}$ which, for notational simplicity, we also denote by $A$.[^17] If we consider** **only one** **Decision Situation $% A $ with $n$ possible outcomes, we can assume that the associated Hilbert space is $n$-dimensional and** **that the eigenvectors of the corresponding observable, which we denote by $|1_{A}\rangle ,|2_{A}\rangle ,...,|n_{A}\rangle ,$ all correspond to different eigenvalues, denoted by $% 1_{A},2_{A},...,n_{A}$ respectively. By convention, the eigenvalue $i_{A}$ will be associated with choice $i.$** ** $$A\left\vert k_{A}\right\rangle =k_{A}\left\vert k_{A}\right\rangle ,\;\;k=1,...,n.$$ As $A\;$is symmetric, there is a unique orthonormal basis of the relevant Hilbert space $\mathcal{H\;}$formed with its eigenvectors. The basis $% \left\{ |1_{A}\rangle ,|2_{A}\rangle ,...,|n_{A}\rangle \right\} $ is the unique orthonormal basis of $\mathcal{H}$ consisting of eigenvectors of $A.$ It is thus possible to represent the agent’s state as a superposition of vectors of this basis: $$\left| \psi \right\rangle \ =\underset{k=1}{\overset{n}{\sum }}\lambda _{k}\left| k_{A}\right\rangle , \label{exp}$$ where $\lambda _{k}\in \mathcal{\mathbb{R}},\forall k\in \{1,...,n\}$ and $% \underset{k=1}{\overset{n}{\sum }}\lambda _{k}^{2}=1$. The Hilbert space can be decomposed as follows $$\mathcal{H=H}_{1_{A}}\mathcal{\oplus },...,\mathcal{\oplus H}_{n_{A}},\;% \mathcal{H}_{i_{A}}\perp \mathcal{H}_{j_{A}},\;i\neq j, \label{exp2}$$where $\oplus $ denotes the direct sum of the subspaces $\mathcal{H}% _{1_{A}},...,\mathcal{H}_{n_{A}}$ spanned by $|1_{A}\rangle ,...,|n_{A}\rangle $ respectively.[^18] Or, equivalently, we can write $I_{\mathcal{H}}=\;P_{1_{A}}+,...,+P_{n_{A}}$ where $P_{i_{A}}$ is the projection operator on $\mathcal{H}_{i_{A}}$ and $I_{\mathcal{H}}$ is the identity operator on $\mathcal{H}.$ A Decision Situation $A$ is an experimental setup and the actual implementation of the experiment is represented by a *measurement* of the associated observable $A$. According to the so-called Reduction Principle* *(see Appendix), the result of such a measurement can only be one of the $n$ eigenvalues of $A$. If the result is $m_{A},$ i.e., the player selects action $m,$ the superposition $\sum \lambda _{i}\left\vert i_{A}\right\rangle $ collapses onto the eigenvector associated with the eigenvalue $m_{A}.$ The initial state $% \left\vert \psi \right\rangle $ is projected into the subspace $\mathcal{H}% _{m_{A}}$(of eigenvectors of $A$ with eigenvalue $m_{A}).$ The probability that the measurement yields the result $m_{A}$ is equal to $\langle m_{A}\left\vert \psi \right\rangle ^{2}=\lambda _{m}^{2},$ i.e., the square of the corresponding coefficient in the superposition. The coefficients themselves are called ‘amplitudes of probability’. They play a key role when studying sequences of measurements (see Section 2.3). As usual, we interpret the probability of $m_{A}$ either as the probability that one agent in state $\left\vert \psi \right\rangle $ selects action $m_{A}$ or as the proportion of the agents who will make the choice $m_{A}$ in a population of many agents, all in the state $\left\vert \psi \right\rangle $.   In our theory an agent is represented by a state. We shall also use the term type (and eigentype) to denote a state corresponding** **to one eigenvector, say $\left\vert m_{A}\right\rangle .\;$An agent in this state is said to be of type $m_{A}.$ An agent in a general state $\left\vert \psi \right\rangle $ can be expressed as a superposition of all eigentypes of the *DS* under consideration.$\ $Our notion of type is closely related to the notion introduced by Harsanyi. Consider a simple choice situation e.g., when an employee faces a menu of contracts. The type captures all the agent’s characteristics (taste, subjective beliefs, private information) of relevance for uniquely predicting the agent’s behavior. In contrast to Harsanyi, we shall not assume that there exists an exhaustive description of the agent that enables us to determine the agent’s choice uniquely and *simultaneously* in all possible Decision Situations . Instead, our types are characterized by an irreducible uncertainty that is revealed when the agent is confronted with a sequence of *DS* (see Section 2.3.2 below for a formal characterization of irreducible uncertainty).   *Remark:* Clearly, when only one *DS* is considered, the above description is equivalent to the traditional probabilistic representation of an agent by a probability vector $(\alpha _{1,....,}\alpha _{n})\;$in which $% \alpha _{k}$ is the probability that the agent will choose action $k_{A}$ and $\alpha _{k}=\lambda _{k}^{2}$ for $k=1,...,n.$ The advantage of the proposed formalism consists in enabling us to study several decision situations and the interaction between them. More than one Decision Situation -------------------------------- When studying more than one *DS*, say $A$ and $B$, the key question is whether the corresponding observables are commuting operators in $% \mathcal{H},$ i.e., whether $AB=BA.$ Whether two *DS* can be represented by two commuting operators or not is an empirical issue. We next study its mathematical implications. ### Commuting Decision Situations Let $A$ and $B$ be two *DS*. If the corresponding observables commute then there is an orthonormal basis of the relevant Hilbert space $\mathcal{H} $ formed by eigenvectors common to both $A$ and $B$. Denote by $\left\vert i\right\rangle $ (for $i=1,...,n)\;$these basis vectors. We have $$A\left\vert i\right\rangle =i_{A}\left\vert i\right\rangle \;\text{and }% B\left\vert i\right\rangle =i_{B}\left\vert i\right\rangle .$$ In general, the eigenvalues can be degenerated (i.e., for some $i$ and $j,$ $_{{}}i_{A}=j_{A}$ or $i_{B}=j_{B}).$[^19] Any normalized vector $\left\vert \psi \right\rangle $ of $% \mathcal{H}$ can be written in this basis: $$\left\vert \psi \right\rangle =\sum_{i}\lambda _{i}\left\vert i\right\rangle ,$$ where $\lambda _{i}\in \mathcal{\mathbb{R}}$, and $\sum_{i}\lambda _{i}^{2}=1.\;$If we measure $A$ first, we observe eigenvalue $i_{A}$ with probability $$p_{A}\left( i_{A}\right) =\sum_{j;j_{A}=i_{A}}\lambda _{j}^{2}. \label{CO1}$$  If we measure $B$ first, we observe eigenvalue $j_{B}$ with probability $% p_{B}\left( j_{B}\right) =\sum_{k;k_{B}=j_{B}}\lambda _{k}^{2}.\;$After $B$ is measured and the result $j_{B}$ is obtained$,$ the state $\left\vert \psi \right\rangle $ is projected into the eigensubspace $\mathcal{E}_{j_{B}}$ spanned by the eigenvectors of $B$ associated with $j_{B}.\;$More specifically, it collapses onto the state: $$\left\vert \psi _{j_{B}}\right\rangle =\frac{1}{\sqrt{\sum_{k;k_{B}=j_{B}}% \lambda _{k}^{2}}}\sum_{k;k_{B}=j_{B}}\lambda _{k}^{{}}\left\vert k\right\rangle$$ (the factor $\frac{1}{\sqrt{\sum_{k;k_{B}=j_{B}}\lambda _{k}^{2}}}$ is necessary to make $\left\vert \psi _{j_{B}}\right\rangle $ a unit vector). When we measure $A$ on the agent in the state $\left\vert \psi _{j_{B}}\right\rangle ,$ we obtain $i_{A}$ with probability $$p_{A}\left( i_{A}|j_{B}\right) =\frac{1}{\sum_{k;k_{B}=j_{B}}\lambda _{k}^{2}% }\sum_{\substack{ k;k_{B}=j_{B} \\ \text{and\ }k_{A}=i_{A}}}\lambda _{k}^{2}.$$ So when we measure first $B$ and then $A,$ the probability of observing the eigenvalue $i_{A}$ is $p_{AB}\left( i_{A}\right) =\sum_{j}p_{B}\left( j_{B}\right) p_{A}\left( i_{A}|j_{B}\right) $: $$\begin{array}{lll} p_{AB}\left( i_{A}\right) & = & \sum_{j}\frac{1}{\sum_{k;k_{B}=j_{B}}\left% \vert \lambda _{k}\right\vert ^{2}}\sum_{k;k_{B}=j_{B}}\lambda _{k}^{2}\sum _{\substack{ _{_{\substack{ l;l_{B}=j_{B}\; \\ \text{and }l_{A}=i_{A}}}} \\ }}\lambda _{l}^{2} \\ & = & \sum_{j}\sum_{\substack{ _{_{\substack{ l;l_{B}=j_{B}\; \\ \text{and\ }l_{A}=i_{A}}}} \\ }}\lambda _{l}^{2}=\sum_{\substack{ _{_{l;l_{A}=j_{A}% \;}} \\ }}\lambda _{l}^{2}.% \end{array}%$$ Hence, $p_{AB}\left( i_{A}\right) =p_{A}\left( i_{A}\right) ,$ $\forall i,\;$and similarly $p_{BA}\left( j_{B}\right) =p_{B}\left( j_{B}\right) ,\forall j.\;$ When dealing with commuting observables it is meaningful to speak of measuring them simultaneously. Whether we measure first $A$ and then $B$ or first $B$ and then $A$, the probability distribution on the joint outcome is $p\left( i_{A}\wedge \text{ }j_{B}\right) =\sum_{_{\substack{ k;k_{B}=j_{B} \\ \text{and}\;k_{A}=i_{A}}}}\lambda _{k}^{2},$ so$\;\left( i_{A\text{ }}% \text{, }j_{B}\right) $ is a well-defined event. Formally, this implies that the two *DS* can be merged into a single *DS.* When we measure it, we obtain a vector as the outcome, i.e., a value in $A\ $and a value in $B$. To each eigenvalue of the merged observable we associate a type that captures all the characteristics of the agent relevant to her choices (one in each *DS*). *Remark:* Note that, as in the case of a single *DS*, for two such commuting *DS* our model is equivalent to a standard (discrete) probability space in which the elementary events are $\left\{ \left( i_{A},j_{B}\right) \right\} $ and$\;p\left( i_{A}\wedge \text{ }j_{B}\right) =\sum_{\substack{ k;k_{B}=j_{B} \\ \text{and}\;k_{A}=i_{A}}}\lambda _{k}^{2}. $ In particular, in accordance with the calculus of probability we see that the conditional probability formula holds: $$p_{AB}(i_{A}\wedge j_{B})=p_{A}(i_{A})\ p_{B}(j_{B}|i_{A}).$$This also means that the type space associated with type characteristics represented by commuting observables is equivalent to the Harsanyi type space. In particular, if the Decision Situations $A$ and $B$ together provide a full characterization of the agent, then all types $i_{A}j_{B}$are mutually exclusive: knowing that the agent is of type $1_{A}2_{B}$ it is sure that she is *not* of type $i_{A}j_{B}$ for $i\neq 1$ and/or $j\neq 2$.   As an example, consider the following two Decision Situations. Let $A$ be the Decision Situation presenting a choice between a week vacation in Tunisia and a week vacation in Italy. And let $B\;$be the choice between buying 1000 euros of shares of** **Bouygues Telecom or of Deutsche Telecom. It is quite plausible that $A$ and $B$ commute, but whether or not this is the case is of course an empirical question. If $A$ and $B$ commute, we expect that** **a decision on portfolio ($B$) will** **not affect the decision-making concerning the vacation ($A)$. And thus the order in which the decisions are made does not matter as in the classical model.   Note finally that the commutativity of the observables does not exclude statistical correlations between observations. To see this, consider the following example in which $A$ and $B\;$each have two degenerated eigenvalues in a four dimensional Hilbert space. Denote by $\left\vert i_{A}j_{B}\right\rangle \;(i=1,2,\;j=1,2)\;$the eigenvector associated with eigenvalues $i_{A}$ of $A$ and $j_{B}$ of $B,\;$and let the state $% \left\vert \psi \right\rangle $ be given by $\left| \psi \right\rangle =\sqrt{\frac{3}{8}}\left| 1_{A}1_{B}\right\rangle +\sqrt{\frac{1}{8}}\left| 1_{A}2_{B}\right\rangle +\sqrt{\frac{1}{8}}\left| 2_{A}1_{B}\right\rangle +\sqrt{\frac{3}{8}}\left| 2_{A}2_{B}\right\rangle $ Then,$\ \ p_{A}\left( 1_{A}|1_{B}\right) =\frac{\frac{3}{8}}{\frac{3}{8}+% \frac{1}{8}}=\frac{3}{4},\ \ p_{A}\left( 2_{A}|1_{B}\right) =\frac{\frac{1}{8% }}{\frac{3}{8}+\frac{1}{8}}=\frac{1}{4}.$ So if we first measure $B$ and find, say, $1_{B},$ it is much more likely (with probability $\frac{3}{4}$) that when measuring $A$ we will find $1_{A}$ rather than $2_{A}\;$(with probability $\frac{1}{4}$)$.$ But the two *interactions* (measurements) do not affect each other, i.e., the distribution of the outcomes of the measurement of $A\;$is the same whether or not we measure $B$ first. ### Non-commuting Decision Situations It is when we consider Decision Situations associated with observables that do not commute that the predictions of our model differ from those of the probabilistic one. In such a context, the quantum probability calculus ($% p(i_{A}\left\vert \psi \right. )=\left\langle i_{A}\right. \left\vert \psi \right\rangle ^{2}$)$\;$generates cross-terms also called interference terms. These cross-terms are the signature of indeterminacy. In the next section, we demonstrate how this feature can capture the phenomenon of cognitive dissonance as well as that of framing.   Consider two Decision Situations $A$ and $B$ with the same number $n$ of possible choices. We shall assume for simplicity that the corresponding observables $A$ and $B$ have non-degenerated eigenvalues $1_{A},2_{A},...,$ $n_{A}$ and $1_{B},2_{B},...,$ $n_{B}$ respectively. Each set of eigenvectors $\left\{ |1_{A}\rangle ,|2_{A}\rangle ,...,|n_{A}\rangle \right\} $ and $\left\{ |1_{B}\rangle ,|2_{B}\rangle ,...,|n_{B}\rangle \right\} $ is an orthonormal basis of the relevant Hilbert space. Let $% \left\vert \psi \right\rangle $ be the initial state of the agent $$\left\vert \psi \right\rangle \ =\underset{i=1}{\overset{n}{\sum }}\lambda _{i}|i_{A}\rangle =\overset{n}{\underset{j=1}{\sum }}\nu _{j}\left\vert j_{B}\right\rangle . \label{statevector}$$ We note that each set of eigenvectors of the respective observables forms a basis of the state space. The multiplicity of alternative basis is a distinguishing feature of this formalism. It implies that there is no single or privileged way to describe (express) the type of the agent. Instead, there exists a multiplicity of equally informative alternative ways to characterize the agent.  We shall now compute the probability for type $i_{A}$ under two different scenarios. In the first scenario we measure $A$ on the agent in state $% \left\vert \psi \right\rangle .$ In the second scenario we first measure $B$ on the agent in state $\left\vert \psi \right\rangle $ and thereafter measure $A$ on the agent in the state resulting from the first measurement. We can write observable $B^{\prime }s$ eigenvector $% |j_{B}\rangle \ $in the basis made of $A^{\prime }$s eigenvectors: $$|j_{B}\rangle =\underset{i=1}{\overset{n}{\sum }}\mu _{ij}|i_{A}\rangle . \label{Bvector}$$Using the expression above we write the last term in equation ([statevector]{}) $$\left\vert \psi \right\rangle \ =\overset{n}{\underset{j=1}{\sum }}\underset{% i=1}{\overset{n}{\sum }}\nu _{j}\mu _{ij}|i_{A}\rangle . \label{state2}$$From expression (\[state2\]), we derive the probability $p_{A}(i_{A})\ $that the agent chooses $i_{A}$ $\ $in the first scenario:** **$% p_{A}(i_{A})=\left( \underset{j=1}{\overset{n}{\sum }}\nu _{j}\mu _{ij}^{{}}\right) ^{2}$**.** In the second scenario, she first plays$% \;B.$ By (\[statevector\]) we see that she selects action $j_{B}$ with probability $\nu _{j}^{2}.$ The state $\left\vert \psi \right\rangle \ $is then projected onto $|j_{B}\rangle .$ When the state is $|j_{B}\rangle ,\ $the probability for $i_{A}$ is given by (\[Bvector\]), it is $\mu _{ij}^{2}.$ Summing up the conditionals, we obtain the (ex-ante) probability for $i_{A}$ when the agent first plays $B$ and then $A:$ $% p_{AB}(i_{A})=$ $\overset{n}{\underset{j=1}{\sum }}\nu _{j}^{2}\mu _{ij}^{2},$ which is, in general, different from $p_{A}(i_{A})=\left( \overset{n}{\underset{j=1}{\sum }}\nu _{j}\mu _{ij}\right) ^{2}.$ Playing first $B$ changes the way $A$ is played. The difference stems from the so-called interference terms $$\begin{aligned} p_{A}(i_{A}) &=&\left( \overset{n}{\underset{j=1}{\sum }}\nu _{j}\mu _{ij}\right) ^{2}=\overset{n}{\underset{j=1}{\sum }}\nu _{j}^{2}\mu _{ij}^{2}+\underset{\text{Interference term}}{\underbrace{2\underset{j\neq j^{\prime }}{\sum }\left[ \left( \nu _{j^{\prime }}\mu _{ij}\right) \left( \nu _{j}^{{}}\mu _{ij^{\prime }}^{{}}\right) \right] \qquad }} \\ &=&p_{AB}(i_{A})+\text{interference term}\end{aligned}$$ The interference term is the sum of cross-terms involving the amplitudes of probability (the Appendix provides a description of interference effects in Physics)**.** Some intuition about interference effects may be provided using the concept of “propensity” due to Popper (1992). Imagine an agent’s mind as a system of propensities to act (corresponding to different possible actions). As long as the agent is not required to choose an action in a given *DS*, the corresponding propensities coexist in her mind; the agent has not made up her mind. A decision situation operates on this state of “hesitation” to trigger the emergence of a single type (of behavior). But as long as alternative propensities are present in the agent’s mind, they affect choice behavior by increasing or decreasing the probability of the choices in the *DS* under investigation.   An illustration of this kind of situation may be supplied by the experiment reported in Knetz and Camerer (2000). The two studied *DS* are the Weak Link (WL) game and the Prisoners’ Dilemma (PD) game.[^20] They compare the distribution of choices in the Prisoners’ Dilemma (PD) game when it is preceded by a Weak Link (WL) game and when only the PD game is being played. Their results show that playing the WL game affects the play of individuals in the PD game. The authors appeal to an informational argument, which they call the precedent effect.[^21] However, they cannot explain the high rate of cooperation (37.5 %) in the last round of the PD game (Table 5, p. 206). Instead, we propose that the WL and the PD are two *DS* that do not commute. In such a case we expect a difference in the distributions of choices in the (last round of the) PD depending on whether or not it was preceded by a play of the WL or another PD game. This is because the type of the agent is being modified by the play of the WL game.     *Remark:* In the case where** ** $A$ and $B$ do not commute, they cannot have simultaneously defined values: the state of the agent is characterized by an** ***irreducible uncertainty*. Therefore, and in contrast with the commuting case, two non-commuting observables cannot be merged into one single observable. There is no probability distribution on the events of the type "to have the value $i_{A}$ for $A$ and the value $j_{B}$ for $B"$. The conditional probability formula does not hold: $$p_{A}(i_{A})\neq \overset{n}{\underset{j=1}{\sum }}p_{B}(j_{B})\ p(i_{A}|j_{B}).$$ Indeed, $\left( \overset{n}{\underset{j=1}{\sum }}\nu _{j}\mu _{ij}\right) ^{2}=p_{A}(i_{A})\neq \overset{n}{\underset{j=1}{\sum }}p_{B}(j_{B})\ p(i_{A}|j_{B})=\overset{n}{\underset{j=1}{\sum }}\nu _{j}^{2}\mu _{ij}^{2}.\ $ Consequently, the choice experiment consisting of asking the agent to select a pair $\left( i_{A},j_{B}\right) $ out the set of alternatives $% A\times B$ is NOT a *DS.* We must here acknowledge a fundamental distinction between the type space of our TI-model (Type Indeterminacy model) and that of Harsanyi. In the Harsanyi type space the (pure) types are all mutually exclusive: the agent is either of type $\theta _{i}$ *or* of type $\theta _{j}$ but she cannot be both!$\ $In the TI-model this is not always the case. When dealing with (complete) non-commuting $\mathit{DS}$[^22], the types associated with respective $% \mathit{DS}\ $are not mutually exclusive: knowing that the agent is of type $% 1_{A},$ which is a full description of her type, it cannot be said that she is *not* of type $1_{B}.\ $The eigentypes of non-commuting *DS* are connected in the sense that the agent can transit from one type to another under the impact of a measurement. When making a measurement of $B$ on the agent of type $1_{A},$ she is projected onto one of the eigenvectors of $\mathit{DS}\ B.$ Her type changes from being $1_{A}$ to being some $j_{B}.$  The Type Indeterminacy model and Social Sciences ================================================ The theory of choice exposed in this paper does not allow for a straightforward comparison with standard choice theory. Significant further elaboration is required. Yet, some implications of the type indeterminacy approach can be explored. First, we shall be interested in comparing the behavior of an agent of indeterminate type with the one of a classical agent in the case where both satisfy the Weak Axiom of Revealed Preference (WARP)[^23]** **which is a basic axiom of rational choice. Then, we will show that this framework can be used to explain two instances of behavioral anomalies that have been extensively studied in the literature. The TI-model and the classical rational man ------------------------------------------- In standard decision theory, it is assumed that an individual has preferences (i.e. a complete ranking or a complete linear ordering) on the universal set of alternatives $X.\ $The individual knows her preferences while the outsiders may not know them. But it is also possible that the individual only knows what she would choose from some limited sets of alternatives and not from the whole set $X$. Thus, a less demanding point of view consists in representing the choice behavior by a choice structure (i.e. a family $\mathcal{B}$ of subsets of the universal set of alternatives $X$  and a choice rule $C$ that assigns a non empty set of chosen elements $% C(A)$ for all $A\in \mathcal{B}$). The link between the two points of view is well-known. From preferences, it is always possible to build a choice structure but the reverse is not always true. For it to be true, the choice structure must display a certain amount of consistency (satisfying the Weak Axiom of Revealed Preference) and the family $\mathcal{B}$ must include all subsets of $X$ of up to three elements[^24]. So, preferences can be revealed by asking the individual to make several choices from subsets of $X.\ $How does this simple scheme changes when we are dealing with an individual whose type is indeterminate? Consider a situation where an individual is invited to make a choice of one item out of a set $A$, $A\subseteq X.$ If this experiment satisfies the first-kindness property, we can consider it to be a measurement represented by an observable. The set of possible outcomes of this experiment is the set $A.$$\ $We also denote the observable by $A$.[^25]   We make two key assumptions on the individual choice behavior: A1. Choices out of a small subsetsatisfy the first-kindness property (the meaning of small will be made precise). A2. Choices out of a small subset respect our Weak Axiom of Revealed Preference (WARP’, see below).   Assumption A1 means that small subsets are associated with *DS,* i.e., the experiment consisting in letting an individual choose an item out of a small subset of item is a measurement. Assumption A2 means that choices from small subsets are rational. The idea behind these assumptions is that an individual can, in her mind, structure any small set of alternatives, i.e., simultaneously compare those alternatives. She may not be able to do that within a big set though. But** **this does not mean that our individual cannot make a choice from a big set. For example, she might use an appropriate sequence of binary comparisons and select the last winning alternative. However, such a compound operator would not in general satisfy the first-kindness property i.e., there may not exist any *DS* representing such an operation. A standard formulation of the WARP can be found in Mas-Colell et al. (1995).[^26] We shall use a stronger version by assuming that $C\left( B\right) ,\ $for any $B\in \mathcal{B},\mathcal{\ }$is a singleton.[^27] For our purpose, it is also useful to formulate the axiom in two parts: Consider two subsets $B,$ $B^{\prime }\in \mathcal{B}\ $such that $B\subset B^{\prime }.$ (a) Let $x,y\in B,\ x\neq y$ with $x\in C\left( B\right) $ then $y\notin C\left( B^{\prime }\right) .\ $ \(b) Let $C\left( B^{\prime }\right) \cap B\neq \varnothing $ then $C\left( B\right) =C\left( B^{\prime }\right) .\medskip $ The intuition for (a) is that as we enlarge the set of items from $B$ to $% B^{\prime }$ a rational decision-maker never chooses from $B^{\prime }$ an item that is available in $B$ but is not chosen. The intuition for (b) is that as we reduce the set of alternatives an item chosen in the large set, is also chosen in the smaller set containing that item and no item previously not chosen becomes chosen.[^28] It can be shown that in the classical context $(b)$ implies $(a)\ $(see Arrow (1959))$.$ In our context where choice experiments can be non-commuting, we do not have such an implication$.$ Moreover, the notion of choice function is not appropriate because it implicitly assumes the commutativity of choice** **(see below).** **Since our purpose is to investigate this issue explicitly, we express the axiom more immediately in terms of (observable) choice behavior in the following way and we call it WARP’: Consider two subsets$\ B,$ $B^{\prime }\in \mathcal{B}$ with $B\subset B^{\prime }$. 1\) Suppose the agent chooses from $B$ some element $x$. In a next subsequent measurement of $B^{\prime }$ the outcome of the choice is not in $B\setminus \{x\}$. 2\) Suppose the agent chooses from $B^{\prime }$ some element $x$ that belongs to $B$. In a next subsequent measurement of $B$ the agent’s choice is $x$. Points (1) and (2) capture the classical intuitions about rational choice behavior associated with (a) and (b) above. A distinction with the standard formulations of WARP is that we do not refer to a choice function but to choice behavior and that our axiom only applies to two subsequent choices. We shall see below that in the classical context where choices commute WARP’ is equivalent to WARP. That is not longer true if we allow for some choices not to commute.   We now investigate the choice behavior of a type indeterminate agent under assumptions A1 and A2 above i.e., under the assumption that choices from small subsets satisfy the property of first-kindness and that they respect WARP’. We consider in turn two cases**.** In the first one, we assume that all the *DS* considered pairwise** **commute. In the second one, we allow for non-commuting *DS*. In the following we define small subsets as subsets of size 3 or less.   *Case 1* The assumption here is that all experiments of choice from small subsets are compatible with each other, i.e., the corresponding observables commute. A first implication of the commutativity of choice experiments is that for any $B\subseteq X,$ the outcome of the choice experiment $B$ only depends on $B$ (whatever was done before the choice in $B$ will always be the same and in particular the outcome does not depend on choices experiments that were performed before)$.$ Therefore, we can express the outcome in terms of *a choice function* $C\left( X\right) :B\rightarrow x$ that associates to each $B\in \mathcal{B}\ $ a chosen item $x\in B.\ $A second consequence of commutativity is that the restriction of WARP’ to two subsequent choices is inconsequential. This is because the outcome of any given choice experiment must be the same in any series of two consecutive experiments. In particular, WARP’ implies the transitivity of choices. This can easily be seen we taking an example on a subset of 3 items and performing choice experiments on pairs in different orders.[^29]**.** So for the case choice experiments commute, WARP’ is equivalent to WARP and we recover the standard results. We know that if $\left( \mathcal{B},\mathcal{\ }C\left( .\right) \right) \ $is a choice structure satisfying WARP’ and defined for $\mathcal{B\ }$including all subsets of $X\ $of up to three elements, then there exists a rational preference relation that rationalizes choice behavior.[^30] It is therefore natural to identify the states (types) of the individual with the preference orders that rank all the elements of $X.\ $The type space $\mathcal{H}$ has dimension $\left\vert X\right\vert !.$ We obtain the well-known classical model.   In this model all *DS* are coarse measurements of the type (the preference order) i.e., their outcomes are degenerated eigenvalues (see Section 2.3.1)$.\ $When all$\ $*DS*$\ $commute and satisfy WARP’, the type indeterminate rational agent behaves in all respects as a classical rational agent. This should not surprise us since we know that when observables commute they can be merged into a single observable and the TI-model is equivalent to the classical model.   *Case 2* We now consider a case where some *DS* do not commute. We need to emphasize that, in contrast with the commuting case, a large amount of non-commuting models are possible. The models differ from each other according to which *DS* commute and which do not. We investigate here a simple example that illustrates interesting issues. Assume the set $X$ consists of four items: $a,b,c$ and $d$. As before, any subset $A\subset X$ consisting of 3 elements or less is associated with a *DS* and we assume that any two consecutive choices made from small subsets respect WARP’. Our non-commuting model is defined by the following assumption: *(nc)* Choices out of $A$ and $B$ commute *if and only if* $% A\cup B\neq X.$ In particular *DS* associated with different triples e.g., a choice out of $\left\{ a,b,c\right\} $ and a choice out of $\left\{ b,c,d\right\} \ $are represented by observables that *do not* commute with each other. So for instance the agent may choose $a$ out of $\left\{ a,b,c\right\} ,$ $b$ out of $\left\{ b,c,d\right\} $ and thereafter $c$out of $\left\{ a,c\right\} .$ This may happened because the order (partially) revealed by the choice in $\left\{ a,b,c\right\} $ has been modified by the performance of the $\left\{ b,c,d\right\} $ choice experiment. $\ $This choice behavior satisfies WARP’ but violates transitivity$.\ $Nevertheless, the satisfaction of WARP’ induces a certain amount of consistency in behavior. In particular, any observable $A\;$with$\ A\subset B$ where $B$ is a triple, commutes with$\ B\ ($since$\ A\cup B\subset X)$. We may perform choice experiments in pairs and in triples so as to elicit a preference relation on each triple *taken separately*. It is therefore natural to identify the types with preferences orders on triples. Since observables representing choice experiments on different triples do not commute with each other, they are *alternative* ways to measure the individual’s preferences. The type space representing the individual is a six dimensional Hilbert space because there are six ways to rank three items. There are four alternative bases spanned by the eigenvectors corresponding to the six possible rankings in each one of the four triples.   In this model the agent only has preferences over triples of items. One may wonder what happens when she must make a choice out of the set $\left\{ a,b,c,d\right\} .$ Many scenarios are possible. Suppose that the individual proceeds by making a series of two measurements. For instance, she first selects from a pair followed by a choice from a triple consisting of the first selected item and the two remaining ones. We can call such a behavior “procedural rational” because she acts *as if* she had preferences over the four items. As an example consider the following line of events and for the ease of presentation we let the *DS* associated with the set $\left\{ a,b,c\right\} $ be denoted $abc$ and similarly for the other $\mathit{DS}.$ Suppose the individual just made a choice in $\left\{ a,b,c\right\} $ and picked up $c.\ $So her initial state is some superposition of type $[c>b>a]$ and type $[c>a>b].$ We now ask her to choose from $\left\{ a,b,c,d\right\} .$ Assume the individual first plays $ab$ then $bcd$ (which by assumption of procedural rationality means that the outcome of $ab\ $is $b)$.[^31]$\ $The choice of $b$ from $\left\{ a,b\right\} \ $changed the type of the individual. In particular, suppose the $ab$ observable is considered as a coarse measurement of $abd\mathit{.\ }$The new type, expressed in the basis of preference orders on $\left\{ a,b,d\right\} \mathit{,}$ is a superposition of $\ $types$\ \left[ b>a>d% \right] ,\ \left[ b>d>a\right] \ $and $\left[ d>b>a\right] .\ $In order to make her choice out of $\left\{ a,b,c,d\right\} ,$ she now plays$\ bcd.$ The result of that last measurement (performed on the individual of the type resulting from the first measurement) is therefore $b$ with positive probability. But this violates the principle of independence of irrelevant alternative (IIA)$.\ $She effectively selects $b$ from $\left\{ a,b,c,d\right\} \ $while she initially picked up $c$ in $\left\{ a,b,c\right\} $ where $b$ was available.** **Yet, it is easy to check that in this example any two consecutive choices satisfy the WARP’ so the choice behavior of our type indeterminate agent is “rational”.     We can now draw some first conclusions from this short exploration. If we demand that the agent’s choice behavior respects the Weak Axiom of Revealed Preference (in our WARP’ formulation) and conforms to procedural rationality, then i\. when all the *DS* associated with subsets of the universal set commute, a type indeterminate agent does not distinguish herself from a classical agent; ii\. when some *DS* associated with subsets of the universal set do not commute with each other, a type indeterminate agent does *not* have a preference order over the universal set $X.$ But she may have well-behaved preferences over subsets of $X.\ $Generally, she does not behave as a classical rational agent.* *In particular, her behavior may exhibit standard phenomena of preference reversal.   In this example the distinction between the classical rational and the type indeterminate rational agent is only due to the non-commutativity of *DS *associated with different subsets of items. This distinctive feature of the TI-model of choice delivers a formulation of bounded rationality in terms of the impossibility to compare and order all items *simultaneously*. Non-commutativity also implies that as the agent makes a choice her type (preferences) changes. A type indeterminate agent does not have a fixed type (preferences). Instead her preferences are shaped in the process of elicitation as proposed by Kahneman and Tversky (2000). Examples -------- In this section we demonstrate how type indeterminacy can explain two well-documented examples of so-called behavioral anomalies. With these examples we want to suggest that one contribution of our approach is to provide a unified framework which can accommodate a variety of behavioral anomalies. These anomalies are currently explained by widely different theories. ### Cognitive dissonance The kind of phenomena we have in mind can be illustrated as follows**.** Numerous studies show that employees in risky industry (like nuclear plants) often neglect safety regulations. The puzzle is that before moving into the risky industry those employees were typically able to estimate the** **risks correctly. They were reasonably averse to risk and otherwise behaved in an ordinary rational manner. Psychologists developed a theory called cognitive dissonance (CD) according to which people modify their beliefs or preferences in response to the discomfort arising from conflicting beliefs or preferences. In the example above, they identify a dissonance as follows. On the one hand, the employee holds an image of himself as a smart person and on the other hand he understands that he deliberately chose to endanger his health (by moving to the risky job), which is presumably not smart. So in order to cancel the dissonance, the employee decides that there is no danger and therefore no need to follow the strict safety regulation. We propose a formal model that is very much in line with psychologists’ theory of cognitive dissonance. We shall compare two scenarios involving a sequence of two non-commuting Decision Situations each with two options. Let $A$ be a *DS* about jobs with options$\;a_{1}$ and $a_{2}$corresponding to taking a job with a dangerous task (adventurous type) and** **respectively staying with the safe routine (habit-prone type).Let $B\;$be a *DS* about the** **willingness to use safety equipement with$\;$choices$\ b_{1}\ $(risk-averse type) and$\ b_{2}\ $(risk-loving type) corresponding to the choice of using and** **respectively not using the safety equipment. *First scenario*: The dangerous task is introduced in an existing context. It is *imposed* on the workers. They are *only* given the choice to use or not to use the safety equipment ($B)$. We write the initial state of the worker in terms of the eigenvectors of observable $A$: $$\left\vert \psi \right\rangle =\lambda _{1}\left\vert a_{1}\right\rangle +\lambda _{2}\left\vert a_{2}\right\rangle ,\ \lambda _{1}^{2}+\lambda _{2}^{2}=1.$$ We develop the eigenvectors of *A* on the eigenvectors of *B:* $$\begin{aligned} \left\vert a_{1}\right\rangle &=&\mu _{11}\left\vert b_{1}\right\rangle +\mu _{21}\left\vert b_{2}\right\rangle , \\ \left\vert a_{2}\right\rangle &=&\mu _{21}\left\vert b_{1}\right\rangle +\mu _{22}\left\vert b_{2}\right\rangle .\end{aligned}$$We now write the state in the basis of the $B$ operator: $$\left\vert \psi \right\rangle =\left[ \lambda _{1}\mu _{11}+\lambda _{2}\mu _{21}\right] \left\vert b_{1}\right\rangle +\left[ \lambda _{1}\mu _{12}+\lambda _{2}\mu _{22}\right] \left\vert b_{2}\right\rangle .$$The probability that a worker chooses to use the safety equipment is $$\begin{aligned} p_{B}\left( b_{1}\right) &=&\ \left\langle b_{1}\right\vert \left. \psi \right\rangle ^{2}=\left[ \lambda _{1}\mu _{11}+\lambda _{2}\mu _{21}\right] ^{2} \label{probb1} \\ &=&\lambda _{1}^{2}\mu _{11}^{2}+\lambda _{2}^{2}\mu _{21}^{2}+2\lambda _{1}\lambda _{2}\mu _{11}\mu _{21}. \notag\end{aligned}$$ *Second scenario:* First $A$ then $B$. The workers choose between taking a new job with a dangerous task or staying with the current safe routine. Those who chose the new job then face the choice between using safety equipment or not. Those who turn down the new job offer are asked to answer a questionnaire about their willingness to use the safety equipment for the case they would be working in the risky industry. The ex-ante probability for observing $b_{1}$ is $$\begin{aligned} p_{BA}\left( b_{1}\right) &=&p_{A}\left( a_{1}\right) p_{B}\left( \left. b_{1}\right\vert a_{1}\right) +p_{A}\left( a_{2}\right) p_{B}\left( \left. b_{1}\right\vert a_{2}\right) \label{probb2} \\ &=&\lambda _{1}^{2}\mu _{11}^{2}+\lambda _{2}^{2}\mu _{21}^{2}. \notag\end{aligned}$$The phenomenon of cognitive dissonance can now be formulated as the following inequality $$p_{BA}\left( b_{1}\right) <p_{B}\left( b_{1}\right) ,$$which occurs in our model when $2\lambda _{1}\lambda _{2}\mu _{11}\mu _{21}>0.$[^32]$\;$We next show that interference effects may be quantitatively significant. *Numerical* *example* Assume for simplicity that $\left\vert \psi \right\rangle =\left\vert b_{1}\right\rangle $ which means that everybody in the first scenario is willing to use the proposed safety equipment. Let$\;prob\left( a_{1}\left\vert \psi \right. \right) =0.25,\;$and$\;prob\left( a_{2}\left\vert \psi \right. \right) =.75$ so $\left\vert \lambda _{1}\right\vert =\sqrt{0.25},\;$and$\;\left\vert \lambda _{2}\right\vert =% \sqrt{0.75}.$ It is possible to show that in this case we have $\left\vert \mu _{11}\right\vert =\sqrt{0.25},\;$and$\;\left\vert \mu _{21}\right\vert =% \sqrt{0.75}.$[^33]$\ $We now compute $p_{B}^{{}}(b_{1})$ $\ $using the formula in (\[probb1\]) and recalling that $\left\vert \psi \right\rangle =\left\vert b_{1}\right\rangle \;$(so$\ p_{B}^{{}}(b_{1})=1):$ $$\begin{aligned} 1 &=&\left\langle b_{1}\right. \left\vert \psi _{{}}\right\rangle ^{2}=\lambda _{1}^{2}\mu _{11}^{2}+\lambda _{2}^{2}\mu _{21}^{2}+2\lambda _{1}\lambda _{2}\mu _{11}\mu _{21} \label{sce1} \\ &=&0.0625+0.5625+2\lambda _{1}\lambda _{2}\mu _{11}\mu _{21}=0.625+2\lambda _{1}\lambda _{2}\mu _{11}\mu _{21}. \notag\end{aligned}$$which implies that the interference effect is positive and equal to $% 1-0.625=0.375$. We note that it amounts to a third of the probability. Under the second scenario the probability for using safety equipment is given by the formula in (\[probb2\]) i.e., it is the same sum as in ([sce1]{}) without the interference term: $$p_{BA}\left( b_{1}\right) =0.625.$$So we see that our TI-model “delivers” cognitive dissonance: $p_{BA}\left( b_{1}\right) <p_{B}\left( b_{1}\right) .$ The key assumption that drives our result is that the choice between jobs and the choice between using or not using the safety equipment are measurements of two incompatible type characteristics (represented by two non-commuting observables). A possible behavioral interpretation is as follows. The job decision appeals to an abstract perception of risk, while the decision to use the safety equipment appeals to an emotional perception of concrete risks. In this interpretation, our assumption is that the two modes of perceptions are incompatible. This is consistent with evidence that shows a gap between perceptions of one and the same issue when the agent is in a “cold” (abstract) state of mind compared to when she is in a “hot” emotional state of mind (see for instance Loewenstein 2005).  *Comments* In their article from 1981 Akerlof and Dickens explain the behavior attributed to cognitive dissonance in terms of a rational choice of beliefs. Highly sophisticated agents choose their beliefs to fit their preferences.[^34] They are fully aware of the way their subjective perception of the world is biased and yet they keep to the wrong views. This approach does explain observed behavior but raises serious questions as to what rationality means. The type indeterminacy approach is consistent with psychologists’ thesis that cognitive dissonance prompts a change in preferences (or attitude). We view its contribution as follows. First, the TI-model provides a model that explains the appearance of cognitive dissonance. Indeed, if coherence is such a basic need, as proposed by L. Festinger and his followers, why does dissonance arise in the first place? In the TI-model dissonance arises when resolving indeterminacy in the first *DS* because of the ‘limitations’ on possible psychological types (see Section 2.3.2). Second, the TI-model features a *dynamic process* such that the propensity to use safety measures is actually altered (reduced) as a consequence of the *act of choice*. This dynamic effect is reminiscent of psychologists’ drive-like property of dissonance which leads to a change in attitude.  ### Framing Effects When alternative descriptions of one and the same decision problem induce different decisions from agents we talk about “framing effects”. Below we attend a well-known experiment which showed that two alternative formulations of the Prisoner Dilemma (the standard presentation in a 2 by 2 matrix and a presentation in a decomposed form, see below) induced dramatically different rates of cooperation. Kahneman and Tversky (2002, p. xiv) address framing effects using a two-steps (nonformal) model of the decision-making process. The first step corresponds to the construction of a representation of the decision situation. The second step corresponds to the evaluation of the choice alternatives. The crucial point is that* the true objects of evaluation are neither objects in the real world nor verbal description of those objects; they are mental representations. *To capture this feature we suggest modelling the “process of constructing a representation” in a way similar to the process of constructing preferences, i.e., as a measurement performed on the state of the agent. This is consistent with psychology that treats attitudes, values (preferences), beliefs, and representations as mental objects of the same kind.  In line with Kahneman and Tversky we describe the process of decision-making as a sequence of two steps. The first step consists of a measurement of the agent’s mental representation of the decision situation.* *The second step corresponds to a measurement of the agent’s preferences. Its outcome is a choice. Note that here** **we depart from standard decision theory. We propose that agents do not always have a *unique* representation of a decision situation, which can be revealed when thinking about it. Instead, uncertainty about what the decision situation actually is about can be resolved in a variety of ways some of which may be reciprocally incompatible. The decision situation itself is, as in standard theory, defined by the monetary payoffs associated with the choices i.e., in a unique way. The *utility* associated with the options generally depends both on the representation and the monetary payoffs. To illustrate this approach we revisit the experiment reported in Pruitt (1970) and discussed in Selten (1998). Two groups of agents are invited to play a Prisoners’ Dilemma. The game is presented to the first group in the usual matrix form with the options labelled $1_{G}$ and$\;2_{G}$ (instead of $C$ and $D,$ presumably to avoid associations with the suggestive terms ‘cooperate’ and ‘defect’): $% \begin{tabular}{|l|l|l|} \hline & $$ & $$ \\ \hline $1\_[G]{}$ & $ [cc]{} & 3\ 3 & $ & $ [cc]{} & 4\ 0 & $ \\ \hline $2\_[G]{}$ & $ [cc]{} & 0\ 4 & $ & $ [cc]{} & 1\ 1 & $ \\ \hline \end{tabular}% \ \ $ The game is presented to the second group in a decomposed form as follows:    $\ \begin{tabular}{|l|l|l|} \hline & For me & For him \\ \hline 1$\_[G]{}$ & $0$ & 3 \\ \hline 2$\_[G]{}$ & 1 & 0 \\ \hline \end{tabular} \ $ The payoffs are computed as the sum of what you take for yourself and what you get from the other player. So for instance if player 1 plays 1$_{G}\ $and player 2 plays 2$_{G}$, player 1 receives $0$ from his own play and $0$ from the other’s play) which sums to 0. Player 2 receives 1 from his own play and 3 from player 1’s play which sums to 4. So we recover the payoffs (0,4) associated with the play of Cooperate for player 1 and Defect for player 2. Game theoretically it should make no difference whether the game is presented in a matrix form or in a decomposed form. Pruitt’s main experimental result is that one observes dramatically more cooperation in the game presented in decomposed form. We now provide a possible TI-model for this situation. Let us consider a two-dimensional state space and a sequence of two measurements. The first measurement determines the mental representation of the *DS.*$\ $We call $A$ the (representation) observable corresponding to the matrix presentation. It has two non-degenerated eigenvalues denoted $\left\vert a_{1}\right\rangle $ and $\left\vert a_{2}\right\rangle .$ Similarly, $B$ is the observable corresponding to the decomposed presentation with two eigenvalues $\left\vert b_{1}\right\rangle $ and $\left\vert b_{2}\right\rangle $. If $\left\vert \psi \right\rangle $ is the initial state of the agent, we can write $\left\vert \psi \right\rangle \ =\alpha _{1}\left\vert a_{1}\right\rangle +\alpha _{2}\left\vert a_{2}\right\rangle $ or $\left\vert \psi \right\rangle \ =\beta _{1}\left\vert b_{1}\right\rangle +\beta _{2}\left\vert b_{2}\right\rangle .\ $The second measurement i.e., the decision observable is unique, we called it$\ G$. $G$ has also two eigenvalues denoted $1_{G}$ and $2_{G}.\;$ For the sake of concreteness we may think of the four alternative mental representations as follows[^35]: $\left\vert a_{1}\right\rangle :\;G$ is perceived as an (artificial) small-stake game; $\left\vert a_{2}\right\rangle :\;G$ is perceived by analogy as a real life PD-like situation (often occurring in a repeated setting). $\left\vert b_{1}\right\rangle :\;G$ is perceived as a test of generosity; $\left\vert b_{2}\right\rangle :\;G$ is perceived as a test of smartness. $\;$ When confronted with a presentation of the $\mathit{DS}$ the agent forms her mental representation of the *DS* which prompts a change in her state from $\left\vert \psi \right\rangle \ $to some $\left\vert a_{i}\right\rangle $ (if the frame is $A)\ $or $\left\vert b_{j}\right\rangle \;i,j=1,2\ ($if the frame is $B)$. The new state can be expressed in terms of the eigenvectors of the decision situation: $% \left\vert a_{i}\right\rangle =\gamma _{1i}\left\vert 1_{G}\right\rangle +\gamma _{2i}\left\vert 2_{G}\right\rangle $ or $\ \left\vert b_{j}\right\rangle =\delta _{1j}\left\vert 1_{G}\right\rangle +\delta _{2j}\left\vert 2_{G}\right\rangle .\;$ We can now formulate the framing effect as the following difference $$p_{GA}\left( i_{G}\right) \neq p_{GB}\left( i_{G}\right) ,\;\text{ }i=1,2.% \text{ } \label{framing}$$Using our result in Section 2.3.2 we get $$p_{GA}\left( 1_{G}\right) =p_{G}\left( 1_{G}\right) -2\alpha _{1}\gamma _{11}\alpha _{2}^{{}}\gamma _{12}^{{}}\text{ and\ }p_{GB}\left( 1_{G}\right) =p_{G}\left( 1_{G}\right) -2\beta _{1}\delta _{11}\beta _{2}^{{}}\delta _{12}^{{}}$$where $p_{G}\left( 1_{G}\right) $ is the probability of choosing 1$\ $in an (hypothetical) unframed situation. So we have a framing effect whenever $% 2\alpha _{1}\gamma _{11}\alpha _{2}^{{}}\gamma _{12}^{{}}\neq 2\beta _{1}\delta _{11}\beta _{2}^{{}}\delta _{12}^{{}}.\ $ The experimental results discussed in Selten (1998) namely that the decomposed presentation induces more cooperation require than $2\alpha _{1}\gamma _{11}\alpha _{2}^{{}}\gamma _{12}^{{}}-2\beta _{1}\delta _{11}\beta _{2}^{{}}\delta _{12}^{{}}>0$. The inequality says that the interference term for $1_{G}$ is larger in the standard matrix presentation $% A$ than in the decomposed form corresponding to the $B$ presentation. In order to better understand the meaning of this difference we shall consider a simple numerical example. Set $\alpha _{1}=\sqrt{.8},\ \alpha _{2}=\sqrt{.2,\ }\beta _{1}=\sqrt{0.4},\ \beta _{2}=\sqrt{0.6}.$ The key variables are the correlations between the “representation types” i.e., $\left\vert a_{i}\right\rangle $ or $\left\vert b_{j}\right\rangle $ and the “game type” $1_{G},$ i.e., the numbers $\gamma _{11},\gamma _{12}^{{}}$ and $\delta _{11},\ \delta _{12}^{{}}.$ We propose that $\gamma _{11}=\sqrt{.3}$ which is interpreted as when the agent views the game as a small stake game she plays $1_{G}$ with probability .3.Similarly $\gamma _{12}=\sqrt{.7}$ which means that when the agent perceives the game by analogy with real life, she “cooperates” with probability .7. In the alternative presentation $B,$ we propose that $\delta _{11}=1,$ i.e., when the game is perceived as a test of generosity, the agent cooperates with probability 1. When the game is perceived as a test of smartness $% \delta _{12}=0$ (because the agent views the play of $1_{G}$ as plain stupid)$.$ Computing these figures, we get $$2\alpha _{1}^{{}}\gamma _{11}\alpha _{2}^{{}}\gamma _{11}-2\beta _{1}\delta _{11}\beta _{2}^{{}}\delta _{12}^{{}}=0.366-0>0$$ In the $A$ presentation the contribution of both the $\left\vert a_{1}\right\rangle $ and the $\left\vert a_{2}\right\rangle \ $type is positive and significant. When the agent is indeterminate both types positively contribute reinforcing each other. In contrast in the $B$ presentation the contribution from $\left\vert b_{2}\right\rangle $ is null so there is no interference between the types. When the agent determines herself i.e., selects a representation, the contribution from indeterminacy is lost and that loss is positive with $A$ while it is null with $B$. Therefore, the probability for $1_{G}\ $when the game is presented in the matrix form is larger (here by .36) than in the game presented in the decomposed form.    *Comments* Selten proposes a ‘bounded rationality’ explanation: players make a superficial analysis and do not perceive the identity of the games presented under the two forms. Our approach is closer to Kahneman and Tversky who suggest that prior to the choice, a representation of the decision situation must be constructed. The TI-model provides a framework for ‘constructing’ a representation such that it delivers framing effects in choice behavior. Of course framing effects can easily be obtained when assuming that the mental images are incomplete or biased. In the TI-model we do not need to appeal to such self-explanatory arguments. In the TI-model framing effects arise as a consequence of (initial) indeterminacy of the agent’s representation of the decision situation. Since alternative (non-commuting) presentations are equally valid and their corresponding representations (eigenvalues) equally informative, highly rational agent can exhibit framing effects. Discussion ========== In this section we briefly discuss some formal features of our model. [^36] Our** **approach to decision-making yields that the type of the agents rather than being exogenously given, emerges as the outcome of the interaction between the agent and the decision situations. This is modelled by letting a decision situation be represented by an operator (observable). Decision-making is modelled as a measurement process. It projects the initial** **state of the agent** **into the subspace of the state (type) space associated with the eigenvalue corresponding to the choice made. Observables may either pairwise commute or not. When the observables commute, the corresponding type space has the properties of the Harsanyi type space. From a formal point of view this reflects the fact that all (pure) types are then mutually exclusive.[^37] When the observables do not commute, the associated pure types are not all mutually exclusive. Instead, an agent who is in a pure state after the measurement of an observable will be in a different pure state after the measurement of another observable that is incompatible with the first one. As a consequence, the type space cannot be associated with a classical probability space and we obtain an irreducible uncertainty in behavior. The Type Indeterminacy model provides a framework where we can deal both with commuting and ** **non-commuting observables. In the TI-model any type (state ) corresponds to a probability measure on the type space which allows to make predictions about the agent’s behavior. It is in this sense that the TI-model generalizes Harsanyi’s approach to uncertainty. The more controversial feature of the TI-model as a framework for describing human behavior is related to the modelling of the impact of measurement on the state i.e., how the type of the agent changes with decision-making. The rules of change are captured in the geometry of the type space and in the projection postulate. It is more than justified to question whether this seemingly very specific process should have any relevance to Social Sciences. It has been shown that the crucial property that gives all its structure to the process of change can be stated as a “minimal perturbation principle”. The substantial content of that principle is that we require that when a coarse *DS* resolves some uncertainty about the type of an agent, the remaining uncertainty is left unaffected. Recall our example in section 3.1 case 2. When the agent chooses an item out of a subset $A$ of 3 items, this prompts a resolution of some uncertainty. The type is projected into the eigenspace spanned by the two orderings consistent with the choice made. The minimal perturbation principle says that uncertainty relative to the ordering of the two remaining items is the same as before. In behavioral terms this can be expressed as follows. When confronted with the necessity to make a choice, the agent only “makes the effort” to select her preferred item while leaving the order relationship between the other items uncertain as initially. It may be argued that the minimal perturbation principle is quite demanding. Returning again to our example, if the mental processes involved in the search for the preferred items fully upset the initial state, the principle is violated. It could also be argued that the mental processes involved in decision-making determine the whole ranking. That would also violate the minimal perturbation principle. This short discussion suggests that selecting a good TI-model requires careful thinking and possibly some trial and error. We do not expect the Type Indeterminacy model to be a fully realistic description of the human behavior rather we propose it as an idealized model of agents characterized by the fact that their type change with decision-making. In particular some features of the TI-model, like the symmetry of the correlation matrix in simple examples may seem very constraining from a behavioral point of view. Consider two *DS* with two options e.g., the Prisoner Dilemma and the Ultimatum game (with option “share fairly” and “share egoistically”). Assuming those *DS* have nondegenerated eigenvalues, the symmetry of correlation matrix means that the probability to play e.g. defect after having played say egoist is exactly the same as the probability to play egoist after having played defect. We do not expect this kind of equality to hold in general.[^38] Nevertheless, keeping in mind some reservations as to its realism, our view is that the Type Indeterminacy model can provide a fruitful framework for analyzing, explaining and predicting human behavior. Clearly much additional work is needed to extend of the TI-model to strategic and repeated decision-making. We are currently exploring this second stage of our research program.   As a final remark it should be emphasized that not all instances of non-commutativity in choice behavior call for Hilbert space modelling. Theories of addiction feature effects of past choices on future preferences. And in standard consumer theory, choices do have implications for future behavior, i.e., when goods are substitutes or complements. But in those cases we *do* expect future preferences to be affected by the choices. The Type Indeterminacy model of decision-making can be useful when we expect choice behavior to be consistent with the standard probabilistic model, because nothing justifies a modification of preferences. Yet, actual behavior contradicts those expectations. [99]{} Akerlof G. A. and T. Dickens (1982) The Economic Consequences of Cognitive Dissonance, *The American Economic Review* 72, 307-319. Ariely D. , G. Prelec and D. Lowenstein “Coherent Arbitrariness”: Stable Demand Curve without Stable Preferences. *Quarterly Journal of Economics*, Feb. 2003, 73-103. Arrow K. J. (1959) “ Rational Choice Functions and Orderings” *Economica*, New Series, Vol 26/102, pp. 121-127. Beltrametti E. G. and G. Cassinelli (1981), The Logic of Quantum Mechanics, Encyclopia of Mathematics and its Applications vol. 15, Addison-Wesley Publishing Company. Benabou R. and J. Tirole (2002), Self-Knowledge and Personal Motivation *Quarterly Journal of Economics*, 117, 871-915. Berthoz A. (2003) La Decision, Odile Jacob, Paris. Birkhoff G. and J. von Neuman (1936) The Logic of Quantum Mechanics *Ann. of Math*. 37, \_823-843. Danilov V. I and A. Lambert-Mogiliansky “ Non-Classical Measurement Theory - A Framework for Behavioral Sciences” Arxiv xxx.lanl.gov/physics/060451 Eisert J., M. Wilkens and M. Lewenstein (1999), Quantum Games and Quantum Strategies  *Phys. Rev. Lett.* 83, 3077. Erev I., G. Bornstein and T. Wallsten (1993) The Negative Effect of Probability Assessment on Decision Quality, *Organizational Behavior and Human Decision Processes* 55, 78-94. Cohen-Tannoudji C., B. Diu and F. Laloe (1973), Mecanique Quantique 1, Herman Editeur des Sciences et des Arts, Paris. Cohen D. W. (1989) An Introduction to Hilbert Space and Quantum Logic, Problem Books in Mathematics, Springer-Verlag, New York. Beltrametti E. G. and G. Cassinelli, (1981), The Logic of Quantum Mechanics, Encyclopedia of Mathematics and its Applications, Ed. G-C Rota, Vol. 15, Addison-Wesley, Massachussets. Ben-Horin D., (1979) Dying for Work: Occupational Cynicism Plagues Chemical Workers in *These Times,* June 27/July 3, 3-24. Feynman R., (1980) La Nature de la Physique, Seuil, Paris. Festinger L., (1957) Theory of Cognitive Dissonance, Stanford University Press, Stanford, CA. Holland S. S. JR. (1995) Orthomodularity in Infinite Dimensions; a Theorem of M. Soler *Bulletin of the American Mathematical Society* 32, 205-234. Kahneman D. and A. Tversky (2000) Choice, Values and Frames, Cambridge Universtity Press, Cambridge. Knez M. and C. Camerer (2000) Increasing Cooperation in Prisoner’s Dilemmas by Establishing a Precedent of Efficiency in Coordination Game, *Organizational Behavior an Human Decision Processes* 82, 194-216. Lowenstein G. (2005) “Hot-Cold Empathy Gaps and Medical Decision-making” *Health Psychology* 24/4 549-556. Gul F. and W. Pesendorfer (2001) ‘Temptation and Self-control’ *Econometrica, *69, 1403-1436. McFadden D. (1999) Rationality of Economists *Journal of Risk and Uncertainty* 19, 73-105 Mc Fadden D. (2000) Economic Choices, Nobel Lecture. Mackey G. W. (2004, first ed. 1963) Mathematical Foundations of Quantum Mechanics, Dover Publication, Mineola New York. Mas-Colell A., M. D. Whinston and J. R. Green (1995), Microeconomic Theory, New York Oxford, Oxford University Press. Popper K. (1992), Un Univers de Propensions, Edition de l’Eclat, Paris. Pruitt D.G. (1970) Reward Structure of Cooperation: the Decomposed Prisoner’s Dilemma Game  *Journal of Personality and Psychology* 7, 21-27. Savage L. J. (1954) The Foundation of Statistics, Wiley, New York. Selten R.(1998) Features of Experimentally Observed Bounded Rationality *European Economic Review* 42/3-5, 413-436. Tversky A. and I. Simonson (1993) Context-Dependent Preferences *Management Sciences* 39, 85-117. Sen A. (1997) “Maximization and the Act of Choice”  *Econometrica* 65, 745-779. Appendix: Elements of Quantum Mechanics ======================================= States and Observables ---------------------- In Quantum Mechanics the state of a system is represented by a vector $% \left\vert \psi \right\rangle $ in a Hilbert space $\mathcal{H}$. According to the superposition principle, every complex linear combination of state vectors is a state vector. A Hermitian operator called an observable is associated to each physical property of the system.  *Theorem 1* A Hermitian operator$\;A$ has the following properties: - Its eigenvalues are real. - Two eigenvectors corresponding to different eigenvalues are orthogonal. - There is an orthonormal basis of the relevant Hilbert space formed with the eigenvectors of $A.$ Let us call $\left| v_{1}\right\rangle ,\left| v_{2}\right\rangle ,...,\left| v_{n}\right\rangle $ the normalized eigenvectors of $A$ forming a basis of $\mathcal{H}$. They are associated with eigenvalues $\alpha _{1,}\alpha _{2},...,\alpha _{n}$, so: $A\left| v_{i}\right\rangle =\alpha _{i}\left| v_{i}\right\rangle $. The eigenvalues can possibly be degenerated, i.e., for some $i$ and $j$, $\alpha _{i}=\alpha _{j\text{.}}$ This means that there is more than one linearly independent eigenvector associated with the same eigenvalue. The number of these eigenvectors defines the degree of degeneracy of the eigenvalue which in turn defines the dimension of the eigensubspace spanned by these eigenvectors. In this case, the orthonormal basis of $\mathcal{H}$ is not unique because it is possible to replace the eigenvectors associated to the same eigenvalue by any complex linear combination of them to get another orthonormal basis. When an observable $A$ has no degenerated eigenvalue there is a unique orthonormal basis of $\mathcal{H}$ formed with its eigenvectors. In this case (see below), it is by itself a Complete Set of Commuting Observables. *Theorem 2* If two observables $A$ and $B$ commute there is an orthonormal basis of $% \mathcal{H}$ formed by eigenvectors common to $A$ and $B$. Let $A$ be an observable with at least one degenerated eigenvalue and $B$ another observable commuting with $A.$ There is no unique orthonormal basis formed by $A$ eigenvectors. But there is an orthonormal basis of the relevant Hilbert space formed by eigenvectors common to $A$ and $B$. By definition, $\left\{ A,B\right\} $ is a Complete Set of Commuting Observables (CSCO) if this basis is unique. Generally, a set of observables $% \left\{ A,B,...\right\} $ is said to be a CSCO if there is a unique orthonormal basis formed by eigenvectors common to all the observables of the set. Measurements ------------ An observable $A\ $is associated to each physical property of a system $S$. Let $\left| v_{1}\right\rangle ,\left| v_{2}\right\rangle ,...,\left| v_{n}\right\rangle $ be the normalized eigenvectors of $A$ associated respectively with eigenvalues $\alpha _{1,}\alpha _{2},...,\alpha _{n}$ and forming a basis of the relevant state space. Assume the system is in the normalized state $\left| \psi \right\rangle $. A measurement of $A$ on $S$ obeys the following rules, collectively called ‘Wave Packet Reduction Principle’ (the Reduction Principle). *Reduction Principle* 1\.  When a measurement of the physical property associated with an observable $A$ is made on a system $S$ in a state $\left\vert \psi \right\rangle $, the result only can be one of the eigenvalues of $A$. 2\. The probability to get the non-degenerated value $\alpha _{i}$ is $% P(\alpha _{i})=\left| \left\langle v_{i}|\psi \right\rangle \right| ^{2}.$ 3\. If the eigenvalue is degenerated then the probability is the sum over the eigenvectors associated with this eigenvalue: $P(\alpha _{i})=\sum \left\vert \left\langle \nu _{i}^{j}|\psi \right\rangle \right\vert ^{2}.$ 4\. If the measurement of $A$ on a system $S$ in the state $\left| \psi \right\rangle $ has given the result $\alpha _{i}$ then the state of the system immediately after the measurement is the normalized projection of $% \left| \psi \right\rangle $ onto the eigensubspace of the relevant Hilbert space associated with $\alpha _{i}$. If the eigenvalue is not degenerated then the state of the system after the measurement is the normalized eigenvector associated with the eigenvalue. If two observables $A$ and $B$ commute then it is possible to measure both simultaneously: the measurement of $A$ is not altered by the measurement of $% B.$ This means that measuring $B$ after measuring $A$ does not change the value obtained for $A$. If we again measure $A$ after a measurement of $B,$ we again get the same value for $A$. Both observables can have a definite value. ### Interferences The archetypal example of interferences in quantum mechanics is given by the famous two-slits experiment.[^39] A parallel beam of photons falls on a diaphragm with two parallel slits and strikes a photographic plate. A typical interference pattern showing alternate bright and dark rays can be seen. If one slit is shut then the previous figure becomes a bright line in front of the open slit. This is perfectly understandable if we consider photons as waves, as it the assumption is in classical electromagnetism. The explanation is based on the fact that when both slits are open, one part of the beam goes through one slit and the other part through the other slit. Then, when the two beams join on the plate, they interfere constructively (giving bright rays) or destructively (giving dark ones), depending on the difference in length of the paths they have followed. But a difficulty arises if photons are considered as particles, as can be the case in quantum mechanics. Indeed, it is possible to decrease the intensity of the beam so as to have only one photon travelling at a time. In this case, if we observe the slits in order to detect when a photon passes through (for example, by installing a photodetector in front of the slits), it is possible to see that each photon goes through only one slit. It is never the case that a photon splits to go through both slits. The photons behave like particles. Actually, the same experiment was done with electrons instead of photons, with the same result. If we do the experiment this way with electrons (observing which slit the electrons go through, i.e., sending light through each slit to see the electrons), we see that each electron goes through just one slit and, in this case, we get no interference. If we repeat the same experiment without observing which slit the electrons pass through then we recover the interference pattern. Thus, the simple fact that we observe which slit the electron goes through destroys the interference pattern (two single slit patterns are observed). The quantum explanation is based on the assumption that when we don’t observe through which slit the electron has gone then its state is a superposition of both states gone through slit 1 and gone through slit 2[^40], while when we observe it, its state collapses onto one of these states. In the first case, the position measurement is made on electrons in the superposed state and gives an interference pattern since both states are manifested in the measurement. In the second case, the position is measured on electrons in a definite state and no interference arises. In other words, when only slit 1 is open we get a spectrum, say $S_{1}$ (and $S_{2\text{ }}$when only slit 2 is open). We expect to get a spectrum $S_{12}$ that sums of the two previous spectra when both slits are open, but this is not the case: $% S_{12}\neq S_{1}+S_{2}.$ [^1]: PSE, (CNRS, EHESS, ENS, ENPC), Ecole Economique de Paris, [email protected] [^2]: CREST-LEI Paris and Center for the Study of Rationality, Hebrew University, [email protected] [^3]: IHPST, UMR 8590 CNRS/Paris 1, and CMLA, UMR 8536 CNRS/ENS Cachan, [email protected] [^4]: We are grateful for helpful comments from V.I. Danilov, J. Dreze, P. Jehiel, D. Laibson, P. Milgrom, W. Pesendorfer, A. Roth, and seminar participants at Harvard, Princeton, and CORE. [^5]: In the classical version of the dilemma, the dominant strategy for both players is to defect and thereby to do worse than if they had both decided to cooperate. In the quantum version, there are a couple of quantum strategies that are both a Nash equilibrium and Pareto optimal and whose payoff is the one of the joint cooperation. [^6]: This approach is closely related to quantum computing. It relies on the use of a sophisticated apparatus to exploit q-bits’ property of entanglement in mixed strategies. [^7]: In this work we borrow the elements of the quantum formalism that concern the measurement process. We will not use the part of the theory theory concerned with the evolution of systems over time. [^8]: The associated concept of irreducible uncertainty, which is the essence of indeterminacy, is formally defined in Section 2 of the paper. [^9]: Leon Festinger is the father of the theory of cognitive dissonance (Festinger 1957). [^10]: The experience was richer. People were divided into two groups offered different amounts of money. For our purpose it is sufficient to focus on a single result. [^11]: All information (beliefs) and strategic considerations are embedded in the definition of the choices. Thus an agent’s play of cooperation in a Prisoner’s Dilemma, is a play of  cooperation given his information (knowledge) about the opponent. [^12]: In Quantum Mechanics the field that is used is the complex numbers field. However, for our purposes the field of real numbers provides the structure and the properties needed (see e.g. Beltrametti and Cassinelli 1981 and Holland 1995). Everything we present in the appendix (Elements of quantum mechanics) remains true when we replace Hermitian operators with real symmetric operators. [^13]: We use the term state to refer to ‘pure state’. Some people use the term state to refer to mixture of pure states. A mixture of pure states combines indeterminacy with elements of incomplete information. They are represented by the** **so called density operators. [^14]: The distributivity condition defining a Boolean space is dropped for a weaker condition called ortho-modularity. The basic structure of the state space is that of a logic, i.e., an orthomodular lattice. For a good presentation of Quantum Logic, a concept introduced by Birkhoff and Von Neuman (1936), and further developed by Mackey (2004, 1963), see Cohen (1989). [^15]: The term first-kind measurement was introduced by Pauli. [^16]: Even standard decision theory implicitely restricts its application to decision problems satisfying the first-kindness property (or that can be derived from such decision problems). In contrast, random utility models do not require choice behavior to satisfy the first-kindness property in the formulation used in this paper. [^17]: Observables in Physics are represented by Hermetian operators because QM is defined over the field of complex numbers. Here, we confine ourselves to the field of real numbers which is why observables are represented by symmetric operators. [^18]: That is, for $i\neq j\ $any vector in $\mathcal{H}_{i_{A}}$ is orthogonal to any vector in$\;\mathcal{H}_{j_{A}}$ and any vector in $\mathcal{H}$ is a sum of $n$ vectors, one in each component space. [^19]: In the argument that we develop in section 3.1, the pure states are linear orders therefore choice experiments are observables with degenerated eigenvalues [^20]: The Weak Link game is a type of coordination game where each player picks an action from a set of integers. The payoffs are defined in such a manner that each player wants to select the minimum of the other players but everyone wants that minimum to be as high as possible. [^21]: The precedent effect hypothesis is as follows: The shared experience of playing the efficient equilibrium in the WL game creates a precedent of efficient play strong enough to (...) lead to cooperation in a finitely repeated PD game, Knetz and Camerer (2000 see p.206). [^22]: We say that a DS is complete when its outcome provides a complete characterization of the agent. [^23]: Samuelson (1947) [^24]: See for example Mas-Colell, Whinston and Green, Microeconomic Theory, p.13. [^25]: The use of the same symbol for sets of items and observables should not confuse the reader. Either the context unambiguously points to the right interpretation or we make it precise. [^26]: "If for some $B\in \mathcal{B}\ $with $x,\ y\in B$ we have $x\in C\left( B\right) ,$ then for any $B^{\prime }\in \mathcal{B}$$\ $with$\ x,y\in B^{\prime }$ and $y\in C\left( B^{\prime }\right) ,$ we must also have $x\in C\left( B^{\prime }\right) ."($Mas-Colell et al. (1995) p.10). [^27]: At this stage of the research we do not want to deal with indifference relations. [^28]: Conditions (a) and (b) are equivalent to C2 and C4 in Arrow (1959). [^29]: Consider the choice set $\left\{ a,b,c\right\} .$ We do the following two series of experiments. In the first series we let the agent choose from $% \left\{ a,b\right\} $, then from $\left\{ b,c\right\} \ $and last from $% \left\{ a,b,c\right\} .$In the second series the agent first choose from $% \left\{ b,c\right\} $ then from $\left\{ a,b\right\} ~$and$\ $last from $% \left\{ a,b,c\right\} $. Assume the outcomes of the two first experiments are $a$ and $\ b\ $then by WRAP’ the third choice maybe either $a,\ $or $% b.\ $In the second series, the outcomes must be (because of commutativity) $% \ b$ and $a\ $respectively.$\ $So the third choice may be either $a$ or $c.$ By commutativity the outcome of $\left\{ a,b,c\right\} $ must be the same in the two series which uniquelly selects $a$ so a choice behavior respecting WRAP’ is transitive when the choices commute. [^30]: See e.g., Mas-Colell et al. (1995) p.13. [^31]: We remind that "playing a *DS"* means performing the corresponding measurement (see section 2.2). [^32]: We note that $p_{BA}\left( b_{1}\right) $ includes the probability of a choice of safety measures *both* in the group that chose the risky job and in the group that chose the safe job. This guarantees that we properly distinguish between the CD effect (change in attitude) and the selection bias. [^33]: This uses the fact that $\left( \begin{tabular}{ll} $\_[11]{}$ & $\_[22]{}$ \\ $\_[21]{}$ & $\_[22]{}$% \end{tabular}% \right) $ is a rotation matrix. [^34]: Akerlof and Dickens allow workers to freely choose beliefs (about risk) so as to optimize utility which includes psychological comfort. The workers are highly rational in the sense that when selecting beliefs, they internalize their effect their own subsequent bounded rational behavior. [^35]: This is only meant as a suggestive illustration. [^36]: For a systematic investigation of the mathematical foundations of the HSM in view of their relevance for social sciences see Danilov and Lambert-Mogiliansky (2005). [^37]: We say that a type is pure when it is obtained as the result of a complete measurement i.e., the measurement of a complete set of commuting observables (CSCO). [^38]: Note that a failure of this inequality to hold may be due to the fact that the *DS* have some degenerated eigenvalues. In fact, this is likely to be the case in Social Sciences since we often deal with rather coarse measurements of the agent’s type. [^39]: See, e.g., Feynman (1965) for a very clear presentation. [^40]: This doesn’t mean that the photon actually went through both slits. This state simply can’t be interpretated from a classical point of view.

ArXiv

--- abstract: 'Microscopic RPA calculations based on the cranked shell model are performed to investigate the quadrupole and octupole correlations for excited superdeformed bands in [$^{190}$Hg]{}, [$^{192}$Hg]{}, and [$^{194}$Hg]{}. The $K=2$ octupole vibrations are predicted to be the lowest excitation modes at zero rotational frequency. At finite frequency, however, the interplay between rotation and vibrations produces different effects depending on neutron number: The lowest octupole phonon is rotationally aligned in [$^{190}$Hg]{}, is crossed by the aligned two-quasiparticle bands in [$^{192}$Hg]{}, and retains the $K=2$ octupole vibrational character up to the highest frequency in [$^{194}$Hg]{}. The $\gamma$ vibrations are predicted to be higher in energy and less collective than the octupole vibrations. From a comparison with the experimental dynamic moments of inertia, a new interpretation of the observed excited bands invoking the $K=2$ octupole vibrations is proposed, which suggests those octupole vibrations may be prevalent in SD Hg nuclei.' address: - 'AECL, Chalk River Laboratories, Chalk River, Ontario K0J 1J0, Canada' - 'Department of Physics, Kyoto University, Kyoto 606–01, Japan' - ' Department of Mathematical Physics, Lund Institute of Technology, Box 118, S-22100, Lund, Sweden' - ' Department of Physics, Kyushu University, Fukuoka 812, Japan' author: - 'Takashi Nakatsukasa[^1]' - Kenichi Matsuyanagi - Shoujirou Mizutori - 'Yoshifumi R. Shimizu' title: | Microscopic Structure of High-Spin Vibrational Excitations\ in Superdeformed $^{190,192,194}$Hg --- Introduction {#sec: intro} ============ Theoretical and experimental studies of collective vibrational states built on the superdeformed (SD) yrast band are open topics of interest in the field of high-spin nuclear structure. Since the large deformation and rapid rotation of SD bands may produce a novel shell structure, we expect that surface vibrations will exhibit quite different features from those found in spherical and normal-deformed nuclei. According to our previous work[@MSM90; @NMM92; @NMM93; @Miz93; @Nak95], low-lying octupole vibrations are more important than quadrupole vibrations when the nuclear shape is superdeformed. Strong octupole correlations in SD states have been also suggested theoretically in Refs.[@DWS90; @Abe90; @HA90; @Bon91; @LDR91; @ND92; @Ska92; @Ska93]. Experimentally, octupole correlations in SD states have been suggested for $^{152}$Dy[@Dag95], $^{193}$Hg[@Cul90] and $^{190}$Hg[@Cro94; @Cro95]. We have reported theoretical calculations corresponding to these data for $^{193}$Hg[@NMM93] and $^{152}$Dy[@Nak95]. In this paper, we discuss the quadrupole and octupole correlations for [$^{190}$Hg]{} (which have been partially reported in Refs.[@Cro95; @Wil95; @Nak95_2]) and for the neighboring SD nuclei $^{192,194}$Hg. We have predicted the low-lying $K=2$ octupole vibrations for SD Hg isotopes $^{190,192,194}$Hg ($E_x\sim 1$MeV)[@NMM93; @Miz93]. These predictions differ from the results of generator-coordinate-method (GCM) calculations[@Ska93] in which the $K=0$ octupole state is predicted to be the lowest in SD [$^{192}$Hg]{} and the excitation energies are significantly higher ($E_x\sim 2$MeV) than in our predictions. Experimentally [@Cro95], the routhians of the lowest octupole state decrease with the rotational frequency, for example from $E'_x\approx 0.7$MeV to 0.3MeV as $\hbar{\omega_{\rm rot}}$ goes from 0.25 to 0.35MeV, therefore to compare the theoretical routhians directly with the experimental ones, we need to calculate them at finite rotational frequency. For this purpose, the cranked shell model extended by the random-phase approximation (RPA) provides us with a powerful tool to investigate collective excitations at high angular momentum. A great advantage of this model is its ability to take into account effects of the Coriolis coupling on the collective vibrational motions in a rapidly rotating system. Since in the normal-deformed nuclei it is known that Coriolis coupling effects are important even for the $3^-$ octupole states[@NV70], one may expect strong Coriolis mixing for high-spin octupole states built on the SD yrast band. On the other hand, our previous calculations suggested weak Coriolis mixing for the lowest octupole state in [$^{192}$Hg]{}[@NMM93] and $^{152}$Dy[@Nak95]. This may be because the angular momentum of the octupole phonon is strongly coupled to the symmetry axis due to the large deformation of the SD shape. Generally speaking, Coriolis mixing is expected to occur more easily in nuclei with smaller deformation. However this expectation may not hold for octupole bands in all SD nuclei because Coriolis mixing depends on the shell structure. In this paper we find a significant difference in the Coriolis mixing between an octupole band in [$^{190}$Hg]{} and the other bands. Another advantage of this model is that it gives us a unified microscopic description of collective states, weakly-collective states, and non-collective two-quasiparticle excitations. A transition of the octupole vibrations into aligned two-quasiparticle bands at high-spin in normal-deformed nuclei has been predicted by Vogel[@Vog76]. In Ref.[@Nak95_2], this transition is discussed in the context of experimental data on rare-earth and actinide nuclei, and a damping of octupole collectivity at high spin was suggested. Since similar phenomena may happen to octupole vibrations in SD states, it is important that our model describes the interplay between collective and non-collective excitations. Recent experimental studies reveal a number of interesting features of excited SD bands in even-even Hg isotopes. In [$^{190}$Hg]{}, almost constant dynamic moments of inertia ${{\cal J}^{(2)}}$ have been observed by Crowell et al.[@Cro94]. Ref.[@Cro95] has established the relative excitation energy of this band and confirmed the dipole character of the decay transitions into the yrast SD band. This band has been interpreted as an octupole vibrational band. Two more excited bands in [$^{190}$Hg]{} have been observed recently by Wilson et al.[@Wil95], one of which shows a sharp rise of ${{\cal J}^{(2)}}$ at low frequency. In [$^{192}$Hg]{}, Fallon et al.[@Fal95] have reported two excited bands which exhibit peaks in ${{\cal J}^{(2)}}$ at high frequency. In contrast with these atypical ${{\cal J}^{(2)}}$ behaviors, two excited bands in [$^{194}$Hg]{} originally observed by Riley et al.[@Ril90] and extended by Cederwall et al.[@Ced95] show a smooth increase with rotational frequency. We show in this paper that this ${{\cal J}^{(2)}}$ behavior can be explained with a single theoretical model which microscopically takes into account shape vibrations and the Coriolis force. The purpose of this paper is to present the RPA method based on the cranked shell model and its ability to describe a variety of nuclear properties including shape vibration at large deformation and high spin. We propose a plausible interpretation for the microscopic structure of excited SD bands in $^{190,192,194}$Hg, and show that octupole bands may be more prevalent than expected in these SD nuclei. Section \[sec: model\] presents a description of the model, in which we stress our improvements to the cranked Nilsson potential and to the coupled RPA method in a rotating system. Section \[sec: detail\] presents details of the calculation in which the pairing and effective interactions are discussed. The results for the excited SD [$^{190}$Hg]{}, [$^{192}$Hg]{}, and [$^{194}$Hg]{}are presented in section \[sec: results\], and compared with the experimental data in section \[sec: comparison\]. The conclusions are summarized in section \[sec: conclusion\]. Theoretical framework {#sec: model} ===================== The theory of the cranked shell model extended by the random-phase approximation (RPA) was first developed by Marshalek[@Mar75] and has been applied to high-spin $\beta$ and $\gamma$ vibrational bands[@EMR80; @SM83; @SM95] and to octupole bands[@RER86; @MSM90; @NMM92; @NMM93; @Miz93; @Nak95]. Since this theory is suitable for describing the collective vibrations built on deformed high-spin states, it is very useful for investigating vibrational motion built on the SD yrast band. The cranked Nilsson potential with the local Galilean invariance ---------------------------------------------------------------- We start with a rotating mean field with a rotational frequency ${\omega_{\rm rot}}$ described by $$\label{sp-potential} h_{\rm s.p.}=h_{\rm Nilsson} + \Gamma_{\rm pair} - {\omega_{\rm rot}}J_x \ + h_{\rm add},$$ where $h_{\rm Nilsson}$ is a standard Nilsson potential defined in single-stretched coordinates, $r'_i=\sqrt{\frac{\omega_i}{\omega_0}} r_i$ and $p'_i=\sqrt{\frac{\omega_0}{\omega_i}} p_i$ ($i=x,y,z$), $$\label{Nilsson} h_{\rm Nilsson}=\left( \frac{\omega_i}{\omega_0} \right) \left( \frac{{{\bf p}'}^2}{2M} +\frac{M\omega_0^2}{2} {{\bf r}'}^2 \right) +v_{ll}\left({{\bf l}'}^2 -\langle {{\bf l}'}^2 \rangle_N\right) +v_{ls}{\bf l}'\cdot {\bf s} \ ,$$ where ${\bf l}'={\bf r}'\times {\bf p}'$. The pairing field $\Gamma_{\rm pair}$ is defined by $$\label{pair-field} \Gamma_{\rm pair} = -\sum_{\tau=n,p}\Delta_\tau \left(P_\tau^\dagger + P_\tau\right) -\sum_{\tau=n,p} \lambda_\tau N_\tau \ ,$$ where $P_\tau = \sum_{k\in\tau,k>0} c_{\bar{k}} c_k$ and $N_\tau=\sum_{k\in\tau} c_k^\dagger c_k$ are the monopole-pairing and number operators, respectively. In section \[sec: detail-MF\], we discuss the details of the pairing field used in the calculations. A standard cranked Nilsson potential has the disadvantage that it overestimates the moments of inertia compared to a Woods-Saxon potential. This problem comes from the spurious velocity-dependence associated with the ${\bf l^2}$-term in the Nilsson potential which is absent for Woods-Saxon. If the mean-field potential is velocity independent, the local velocity distribution in the rotating nucleus remains isotropic in velocity space, which means that the flow pattern becomes the same as for a rigid-body rotation[@BM75]. However, in the cranked Nilsson potential, this isotropy of the velocity distribution is significantly broken due to the ${\bf l^2}$-term. Thus the Coriolis force introduces a spurious flow in the rotating coordinate system, proportional to the rotational frequency. This spurious effect can be compensated by an additional term that restores the local Galilean invariance. This additional term is obtained by substituting (the local Galilean transformation) $${\bf p} \longrightarrow {\bf p}-M\left(\mbox{\boldmath ${\omega_{\rm rot}}$}\times {\bf r}\right)\ ,$$ in the ${\bf ls}$- and ${\bf l^2}$-terms of the Nilsson potential. This prescription was suggested by Bohr and Mottelson[@BM75], and developed by Kinouchi and Kishimoto[@Kin88]. For a momentum-dependent potential $V({\bf r},{\bf p})$, $$\begin{aligned} V({\bf r},{\bf p}) + h_{\rm add} &=& V\left({\bf r},{\bf p} -M\left(\mbox{\boldmath ${\omega_{\rm rot}}$}\times {\bf r}\right)\right) \ ,\\ &\approx& V({\bf r},{\bf p}) - {\omega_{\rm rot}}M \left( y \frac{\partial}{\partial p_z} - z \frac{\partial}{\partial p_y}\right) V({\bf r},{\bf p})\ ,\\ &=& V({\bf r},{\bf p}) + \frac{i}{\hbar} {\omega_{\rm rot}}M \left( y \left[ z,V\right] - z \left[ y,V\right] \right) \ ,\end{aligned}$$ where we assume uniform rotation around the $x$-axis, $\mbox{\boldmath ${\omega_{\rm rot}}$}=({\omega_{\rm rot}},0,0)$. Following this prescription, the additional term $h_{\rm add}$ in eq.(\[sp-potential\]) is obtained for the Nilsson potential (\[Nilsson\]), $$\label{additional-term} h_{\rm add}=-\frac{{\omega_{\rm rot}}}{\sqrt{\omega_y\omega_z}}\left\{ v_{ll} \left(2M\omega_0 {{\bf r}'}^2 - \hbar\left(N_{\rm osc}+\frac{3}{2} \right)\right) l'_x +v_{ls}M\omega_0\left({{\bf r}'}^2 s_x - r'_x \left({\bf r}'\cdot {\bf s}\right)\right)\right\} \ .$$ Note that the term proportional to $\left( N_{\rm osc}+\frac{3}{2} \right)$ in eq.(\[additional-term\]) comes from the velocity-dependence of $\langle {{\bf l}'}^2 \rangle_N$ in eq.(\[Nilsson\]). This result, eq.(\[additional-term\]), has been applied to the SD bands in $^{152}$Dy [@Nak95] where the single-particle routhians were found to be very similar to those obtained by using the Woods-Saxon potential. In Fig. \[dy152\_J\], moments of inertia for SD $^{152}$Dy calculated with and without the additional term (\[additional-term\]) are displayed. Since the effects of the mixing among the major oscillator shells $N_{\rm osc}$ are neglected in calculating our routhians, kinematic (${{\cal J}^{(1)}}$) and dynamic (${{\cal J}^{(2)}}$) moments of inertia are obtained by adding the contributions of the $N_{\rm osc}$-mixing effects to the values calculated without them: $$\begin{aligned} {{\cal J}^{(1)}}&=& \frac{\langle J_x \rangle}{{\omega_{\rm rot}}} + {\it\Delta}{\cal J}_{\rm Inglis} \ ,\\ {{\cal J}^{(2)}}&=& \frac{d\langle J_x \rangle}{d{\omega_{\rm rot}}} + {\it\Delta}{\cal J}_{\rm Inglis} \ ,\\ {\it\Delta}{\cal J}_{\rm Inglis} &=& {\cal J}_{\rm Inglis}-{\cal J}_{\rm Inglis}^{{\it\Delta}N=0} = 2 \sum_{n({\it\Delta}N=2)} \frac{|{\langle n |}J_x {| 0 \rangle}|^2}{E_n-E_0} \ ,\end{aligned}$$ where ${\it\Delta}{\cal J}_{\rm Inglis}$ is difference between the Inglis moments of inertia with and without the ${\it\Delta}N_{\rm osc}=2$ contributions[@SVB90]. The ${{\cal J}^{(1)}}$ and ${{\cal J}^{(2)}}$ values calculated with the additional term are very close to the rigid-body value at low frequency, which means that the spurious effects of the ${\bf l}^2$-term have been removed. Note that the abscissa of Fig. \[dy152\_J\] corresponds to the “bare” rotational frequency without renormalization. The drastic reduction of ${{\cal J}^{(1)}}$ and ${{\cal J}^{(2)}}$ at high frequency is corrected by the additional term, and this is seen to be important in reproducing the experimental ${{\cal J}^{(2)}}$-behavior of the yrast SD band. The RPA in the rotating frame ----------------------------- The residual interactions are assumed to be in a separable form, $$H_{\rm int} = -\frac{1}{2} \sum_{\rho,\alpha} \chi_\rho R_\rho^\alpha R_\rho^\alpha \ , \label{residual-int}$$ where $R_\rho^\alpha$ are one-body Hermitian operators, and $\chi_\rho$ are coupling strengths. The indices $\alpha$ indicate the signature quantum numbers ($\alpha=0,1$) and $\rho$ specifies other modes. In this paper, we take as $R_\rho^\alpha$ the monopole pairing and the quadrupole operators for positive-parity states, and the octupole and the isovector dipole operators for negative-parity states (see eq.(\[operator\_R\])). Since the $K$-quantum number is not conserved at finite rotational frequency, it is more convenient to make the multipole operators have good signature quantum numbers. In general, the Hermitian multipole (spin-independent) operators with good signature quantum numbers are constructed by $$Q_{\lambda K}^\alpha = \frac{i^{\lambda+\alpha+K}}{\sqrt{2(1+\delta_{K0})}} \left( r^\lambda Y_{\lambda K} + (-)^{\lambda+\alpha} r^\lambda Y_{\lambda -K} \right) \quad\quad (K\ge 0) \label{multipole-op}$$ where the spherical-harmonic functions $Y_{\lambda K}$ are defined with respect to the symmetry ($z$-) axis. All multipole operators are defined in doubly-stretched coordinates, ($r^{''}_i=\frac{\omega_i}{\omega_0} r_i$), which can be regarded as an improved version of the conventional multipole interaction. Sakamoto and Kishimoto[@SK89] have shown that at the limit of the harmonic-oscillator potential (at ${\omega_{\rm rot}}=0$), it guarantees nuclear self-consistency[@BM75], restoration of the symmetry broken in the mean field, and separation of the spurious solutions. The coupling strengths $\chi_\rho$ should be determined by the self-consistency condition between the density distribution and the single-particle potential (see section \[sec: detail-RPA\] for details). To describe vibrational excitations in the RPA theory, we must define the [*quasiparticle vacuum*]{} on which the vibrations are built. The observed moments of inertia ${{\cal J}^{(2)}}$ of the yrast SD bands smoothly increase in the A=190 region, which suggests that the internal structure also smoothly changes as a function of the frequency ${\omega_{\rm rot}}$. Therefore the [*adiabatic representation*]{}, in which the quasiparticle operators are always defined with respect to the yrast state ${| {\omega_{\rm rot}}\rangle}$, is considered to be appropriate in this work. In terms of quasiparticles, the Hamiltonian of eq.(\[sp-potential\]) can be diagonalized (by the general Bogoliubov transformation) as $$h_{\rm s.p.}=\mbox{const.} + \sum_\mu \left( E_\mu a_\mu^\dagger a_\mu \right) + \sum_{\bar\mu} \left( E_{\bar\mu} a_{\bar\mu}^\dagger a_{\bar\mu} \right) \ ,$$ with $$a_\mu {| {\omega_{\rm rot}}\rangle} = a_{\bar\mu} {| {\omega_{\rm rot}}\rangle} = 0 \ ,$$ where ($a_\mu^\dagger, a_{\bar\mu}^\dagger$) represent the quasiparticles with signature $\alpha=(1/2,-1/2)$, respectively. The excitation operators of the RPA normal modes ${X_n^\alpha}^\dagger$ ($\alpha=0,1$) are defined by $$\begin{aligned} {X_n^0}^\dagger &=& \sum_{\mu\bar\nu} \left\{ \psi_n^0(\mu\bar\nu) a_\mu^\dagger a_{\bar\nu}^\dagger +\varphi_n^0(\mu\bar\nu) a_{\bar\nu} a_\mu \right\} \ ,\\ \label{normal-mode-0} {X_n^1}^\dagger &=& \sum_{\mu < \nu} \left\{ \psi_n^1(\mu\nu) a_\mu^\dagger a_\nu^\dagger +\varphi_n^1(\mu\nu) a_\nu a_\mu \right\} +\sum_{\bar{\mu} < \bar{\nu}} \left\{ \psi_n^1(\bar{\mu}\bar{\nu}) a_{\bar\mu}^\dagger a_{\bar\nu}^\dagger +\varphi_n^1(\bar{\mu}\bar{\nu}) a_{\bar\nu} a_{\bar\mu} \right\} \ , \label{normal-mode-1}\end{aligned}$$ where indices $n$ specify excited states and $\psi_n^\alpha(\mu\nu)$ ($\varphi_n^\alpha(\mu\nu)$) are the RPA forward (backward) amplitudes. Quasiparticle-scattering terms such as $a_\mu^\dagger a_\nu$ are regarded as higher-order terms in the boson-expansion theory and are neglected in the RPA[^2]. The equation of motion and the normalization condition in the RPA theory, $$\begin{aligned} \left[ h_{\rm s.p.}+H_{\rm int}, {X_n^\alpha}^\dagger \right]_{\rm RPA} &=& \hbar\Omega_n^\alpha {X_n^\alpha}^\dagger \ , \label{RPA-eom}\\ \left[ X_n^\alpha, {X_n^\alpha}^\dagger \right]_{\rm RPA} &=& \delta_{nn'} \ , \label{RPA-norm}\end{aligned}$$ are solved with the following multi-dimensional response functions: $$S_{\rho\rho'}^\alpha (\Omega) = \sum_{\gamma\delta} \left\{ \frac{{R_\rho^\alpha(\gamma\delta)}^* R_{\rho'}^\alpha(\gamma\delta)} {E_\gamma+E_\delta-\hbar\Omega} +\frac{R_\rho^\alpha(\gamma\delta) {R_{\rho'}^\alpha(\gamma\delta)}^*} {E_\gamma+E_\delta+\hbar\Omega} \right\} \ ,$$ where $(\gamma\delta)=(\mu\bar\nu)$ for $\alpha=0$ states, and $(\gamma\delta)=(\mu<\nu),(\bar\mu<\bar\nu)$ for $\alpha=1$ states. The two-quasiparticle matrix elements $R_\rho^\alpha(\gamma\delta)$ are defined by $R_\rho^\alpha(\gamma\delta)\equiv{\langle {\omega_{\rm rot}}|}a_\delta a_\gamma R_\rho^\alpha {| {\omega_{\rm rot}}\rangle}$. Let us denote the transition matrix elements between the RPA excited states ${| n \rangle}$ and the yrast state as $$t_\rho^\alpha (n) \equiv t_n\left[ R_\rho^\alpha \right] \equiv {\langle {\omega_{\rm rot}}|} R_\rho^\alpha {| n \rangle} = {\langle {\omega_{\rm rot}}|} \left[ R_\rho^\alpha, {X_n^\alpha}^\dagger \right] {| {\omega_{\rm rot}}\rangle} = \left[ R_\rho^\alpha, {X_n^\alpha}^\dagger \right]_{\rm RPA} \ .$$ Then, the equation of motion (\[RPA-eom\]) is equivalent to $$t_\rho^\alpha (n) = \sum_{\rho'} \chi_\rho^\alpha S_{\rho\rho'}^\alpha(\Omega) t_{\rho'}^\alpha (n) \ . \label{RPA-eq}$$ RPA solutions (eigen-energies) $\hbar\Omega_n$ are obtained by solving the equation, $$\det \left( S_{\rho\rho'}^\alpha (\Omega) - \frac{1}{\chi_\rho}\delta_{\rho\rho'} \right) = 0\ , \label{RPA_dispersion_eq}$$ which corresponds to the condition that eq.(\[RPA-eq\]) has a non-trivial solution ($t_\rho^\alpha (n)\neq 0$). Each RPA eigenstate is characterized by the corresponding forward and backward amplitudes which are calculated as $$\psi_n^\alpha(\gamma\delta) = \frac{\sum_\rho \chi_\rho^\alpha t_\rho^\alpha (n) R_\rho^\alpha(\gamma\delta)} {E_\gamma + E_{\delta} -\hbar\Omega_n}\ ,\quad \varphi_n^\alpha(\gamma\delta) = \frac{-\sum_\rho \chi_\rho^\alpha t_\rho^\alpha (n) {R_\rho^\alpha(\gamma\delta)}^*} {E_\gamma + E_\delta +\hbar\Omega_n}\ ,$$ and satisfies the normalization condition (\[RPA-norm\]). The transition matrix elements ${\langle {\omega_{\rm rot}}|} Q {| n \rangle}$ of any one-body operator $Q$ can be expressed in terms of these amplitudes $\psi_n$ and $\varphi_n$. $$\begin{aligned} t_n\left[ Q \right] &\equiv& {\langle {\omega_{\rm rot}}|} Q {| n \rangle} \nonumber \\ &=& \sum_{\gamma\delta}\left\{ {Q(\gamma\delta)}^* \psi_n(\gamma\delta) - Q(\gamma\delta) \varphi_n(\gamma\delta) \right\} \ . \label{Q-amplitude}\end{aligned}$$ The phase relation between the matrix elements $Q(\gamma\delta)$ and the amplitudes ($\psi_n(\gamma\delta),\varphi_n(\gamma\delta)$) is very important, because it determines whether the transition matrix element $t_n\left[ Q \right]$ is coherently enhanced or canceled out after the summation in eq.(\[Q-amplitude\]). For instance, a collective quadrupole vibrational state has a favorable phase relation for the quadrupole operators. Therefore, it gives large matrix elements for the $E2$ operators, while for the M1 operators, the contributions are normally canceled out after the summation. Finally we obtain a diagonal form of the total Hamiltonian in the rotating frame by means of the RPA theory, $$H' = h_{\rm s.p.} + H_{\rm int} \approx \mbox{const.} + \sum_{n,\alpha} \hbar\Omega_n^\alpha {X_n^\alpha}^\dagger X_n^\alpha \ .$$ It is worth noting that since the effect of the cranking term on the quasiparticles depends on rotational frequency, the effects of Coriolis coupling on the RPA eigenstates are automatically taken into account. Details of calculations {#sec: detail} ======================= The mean-field parameters and the improved quasiparticle routhians {#sec: detail-MF} ------------------------------------------------------------------ We adopt standard values for the parameters $v_{ll}$ and $v_{ls}$[@BR85] and use different values of the oscillator frequency $\omega_0$ for neutrons and protons in the Nilsson potential (\[Nilsson\]) in order to ensure equal root-mean-square radii[@BK68]. $$\label{oscillator_frequency} \omega_0 \longrightarrow \left\{ \begin{array}{ll} \left(\frac{2N}{A}\right)^{1/3} \omega_0\ , & \mbox{for neutrons} \ ,\\ \left(\frac{2Z}{A}\right)^{1/3} \omega_0\ , & \mbox{for protons} \ , \end{array} \right.$$ where $\hbar\omega_0=41A^{-1/3}$MeV. The quadrupole deformation $\epsilon$ is determined by minimizing the total routhian surface (TRS), and the strength for the monopole pairing interaction $G$ is taken from the prescription of Ref.[@Bra72] with the average pairing gap $\tilde\Delta = 12A^{-1/2}$MeV and the cut-off parameter of the pairing model space $\Lambda = 1.2\hbar\omega_0$. In principle the pairing gaps ($\Delta_n,\Delta_p$) and the chemical potentials ($\lambda_n,\lambda_p$) should be calculated self-consistently satisfying the usual BCS conditions at each rotational frequency: $$\begin{aligned} G_\tau {\langle {\omega_{\rm rot}}|}P_\tau {| {\omega_{\rm rot}}\rangle} &=& \Delta_\tau\ , \label{gap_equation}\\ {\langle {\omega_{\rm rot}}|} N_\tau {| {\omega_{\rm rot}}\rangle} &=& N (Z) \quad\mbox{ for } \tau=n(p)\ . \label{number_equation}\end{aligned}$$ However, the mean-field treatment of the pairing interaction predicts a sudden collapse of the proton pairing gap at $\hbar{\omega_{\rm rot}}\approx0.3\mbox{MeV}$ and of the neutron gap at $\hbar{\omega_{\rm rot}}\approx0.5\mbox{MeV}$. This transition causes a singular behavior in the moments of inertia which is inconsistent with experimental observations. It arises from the poor treatment of number conservation, and such sudden transitions should not occur in a finite system like the nucleus. In this paper we have therefore adopted the following phenomenological prescription for the pairing correlations at finite frequency[@Wys90]: $$\label{phenom_pairing} \Delta_\tau(\omega) = \left\{ \begin{array}{ll} \Delta_\tau(0) \left(1-\frac{1}{2} \left( \frac{\omega}{\omega_c} \right)^2\right), & \mbox{for }\omega < \omega_c, \\ \frac{1}{2}\Delta_\tau({0}) \left( \frac{\omega_c}{\omega} \right)^2, & \mbox{for }\omega > \omega_c, \end{array} \right.\ .$$ The chemical potentials are calculated with eq.(\[number\_equation\]) at each rotational frequency. The parameters $\Delta(0)=0.8$ (0.6) MeV and $\hbar\omega_c=0.5$ (0.3) MeV for neutrons (protons) are used in common for $^{190,192,194}$Hg. The quadrupole deformation $\epsilon=0.44$ is used in the calculations. For simplicity, we assume the deformation to be constant with rotational frequency, and neglect hexadecapole deformation. The equilibrium deformation and pairing gaps have been determined at ${\omega_{\rm rot}}=0$, with the truncated pairing model space $\Lambda=1.2\hbar\omega_0$. Then, the pairing force strengths $G_\tau$ are adjusted so as to reproduce the pairing gap of eq.(\[phenom\_pairing\]) in the whole model space. The experiments[@Cro94; @Cro95] have reported a sharp rise of ${{\cal J}^{(2)}}$ moments of inertia for the yrast SD band in [$^{190}$Hg]{}at $\hbar{\omega_{\rm rot}}\approx 0.4 \mbox{MeV}$. This rise was reproduced in the cranked Woods-Saxon calculations[@Dri91] and results from a crossing between the yrast band and the aligned $\nu(j_{15/2})^2$ band, however, the predicted crossing frequency was lower ($\hbar{\omega_{\rm rot}}\approx 0.3\mbox{MeV}$) than in the experiment. Our Nilsson potential without the additional term (\[additional-term\]) indicates the same disagreement. In order to demonstrate the effects of the term $h_{\rm add}$ on the routhians, we present in Fig. \[hg190\_routh\_comp\] the quasiparticle routhians for [$^{190}$Hg]{} with $h_{\rm add}$, without $h_{\rm add}$, and for the standard Woods-Saxon potential ($\beta_2=0.465$, $\beta_4=0.055$). By including, $h_{\rm add}$, the correct frequency is reproduced. This term affects the proton routhians: for example, the alignment of the intruder $\pi j_{15/2} (\alpha=-1/2)$ orbit is predicted to be $i\approx 6.5\hbar$ without $h_{\rm add}$ and this orbit becomes the lowest at $\hbar{\omega_{\rm rot}}\geq 0.37\mbox{MeV}$. The alignment is significantly reduced ($i\approx 4\hbar$) with $h_{\rm add}$. The behavior of high-$N$ intruder orbits in the proton routhians are similar to those in the Woods-Saxon potential. It is worth noting that the conventional renormalization in the Nilsson potential scales the rotational frequency for all orbits, while eq.(\[additional-term\]) renormalizes alignment in a different way depending on the spurious effect on each orbit. The residual interactions and the RPA {#sec: detail-RPA} ------------------------------------- We adopt the following operators as $R_\rho^\alpha$ in the residual interactions (\[residual-int\]). $$\label{operator_R} \begin{array}{lllllll} P_+ & P_- & Q_{20}^0 & Q_{21}^\alpha & Q_{22}^\alpha & & \mbox{for positive-parity states}\ ,\\ Q_{30}^1 & Q_{31}^\alpha & Q_{32}^\alpha & Q_{33}^\alpha & \tilde\tau_3 Q_{10}^1 & \tilde\tau_3 Q_{11}^\alpha & \mbox{for negative-parity states}\ , \end{array}$$ where $\tilde\tau_3 = \tau_3 -(N-Z)/A$ which is needed to guarantee the translational invariance. Here, the operators $Q_{\lambda K}^\alpha$ are defined by eq.(\[multipole-op\]) in the doubly-stretched coordinates, and $P_\pm$ are defined by $$\begin{aligned} P_+ &=& \frac{1}{\sqrt{2}}\left(\tilde{P} + \tilde{P^\dagger}\right)\ ,\\ P_- &=& \frac{i}{\sqrt{2}}\left(\tilde{P} - \tilde{P^\dagger}\right)\ ,\end{aligned}$$ where $\tilde{P}=P-{\langle {\omega_{\rm rot}}|}P{| {\omega_{\rm rot}}\rangle}$. Note that the $K=0$ quadrupole (octupole) operator $Q_{20}$ ($Q_{30}$) has a unique signature $\alpha=0$ ($\alpha=1$), which corresponds to the fact that $K=0$ bands have no signature partners. Since we use the different oscillator frequency $\omega_0$ for neutrons and protons in the Nilsson potential (see eq.(\[oscillator\_frequency\])), we use the following modified doubly-stretched multipole operators for the isoscalar channels: $$Q_{\lambda K}^\alpha \longrightarrow \left\{ \begin{array}{ll} \left(\frac{2N}{A}\right)^{2/3} Q_{\lambda K}^\alpha\ , & \mbox{for neutrons}\ ,\\ \left(\frac{2Z}{A}\right)^{2/3} Q_{\lambda K}^\alpha\ , & \mbox{for protons}\ . \end{array} \right.$$ This was originally proposed by Baranger and Kumar[@BK68] for quadrupole operators. Recently Sakamoto[@Sak93] has generalized it for an arbitrary multipole operator and proved that by means of this scaling the translational symmetry is restored in the limit of the harmonic-oscillator potential. In addition, for the collective RPA solutions this treatment makes the transition amplitudes of the electric operators approximately $Z/A$ of those of the mass operators, in the same way as in the case of the static quadrupole moments[@SM95]. We use the pairing force strengths $G_\tau$ reproducing the pairing gaps of eq.(\[phenom\_pairing\]). For the isovector dipole coupling strengths, we adopt the standard values in Ref.[@BM75], $$\chi_{1K}=-\frac{\pi V_1}{A\langle (r^2)^{''} \rangle_0}\ ,$$ with $A\langle (r^2)^{''} \rangle_0 = \langle \sum_k^A (r_k^2)^{''} \rangle_0$ and $V_1 = 130$MeV. The self-consistent values for the coupling strengths $\chi_{\lambda K}$ of the isoscalar quadrupole and octupole interactions can be obtained for the case of the anisotropic harmonic-oscillator potential[@SK89; @Sak93]: $$\begin{aligned} \chi_{2K}^{\rm HO} &=& \frac{4\pi M\omega_0^2}{5A\langle(r^2)^{''}\rangle}\ ,\\ \chi_{3K}^{\rm HO} &=& {4\pi \over 7} M\omega^2_0 \biggl\{A\langle (r^4)^{''} \rangle + {2 \over 7}(4 - K^2) A\langle (r^4P_2)^{''}\rangle \nonumber\\ &+& {1 \over 84}(K^2(7K^2 - 67) + 72) A\langle (r^4P_4)^{''} \rangle \biggr\}^{-1}\ ,\end{aligned}$$ with $$A\langle (r^n P_\l )\rangle \equiv \left(\frac{2N}{A}\right)^{2/3} \langle \sum_k^N (r_k)^n P_\l \rangle_0 + \left(\frac{2Z}{A}\right)^{2/3} \langle \sum_k^Z (r_k)^n P_\l \rangle_0\ .$$ A large model space has been used for solving the coupled RPA equations, including seven major shells with $N_{\rm osc}=3\sim 9$ $(2\sim 8)$ for neutrons (protons) in the calculations of positive-parity states, and nine major shells with $N_{\rm osc}=2\sim 10$ $(1\sim 9)$ for the negative-parity states. The mesh of the rotational frequency for the calculations has been chosen as ${\it\Delta}\hbar{\omega_{\rm rot}}=0.01 \mbox{MeV}$ which is enough to discuss the properties of band crossing and Coriolis couplings. Since our mean-field potential is not the simple harmonic oscillator, we use scaling factors $f_\lambda$ as $$\chi_{\lambda K} = f_\lambda \cdot \chi_{\lambda K}^{\rm HO}\ , \label{coupling-strength}$$ for the isoscalar interactions with $\lambda=2$ and 3. These factors are determined by the theoretical and experimental requirements: As for the octupole interactions, we have the experimental routhians for the lowest octupole vibrational state in SD [$^{190}$Hg]{}[@Cro95]. We assume the common factor $f_3$ for all $K$-values and fix it so as to reproduce these experimental data. In this case $f_3=1$ can nicely reproduce the experimental routhians[^3], and we use the same value for [$^{192}$Hg]{} and [$^{194}$Hg]{}. For the quadrupole interactions, we determine it so as to reproduce the zero-frequency (Nambu-Goldstone) mode for $K=1$ at ${\omega_{\rm rot}}=0$ and use the same value for $K=0$ and 2. $f_2=1.007$, 1.005, and 1.005 are obtained for [$^{190}$Hg]{}, [$^{192}$Hg]{}, and [$^{194}$Hg]{}, respectively, by using the adopted model space. The fact that these values of $f_\lambda$ are close to unity indicates that the size of the adopted model space is large enough. According to systematic RPA calculations for the low-frequency $\beta$, $\gamma$, and octupole states in medium-heavy deformed nuclei, we have found that the values of $f_\lambda$ reproducing the experimental data are very close to unity for the Nambu-Goldstone mode, the $\gamma$ and octupole vibrational states. On the other hand, those values are quite different from unity for the $\beta$ vibrational states. This may be associated with the simplicity of the monopole pairing interaction. Since we can not find the realistic force strength $\chi_{20}$ for SD states, we do not discuss the property of the $\beta$ vibrations in this paper. The results of numerical calculations {#sec: results} ===================================== Quasiparticle routhians ----------------------- In this section we present calculated quasiparticle routhians in the improved cranked Nilsson potential and discuss their characteristic feature. In Fig. \[neutron\_routhians\] we compare the neutron quasiparticle routhians for $^{190,192,194}$Hg. The proton routhians of [$^{190}$Hg]{} are shown above in Fig. \[hg190\_routh\_comp\] and are almost identical for [$^{192}$Hg]{} and [$^{194}$Hg]{}. The calculations show the strong interaction strength between the $\pi($\[642 5/2\]$)^2$ configuration (for simplicity we denote these orbits by $\pi 6_1$ and $\pi 6_2$ in the following) and the yrast configuration which may contribute to the smooth increase of the yrast ${{\cal J}^{(2)}}$ moments of inertia. On the other hand, the interaction of $\nu$\[761 3/2\] orbits ($\nu 7_1$ and $\nu 7_2$ in the following) strongly depends on the chemical potential (neutron number): The interaction is strongest in [$^{194}$Hg]{}, and weakest in [$^{190}$Hg]{}. This is qualitatively consistent with the experimental observation of the yrast ${{\cal J}^{(2)}}$ moments of inertia and the experimental quasiparticle routhians in $^{191,193}$Hg[@Joy94; @Car95]. The characteristic features of the high-$N$ intruder orbits are similar to those of the Woods-Saxon potential, except the alignments of $\nu 7_1$ and $\nu 7_2$ orbits which are, respectively, $i\approx 3\hbar$ and $2\hbar$ in ours while $i\approx 4\hbar$ and $3\hbar$ in Woods-Saxon’s. This results in the different crossing frequency between the ground band and the $\nu (j_{15/2})^2$ band, as discussed in section \[sec: detail-MF\]. The observed crossing in [$^{190}$Hg]{} and the quasiparticle routhians in $^{191,193}$Hg seem to favor our results. There are some other minor differences concerning the position of each orbit in the Nilsson and in the Woods-Saxon potential. However, these differences do not seriously affect our main conclusions because the collective RPA solutions are not sensitive to the details of each orbit. The octupole vibrations {#sec: octupole-results} ----------------------- Here, we discuss the negative-parity excitations in SD $^{190,192,194}$Hg. We have solved the RPA dispersion equation (\[RPA\_dispersion\_eq\]) and have obtained all low-lying solutions ($E'_x \leq 2$MeV). The excitation energies and the $B(E3)$ values calculated at ${\omega_{\rm rot}}=0$ are listed in Table \[oct-E-E3\]. This result shows that $K=2$ octupole states are the lowest for these Hg isotopes, which is consistent with our previous results[@NMM93; @Miz93]. The $B(E3; 0^+ \rightarrow 3^-,K)$ are calculated by using the strong coupling scheme[@BM75] neglecting effects of the Coriolis force. Absolute values of $B(E3)$s cannot be taken seriously because they depend on the adopted model space and are very sensitive to the octupole coupling strengths $\chi_{3K}$: For instance, if we use $f_3=1.05$ instead of $f_3=1$ in eq.(\[coupling-strength\]), the $B(E3)$ increase by about factor of two while the reduction of their excitation energy is about 15%. In addition, the effects of the Coriolis coupling tend to concentrate the $B(E3)$ strengths onto the lowest octupole states[@NV70]. At ${\omega_{\rm rot}}=0$, the lowest $K=2$ octupole states exhibit almost identical properties in $^{190,192,194}$Hg. However they show different behavior as functions of ${\omega_{\rm rot}}$ as shown in Figs. \[oct\_hg190\], \[oct\_hg192\], and \[oct\_hg194\], respectively. All RPA solutions, including non-collective solutions as well as collective vibrational ones, are presented in these figures. The size of the circle on the plot indicates the magnitude of the $E3$ transition amplitudes between an RPA solution and the yrast state. The $(K,\alpha)=(2,1)$ octupole state in [$^{190}$Hg]{} has significant Coriolis mixing and the octupole phonon is aligned along the rotational axis at higher frequency. This is caused by the relatively close energy spacing between the $K=2$ and the $K=0, 1$ octupole states in this nucleus. These low-$K$ members of the octupole multiplet are calculated to lie much higher in [$^{192}$Hg]{} and [$^{194}$Hg]{}, which reduces the Coriolis mixing in these nuclei. As a result of these phonon alignments, the experimental routhians for Band 2 in [$^{190}$Hg]{} are nicely reproduced by the lowest $\alpha=1$ octupole state. It should be emphasized that although the excitation energy at one frequency point can be obtained by adjusting the octupole-force strengths, the agreement over the whole frequency region is not trivial. Since there is no $K=0$ octupole state in the signature $\alpha=0$ sector, the Coriolis mixing is much weaker for the lowest $(K,\alpha)=(2,0)$ octupole state. The calculation predicts that this state is crossed by the negative-parity two-quasiparticle band $\nu(7_1\otimes[642\ 3/2])_{\alpha=0}$ at $\hbar{\omega_{\rm rot}}\approx 0.27\mbox{MeV}$. In [$^{192}$Hg]{}, the same kind of crossing is seen for both signature partners of the $K=2$ octupole bands. We can clearly see, for the lowest excited state in each signature sector, the transition of the internal structure from collective octupole states (large circles in Fig. \[oct\_hg192\]) to non-collective two-quasineutrons (small circles). The two-quasineutron configurations which cross the octupole vibrational bands are $7_1\otimes[642\ 3/2](\alpha=-1/2)$ for $\alpha=1$ and $7_1\otimes[642\ 3/2](\alpha=1/2)$ for $\alpha=0$. The crossing frequency is lower for the $\alpha=1$ band due to signature splitting of the $\nu$\[642 3/2\] orbits. In contrast to $^{190,192}$Hg, the $K=2$ octupole bands in [$^{194}$Hg]{} indicate neither the signature splitting nor the crossings. The routhians are very smooth up to the highest frequency. This is because the neutron orbits $7_1$ and $7_2$ have a “hole” character and their interaction strengths with the negative-energy orbits become larger with increasing neutron numbers (see Fig. \[neutron\_routhians\]). Therefore these orbits go to higher energy and the energies of the two-quasiparticle bands $\nu(7_1\otimes[642\ 3/2])$ never come lower than the $K=2$ octupole bands even at the highest frequency. These properties of the $K=2$ octupole vibrations come from the effects of the Coriolis force and from the chemical-potential dependence of the aligned two-quasiparticle bands. In order to reproduce these rich properties of the collective vibrations at finite frequency, a microscopic model, which can describe the interplay between the Coriolis force and the correlations of shape fluctuations, is needed. The $\gamma$ vibrations {#sec: gamma-results} ----------------------- In this section we present results for the $\gamma$-vibrational states built on the SD yrast band. As mentioned in section \[sec: detail-RPA\], we do not discuss the property of the $\beta$ band since it is difficult to determine a reliable value of the coupling strength $\chi_{20}$ for the K=0 channel of the quadrupole interaction. The properties of $\gamma$ bands at ${\omega_{\rm rot}}=0$ are listed in Table \[gamma-E-E2\]. The excitation energies of $\gamma$ vibrations are predicted to be higher than the $K=2$ octupole vibrations by 200–350 keV. It is known that calculations using the full model space considerably overestimate the $B(E2)$ values. In Ref.[@SM95], it has been shown that the three $N_{\rm osc}$-shells calculation reproduces the experimental values very well. If we use the model space $N_{\rm osc}=5\sim 7$ ($4\sim 6$) for neutrons (protons), then the $B(E2)$ values in the table decrease by about factor 1/3. The collectivity of the $\gamma$ vibrations turns out to be very weak in these SD nuclei. Figs. \[quad\_hg190\], \[quad\_hg192\], and \[quad\_hg194\] illustrate the excitation energy of $\gamma$ vibrations as functions of the rotational frequency for [$^{190}$Hg]{}, [$^{192}$Hg]{}, and [$^{194}$Hg]{}, respectively. The unperturbed two-quasiparticle routhians are also depicted by solid (neutrons) and dashed (protons) lines. Since the $K$ quantum number is not a conserved quantity at finite rotational frequency, we have defined the solutions with the large $K=2$ $E2$ transition amplitude as the $\gamma$ vibrations. As seen in the figure, they lose their vibrational character by successive crossings with many two-quasiparticle bands and become the dominant two-quasiparticle states at high frequency. The reduction of collectivity is more rapid for the $\alpha=0$ $\gamma$ vibrations, because the two-quasiparticle states come down more quickly in the $\alpha=0$ sector. Similar crossings occur for the $K=2$ octupole bands in [$^{192}$Hg]{} (see Fig. \[oct\_hg192\]), however, the crossing frequency is much higher than that of the $\gamma$ bands. This is because the excitation energies of the octupole bands are relatively lower than those of the $\gamma$ bands. The predicted properties of $\gamma$ vibrations are different from those in Ref.[@Gir92]. In the frequency region ($0.15\leq\hbar{\omega_{\rm rot}}\leq 0.4\mbox{MeV}$) where the excited SD bands are observed in experiments, the $\gamma$ bands are predicted to be higher than both the $K=2$ octupole bands and the lowest two-quasiparticle states. Therefore the experimental observation of the $\gamma$ vibrations is expected to be more difficult than that of the octupole bands. Comparison with experimental data {#sec: comparison} ================================= In this section, we compare the results obtained in the previous section with the available experimental data for the excited SD bands in $^{190,192,194}$Hg. The routhians relative to the yrast SD band have been observed only for Band 2 in [$^{190}$Hg]{} and the comparison with our calculated routhians has been done in the section \[sec: octupole-results\]. The excitation energies of the other bands are not known. Therefore, in order to compare our theory with experimental data, we have calculated the dynamic moments of inertia ${{\cal J}^{(2)}}$. The ${{\cal J}^{(2)}}$ of the excited bands are calculated as $${{\cal J}^{(2)}}(\omega) = {{\cal J}^{(2)}}_0(\omega) + \frac{di}{d\omega} = {{\cal J}^{(2)}}_0(\omega) - \frac{d^2 E'_x}{d\omega^2} \ , \label{J2-inertia}$$ where ${{\cal J}^{(2)}}_0$ denotes the dynamic moments of inertia for the yrast SD bands (RPA vacuum), and $i$ and $E'_x$ are the calculated alignments and routhians relative to the yrast band, respectively. The ${{\cal J}^{(2)}}_0$ values of the yrast SD bands are taken from the experiments and approximated by the Harris expansion, $${{\cal J}^{(2)}}_0(\omega) = J_0 + 3 J_1 \omega^2 + 5 J_2 \omega^4 \ . \label{Harris}$$ It is known that the effect of pairing fluctuations is important in reproducing the moments of inertia at high spin. However, since our model provides us with relative quantities (excitation energy, alignment, etc.) between the excited bands and the yrast band, it is not critical if we neglect the pairing fluctuations. In other words, the fluctuations are included in the experimental ${{\cal J}^{(2)}}_0$ of eq.(\[J2-inertia\]). The lower the excitation energy of an excited band relative to the yrast SD band, the more strongly will it be populated. In experiments, the SD bands are populated at high frequency, thus, it is the excitation energy in the feeding region at high frequency that is relevant in this problem. We list in Table \[Ex\] the calculated excitation energies of the low-lying positive- and negative-parity states at $\hbar{\omega_{\rm rot}}=0.4\mbox{MeV}$. In [$^{190}$Hg]{} three excited SD bands (Bands 2, 3 and 4) have been observed[@Cro94; @Cro95; @Wil95]. Band 2 has been assigned as the lowest octupole band[@Cro94; @Cro95] because of its strong decays into the yrast SD band. According to our calculations, in addition to this octupole band ($\alpha=1$), the aligned two-quasineutron bands come down at high frequency. We assign Band 4 at high frequency as the $\nu (7_1 \otimes [642\ 3/2])_{\alpha=0}$ because this negative-parity two-quasineutron state is crossed by the $\alpha=0$ octupole band at $\hbar{\omega_{\rm rot}}\approx 0.26\mbox{MeV}$ which may correspond to the observed sharp rise of ${{\cal J}^{(2)}}$ at $\hbar{\omega_{\rm rot}}\approx 0.23\mbox{MeV}$ (Fig. \[oct\_hg190\]). The positive-parity $\nu (7_1 \otimes 7_2)_{\alpha=0}$ state is also relatively low-lying at high frequency. Since this band does not show any crossing at $\hbar{\omega_{\rm rot}}> 0.12\mbox{MeV}$ in the calculations, this may be a good candidate for Band 3 (Fig. \[quad\_hg190\]). In [$^{192}$Hg]{} two excited SD bands (Bands 2 and 3) have been observed[@Fal95] and both bands exhibit a bump in ${{\cal J}^{(2)}}$ at $\hbar{\omega_{\rm rot}}\approx 0.3$ (Band 2) and $0.33\mbox{MeV}$ (Band 3). We assume these bands correspond to $\nu(7_1\otimes [642\ 3/2])_{\alpha=0,1}$ at high frequency. This two-quasineutron configuration for Band 2 is the same as that suggested in Ref.[@Fal95]. However our theory predicts a different scenario at low spin: This band is crossed by the octupole band ($\alpha=1$) at $\hbar{\omega_{\rm rot}}\approx 0.3\mbox{MeV}$. Thus, Band 2 is interpreted as an $\alpha=1$ octupole vibrational band in the low-frequency region ($\hbar{\omega_{\rm rot}}<0.3\mbox{MeV}$). In the same way, the bump in ${{\cal J}^{(2)}}$ in Band 3 is interpreted as a crossing between $\nu(7_1\otimes [642\ 3/2])_{\alpha=0}$ and the $\alpha=0$ octupole vibrational band (Fig. \[oct\_hg192\]). For high frequencies, the positive-parity $\nu(7_1 \otimes [512\ 5/2])$ state is calculated to lie almost at the same energy as the lowest $\alpha=0$ negative-parity state. However no crossing is predicted for the $\alpha=1$ state at $\hbar{\omega_{\rm rot}}> 0.15\mbox{MeV}$ but many crossings are predicted for the $\alpha=0$ state (Fig. \[quad\_hg192\]). Both properties are incompatible with the observed features. In [[$^{194}$Hg]{}]{}, two excited SD bands (Bands 2 and 3) have been observed[@Ril90; @Ced95]. In contrast to [$^{192}$Hg]{}, the observed dynamic moments of inertia ${{\cal J}^{(2)}}$ do not show any singular behavior and are more or less similar to those of the yrast band. Bands 2 and 3 have been interpreted as signature partners because the $\gamma$-ray energies of Band 3 are observed to lie mid-way between those of Band 2 and furthermore the bands have similar intensity[@Ril90]. From these observations and the excitation energies listed in Table \[Ex\], we assume that both bands correspond to $K=2$ octupole vibrations ($\alpha=0$, 1), which are calculated to be the lowest excited states (Fig. \[oct\_hg194\]). Any other assignment faces serious difficulties: (i) The positive-parity two-quasiparticle configurations listed in Table \[Ex\] have no signature partners. (ii) The other low-lying two-quasiparticle states occupy $\nu 7_1$ or $\pi 6_1$ orbits. Now the increase in ${{\cal J}^{(2)}}$ for the yrast SD band is partially attributed to the alignment of these high-$j$ intruder orbits and, since the blocking effect of the quasiparticles prevents any alignment due to band crossings involving these orbits, the lack of alignment should produce an ${{\cal J}^{(2)}}$ curve quite different from those of the yrast SD band. (iii) The configuration $\nu ([512\ 5/2]\otimes [624\ 9/2])$ suggested in Ref.[@Ril90] has the problem with its magnetic property, which has been recently pointed out in Ref.[@SRA95]. If this configuration is the $K^\pi=7^-$, then strong M1 transitions between the signature partners should have been observed. The energy of the $K^\pi=2^-$ configuration is certainly lowered by octupole correlations. In our calculations, however, this configuration accounts for only 20% of all components constituting the octupole vibration. The $\gamma$ vibrations are calculated to be much higher and crossed by several two-quasiparticle bands (Fig. \[quad\_hg194\]). Therefore, we believe the octupole vibration is the best candidate[^4]. Assuming the above configurations, the dynamic moments of inertia ${{\cal J}^{(2)}}$ are calculated with eq.(\[J2-inertia\]), and compared with the experimental data (Fig. \[J2\]). In [$^{190}$Hg]{}, the characteristic features are well reproduced for Bands 2 and 4; the constant ${{\cal J}^{(2)}}$ of Band 2 (the $\alpha=1$ octupole vibration), and the bump of Band 4 (the crossing between the $\alpha=0$ octupole vibration and the aligned two-quasineutron band) are reproduced although the crossing frequency is smaller in the experiment. For Band 3, the high ${{\cal J}^{(2)}}$ values at low spin are well accounted for by the alignment-gain of the two-quasineutron state. However the calculation predicts the lack of alignment due to the blocking of $N=7$ orbits at $\hbar{\omega_{\rm rot}}>0.25\mbox{MeV}$, which makes the ${{\cal J}^{(2)}}$ smaller than those of the yrast band. In [$^{192}$Hg]{}, the bumps of ${{\cal J}^{(2)}}$ are nicely reproduced in the calculations, which correspond to the crossings between $K=2$ octupole vibrations and the aligned two-quasineutron bands in each signature partner. The alignment gain ${\it \Delta}i$ before and after crossing for Band 2 is ${\it\Delta} i\approx 2\hbar$ which is comparable to the experimental value ${\it\Delta} i_{\rm exp}\approx 2.6\hbar$[@Fal95]. The agreement is less satisfactory in [$^{194}$Hg]{}. The calculated ${{\cal J}^{(2)}}$ are lower than the experimental data for $0.2\leq\hbar{\omega_{\rm rot}}\leq 0.35\mbox{MeV}$ (similar disagreement can be seen for Band 3 in [$^{192}$Hg]{}). This effect comes from the blocking effect mentioned above, associated with the $\nu 7_1$, $\nu 7_2$, $\pi 6_1$ and $\pi 6_2$ orbits. In the RPA (Tamm-Dancoff) theory (neglecting the backward amplitudes), the octupole vibrations are described by superposition of two-quasiparticle excitations, $${| \rm oct.vib. \rangle} = \sum_{\gamma\delta} \psi(\gamma\delta){| \gamma\delta \rangle}\ ,$$ where ${| \gamma\delta \rangle}=a_\gamma^\dagger a_\delta^\dagger {| {\omega_{\rm rot}}\rangle}$. Some of these components ${| \gamma\delta \rangle}$ associated with the particular orbits ($\nu 7_1$, $\nu7_2$, $\pi 6_1$ and $\pi6_2$) show significant lack of alignment. However, if the octupole vibrations are collective enough, the amplitudes $\psi(\gamma\delta)$ are distributed over many two-quasiparticle excitations ${| \gamma\delta \rangle}$. Thus, each amplitude becomes small and blocking effects may be canceled. In order to demonstrate this “smearing” effect of collective states, we use a slightly stronger octupole force, $f_3=1.05$ in eq.(\[coupling-strength\]), and carry out the same calculations for [$^{194}$Hg]{}. The results are shown in Fig. \[J2\_hg194\]. The higher coupling strengths make the octupole vibrations more collective and the experimental data are better reproduced. Perhaps the collectivity of these octupole vibrations was underestimated in the calculations with $f_3=1$. Finally we should mention the decays from the octupole bands to the yrast SD band. We have assigned all observed excited SD bands (except Band 3 in [$^{190}$Hg]{}) as octupole vibrational bands (at least in the low-spin region). However, strong dipole decays into the yrast band have been observed only for Band 2 in [$^{190}$Hg]{}. Although this seems to contradict our proposals, in fact our calculations provide us with a qualitative answer. Let us discuss the relative $B(E1; {\rm oct}\rightarrow {\rm yrast})$ values. Using the $E1$ recoil charge ($-Ze/A$ for neutrons and $Ne/A$ for protons), then the $B(E1)$ values at $\hbar{\omega_{\rm rot}}=0.25\mbox{MeV}$ are calculated to be small for all the $K=2$ octupole bands except for the $\alpha=1$ (Band 2) in [$^{190}$Hg]{}: With the scaling factors $f_3=1\sim 1.08$ in eq.(\[coupling-strength\]), the calculation suggests $B(E1)\approx 10^{-7}$ W.u. for the $(K,\alpha)=(2,0)$ octupole bands, and $B(E1)\approx 10^{-8}\sim 10^{-6}$ W.u. for the $(K,\alpha)=(2,1)$ bands. The $B(E1)$ for Band 2 in [$^{190}$Hg]{} is predicted to be larger than these values by 1 – 2 orders of magnitude, $B(E1)\approx 10^{-6}\sim 10^{-4}$W.u. Although the absolute values are very sensitive to the parameters used in the calculation, the $E1$ strengths of Band 2 in [$^{190}$Hg]{} are always much larger than those for the other bands. To clarify the reason for this $E1$ enhancement in this particular band, we display the $E3$ amplitudes ($K=0$, 1, 2 and 3) of these octupole states as functions of frequency in Fig. \[E3\_amplitudes\]. As mentioned in section  \[sec: octupole-results\], the Coriolis mixing is completely different between Band 2 in [$^{190}$Hg]{} and the others: The former has significant Coriolis mixing at finite frequency while the latter retains the dominant $K=2$ character up to very high spin. Since the $K=2$ octupole components can not carry any $E1$ strength, the strong $E1$ transition amplitudes come from Coriolis coupling, namely the mixing of the $K=0$ and 1 octupole components. Therefore, the observed decay property does not contradict our interpretation. Conclusions {#sec: conclusion} =========== The microscopic structure of the $\gamma$ and the octupole vibrations built on the SD yrast bands in $^{190,192,194}$Hg were investigated with the RPA based on the cranked shell model. The $K=2$ octupole vibrations are predicted to lie lowest. To reproduce the characteristic features of the experimental data it was essential to include octupole correlations and the effect of rapid rotation explicitly. From the calculations, we assigned the following configurations to the observed excited bands: ---------------- ---------- ---------------------------------------------------------------------------------- [$^{190}$Hg]{} Band 2 : the rotationally-aligned $\alpha=1$ octupole vibration. Band 3 : the two-quasineutron band $\nu(7_1 \otimes 7_2)$. Band 4 : the $(K,\alpha)=(2,0)$ octupole vibration at low spin, the two-quasineutron band $\nu(7_1 \otimes [642\ 3/2])_{\alpha=0}$ at high spin. [$^{192}$Hg]{} Band 2 : the $(K,\alpha)=(2,1)$ octupole vibration at low spin, the two-quasineutron band $\nu(7_1 \otimes [642\ 3/2])_{\alpha=1}$ at high spin. Band 3 : the $(K,\alpha)=(2,0)$ octupole vibration at low spin, the two-quasineutron band $\nu(7_1 \otimes [642\ 3/2])_{\alpha=0}$ at high spin. [$^{194}$Hg]{} Band 2 : the $(K,\alpha)=(2,0)$ octupole vibration. Band 3 : the $(K,\alpha)=(2,1)$ octupole vibration. ---------------- ---------- ---------------------------------------------------------------------------------- With these assignments, most of the experimentally observed features were well accounted for in our theoretical calculations. The Coriolis force makes the lowest octupole state in [$^{190}$Hg]{} align along the rotational axis, while this effect is predicted to be very weak for other octupole states. This is due to the relatively low excitation energy of the $K=0$ ($\alpha=1$) octupole state in [$^{190}$Hg]{}, in which the close spacing in energy of the octupole multiplet makes the Coriolis mixing easier. This aligned octupole phonon in [$^{190}$Hg]{} reproduces the observed behavior for Band 2. Our interpretation for the excited SD bands in [$^{192}$Hg]{} solves a puzzle mentioned in Ref. [@Fal95] in which Band 2 was assigned as the two-quasineutron excitation $\nu (7_3\otimes [642\ 3/2])$. The bump in the ${{\cal J}^{(2)}}$ curve was considered to be associated with a crossing between the $\nu 7_1$ and $\nu$\[512 5/2\] orbits. According to this assignment, we expect similar properties for the observed crossing in [$^{192}$Hg]{} and $^{193}$Hg, and the difference of crossing frequencies and alignment gains was a puzzle. This is no longer a puzzle in our interpretation because the microscopic structure of Band 2 is the octupole vibration (before the crossing). Due to the correlation-energy gains, the excitation energies of the octupole vibrations should be lower than the unperturbed two-quasiparticle states. Therefore it is natural that the observed crossing frequency is larger than the one predicted by the quasiparticle routhians without the octupole correlations. Our interpretation also solves some difficulties in [$^{194}$Hg]{}: The smooth ${{\cal J}^{(2)}}$ behavior of Bands 2 and 3 can be explained by the “smearing” effect of the collective states. The non-observation of the expected strong M1 transitions between Bands 2 and 3[@SRA95] is solved by substituting the $K=2$ octupole vibrations for the two-quasineutron states $\nu ([512\ 5/2]\otimes [624\ 9/2])$, because the octupole correlations lower the $K=2$ configurations and the summation of many two-quasiparticle ($M1$) matrix elements may be destructive (see discussion below eq.(\[Q-amplitude\])). The enhanced $E1$ transitions from the octupole states to the yrast SD band are expected only for Band 2 in [$^{190}$Hg]{}. This comes about because the other octupole states do not have strong Coriolis mixing and keep their $K=2$ character even at high frequency. This agrees with experimental observations. Although most of the observed properties are explained by our calculations, there remain some unsolved problems in [$^{190}$Hg]{} and [$^{192}$Hg]{}. For [$^{190}$Hg]{}, according to the calculations with constant pairing gaps reported in Ref.[@Nak95_2], it is suggested that Band 4 may correspond to the $(K,\alpha)=(1,0)$ octupole band which is predicted to be crossed by the two-quasineutron band $\nu(7_1\otimes [642\ 3/2])_{\alpha=0}$ at $\hbar{\omega_{\rm rot}}\approx 0.21\mbox{MeV}$. Because of the phenomenological treatment for the pairing gaps at finite frequency, it is difficult to deny this possibility. The experimental intensity of Band 3 raises another ambiguity: Since it is much weaker than Bands 2 and 4, it might be associated with a higher-lying configurations[@Wil95]. For [$^{192}$Hg]{}, our calculations predict no signature splitting for the lowest octupole bands at $\hbar{\omega_{\rm rot}}\leq 0.25\mbox{MeV}$. Therefore one may expect $\gamma$-ray energies typical of the signature-partner pair for Bands 2 and 3 similar to that in [$^{194}$Hg]{}, which is different from what is observed[@Fal95]. Improvement of the pairing interactions (fluctuations, quadrupole pairing) might solve these problems as well as enable us to perform reliable RPA calculations for $\beta$ vibrations. Theoretical study of octupole vibrations carrying large $E1$ strengths would be of great interest, because this could offer direct experimental evidence. An improved version of calculations for $E1$ strengths of high-spin octupole bands are in progress, taking into account the restoration of translational and Galilean invariance. The $K=0$ octupole vibration in $^{152}$Dy has been predicted in Ref.[@Nak95] and its decay into the yrast band has been suggested [@Dag95]. Strong $E1$ transition probabilities have been suggested by Skalski[@Ska94] for $K=0$ octupole states in the A=190 region. Therefore, the search for low-lying low-$K$ octupole vibrations is an important subject for the future. We would like to acknowledge W. Nazarewicz for discussions and suggestions for this paper. One of authors (T.N.) also thank B. Crowell, P. Fallon, J.F. Sharpey-Schafer, J. Skalski and A.N. Wilson for valuable discussions. Three of us (T.N., K.M. and Y.R.S.) thank the Institute for Nuclear Theory at the University of Washington for its hospitality and the Department of Energy for partial support during the completion of this work. [99]{} S. Mizutori, Y.R. Shimizu and K. Matsuyanagi, Prog. Theor. Phys. [**83**]{}, 666 (1990); [**85**]{}, 559 (1991); [**86**]{}, 131 (1991). T. Nakatsukasa, S. Mizutori and K. Matsuyanagi, Prog. Theor. Phys. [**87**]{}, 607 (1992). T. Nakatsukasa, S. Mizutori and K. Matsuyanagi, Prog. Theor. Phys. [**89**]{}, 847 (1993). S. Mizutori, T. Nakatsukasa, K. Arita, Y.R. Shimizu and K. Matsuyanagi, Nucl. Phys. [**A557**]{}, 125c (1993). T. Nakatsukasa, K. Matsuyanagi, S. Mizutori and W. Nazarewicz, Phys. Lett. [**B343**]{}, 19 (1995). J. Dudek, T.R. Werner and Z. Szymański, Phys. Lett. [**B248**]{}, 235 (1990). S. Åberg, Nucl. Phys. [**A520**]{}, 35c (1990). J. Höller and S. Åberg, Z. Phys. [**A336**]{}, 363 (1990). P. Bonche, S.J. Krieger, M.S. Weiss, J. Dobaczewski, H. Flocard and P.-H. Heenen, Phys. Rev. Lett. [**66**]{}, 876 (1991). Xunjun Li, J. Dudek and P. Romain, Phys. Lett. [**B271**]{}, 281 (1991). W. Nazarewicz and J. Dobaczewski, Phys. Rev. Lett. [**68**]{}, 154 (1992). J. Skalski, Phys. Lett. [**B274**]{}, 1 (1992). J. Skalski, P.-H. Heenen, P. Bonche, H. Flocard and J. Meyer, Nucl. Phys. [**A551**]{}, 109 (1993). P.J. Dagnall et al., Phys. Lett. [**B335**]{}, 313 (1995). D.M. Cullen et al., Phys. Rev. Lett. [**65**]{}, 1547 (1990). B. Crowell et al., Phys. Lett. [**B333**]{}, 320 (1994). B. Crowell et al., Phys. Rev. [**C51**]{}, R1599 (1995). A.N. Wilson et al., Proceedings of European Physical Society XV Divisional Conference, St Petersburg, Russia, April 1995, [*Low Energy Nuclear Dynamics*]{} (LEND 95), (World Scientific), in press. T. Nakatsukasa, Act. Phys. Pol. B, in press; Proceedings of XXIV Mazurian Lakes School of Physics, Piaski, Poland, August 1995 (Preprint TASCC-P-95-26). K. Neergård and P. Vogel, Nucl. Phys. [**A145**]{}, 33 (1970); [**A149**]{}, 209 (1970); [**A149**]{}, 217 (1970). P. Vogel, Phys. Lett. [**B66**]{}, 431 (1976). P. Fallon et al., Phys. Rev. [**C51**]{}, R1609 (1995). M.A. Riley et al., Nucl. Phys. [**A512**]{}, 178 (1990). B. Cederwall et al., Phys. Rev. Lett. [**72**]{}, 3150 (1995). E.R. Marshalek, Phys. Rev. [**C11**]{}, 1426 (1975); Nucl. Phys. [**A266**]{}, 317 (1976). J.L. Egido, H.J. Mang and P. Ring, Nucl. Phys. [**A339**]{}, 390 (1980). Y.R. Shimizu and K. Matsuyanagi, Prog. Theor. Phys. [**79**]{}, 144 (1983); [**72**]{}, 799 (1984). Y.R. Shimizu and K. Matsuzaki, Nucl. Phys. [**A588**]{}, 559 (1995). I.M. Robledo, J.L. Egido and P. Ring, Nucl. Phys. [**A449**]{}, 201 (1986). A. Bohr and B.R. Mottelson, [*Nuclear Structure*]{}, Vol.2, (Benjamin, New York,1975). S. Kinouchi, Ph.D.Thesis (1988), Univ. of Tsukuba. Y.R. Shimizu, E. Vigezzi and R.A. Broglia, Nucl. Phys. [**A509**]{}, 80 (1990). H. Sakamoto and T. Kishimoto, Nucl. Phys. [**A501**]{}, 205 (1989). T. Bengtsson and I. Ragnarsson, Nucl. Phys. [**A436**]{}, 14 (1985). M. Baranger and K. Kumar, Nucl. Phys. [**A110**]{}, 490 (1968). M. Brack, J. Damgaard, A.S. Jensen, H.C. Pauli, V.M. Strutinsky and C.Y. Wong, Rev. Mod. Phys. [**44**]{}, 320 (1972). R. Wyss, W. Satula, W. Nazarewicz and A. Johnson, Nucl. Phys. [**A511**]{}, 324 (1990). M.W. Drigert et al., Nucl. Phys. [**A530**]{}, 452 (1991). H. Sakamoto, Nucl. Phys. [**A557**]{}, 583c (1993). M.J. Joyce et al., Phys. Lett. [**B340**]{}, 150 (1994). M.P. Carpenter et al., Phys. Rev. [**C51**]{}, 2400 (1995). M. Girod, J.P. Delaroche, J. Libert and I. Deloncle, Phys. Rev. [**C45**]{}, R1420 (1992). P.B. Semmes, I. Ragnarsson and S. Åberg, Phys. Lett. [**B345**]{}, 185 (1995). J. Skalski, Phys. Rev. [**C49**]{}, 2011 (1994). -------------------------- ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ K=0 K=1 K=2 K=3 K=0 K=1 K=2 K=3 K=0 K=1 K=2 K=3 E \[ MeV \] 1.37 1.45 1.20 1.52 1.55 1.58 1.18 1.53 1.83 1.62 1.14 1.53 $B(E3)/B(E3)_{\rm s.p.}$ 6.6 11.9 10.0 1.0 7.6 10.1 10.1 0.8 11.5 11.2 10.2 0.7 -------------------------- ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ ------ : Calculated excitation energies of octupole vibrations and $B(E3; 0^+ \rightarrow 3^-,K)$ values estimated using the strong coupling scheme for SD $^{190,192,194}$Hg. []{data-label="oct-E-E3"} [$^{190}$Hg]{} [$^{192}$Hg]{} [$^{194}$Hg]{} -------------------------- ---------------- ---------------- ---------------- E \[ MeV \] 1.39 1.50 1.45 $B(E2)/B(E2)_{\rm s.p.}$ 2.7 3.0 3.8 : Calculated excitation energies of $\gamma$ vibrations and $B(E2; 0^+ \rightarrow 2^+,K=2)$ values estimated using the strong coupling scheme for SD $^{190,192,194}$Hg. []{data-label="gamma-E-E2"} ---------------- ------------------ ------------------------------------------ --------------------------------------------- ------------------------------------------ ------------------------------------------ The lowest The second The lowest The second [$^{190}$Hg]{} $E_x'$ \[ keV \] 113 389 $\approx 0$ 256 config. $\nu(7_1 \otimes 7_2)_{\alpha=0}$ $\nu(7_1 \otimes [505\ 11/2])_{\alpha=0,1}$ $(\mbox{oct.vib.})_{\alpha=1}$ $\nu(7_1 \otimes [642\ 3/2])_{\alpha=0}$ exp. Band 3 Band 2 Band 4 [$^{192}$Hg]{} $E_x'$ \[ keV \] 611 611 441 632 config. $\nu(7_1 \otimes [512\ 5/2])_{\alpha=1}$ $\nu(7_1 \otimes [512\ 5/2])_{\alpha=0}$ $\nu(7_1 \otimes [642\ 3/2])_{\alpha=1}$ $\nu(7_1 \otimes [642\ 3/2])_{\alpha=0}$ exp. Band 2 Band 3 [$^{194}$Hg]{} $E_x'$ \[ keV \] 857 892 738 759 config. $\nu([514\ 7/2])^2_{\alpha=0}$ $\pi([530\ 1/2])^2_{\alpha=0}$ $(\mbox{oct.vib.})_{\alpha=0}$ $(\mbox{oct.vib.})_{\alpha=1}$ exp. Band 2 Band 3 ---------------- ------------------ ------------------------------------------ --------------------------------------------- ------------------------------------------ ------------------------------------------ : The lowest and the second lowest configurations at $\hbar{\omega_{\rm rot}}=0.4\mbox{MeV}$ in each parity sector. The proposed assignments of the observed excited SD bands are also shown. The excitation energies of the negative-parity two-quasineutron states, 256keV for [$^{190}$Hg]{} and 441 and 632keV for [$^{192}$Hg]{}, contain very weak octupole correlations. The corresponding unperturbed two-quasineutron energies are 261, 460 and 635 keV, respectively. []{data-label="Ex"} [^1]: E-mail : [[email protected]]{} [^2]: In the following, the notation $[A,B]_{\rm RPA}$ means that we neglect these higher order terms in calculating the commutator between $A$ and $B$. [^3]: This value depends on the treatment of the pairing gaps at finite frequency. If we use constant pairing gaps against ${\omega_{\rm rot}}$ we get the best value $f_3=1.05$. [^4]: The signature for Bands 2 and 3 is determined by following the spin assignment in Ref.[@Ril90].

ArXiv

--- abstract: 'Application developers often place executable assertions – equipped with program-specific predicates – in their system, targeting programming errors. However, these detectors can detect data errors resulting from transient hardware faults in main memory as well. But while an assertion reduces silent data corruptions (SDCs) in the program state they check, they add runtime to the target program that increases the attack surface for the remaining state. This article outlines an approach to find an optimal subset of assertions that minimizes the SDC count, without the need to run fault-injection experiments for every possible assertion subset.' author: - bibliography: - 'bib/macros-short.bib' - 'bib/all.bib' - 'local.bib' title: 'DETOx: Towards Optimal Software-based Soft-Error Detector Configurations [^1]' --- Introduction ============ With continuously shrinking semiconductor structure sizes and lower supply voltages, the per-device susceptibility to transient hardware faults is on the rise. A class of countermeasures with growing popularity is *software-implemented* hardware fault tolerance (SIHFT), which avoids expensive hardware mechanisms and can be applied application/specifically. However, SIHFT can, against intuition, cause more harm than good: its overhead in execution time and memory space also increases the system’s figurative “attack surface”. In consequence, this phenomenon can diminish all gains from detected or corrected errors by the increased possibility of being struck by radiation in the first place [@schirmeier:15:dsn]. One class of SIHFT measures are *executable assertions* capable of detecting data errors. These statements or code sequences check statically-known state invariants in specifically chosen points of a workload at runtime, and *“can detect errors in input data and prevent error propagation”* [@saib:1977:executable-assertions]. Although primarily used as a means to complement unit tests – aiming at detecting programming errors during the development process especially of safety-critical software, – assertions can also detect data errors caused by hardware faults, or even be specifically designed for this purpose. While, for example, Hiller et al. [@hiller:2002:placement-executable-assertions] identify good placements and adequate predicates for executable assertions, we assume the target workload is already equipped with a set of assertions. We explore the tradeoff between each assertion’s capability to detect data errors, and its runtime cost that increases the liveness of critical data in variables residing in memory during their execution. We also assume that data errors may be detected by more than one assertion, and aim at finding a subset of assertions – an assertion *configuration* – that catches most errors but minimizes the total *silent data corruption* (SDC) count. The Attack-Surface Tradeoff =========================== We assume a fault model of uniformly distributed, independent and transient single-bit flips in main memory, modeled as originating from *direct* influences of ionizing radiation [@sridharan:2013:fengshui-s]. Under this assumption, live data stored in a variable in memory has a corruption probability that is proportional to its lifetime during the workload run and its memory size [@schirmeier:15:dsn] – or, in other words, proportional to its *area* in a fault-space diagram. The X-axis of the example diagram in shows the workload’s runtime, the Y-axis all bits in memory. In the example, *var2* holds live data throughout the complete runtime, and a fault-injection (FI) campaign injecting faults at all possible time/memory-bit locations yields the information that bit-flips in this variable during this time frame lead to SDCs (color-coded *gray*). The gray-colored area contributes to the total SDC count of the workload. ### Error Detection and Additional Runtime {#error-detection-and-additional-runtime .unnumbered} Similarly, *var1* holds live data, but is checked against an application-specific predicate by executable assertions 1 and 2 at runtime. Hence, corruptions occurring in *var1* before the point in time when assertion 2 runs are *detected* (color-coded *green*). However, both assertions entail a runtime cost: The instructions executed for each assertion extend the workload runtime (by $d_{1}$ respectively $d_{2}$). Thereby, these assertions prevent corruptions in *var1* from causing an SDC, but increase the lifetime of the data stored in *var2* – which in turn enlarges the grey-colored area, or the total SDC count. In consequence, depending on each assertion’s balance between SDC *reduction* (by turning otherwise gray *SDC* areas into green *detected* results) and SDC *increase* (by increasing the lifetime of other variables), it may influence the workload’s total SDC count positively or negatively. However, in the example in , assertion 1 detects a subset of the data errors in *var1* that assertion 2 detects as well. Thus, it seems advisable to generate a workload variant *without* assertion 1, yielding a different fault-space diagram () with a lower total runtime and, hence, a lower (*gray*) SDC count originating from *var2*, but still most data errors in *var1* being detected by assertion 2. Unfortunately, assertions cannot be considered independently: Depending on the assertion *configuration* – i.e., other assertions being compiled in or out of a specific workload variant – an assertion may increase the fault-space area of a protected (yielding green *detected* results) or an unprotected variable (gray *SDC* results). This interdependency makes it necessary to consider the total SDC count resulting from *every possible* assertion configuration. ### Approach: FI-based Result Calculation {#approach-fi-based-result-calculation .unnumbered} Considering real-world workloads with the number of $N$ different executable assertions in the hundreds or thousands makes clear, that generating a workload variant for each of the $2^{N}$ possible assertion configurations is generally infeasible. Running an FI campaign measuring the total SDC count for each of these workload variants even exponentiates this problem. Instead, our approach is based on FI results from a single workload variant with *all* assertions enabled. We exhaustively scan the single-bit flip fault space of this workload (using advanced pruning methods within the FAIL[\*]{} FI tool [@schirmeier:15:edcc-s]) and record a result – for example, *No Effect* or *SDC* – for each fault-space coordinate. If an FI experiment observes that one of the assertions detected a data error, we do not abort the experiment and record *detected* (*green* in ), but record *which* assertion triggered *and let the experiment continue running*. So, for example, an FI experiment injecting in *var1* at some point in time before assertion 1 runs () would record assertion 1 to have detected the error, continue running, then also record assertion 2, and finally record that an SDC occurred – because it *would* have occurred if none of the two assertions had been in place. Based on this result data, our DETOx tool prototype can *calculate* the SDC count for an arbitrary assertion configuration. Removing, e.g., assertion 2 from the example in requires 1) *subtracting* all gray SDC areas between $t_{2,1}$ and $t_{2,2}$ from the total SDC count, and 2) *“re-dyeing”* all remaining areas that were only detected by assertion 2 to the final result recorded in the FI experiments. Sorting Example =============== Using our approach we evaluated the configurations of a simple sorting program taking 24 input elements with two assertions: the first one checks for ascending order of two swapped elements, while the second repeats the same test on the complete array after sorting. The following table shows the FI-obtained SDC count for the all-enabled (“11”) variant, and the predicted results for all other configurations. Assertion configurations are represented as a bit vector, where the bits indicate whether either assertion was enabled/disabled. **00** **01** **10** **11** ---------------------------------------------- ------------- ------------- ------------- -------- [$\textrm{{SDC}}_{\textrm{{prediction}}}$]{} **2315851** **1395495** **2318461** [2319929]{} [1393233]{} [2324435]{} [error]{} [-0.176%]{} [+0.162%]{} [-0.257%]{} To quantify the prediction quality, we ran FI campaigns for all other configurations besides “11” as a ground truth for comparison – information that would usually not be available, shown in the [$\textrm{{SDC}}_{\textrm{{reality}}}$ row. ]{}In this example, the SDC-count predictions are accurate to within $0.3$%. The optimal, lowest SDC-count configuration is “01”, i.e., only the second assertion gets enabled. Conclusions =========== To conclude, our approach allows for fast and cheap exploration of the assertion-configuration space, and is based on FI results of a single, all-enabled configuration. This allows searching for optimal configurations in future work, for example using genetic algorithms or optimization techniques such as integer-linear programming. [^1]: This work was supported by the German Research Council (DFG) focus program SPP 1500 under grant SP 968/5-3.

ArXiv

--- abstract: 'The zero sets of the Bergman space $A^p_\om$ induced by either a radial weight $\om$ admitting a certain doubling property or a non-radial Bekollé-Bonami type weight are characterized in the spirit of Luecking’s results from 1996. Accurate results obtained en route to this characterization are used to generalize Horowitz’s factorization result from 1977 for functions in $A^p_\om$. The utility of the obtained factorization is illustrated by applications to integration and composition operators as well as to small Hankel operator induced by a conjugate analytic symbol. Dominating sets and sampling measures for the weighted Bergman space $A^p_\om$ induced by a doubling weight are also studied. Several open problems related to the scheme of the paper are posed.' address: - 'University of Eastern Finland, P.O.Box 111, 80101 Joensuu, Finland' - 'University of Eastern Finland, P.O.Box 111, 80101 Joensuu, Finland' author: - Taneli Korhonen - Jouni Rättyä title: 'Zero sequences, factorization and sampling measures for weighted Bergman spaces' --- [^1] Introduction and main results {#sec:Intro} ============================= Let ${\mathcal{H}}({\mathbb{D}})$ denote the space of analytic functions in the unit disc ${\mathbb{D}}=\{z\in{\mathbb{C}}:|z|<1\}$ of the complex plane ${\mathbb{C}}$. A function $\omega:{\mathbb{D}}\to[0,\infty)$, integrable over ${\mathbb{D}}$, is called a weight. It is radial if $\omega(z)=\omega(|z|)$ for all $z\in{\mathbb{D}}$. For $0<p<\infty$ and a weight $\omega$, the weighted Bergman space $A^p_\omega$ consists of $f\in{\mathcal{H}}({\mathbb{D}})$ such that $$\|f\|_{A^p_\omega}^p=\int_{\mathbb{D}}|f(z)|^p\omega(z)\,dA(z)<\infty,$$ where $dA(z)=\frac{dx\,dy}{\pi}$ is the normalized Lebesgue area measure on ${\mathbb{D}}$. As usual, $A^p_\alpha$ stands for the classical weighted Bergman space induced by the standard radial weight $\omega(z)=(1-|z|^2)^\alpha$, where $-1<\alpha<\infty$. For $f\in{\mathcal{H}}({\mathbb{D}})$ and $0<r<1$, set $$\begin{split} M_p(r,f)&=\left(\frac{1}{2\pi}\int_{0}^{2\pi} |f(re^{it})|^p\,dt\right)^{1/p},\quad 0<p<\infty, \end{split}$$ and $M_\infty(r,f)=\max_{|z|=r}|f(z)|$. For $0<p\le\infty$, the Hardy space $H^p$ consists of $f\in {\mathcal{H}}(\mathbb D)$ such that $\|f\|_{H^p}=\sup_{0<r<1}M_p(r,f)<\infty$. In this paper we are mainly interested in zero-sequences, factorization, dominating sets and sampling measures for the Bergman space $A^p_\om$ induced by either a non-radial weight belonging to a kind of Bekollé-Bonami class or a radial weight admitting a certain doubling property. Our studies of zeros and factorization go hand-in-hand and dominating sets and sampling measures are almost equally interrelated. While our results on zeros and factorization improve and generalize certain results in the literature, our findings on dominating sets and sampling measures are less complete but offer a somewhat new approach to these topics and also give arise to several open problems. We begin with zeros and factorization. Let $\om:{\mathbb{D}}\to[0,\infty)$ be a weight, $0 < p < \infty$ and $f \in A^p_\om$ such that $f\not\equiv0$. Then a sequence $Z \subset {\mathbb{D}}$ is called the *zero set (or sequence)* of $f$, denoted by ${\mathcal{Z}}(f)$, if $f(a) = 0$ for all $a \in Z$, counting multiplicities, and $f(a) \neq 0$ for all $a\in{\mathbb{D}}\setminus Z$. A set $Z$ is called a *zero set* for $A^p_\om$ if there exists a nonzero function $f \in A^p_\om$ such that $Z = {\mathcal{Z}}(f)$. The set of all zero sets of $A^p_\om$ is denoted by ${\mathcal{Z}}(A^p_\om)$. Zeros of functions in Hardy spaces are neatly characterized by the Blaschke condition and each Hardy-function $f$ admits the well-known inner-outer factorization, where the inner part is a product of a singular inner function and a Blaschke product containing all the zeros of $f$ and the non-vanishing outer part has the same norm as $f$ [@Duren1970]. The situation of Bergman spaces is completely different because the distribution of zeros of functions in Bergman spaces is not that well understood neither such an efficient factorization as in the Hardy space case is known. Even if the geometric distribution of zeros is not completely described, the difference between known sufficient and necessary conditions for a sequence to be a zero set for $A^p_\alpha$ is small. Probably the most commonly known results in this context are due to Luecking, Korenblum, Hedenmalm, Horowitz and Seip. Horowitz [@Horzeros; @Horzeros1; @Horzeros2] studied unions, subsets and dependence on $p$ of the zero sets of functions in $A^p_\alpha$ and obtained accurate information on certain sums involving the moduli of zeros. Some of these results were generalized to certain $A^p_\om$ in [@PR2014Memoirs]. The studies by Korenblum [@Kor], Hedenmalm [@HedStPeter] and Seip [@S1; @S2] employ methods based on the use of densities defined in terms of partial Blaschke sums, Stolz star domains and Beurling-Carleson characteristic of the corresponding boundary set, and yield more complete results. Luecking [@L1996] gave a description of zero sets of $A^p_\alpha$ in terms of certain auxiliary functions induced by the zeros. Even if this characterization does not reveal the geometric distribution very transparently, it yields accurate information on the subsets of zero sets. Horowitz [@HorFacto] also established a useful factorization theorem for functions in $A^p_\alpha$, but this factorization does not allow to take one of factors non-vanishing, and thus does not behave equally well as the inner-outer factorization in Hardy spaces. This factorization result was generalized to some $A^p_\om$ in [@HorSchna; @PR2014Memoirs]. We start with employing Luecking’s [@L1996] approach to describe $A^p_\om$ zero sets when $\om$ belongs to a kind of Bekollé-Bonami class. For $1<q<\infty$, write $\om\in B_q$ if the weight $\om$ is (almost everywhere) strictly positive and $$B_q(\om)=\sup_{S}\left(\frac{1}{|S|^2}\int_{S}\om(z)(1-|z|^2)^{2q}\,dA(z)\right) \left(\frac{1}{|S|^2}\int_{S} \om(z)^{-\frac1{q-1}}\,dA(z)\right)^{q-1} <\infty,$$ where the supremum is taken over all Carleson squares $S\subset{\mathbb{D}}$, and $|S|$ denotes the Euclidean area of $S$. Denote $B_\infty = \bigcup_{q>1}B_q$ for short. Recall that each Carleson square is of the form $$S(z)=\left\{re^{i\theta}\in{\mathbb{D}}:|z|<r<1,\,|\arg ze^{-i\theta}|<\frac{1-|z|}2\right\},\quad z\in{\mathbb{D}}\setminus\{0\}.$$ In the beginning of Section \[sec:ProofThm1\] we briefly analyze the classes $B_q$. In particular, we show that $B_p\subsetneq B_q$ for $1<p<q<\infty$ and discuss their Kerman-Torchinsky properties. Following Luecking [@L1996], for a sequence $Z \subset {\mathbb{D}}$, we use the notation $$\begin{split} \psi_Z(z)&= \prod_{a\in Z} \overline{a}\frac{a-z}{1-\overline{a}z} \exp\left(1-\overline{a}\frac{a-z}{1-\overline{a}z}\right), \quad z \in {\mathbb{D}},\\ W_Z(z)&= e^{k_Z(z)},\quad k_Z(z)= \frac{|z|^2}2 \sum_{a \in Z}\frac{\left(1-|a|^2\right)^2}{|1-\overline{a}z|^2}, \quad z \in {\mathbb{D}}. \end{split}$$ The first of our main results on Bergman zero sets is a generalization of [@L1996 Theorem 3] and reads as follows. \[theo:zeroset\] Let $0 < p < \infty$, $\om \in B_\infty$ and $Z$ be a sequence in ${\mathbb{D}}$. Then the following statements are equivalent: - $Z \in {\mathcal{Z}}(A^p_\om)$; - Any subsequence of $Z$ belongs to ${\mathcal{Z}}(A^p_\om)$; - $\sum_{a \in Z} (1-|a|)^2 < \infty$ and there is a nowhere zero function $F \in {\mathcal{H}}({\mathbb{D}})$ such that $FW_Z \in L^p_\om$; - $\sum_{a \in Z} (1-|a|)^2 < \infty$ and there is a nonzero function $F \in {\mathcal{H}}({\mathbb{D}})$ such that $FW_Z \in L^p_\om$. Moreover, if *(a)* is true, then the mapping $f \mapsto f/\psi_Z$ is a continuous isomorphism from $\left\{f \in A^p_\om : Z \subset {\mathcal{Z}}(f)\right\}$ onto $\left\{F \in {\mathcal{H}}({\mathbb{D}}) : FW_Z \in L^p_\om\right\}$. If $F$ is that of Theorem \[theo:zeroset\] (c), and $h = -p\log|F|$, then $h$ is harmonic and $$|FW_Z|^p = |F|^pW_Z^p = e^{p\log|F|}e^{pk_Z} = \exp(pk_Z - h).$$ Therefore the equivalence of (a) and (c) in Theorem \[theo:zeroset\] leads to the following result. \[theo:zeroset2\] Let $0 < p < \infty$, $\om \in B_\infty$ and $Z$ be a sequence in ${\mathbb{D}}$. Then $Z \in {\mathcal{Z}}(A^p_\om)$ if and only if there exists a harmonic function $h$ such that $$\int_{\mathbb{D}}\exp(p k_Z(z) - h(z)) \om(z)\,dA(z) < \infty.$$ By performing a certain perturbation on a zero set, in this case moving the points closer to the boundary, it becomes a zero set for some other weighted Bergman space. A sequence $(z_n)$ in ${\mathbb{D}}$ is separated or equivalently uniformly discrete if $\inf_{k\ne n}\r_p(z_k,z_n)=\inf_{k\ne n}|\vp_{z_k}(z_n)|>0$, where $\varphi_a(z) = \frac{a-z}{1-\overline{a}z}$ is the standard automorphism of the unit disc. \[coro:zeroPerturb\] Let $0<p<\infty$ and $\om\in B_\infty$. Let $Z$ be a zero set for $A^p_\om$ such that it is a finite union of separated sequences. Let $0 < \g < 1$ and suppose that there exists another set $Z'$ and a one-to-one correspondence $\s : Z \to Z'$ such that $1-|\s(a)|^2 = \g\left(1-|a|^2\right)$ and $\rho_p(a,\s(a))$ is uniformly bounded away from 1 on $Z$. Then $Z'$ is a zero set for $A^{p/\g}_\om$. To deduce Corollary \[coro:zeroPerturb\], note first that since $Z$ is a zero set for $A^p_\om$ by the hypothesis, there exists a harmonic function $h$ such that $\exp(pk_Z-h)$ is integrable with respect to $\om dA$ by Corollary \[theo:zeroset2\]. Therefore it suffices to find a harmonic majorant $g$ of $k_{Z'}/\g - k_Z$. Indeed, if such $g$ exists, then $pk_{Z'}/\g-(pg+h) \le pk_Z - h$, and hence $\exp(\frac{p}{\g}k_{Z'}-(pg+h))$ is integrable with respect to $\om dA$, and consequently $Z'$ is a zero set for $A^{p/\g}_\om$ by Corollary \[theo:zeroset2\]. A function $g$ with desired properties is constructed in the proof of [@L2000 Theorem 4]. Horowitz [@Horzeros; @HorFacto] also obtained some results in the spirit of Corollary \[coro:zeroPerturb\] describing how the zero sets depend on the parameter $p$. Some of those results were generalized for certain $A^p_\om$ in [@PR2014Memoirs], and can be further improved by applying Luecking’s approach in studying zero sets. In particular, we note that, by applying Proposition \[prop:zeroseq\] below instead of [@HorFacto Lemma 3], Theorem 4 and Corollary 2 in [@HorFacto] can be generalized to the case $\om\in B_\infty$ with only minor modifications to the original proofs. We now turn to consider factorization of functions in $A^p_\om$. By refining Horowitz’ original probabilistic argument by Luecking’s method to study the zero sets, and then adopting the whole reasoning to the class of weights we are interested in we derive the following factorization result. \[theo:factorization\] Let $0<p<\infty$ and $\om\in B_\infty$ such that the polynomials are dense in $A^p_\om$. Let $f\in A^p_\om$ and $0<p_1,p_2<\infty$ such that $p^{-1}=p_1^{-1}+p_2^{-1}$. Then there exist $f_1\in A^{p_1}_\om$ and $f_2\in A^{p_2}_\om$ such that $f=f_1f_2$ and $$\|f_1\|_{A^{p_1}_\om}^{p}\|f_2\|_{A^{p_2}_\om}^{p} \le\frac{p}{p_1}\|f_1\|_{A^{p_1}_\om}^{p_1}+\frac{p}{p_2}\|f_2\|_{A^{p_2}_\om}^{p_2} \le C\|f\|_{A^p_\om}^p\le C\|f_1\|_{A^{p_1}_\om}^{p}\|f_2\|_{A^{p_2}_\om}^{p}$$ for some constant $C=C(\om)>0$. The density of polynomials is needed as a hypothesis only to guarantee the existence of a dense family of functions with finitely many zeros. Any other requirement implying this property would suffice here. The question of when polynomials are dense in $A^p_\om$ is an old problem and remains unsolved in general, see for example [@PR2014Memoirs Section 1.5] for basic information and relevant references. However, in certain special cases the density of polynomials can be deduced from an atomic decomposition. For example, [@L1985/2 Theorem 4.1] offers such a decomposition for functions in certain weighted Bergman spaces induced by non-radial weights in terms of the kernel functions of the standard weighted Bergman spaces. Now that these kernels are analytic beyond the boundary of the unit disc, this in turn yields the density of polynomials. Another relevant reference regarding atomic decomposition in the non-radial case is [@Constantin]. The argument used in these studies does not work for the class $B_\infty$, and that is understandable because $B_\infty$ contains weights that induce very small Bergman spaces. We next discuss a specific step in the proofs of the results above and then turn to consider zeros and factorization for $A^p_\om$ induced by a radial weight. The key ingredient in the proofs of Theorems \[theo:zeroset\] and \[theo:factorization\] is Proposition \[prop:zeroseq\] which states that $$h(z) = \frac{|f(z)|} {\prod_{a \in Z} \left\{\left|\frac{a-z}{1-\overline{a}z}\right| \exp\left[\frac12\left(1-\left|\frac{a-z}{1-\overline{a}z}\right|^2\right)\right]\right\}},$$ induced by $Z\subset \mathcal{Z}(f)$, satisfies $\|f\|_{A^p_\om}\asymp \|h\|_{L^p_\om}$ for each $f \in A^p_\om$. The proof of this fact eventually boils down to showing that $$R(f)(z)= \int_{\mathbb{D}}f(w)\frac{\left(1-|z|^2\right)^2}{|1-\overline{z}w|^4}\,dA(w),\quad z\in{\mathbb{D}},$$ is a bounded operator from $L^q_\om$ into itself for some $q>1$. This is in turn equivalent to the boundedness of the Bergman projection $$P_\a(f)(z) = \int_{\mathbb{D}}\frac{f(w)}{(1-z\overline{w})^{2+\a}} \left(1-|w|^2\right)^\a\,dA(w),\quad z\in{\mathbb{D}},$$ for $\alpha=2$ on certain $L^q$-spaces, and therefore this step is done at once by using the classical characterization of the one-weight inequality for the Bergman projection by Bekollé and Bonami [@B1981; @BB1978]. This yields the hypothesis $\om\in B_\infty$ in Theorems \[theo:zeroset\] and \[theo:factorization\], the proofs of which are presented in Section \[sec:ProofThm1\]. The defect in the hypothesis $\om\in B_\infty$ is that it does not allow $\om$ to vanish in a set of positive measure, neither $\om$ may be small in a relatively large part of each outer annulus of ${\mathbb{D}}$. Our next goal is to show that by restricting our consideration to radial weights we can do better and no longer need to require such smoothness. To do this, let $\om$ be a radial weight such that $\widehat{\om}(z)=\int_{|z|}^1\om(s)\,ds>0$ for all $z\in{\mathbb{D}}$, for otherwise $A^p_\om={\mathcal{H}}({\mathbb{D}})$. A radial weight $\om$ belongs to the class ${\widehat{\mathcal{D}}}$ if there exists a constant $C=C(\om)\ge1$ such that the doubling inequality $\widehat{\om}(r)\le C\widehat{\om}(\frac{1+r}{2})$ is valid for all $0\le r <1$. If there exist $K=K(\om)>1$ and $C=C(\om)>1$ such that $\widehat{\om}(r)\ge C\widehat{\om}\left(1-\frac{1-r}{K}\right)$ for all $0\le r<1$, then $\om\in{\widecheck{\mathcal{D}}}$. Additionally, we write ${\mathcal{D}}={\widehat{\mathcal{D}}}\cap{\widecheck{\mathcal{D}}}$. The classes of weights ${\widehat{\mathcal{D}}}$ and ${\mathcal{D}}$ emerge from fundamental questions in operator theory: recently Peláez and Rättyä [@PR2017Proj] showed that the weighted Bergman projection $P_\om$, induced by a radial weight $\om$, is bounded from $L^\infty$ to the Bloch space $\mathcal{B} = \{f \in \mathcal{H}({\mathbb{D}}) : \sup_{z\in{\mathbb{D}}} |f'(z)|(1-|z|) < \infty\}$ if and only if $\om \in {\widehat{\mathcal{D}}}$, and further, it is bounded and onto if and only if $\om \in {\mathcal{D}}$. For further information on these classes of weights, see [@PelSum14; @PR2014Memoirs; @PR2015Embedding; @PR2016TwoWeight]. Moreover, since weights in ${\widehat{\mathcal{D}}}$ may very well vanish in a set of positive measure, ${\widehat{\mathcal{D}}}$ is not contained in $B_\infty$. Assume $f\in A^p_\omega$, where $\om\in{\widehat{\mathcal{D}}}$. Then $$\begin{split} \|f\|_{A^p_\omega}^p&\ge\int_{{\mathbb{D}}\setminus D\left(0,\frac{1+r}{2}\right)}|f(z)|^p\omega(z)\,dA(z) \gtrsim M_p^p\left(\frac{1+r}{2},f\right)\widehat{\om}\left(r\right),\quad r\to1^-, \end{split}$$ from which the well known inequality $M_\infty(r,f)\lesssim M_p(\frac{1+r}{2},f)(1-r)^{-1/p}$ yields $$ M^p_\infty(r,f)\lesssim\frac{\|f\|_{A^p_\omega}^p}{(1-r)\widehat{\om}\left(r\right)},\quad r\to1^-.$$ Now that $\om\in{\widehat{\mathcal{D}}}$, there exist $C=C(\om)>0$ and $\b=\b(\om)>0$ such that $\widehat{\om}\left(r\right)\ge C\widehat{\om}(0)\left(1-r\right)^\b$ by Lemma \[Lemma:weights-in-D-hat\](ii) below. Hence $ |f(z)|\lesssim\|f\|_{A^p_\omega}(1-|z|)^{-\frac{\b+1}{p}} $ for all $z\in{\mathbb{D}}$, and it follows that $\log|f|\in L^1$. Therefore the reasoning to be used in the proof of Proposition \[prop:zeroseq\] can be applied in this case also. However, as explained above, the argument relies on the characterization of the one-weight inequality for the Bergman projection by Bekollé and Bonami, which guarantees the boundedness of the auxiliary operator $R:L^q_\om\to L^q_\om$ under the hypothesis $\om\in B_q\subset B_\infty$. But what is actually needed here is to show that $$\label{eq:step-radial-case} \int_{\mathbb{D}}\left(\int_{\mathbb{D}}|f(\z)|^\frac{p}{q}\frac{(1-|z|^2)^2}{|1-\overline{z}\zeta|^4}\,dA(\z)\right)^q\om(z)\,dA(z)\lesssim\|f\|_{A^p_\om}^p,\quad f\in A^p_\om,$$ for a sufficiently large $q=q(\om)>p$. The functions involved being analytic, this inequality is better controllable by using Carleson embedding theorems [@PR2015Embedding] rather than weight inequalities for $L^p$-functions [@B1981; @BB1978]. By using this approach we will prove the statement of Proposition \[prop:zeroseq\] for $\om\in{\widehat{\mathcal{D}}}$, and deduce the following result. \[Theorem:D-hat\] The statements in Theorems \[theo:zeroset\] and \[theo:factorization\] and Corollaries \[theo:zeroset2\] and \[coro:zeroPerturb\] are valid when the hypothesis $\om\in B_\infty$ is replaced by $\om\in{\widehat{\mathcal{D}}}$. Details of the deduction yielding this theorem are given in Section \[Sec:radial-zeros-factorization\]. The reason why this approach does not yield a better result in the non-radial case is that known Carleson embedding theorems impose growth and/or smoothness restrictions to the weight. In particular, [@Constantin2 Theorem 3.1] states that if $P^+_\eta:L^q_\om\to L^q_\om$ is bounded for some $q>1$ and $\eta>-1$, then $A^p_\om$ is continuously embedded into $L^p_\mu$ if and only if $\mu(\Delta(z,r))\lesssim\om(\Delta(z,r))$ for all $z\in{\mathbb{D}}$. Each radial weight $(1-|z|)^{-1}\left(\log\frac{e}{1-|z|}\right)^{-\b}$ with $\b>1$ belongs to $B_\infty$, but does not satisfy the Bekollé-Bonami condition, and thus this approach does not give us anything better than what we know already. This and many other possible applications, some of which will appear later in this paper, suggest that it would be desirable to obtain new information on the embedding $A^p_\om\subset L^p_\mu$ when $\om$ is non-radial and $\mu$ is a positive Borel measure on ${\mathbb{D}}$. In the radial case the hypothesis on the density of polynomials in Theorem \[theo:factorization\] is of course always satisfied and can thus be omitted. The statement in Theorem \[theo:factorization\] for $\om\in B_\infty\cup{\widehat{\mathcal{D}}}$ significantly improves the main result in [@PR2014Memoirs Chapter 3] and Horowitz’ original result as well because now the factorization is available for the whole class ${\widehat{\mathcal{D}}}$ and the constant $C$ appearing in the inequality for the norms is independent of the parameters $p$, $p_1$ and $p_2$. We mention three immediate consequences of our results so far. The factorization given in Theorem \[Theorem:D-hat\] shows that the statement in [@PR2014Memoirs Theorem 4.1(iv)] concerning the integration operator $T_g(f)(z)=\int_0^z f(\z)g'(\z)\,d\z$ induced by $g\in\mathcal{H}({\mathbb{D}})$ is valid for $\om\in{\widehat{\mathcal{D}}}$. The theorem states that $T_g:A^p_\om\to A^q_\om$ is bounded in the upper triangular case $0<q<p<\infty$ if and only if $g\in A^s_\om$ with $\frac1s=\frac1q-\frac1p$. Also the Aleman-Sundberg question, discussed in more detail in [@PR2014Memoirs p. 38], on subsets of zero-sets has an affirmative answer when $\om\in{\widehat{\mathcal{D}}}$ by Theorem \[theo:zeroset\] because in this case each subset of a zero set is always a zero set. Moreover, the statement in [@PelSum14 Proposition 7.5] concerning bounded and compact composition operators acting between different weighted Bergman spaces is valid under the hypotheses $\om\in B_\infty\cup{\widehat{\mathcal{D}}}$ and the density of polynomials in $A^p_\om$ by Theorems \[theo:factorization\] and \[Theorem:D-hat\]. The proposition says that if $0<p,q<\infty$, $n\in{\mathbb{N}}$ and $\nu$ is any weight, then the composition operator $C_\vp$, defined by $C_\vp(f)=f\circ\vp$, is bounded (resp.compact) from $A^p_\om$ to $A^q_\nu$ if and only if $C_\vp:A^{np}_\om\to A^{nq}_\nu$ is bounded (resp.compact). Therefore one may assume that the both parameters $p$ and $q$ are greater than two and this often simplifies some arguments. To finish our consideration of zeros and factorization, we give an application of the obtained factorization in the study of the small Hankel operator induced by a conjugate analytic symbol in the upper triangular case. Let $\om$ be a weight such that the reproducing kernels $B^\om_z$ of $A^2_\om$ exist, that is, $f(z)=\langle f,B^\om_z\rangle_{A^2_\om}$ for all $z\in{\mathbb{D}}$ and $f\in A^2_\om$. For $f\in{\mathcal{H}}({\mathbb{D}})$ consider the small Hankel operator $$h^\om_{\overline{f}}(g)(z)=\overline{P_\om}(\overline{f}g)=\int_{\mathbb{D}}\overline{f(\z)}g(\z)B^\om_z(\z)\om(\z)\,dA(\z),\quad z\in{\mathbb{D}}.$$ \[corollary:Hankel\] Let $1<q<p<\infty$ and $1<s<\infty$ such that $\frac1s=\frac1q-\frac1p$, and let $\om\in{\widehat{\mathcal{D}}}$. Then $h^\om_{\overline{f}}:A^p_\om\to\overline{A^q_\om}$ is bounded if and only if $f\in A^s_\om$. This corollary combined with our earlier observation on $T_g$ acting from $A^p_\om$ to $A^q_\om$ confirms the well known phenomenon that in many cases the boundedness of the integration operator and the small Hankel operator are characterized by the same condition. Corollary \[corollary:Hankel\] extends the results of Pau and Zhao [@PauZhao] to the case of Bergman spaces of one complex variable induced by doubling weights. Hankel, integration and composition operators are among the most studied objects in operator theory of analytic function spaces and the literature concerning the subject is vast. Since none of these operators is in the main focus of the present paper, we invite the reader to see [@CowenMac95; @Shapiro93] for the theory of composition operators, and [@PR2016Toeplitz] for the case ${\widehat{\mathcal{D}}}$, [@Aleman; @AlemanCima] for integration, and [@Peller2003; @Zhu] for Hankel. For some recent developments on Hankel operators in Bergman spaces, see [@PauZhaoZhu] and references therein. Our next goal is to study dominating sets for $A^p_\om$ in order to obtain a sufficient condition for a positive Borel measure $\mu$ on ${\mathbb{D}}$ to be a sampling measure for $A^p_\om$. A measurable set $G=G(f)\subset{\mathbb{D}}$ is called a dominating set for $f\in A^p_\om$ if there exists $\d=\d(G)>0$ such that $$\int_G|f(z)|^p\om(z)\,dA(z)\ge\d\|f\|_{A^p_\om}^p.$$ If this inequality is valid for all $f\in A^p_\om$, then $G$ is called a dominating set for $A^p_\om$. To state the results, some more notation is needed. The reproducing kernels of the Hilbert space $A^2_\om$ induced by a radial weight $\om$ are given by $$B^\om_z(\z)=\sum_{n=0}^\infty\frac{(\overline{z}\z)^n}{2\om_{2n+1}},\quad z,\z\in{\mathbb{D}},$$ where $\om_x=\int_0^1s^{x}\om(s)\,ds$ for all $-1<x<\infty$, and each $f\in A^1_\om$ satisfies $$\label{eq:reproducing-formula} f(z)=\langle f,B^\om_z\rangle_{A^2_\om}=\int_{\mathbb{D}}f(\z)\overline{B_z^\om(\z)}\om(\z)\,dA(\z),\quad z\in{\mathbb{D}}.$$ Write $$K_z^\om(\z)=\frac{|B_z^\om(\z)|^2}{\|B_z^\om\|^2_{A^2_\om}},\quad z,\z\in{\mathbb{D}}.$$ Our first result on dominating sets reads as follows. \[theo:dominating-set-for-f\] Let $0<q<p<\infty$ and $\om\in{\widehat{\mathcal{D}}}$. Then there exists a constant $C=C(\om,q)>0$ such that $$\label{eq:PointEval} |f(z)|^q \le C \int_{\mathbb{D}}|f(\z)|^q K_z^\om(\z) \om(\z)\,dA(\z),\quad f\in A^q_\om,\quad z \in {\mathbb{D}}.$$ Moreover, if $f\in A^p_\om$ and $E = E(\e,q,f)$ is the set of points $z\in{\mathbb{D}}$ for which $$\label{eq:BadPoints} |f(z)|^q \leq \e \int_{\mathbb{D}}|f(\z)|^q K^\om_z(\z)\om(\z)\,dA(\z),$$ then there exists a constant $C = C(p/q,\om)$ such that $$\label{eq:BadPointsEstimate} \int_E |f(z)|^p\om(z)\,dA(z) \le (C\e)^{p/q} \|f\|^p_{A^p_\om}.$$ Therefore $\e > 0$ may be chosen such that $G={\mathbb{D}}\setminus E$ satisfies $$ \int_G |f(z)|^p\om(z)\,dA(z) \ge \frac12 \|f\|^p_{A^p_\om},\quad f\in A^p_\om,$$ and thus $G$ is a dominating set for $f\in A^p_\om$ if $\e>0$ is sufficiently small. The inequality with $C=1$ is easy to establish if $q\ge1$. Namely, an application of the reproducing formula to $fB^\om_z$ gives $$f(z)=\int_{\mathbb{D}}f(\z)K^\om_z(\z)\om(\z)\,dA(\z),\quad z\in{\mathbb{D}}.$$ This gives the assertion for $q=1$, which together with Hölder’s inequality can be used to establish the claim in the case $q>1$. This kind of reasoning does not seem to work for $0<q<1$, and we will argue differently; we first use the subharmonicity of $|f|^q$ together with the fact that $|B^\om_z(\z)|\asymp B^\om_z(z)$ for all $\z\in\Delta(z,r)$ and $z\in{\mathbb{D}}$ if $r=r(\om)\in(0,1)$ is sufficiently small [@PRS2015Berezin Lemma 8], and then employ a proof of a Carleson embedding theorem for specific subharmonic functions [@PelSum14 Theorem 3.3]. The estimate can be obtained quite easily for $\om\in\mathcal{D}\subset{\widehat{\mathcal{D}}}$ by using the boundedness of the maximal Bergman projection $P^+_\om$ on $L^p_\om$ for each $p>1$, but the general case $\om\in{\widehat{\mathcal{D}}}$ is more laborious and relies on the $L^p$-estimates of the kernel functions $B_z^\om$ given in [@PR2016TwoWeight]. The special case of Theorem \[theo:dominating-set-for-f\] concerning standard weighted Bergman spaces can be found in [@L2000 Lemma 2]. The proof there is different and does not carry over to the situation of Theorem \[theo:dominating-set-for-f\]. The proof of in Theorem \[theo:dominating-set-for-f\] does not work for $p=q$. In that case, it is natural to replace the the right hand side of  by an average over a subset of a pseudohyperbolic disc centered at $z$. To state the result some notation is needed. Let $r\in(0,1)$, $\nu$ a positive Borel measure on ${\mathbb{D}}$ and $E(z)\subset\Delta(z,r)$ such that $\nu(E(z))>0$ for all $z\in{\mathbb{D}}$. Define $$Q(f)(z)=\frac1{\nu(E(z))} \int_{E(z)} |f(\z)|^p\nu(\z)\,dA(\z), \quad z\in{\mathbb{D}},$$ and $E=E(f,\nu)=\{z\in{\mathbb{D}}:|f(z)|^p\le\e Q(f)(z)\}$. \[theo:dominating-set-for-f-2\] Let $0<p<\infty$, $0<r<1$ and $\om\in{\widehat{\mathcal{D}}}$. Then there exists a constant $C=C(r,\om)>0$ such that $$\int_E|f(z)|^p\om(z)\,dA(z)\le C\e\|f\|^p_{A^p_\om},$$ and thus ${\mathbb{D}}\setminus E$ is a dominating set for $f\in A^p_\om$ if $\e=\e(r,\om)>0$ is sufficiently small. Special cases of Theorem \[theo:dominating-set-for-f-2\] with $\nu$ being the Lebesgue measure can be found in [@L1981 Lemmas 2 and 3]. Our proof is different from these results and relies on Carleson measures. Since our main results so far concern also certain non-radial weights, it is reasonable to discuss that aspect of Theorems \[theo:dominating-set-for-f\] and \[theo:dominating-set-for-f-2\] as well. The first obstruction in the proof of Theorem \[theo:dominating-set-for-f\] for non-radial weights is the pointwise estimate $|B^\om_z(\z)|\asymp B^\om_z(z)$ which does not have a known sufficiently general non-radial extension. The second problem arises with Carleson measures, and finally the lack of satisfactory norm estimates for kernel functions prevents our reasoning from carrying over to the non-radial case all together. The situation of Theorem \[theo:dominating-set-for-f-2\] is better because of [@L1984 Lemmas 1 and 2] and the proof of [@L1985/2 Theorem 3.9], though the weight $\om$ in [@L1984] is rather particular (but the domain lies in ${\mathbb{C}}^n$ and is quite general). A careful inspection of the proof of [@PR2015Embedding Theorem 9] shows that the argument used to obtain Theorem \[theo:dominating-set-for-f-2\] works for weights $\om$ that are doubling in Carleson squares, denoted by $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$ and defined in detail below, if $N:A^p_\om\to L^p_\om$ is bounded. This raises the question if the maximal operator $N(f)(z)=\sup_{\z\in\G(z)}|f(\z)|$, where $$\label{eq:gammadeuintro} \Gamma(\z)=\left\{z\in {\mathbb{D}}:\,|\t-\arg z|<\frac12\left(1-\frac{|z|}{r}\right)\right\},\quad \z=re^{i\theta}\in{\mathbb{D}}\setminus\{0\},$$ are non-tangential approach regions with vertexes inside the disc [@PR2014Memoirs Chapter 4.1], is bounded from $A^p_\om$ to $L^p_\om$ when $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$. Unfortunately, we do not know the answer to this question. We next proceed to study dominating sets for the whole space $A^p_\om$. For $1<q<\infty$ and a (almost everywhere) positive weight $\om$, write $\om\in C_q$ if for some some (equivalently for each) $r\in(0,1)$, there exists a constant $C=C(q,r,\om)>0$ such that $$\left(\int_{\Delta(z,r)}\om(z)\,dA(z)\right)^\frac1q\left(\int_{\Delta(z,r)}\om(z)^{-\frac{q'}{q}}\,dA(z)\right)^\frac1{q'}\le C |\Delta(z,r)|,\quad z\in{\mathbb{D}},$$ and set $C_\infty=\cup_{q>1}C_q$. Luecking [@L1985/2 Theorem 3.9] showed that if $G\subset{\mathbb{D}}$ is measurable, $0<p<\infty$ and $\om\in C_\infty$ such that $$\label{Eq:Luecking-dominating} |G\cap \Delta(z,r)|\ge\delta|\Delta(z,r)|,\quad z\in{\mathbb{D}},$$ for some $\d>0$ and $r\in(0,1)$, then $G$ is a dominating set for $A^p_{\om}$. This condition is equivalent to the existence of a constant $\delta_0=\delta_0>0$ such that $|G\cap S|>\delta_0|S|$ for all Carleson squares $S$. One can also replace the pseudohyperbolic disc $\Delta(a,r)$ by a suitable Euclidean disc, for example, $D(a, \eta(1-|a|))$ for a fixed $0<\eta<1$ would work here. For the proofs of these equivalences, see [@L1981]. The condition is known to be also a necessary condition for $G$ to be a dominating set for $A^p_{\om}$ if $\om$ is one of the standard weights by [@L1981]. The existing literature does not offer results concerning the converse statement of the above-mentioned result in the non-radial case. We next turn our attention to this matter, and to do it we write $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$ if there exists $C=C(\om)>0$ such that $\om(S(a))\le C\om(S(\frac{1+|a|}{2}e^{i\arg a}))$ for all $a\in{\mathbb{D}}\setminus\{0\}$. It is easy to see that each $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$ satisfies $\om(S(a'))\le C(C+1)\om(S(a))$ for all $a,a'\in{\mathbb{D}}\setminus\{0\}$ with $|a'|=|a|$ and $\arg a'=\arg a\pm(1-|a|)$. Therefore $\om(S(a))\lesssim\om(S(b))$ whenever $|b|=\frac{1+|a|}{2}$ and $S(b)\subset S(a)$. Moreover, it is obvious that radial weights in ${\widehat{\mathcal{D}}}({\mathbb{D}})$ form the class ${\widehat{\mathcal{D}}}$. \[theo:dominating-sets-non-radial\] Let $0<p<\infty$ and $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$. If $G$ is a dominating set for $A^p_\om$, then there exists a constant $\delta > 0$ such that $$\label{eq:delta-condition-set-G-non-radial} \om(G\cap S)>\delta\om(S)$$ for all Carleson squares $S$. The proof of Theorem \[theo:dominating-sets-non-radial\] is based on characterizations of weights in ${\widehat{\mathcal{D}}}({\mathbb{D}})$ given in Lemma \[Lemma:test-functions-non-radial\] below and appropriately chosen test functions. By combining Theorem \[theo:dominating-sets-non-radial\] with [@L1985/2 Theorem 3.9] and imposing severe additional hypothesis on $\om$ one can certainly obtain a characterization of dominating sets in the non-radial case but because of these extra assumptions the resulting description is far from being satisfactory. The approach involving the Lebesgue measure and yielding is natural and has been efficiently used in [@L1981], [@L1984] and [@L1985/2], but it seems that the arguments used there are not adoptable as such to prove to be a sufficient condition. It is of course equally natural to measure the set $G$ as in by using the weight $\om$ itself that induces the space. Moreover, the studies on Carleson measures [@PelSum14], [@PR2014Memoirs], [@PR2015Embedding], [@PelRatSie2015] strongly support the use of Carleson squares instead of pseudohyperbolic discs as testing sets, at least when $\om$ induces a very small weighted Bergman space. It is also worth noticing that the hypothesis $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$ allows $\om$ to vanish in a relatively large part of each outer annulus of ${\mathbb{D}}$, meanwhile weights in $C_\infty$ may not have this property because of the negative power $-\frac{q'}{q}$ appearing in the definition of $C_q$. Therefore a complete solution to the question of when a set $G$ is a dominating set for $A^p_\om$ remains as an open problem in both non-radial weight classes $C_\infty$ and ${\widehat{\mathcal{D}}}({\mathbb{D}})$. Finally, we turn our attention to sampling. A positive Borel measure on ${\mathbb{D}}$ is a sampling measure for $A_\om^p$ if $$\int_{\mathbb{D}}|f(z)|^p \,d\mu(z) \asymp \|f\|_{A_\om^p}^p, \quad f \in A_\om^p.$$ The measures $\mu$ satisfying the inequality ”$\lesssim$” are the $p$-Carleson measures for $A_\om^p$. These measures in the case $\om\in{\widehat{\mathcal{D}}}$ have been studied in [@PR2014Memoirs; @PR2015Embedding; @PelRatSie2015], and can be characterized in terms of the weighted maximal function $$M_\om(\mu)(z) = \sup_{S\ni z}\frac{\mu(S)}{\om(S)}, \quad z \in {\mathbb{D}}:$$ $\mu$ is a $p$-Carleson measure for $A_\om^p$ if and only if $M_\om(\mu) \in L^\infty$, and $\|Id\|_{A^p_\om\to L^p_\mu}^p\asymp\|M_\om(\mu)\|_{L^\infty}$. Therefore these measures are independent of $p$ for each $\om\in{\widehat{\mathcal{D}}}$. Before stating our result, some more notation is needed. For a positive Borel measure $\mu$ on ${\mathbb{D}}$ and $r\in(0,1)$, let $k_r(z) = \mu(\Delta(z,r))/\om(\Delta(z,r))$ for all $z\in{\mathbb{D}}$. The next theorem describes a condition sufficient to guarantee that a positive Borel measure $\mu$ is a sampling measure for $A^p_\om$. Recall that the hypothesis $\mu(\Delta(a,r))\lesssim\om(\Delta(a,r))$ characterizes Carleson measures for certain $A^p_\om$ as mentioned in the paragraph just after Theorem \[Theorem:D-hat\]. \[theo:sampling-suff\] Let $0<p<\infty$, $\e>0$ and either $\om\in{\widehat{\mathcal{D}}}$ such that $\om>0$ almost everywhere on ${\mathbb{D}}$, and $\mu$ a $p$-Carleson measure for $A^p_\om$, or $\om\in C_\infty$ and $\mu(\Delta(a,r))\lesssim\om(\Delta(a,r))$ for all $a\in{\mathbb{D}}\setminus\{0\}$. Then there exists an $r \in (0,1)$ such that $\mu$ is a sampling measure for $A^p_\om$ whenever the set $G = \{z \in {\mathbb{D}}: k_r(z) > \e\|M_\om(\mu)\|_{L^\infty}\}$ is a dominating set for $A^p_\om$. The proof of Theorems \[theo:sampling-suff\] follows the ideas of Luecking [@L1985], but a crucial step in the case of ${\widehat{\mathcal{D}}}$ relies on the characterization of Carleson measures. Additionally, the presence of the weight $\om$ also makes the use of convenient changes of variables and automorphisms difficult and thus forces us to make some more delicate observations. One can readily see from the proof that in the case of ${\widehat{\mathcal{D}}}$ one may omit the extra hypothesis on the positivity of $\om$ by replacing $k_r$ by $k^\star_r(z)=\mu(\Delta(z,r))/\om(S(z))$. In this case the set $G$ may become essentially smaller and thus it being dominating set would be a stronger hypothesis. Let $(\mu_n)$ be a sequence of measures on ${\mathbb{D}}$. We say that $(\mu_n)$ converges weakly to a measure $\mu$, denoted by $\mu_n \rightharpoonup \mu$, if $$\int_{\mathbb{D}}h(z)\,d\mu_n(z) \to \int_{\mathbb{D}}h(z)\,d\mu(z)$$ for all $h$ in the class $C_c({\mathbb{D}})$ of nonnegative continuous compactly supported functions in ${\mathbb{D}}$. The following result is a generalization of [@L2000 Theorem 1], and completes our study of sampling measures. \[theo:w-conv-meas\] Let $0<p<\infty$ and $\om\in{\widehat{\mathcal{D}}}$, and let $(\mu_n)$ be a sequence of $p$-Carleson measures for $A^p_\om$ such that $\sup_n\|M_\om(\mu_n)\|_{L^\infty}< \infty$. Then $(\mu_n)$ has a weakly convergent subsequence. Further, if $\mu_n \rightharpoonup \mu$, then $$\label{eq:w-conv-meas} \lim_{n \to \infty} \int_{\mathbb{D}}|f(z)|^p \,d\mu_n(z) = \int_{\mathbb{D}}|f(z)|^p \,d\mu(z),\quad f \in A_\om^p,$$ and $\mu$ is a $p$-Carleson measure for $A^p_\om$ with $\|Id\|_{A^p_\om \to L^p_\mu}\le\liminf_{n\to\infty}\|Id\|_{A^p_\om\to L^p_{\mu_n}}$. Furthermore, if $\mu_n$ are sampling measures with sampling constants at most $\Lambda > 0$, then $\mu$ is also a sampling measure with a sampling constant at most $\Lambda$. The proof of Theorem \[theo:w-conv-meas\] follows the lines of that of [@L2000 Theorem 1], but the crucial step which differs from the original argument relies on Carleson embedding for tent spaces given in [@PR2015Embedding Theorem 9]. These tent spaces of measurable functions are defined by using the maximal function $N(f)$. Luecking [@L2000] characterized the sampling measures for Bergman spaces induced by standard weights. However, the methods used there do not generalize for weights in the class ${\widehat{\mathcal{D}}}$ because the weights in ${\widehat{\mathcal{D}}}$ can be such that compositions of functions in $A^p_\om$ with Möbius transformations cannot be controlled in norm. Thus, the problem of characterizing sampling measures for $A^p_\om$ when $\om \in {\widehat{\mathcal{D}}}$ remains open. However, the closely related sampling sequences were characterized by Seip [@Seip] in small weighted Bergman spaces induced by weights admitting a pointwise doubling condition in $\om$ instead of $\widehat{\om}$. Zeros and factorization when $\om\in B_\infty$ {#sec:ProofThm1} ============================================== We begin with briefly analyzing the classes $B_q$ and $B_\infty$. For $-1<\alpha<\infty$, write $dA_\alpha(z)=(1-|z|^2)^\alpha\,dA(z)$ for short. Bekollé and Bonami [@B1981; @BB1978] showed that for $1<p<\infty$, $P_\alpha:L^p_{\om_{[\alpha]}}\to L^p_{\om_{[\alpha]}}$ is bounded if and only if $$BB_{p,\alpha}(\om)=\sup_{S}\left(\frac{1}{A_\alpha(S)}\int_{S}\om(z)\,dA_\a(z)\right) \left(\frac{1}{A_\alpha(S)}\int_{S}\om(z)^{-\frac1{p-1}}\,dA_{\a}(z)\right)^{p-1} <\infty,$$ where $A_\alpha(E)=\int_EdA_\alpha$ for each measurable set $E\subset{\mathbb{D}}$. Denote by $BB_{p,\a}$ the set of these weights, and write $BB_{\infty,\a}=\cup_{1<p<\infty}BB_{p,\a}$. By comparing this to the definition of $B_q$, we see that $\om\in B_q$ if and only if $\om_{[2q-2]}\in BB_{q,2}$. Moreover, it is known that $\om\in BB_{\infty,\a}$ if and only if there exist $\delta\in(0,1)$ and $C>0$ such that $$\frac{A_\alpha(E)}{A_\alpha(S)}\le C\left(\frac{\om(E)}{\om(S)}\right)^\d,\quad E\subset S,$$ for all Carleson squares $S$. This condition corresponds to the characterization of the restricted weak-type inequality for the Hardy-Littlewood maximal operator by Kerman and Torchinsky [@KT1980]. For more on $A_\infty$-conditions, see [@DMO2016] and the references therein. Let $\om$ be an almost everywhere strictly positive weight. Then the following assertions hold: - $B_p\subsetneq B_q$ for $1<p<q<\infty$; - $\om\in B_q$ if and only if $W_{q,\om}\in B_{q'}$, where $W_{q,\om}(z)=(\om(z)^\frac1q(1-|z|)^2)^{-q'}$; - If $\om\in B_q$, then $$\frac{A_2(E)}{|S|^2}\le B_q(\om)^\frac1q\left(\frac{\om_{[2q]}(E)}{\om_{[2q]}(S)}\right)^\frac1q,\quad E\subset S\subset{\mathbb{D}};$$ - If there exist $q>1$, $\d\in(\frac1q,1)$ and $C>0$ such that $$\label{3} \frac{A_2(E)}{|S|^2}\le C\left(\frac{\om_{[2q]}(E)}{\om_{[2q]}(S)}\right)^\d,\quad E\subset S\subset{\mathbb{D}},$$ then $\om\in B_q$; - $B_q(\om)\lesssim BB_{q,0}(\om_{[\alpha]})$ for all $\alpha\le2q$. Hölder’s inequality and the inequality $(1-|z|)^2\lesssim|S|$ for $z\in S$ imply $B_q(\om)\lesssim B_p(\om)$ for $1<p<q<\infty$, and thus $B_p\subset B_q$. The inclusion is seen to be strict by considering standard power weight $(1-|z|)^\alpha$ with $p<\a+1<q$. Thus (i) is satisfied. Moreover, since $W_{q,\om}(z)(1-|z|^2)^{2q'}=\om(z)^{-\frac1{q-1}}$ and $W_{q,\om}(z)^{-\frac1{q'-1}}=\om(z)(1-|z|^2)^{2q}$, the assertion (ii) follows by the definition of $B_q$. To prove (iii), assume $\om\in B_q$. Then, for each $E\subset S$, Hölder’s inequality and the definition of $B_q$ give $$\begin{split} A_2(E) &\le\left(\int_E\om_{[2q]}(z)\,dA(z)\right)^\frac1q\left(\int_S\om(z)^{-\frac{q'}{q}}\,dA(z)\right)^\frac1{q'}\\ &\le B_q(\om)^\frac1q|S|^2\left(\frac{\int_E\om_{[2q]}(z)\,dA(z)}{\int_S\om_{[2q]}(z)\,dA(z)}\right)^\frac1q, \end{split}$$ and thus $$\frac{A_2(E)}{|S|^2}\le B_q(\om)^\frac1q\left(\frac{\int_E\om_{[2q]}(z)\,dA(z)}{\int_S\om_{[2q]}(z)\,dA(z)}\right)^\frac1q$$ as claimed. To see (iv), assume and write $E_\lambda=\{z\in S:\om_{[2q-2]}(z)<1/\lambda\}$ for all $\lambda>0$. Then $$\lambda\om_{[2q]}(E_\lambda)\le A_2(E_\lambda)\le C|S|^2\left(\frac{\om_{[2q]}(E_\lambda)}{\om_{[2q]}(S)}\right)^\d,$$ and hence $$\om_{[2q]}(E_\lambda)\le C^\frac1{1-\delta}\lambda^{-\frac{1}{1-\d}}\left(\frac{|S|^2}{\om_{[2q]}(S)^\d}\right)^\frac1{1-\d}.$$ Therefore, by denoting $M=|S|^2/\om_{[2q]}(S)$, we deduce $$\begin{split} \int_S\om(z)^{-\frac1{q-1}}\,dA(z) &=\int_S\frac{\om_{[2q]}(z)}{\left(\om_{[2q-2]}(z)\right)^{q'}}\,dA(z) =q'\int_0^\infty\lambda^{q'-1}\om_{[2q]}(E_\lambda)\,d\lambda\\ &=q'\left(\int_0^M+\int_M^\infty\right)\lambda^{q'-1}\om_{[2q]}(E_\lambda)\,d\lambda\\ &\le\om_{[2q]}(S)M^{q'}+C^\frac1{1-\d}\left(\frac{|S|^2}{\om_{[2q]}(S)^\d}\right)^\frac1{1-\d}q'\int_M^\infty\lambda^{q'-1-\frac1{1-\d}}\,d\lambda\\ &=\frac{|S|^{2q'}}{\left(\om_{[2q]}(S)\right)^{q'-1}}+\frac{C^\frac1{1-\d}q'}{\frac1{1-\d}-q'}\left(\frac{|S|^2}{\om_{[2q]}(S)^\d}\right)^\frac1{1-\d}M^{q'-\frac1{1-\d}}\\ &=\left(1+\frac{q(1-\d)C^{\frac1{1-\d}}}{q\d-1}\right)\frac{|S|^{2q'}}{\left(\om_{[2q]}(S)\right)^{q'-1}} \end{split}$$ and it follows that $\om\in B_q$. To see (v), let $q\ge\frac{\a}{2}\ge0$. Then $$\begin{split} &\om_{[2q]}(S) \left(\int_{S} \om(z)^{-\frac1{q-1}}\,dA(z)\right)^{q-1}\\ &=\left(\int_{S}\om_{[\a]}(z)(1-|z|^2)^{2q-\a}\,dA(z)\right) \left(\int_{S} \om_{[\a]}(z)^{-\frac1{q-1}}(1-|z|^2)^\frac{\a}{q-1}\,dA(z)\right)^{q-1}\\ &\le\left(\int_{S}\om_{[\a]}(z)\,dA(z)\right) \left(\int_{S} \om_{[\a]}(z)^{-\frac1{q-1}}\,dA(z)\right)^{q-1}(1-|a|^2)^{2q}, \end{split}$$ and the assertion follows. The following key proposition is a direct generalization of [@L1996 Theorem 2]. \[prop:zeroseq\] Let $0<p<\infty$ and $\om \in B_\infty$. Let $f \in {\mathcal{H}}({\mathbb{D}})$ and $Z \subset {\mathcal{Z}}(f)$. Then the function $$h(z) = \frac{|f(z)|} {\prod_{a \in Z} \left\{\left|\frac{a-z}{1-\overline{a}z}\right| \exp\left[\frac12\left(1-\left|\frac{a-z}{1-\overline{a}z}\right|^2\right)\right]\right\}},\quad z\in{\mathbb{D}},$$ belongs to $L^p_\om$ if and only if $f \in A^p_\om$. Moreover, there exists a constant $C=C(\om)>0$ such that $$\|f\|_{A^p_\om}^p\le\|h\|_{L^p_\om}^p\le C\|f\|_{A^p_\om}^p, \quad f \in A^p_\om.$$ Since $1-x^2 < 2\log\frac1x$ for $x \in (0,1)$, each factor in the denominator of $h$ is less than $1$. Thus the ”only if” part along with the first inequality is trivial, and for the converse, it suffices to consider the case where $Z = {\mathcal{Z}}(f)$. To see the second inequality, we start by constructing the denominator of $h$. For $f \in {\mathcal{H}}({\mathbb{D}})$ with $f(0) \neq 0$ and the zero sequence ${\mathcal{Z}}(f)$, Jensen’s formula gives $$\label{eq:Jensen} \log|f(0)| + \sum_{a \in {\mathcal{Z}}(f)}\log\frac{r}{|a|}\chi_{[|a|,1)}(r) = \frac1{2\pi} \int_0^{2\pi} \log\left|f\left(re^{i\theta}\right)\right|\,d\theta.$$ If $f\in A^p_\om$ with $\om\in B_\infty$, then $\log|f|$ is area integrable. Indeed, since $\om \in B_\infty$, there exists $q=q(\om)> 1$ such that $\om \in B_q$, and hence $$\begin{split} \int_{\mathbb{D}}\log|f(z)|\,dA(z) &= \int_{\mathbb{D}}\log\left(|f(z)|\om(z)^{1/p}\right)\,dA(z) - \int_{\mathbb{D}}\log\om(z)^{1/p}\,dA(z) \\ &\leq \frac1p\|f\|_{A^p_\om}^p + \frac1p \int_{\mathbb{D}}\log\frac1{\om(z)}\,dA(z) \\ &\leq \frac1p\|f\|_{A^p_\om}^p + \frac{q-1}p \int_{\mathbb{D}}\om(z)^{-1/(q-1)}\,dA(z) \end{split}$$ because $\log x \leq \frac1\d x^\d$ for all $x,\d > 0$. But the last integral is convergent because $\om \in B_q$, and hence $\log|f| \in L^1$. Therefore, as in [@L1996 p. 348], an integration of with respect to $2r\,dr$ now gives $$\log|f(0)| + \sum_{a \in {\mathcal{Z}}(f)}\left(\log\frac{1}{|a|} - \frac12\left(1-|a|^2\right)\right) = \int_{\mathbb{D}}\log|f(w)|\,dA(w),$$ and applying this to $w \mapsto f\left(\vp_z(w)\right)$ yields $$ \log|f(z)| + \sum_{a \in {\mathcal{Z}}(f)}\left[\log\frac{1}{|\vp_a(z)|} - \frac12\left(1-\left|\vp_a(z)\right|^2\right)\right] = \int_{\mathbb{D}}\log|f(w)| \left|\vp_z'(w)\right|^2\,dA(w)$$ for $z \notin {\mathcal{Z}}(f)$. Exponentiating this and applying Jensen’s inequality then gives $$\label{eq:gCalc} \frac{|f(z)|} {\prod_{a \in {\mathcal{Z}}(f)} \left\{\left|\vp_a(z)\right| \exp\left[\frac12\left(1-\left|\vp_a(z)\right|^2\right)\right]\right\}} \leq \left(\int_{\mathbb{D}}|f(w)|^\d \left|\vp_z'(w)\right|^2\,dA(w)\right)^{1/\d}$$ for any $\d>0$. We next consider the linear integral operator $$R(f)(z)=\int_{\mathbb{D}}f(w)\frac{\left(1-|z|^2\right)^2}{|1-\overline{z}w|^4}\,dA(w),\quad z\in{\mathbb{D}},$$ appearing on the right-hand side of , and will show its boundedness on $L^q_\om$. To do this, write $f_{-2}(z) = f(z)\left(1-|z|^2\right)^{-2}$ and $\om_{[x]}(z) = \om(z)\left(1-|z|^2\right)^{x}$ for short. Then $\|f\|_{L^q_\om} = \left\|f_{-2}\right\|_{L^q_{\om_{[2q]}}}$ and $\|R(f)\|_{L^q_\om} = \left\|P^+_2\left(f_{-2}\right)\right\|_{L^q_{\om_{[2q]}}}$, where $$P^+_\a(f)(z) = \int_{\mathbb{D}}\frac{f(w)}{|1-z\overline{w}|^{2+\a}} \left(1-|w|^2\right)^\a\,dA(w),\quad z\in{\mathbb{D}},$$ is the maximal Bergman projection. Hence, $R : L^q_\om \to L^q_\om$ is bounded if and only if $P^+_2 : L^q_{\om_{[2q]}} \to L^q_{\om_{[2q]}}$ is bounded, and the corresponding operator norms are equal. By the characterization of Bekollé and Bonami [@B1981; @BB1978], this is equivalent to $$\int_{S(a)} \om_{[2q-2]}(z)(1-|z|)^2\,dA(z) \left(\int_{S(a)} \om_{[2q-2]}(z)^{-\frac1{q-1}}(1-|z|)^2\,dA(z)\right)^{q-1} \lesssim (1-|a|)^{4q},\quad a\in{\mathbb{D}}\setminus\{0\},$$ that is, $$\int_{S(a)} \om(z)(1-|z|)^{2q}\,dA(z) \left(\int_{S(a)} \om(z)^{-\frac1{q-1}}\,dA(z)\right)^{q-1} \lesssim (1-|a|)^{4q},\quad a\in{\mathbb{D}}\setminus\{0\}.$$ Thus, $R : L^q_\om \to L^q_\om$ is bounded if and only if $\om \in B_q$. Moreover, $\|R\|_{L^q_\om \to L^q_\om}\le CB_q(\om)^{\max\{1,\frac1{q-1}\}}$, where $C=C(\om,q)$, by [@PottRegueraJFA13 Theorem 1.5]. Let now $Z = {\mathcal{Z}}(f)$. Clearly, $f \in A^p_\om$ implies $|f|^\d \in L^{p/\d}_\om$. Choose $\d = p/q$, so that $p/\d = q$. By the boundedness of $R$ and , we obtain $$\|h\|_{L^p_\om}^p\le\left\|R\left(|f|^\frac{p}{q}\right)^{\frac{q}{p}}\right\|_{L^p_\om}^p = \left\|R\left(|f|^\frac{p}{q}\right)\right\|^q_{L^{q}_\om}\le C^qB_q(\om)^{\max\{q,q'\}}\|f\|^p_{A^p_\om},$$ where $C=C(q,\om)$. Since $q=q(\om)$, the norm estimate we are after follows. [Theorem \[theo:zeroset\]]{} We start the proof by using the function $h$ in Proposition \[prop:zeroseq\] to find the analytic function $F$ in cases (c) and (d) of the theorem. To do this, we first make some observations about the functions $\psi$, $k$ and $W$ defined in Section \[sec:Intro\]. Standard estimates using the power series of the exponential show that $$\left|\overline{a}\frac{a-z}{1-\overline{a}z} \exp\left(1-\overline{a}\frac{a-z}{1-\overline{a}z}\right) - 1\right| = O\left((1-|a|)^2\right), \quad |a| \to 1^-,$$ uniformly on compact subsets of ${\mathbb{D}}$, and thus $\psi\in{\mathcal{H}}({\mathbb{D}})$ if $\sum_{a \in Z} (1-|a|)^2 < \infty$. This sum converges for $Z \subset {\mathcal{Z}}(f)$ whenever $\log |f|$ is integrable on ${\mathbb{D}}$. This, in turn, is true for any $f \in A^p_\om$ with $\om\in B_\infty$ by the proof of Proposition \[prop:zeroseq\]. A direct calculation shows that $$\begin{aligned} {\textrm{Re}\,}\left(1-\overline{a}\frac{a-z}{1-\overline{a}z}\right) &= \frac12\left(1-|a|^2\right) + \frac12\left(1-\left|\frac{a-z}{1-\overline{a}z}\right|^2\right) + \frac{|z|^2}2\frac{\left(1-|a|^2\right)^2}{|1-\overline{a}z|^2},\end{aligned}$$ and therefore $$\begin{aligned} |\psi_Z(z)| &= \prod_{a \in Z} |a|e^{\left(1-|a|^2\right)/2} \prod_{a \in Z}\left\{\left|\vp_a(z)\right| \exp\left[\frac12\left(1-\left|\vp_a(z)\right|^2\right)\right]\right\} \exp\left[\frac{|z|^2}2 \sum_{a \in Z} \frac{\left(1-|a|^2\right)^2}{|1-\overline{a}z|^2}\right],\end{aligned}$$ see [@L1996] for details. The last exponential is now exactly the function $W_Z$ defined in the first section, and the second product is the denominator of $h$ in Proposition \[prop:zeroseq\]. The function $h$ can thus be written as $$h(z) = C\left|\frac{f(z)}{\psi_Z(z)}\right|W_Z(z),$$ where $C = \prod_{a \in Z} |a|e^{\left(1-|a|^2\right)/2}$. We will soon see that $f/\psi_Z$ is the function $F$ we are after. We first show the equivalence between (a) and (c). The calculations above together with Proposition \[prop:zeroseq\] show that if $f \in A^p_\om$ and $Z = {\mathcal{Z}}(f)$, then $\sum_{a \in Z} (1-|a|)^2 < \infty$ and $FW_Z = C^{-1}h \in L^p_\om$, where $F = f/\psi_Z$ has no zeros. Conversely, if $F$ is a nowhere zero analytic function with $FW_Z \in L^p_\om$, $\sum_{a \in Z} (1-|a|)^2 < \infty$, then $f = F\psi_Z \in A^p_\om$ because $|\psi_Z(z)| \leq W_Z(z)$ for all $z$, and clearly ${\mathcal{Z}}(f) = Z$. Since considering a subsequence $Z'$ of $Z$ instead of $Z$ will only decrease the values of $k_Z$ and $W_Z$, it is clear that (a) and (b) are equivalent; (a) $\Rightarrow$ (c) $\Rightarrow$ (b) $\Rightarrow$ (a). Now, if $F$ is any (nonzero) analytic function with $FW_{Z'} \in L^p_\om$, $\sum_{a \in Z'} (1-|a|)^2 < \infty$, then $f = F\psi_{Z'} \in A^p_\om$ and $Z' \subset {\mathcal{Z}}(f)$. The equivalence of (a) and (b) now gives $Z' \in {\mathcal{Z}}(A^p_\om)$, and thus (d) implies (a). Since (c) is a special case of (d), the equivalence part of the theorem is proved. Finally, because of the part of the proof considering the function $h$, it is clear that the last statement of the theorem is simply a restatement of Proposition \[prop:zeroseq\]. [Theorem \[theo:factorization\]]{} The first and the last inequality are obvious, so it suffices to prove the middle one. For $0<p<q<\infty$, $\om\in B_\infty$ and $f\in A^p_\om$, consider the function $$g(z)=|f(z)|^p\prod_{z_k\in\mathcal{Z}(f)}\frac{1-\frac{p}{q}+\frac{p}{q}|\vp_{z_k}(z)|^q}{|\vp_{z_k}(z)|^p},\quad z\in{\mathbb{D}},$$ defined in [@HorFacto Lemma 2], and let $h$ be as in Proposition \[prop:zeroseq\] with $Z=\mathcal{Z}(f)$. Let $n(r,f)$ denote the number of zeros of $f$ in $D(0,r)$, counted according to multiplicity, and $$N(r,f)=\int_0^r\frac{n(s)-n(0,f)}{s}\,ds+n(0,f)\log r$$ be the integrated counting function. Two integrations by parts and Jensen’s formula show that $$\begin{split}\sum_{z_k\in\mathcal{Z}(f)}\log\left(\frac{1-\frac{p}{q}+\frac{p}{q}|z_k|^q}{|z_k|^p}\right) &=\int_0^1\log\left(\frac{1-\frac{p}{q}+\frac{p}{q}r^q}{r^p}\right)\,dn(r) =\int_0^1\frac{(\frac{q}{p}-1)(1-r^q)}{\frac qp-1+r^q}\frac{pn(r)}{r}\,dr\\ &=-\int_0^1pN(r)\,du(r) =\int_{\mathbb{D}}\log|f(w)|^p\,d\sigma(w)-\log|f(0)|^p, \end{split}$$ where $$d\sigma(w)=-u'(|w|)\frac{dA(w)}{2|w|},\quad u(r)=\frac{(\frac{q}{p}-1)(1-r^q)}{\frac{q}{p}-1+r^q},$$ is a positive measure of unit mass on ${\mathbb{D}}$. Hence $$\begin{split} \log(g(0))&=\log\left(|f(0)|^p\prod_{z_k\in\mathcal{Z}(f)}\left(\frac{1-\frac{p}{q}+\frac{p}{q}|z_k|^q}{|z_k|^p}\right)\right)\\ &=\log|f(0)|^p+\sum_{z_k\in\mathcal{Z}(f)}\log\left(\frac{1-\frac{p}{q}+\frac{p}{q}|z_k|^q}{|z_k|^p}\right) =\int_{\mathbb{D}}\log|f(w)|^p\,d\sigma(w). \end{split}$$ Replacing now $f$ by $f\circ\vp_z$ and using [@PR2014Memoirs (3.9)] to pass from $d\s$ to $dA$, we obtain $$ \begin{split} \log(g(z)) &=\int_{\mathbb{D}}\log|f(\vp_z(w))|^p\,d\sigma(w) \leq \int_{\mathbb{D}}\log |f(\vp_z(w))|^p\,dA(w) \\ &=p\int_{\mathbb{D}}\log |f(\vp_z(w))|\,dA(w) = p\log h(z) = \log h(z)^p, \quad z \notin{\mathcal{Z}}(f). \end{split}$$ Thus $g \leq h^p$, from which Proposition \[prop:zeroseq\] gives $\|g\|_{L^1_\om}\le C\|f\|_{A^p_\om}^p$ for some constant $C=C(\om)>0$. This is the statement of [@PR2014Memoirs Lemma 3.3] with the difference that now we have better control over the constant appearing on the right-hand side of the inequality and $\om$ is only required to belong to $B_\infty$. By following the proof of [@PR2014Memoirs Theorem 3.1], which in turn follows Horowitz’ original probabilistic argument, now gives the assertion of Theorem \[theo:factorization\] for functions $f$ with finitely many zeros. To complete the argument used in the said proof, it suffices to show that every norm-bounded family in $A^p_\om$ with $\om\in B_\infty$ is a normal family of analytic functions. To see this, let $x>1$ such that $\om\in B_x$. Then the subharmonicity and Hölder’s inequality yield $$\begin{split} |f(z)|^\frac{p}{x}&\lesssim\frac{1}{(1-|z|)^2}\int_{\Delta(z,r)}|f(\z)|^\frac{p}{x}\om(\z)^\frac1x\om(\z)^{-\frac1x}\,dA(\z)\\ &\le\frac{1}{(1-|z|)^2}\left(\int_{\Delta(z,r)}|f(\z)|^p\om(\z)\,dA(\z)\right)^\frac1x \left(\int_{\Delta(z,r)}\om(\z)^{-\frac{x'}x}\,dA(\z)\right)^\frac1{x'}\\ &\le\frac{1}{(1-|z|)^2}\|f\|_{A^p_\om}^\frac{p}x\left(\int_{{\mathbb{D}}}\om(\z)^{-\frac{x'}x}\,dA(\z)\right)^\frac1{x'}, \end{split}$$ where the constant of comparison depends only on the fixed $r\in(0,1)$. Since $\om\in B_x$, the last integral is finite, and Montel’s theorem shows that every norm-bounded family in $A^p_\om$ is a normal family of analytic functions. With this guidance we consider Theorem \[theo:factorization\] proved. Zeros and factorization when $\om\in{\widehat{\mathcal{D}}}$ {#Sec:radial-zeros-factorization} ============================================================ We will need the following technical auxiliary result [@PR2015Embedding Lemma 1]. \[Lemma:weights-in-D-hat\] Let $\om$ be a radial weight. Then the following statements are equivalent: - $\om\in{\widehat{\mathcal{D}}}$; - There exist $C=C(\om)>0$ and $\b=\b(\om)>0$ such that $$\begin{split} \widehat{\om}(r)\le C\left(\frac{1-r}{1-t}\right)^{\b}\widehat{\om}(t),\quad 0\le r\le t<1; \end{split}$$ - There exist $C=C(\om)>0$ and $\gamma=\gamma(\om)>0$ such that $$\begin{split} \int_0^t\left(\frac{1-t}{1-s}\right)^\g\om(s)\,ds \le C\widehat{\om}(t),\quad 0\le t<1. \end{split}$$ [Theorem \[Theorem:D-hat\]]{} As explained in the introduction we begin with modifying the proof of Proposition \[prop:zeroseq\] so that it covers the case $\om\in{\widehat{\mathcal{D}}}$. This boils down to showing for sufficiently large $q=q(\om)>\max\{p,1\}$. To prove , let $q\ge2$ and $k(\z)=(1-|\z|)^\e$, where $\e<1-1/q$ is fixed. Writing $I(f)$ for the left-hand side of and using Hölder’s inequality and Fubini’s theorem, we have $$\begin{split} I(f)&=\int_{\mathbb{D}}\left(\int_{\mathbb{D}}|f(\z)|^\frac{p}{q}\frac{k(\z)}{|1-\overline{z}\z|^2}\frac{k^{-1}(\z)}{|1-\overline{z}\z|^2}\,dA(\z)\right)^q\om_{[2q]}(z)\,dA(z)\\ &\le\int_{\mathbb{D}}|f(\z)|^pk(\z)^q\left(\int_{\mathbb{D}}\frac{\om_{[2q]}(z)}{|1-\overline{z}\z|^{2q}} \left(\int_{\mathbb{D}}\frac{dA(u)}{k(u)^{q'}|1-\overline{z}u|^{2q'}}\right)^{q-1}dA(z)\right)dA(\z). \end{split}$$ Since $\e<1-1/q$, $$\begin{split} \int_{\mathbb{D}}\frac{\om_{[2q]}(z)}{|1-\overline{z}\z|^{2q}} \left(\int_{\mathbb{D}}\frac{dA(u)}{k(u)^{q'}|1-\overline{z}u|^{2q'}}\right)^{q-1}dA(z) &\asymp\int_{\mathbb{D}}\frac{\om_{[2q-2-\e q]}(z)}{|1-\overline{z}\z|^{2q}}\,dA(z)\\ &\asymp\int_0^1\frac{\om(s)(1-s)^{2q-2-\e q}}{(1-|\z|s)^{2q-1}}\,ds, \end{split}$$ and therefore $I(f)\lesssim\|f\|_{L^p_\mu}^p$, where $$\begin{split} d\mu(\z)=(1-|\z|)^{\e q}\left(\int_0^{|\z|}\frac{\om(s)}{(1-s)^{1+\e q}}\,ds +(1-|\z|)^{1-2q}\int_{|\z|}^1\om_{[2q-2-\e q]}(s)\,ds\right)dA(\z). \end{split}$$ To establish , it now suffices to show that $\mu$ is a $p$-Carleson measure for $A^p_\om$, that is, $\mu(S)\lesssim\om(S)$ for all Carleson squares $S$ by [@PR2015Embedding Theorem 1]. By Fubini’s theorem and the inequality $\e q-2q+1<-q<-1$, the term corresponding to the second summand satisfies $$\begin{split} \int_r^1(1-t)^{\e q+1-2q}\int_{t}^1\om_{[2q-2-\e q]}(s)\,ds\,dt \le\int_r^1\om_{[2q-2-\e q]}(s)\int_0^s(1-t)^{\e q-2q+1}\,dt\,ds \lesssim\widehat{\om}(r) \end{split}$$ while a similar reasoning (divide the integral from $0$ to $t$ at $r$) for the first term shows that $$\begin{split} \int_r^1(1-t)^{\e q}\left(\int_0^{t}\frac{\om(s)}{(1-s)^{1+\e q}}\,ds\right)dt \lesssim(1-r)^{\e q+1}\int_0^{r}\frac{\om(s)}{(1-s)^{1+\e q}}\,ds+\widehat{\om}(r). \end{split}$$ By choosing $\e q$ sufficiently large, Lemma \[Lemma:weights-in-D-hat\](iii) gives what we want. Therefore we have proved the statement of Proposition \[prop:zeroseq\] for $\om\in{\widehat{\mathcal{D}}}$. The proofs of Theorems \[theo:zeroset\] and \[theo:factorization\] now work as such and thus the statements of these results as well as Corollaries \[theo:zeroset2\] and \[coro:zeroPerturb\] are valid for $\om\in{\widehat{\mathcal{D}}}$. [Corollary \[corollary:Hankel\]]{} For each $p>1$ and $\om\in{\mathcal{D}}$, the dual of $A^p_\om$ can be identified with $A^{p'}_\om$ via the $A^2_\om$-pairing $\langle f,g\rangle_{A^2_\om}=\lim_{r\to1^-}\int_{\mathbb{D}}f_r\overline{g_r}\om\,dA$, where $f_r(z)=f(rz)$, by the proofs of [@PR2016TwoWeight Theorem 6 and Corollary 7]. Therefore the boundedness of $h^\om_{\overline{f}}:A^p_\om\to\overline{A^q_\om}$ is equivalent to $$\label{Equation:hankel} \begin{split} |\langle h^\om_{\overline{f}}(g),h\rangle_{A^2_\om}| &=\left|\lim_{r\to1^-}\int_{\mathbb{D}}\overline{f(z)}g(z)\overline{h(r^2z)}\om(z)\,dA(z)\right|\\ &\lesssim\|g\|_{A^p_\om}\|h\|_{L^{q'}_\om},\quad g\in A^p_\om,\quad h\in \overline{A^{q'}_\om}. \end{split}$$ If $f\in A^s_\om$ with $\frac1s=\frac1q-\frac1p$, then Hölder’s inequality implies $|\langle h^\om_{\overline{f}}(g),h\rangle_{A^2_\om}|\le\|g\|_{A^p_\om}\|h\|_{L^{q'}_\om}\|f\|_{A^s_\om}$, and hence $h^\om_{\overline{f}}:A^p_\om\to\overline{A^q_\om}$ is bounded and $\|h^\om_{\overline{f}}\|_{A^p_\om\to\overline{A^q_\om}}\le\|f\|_{A^s_\om}$. Conversely, assume that $h^\om_{\overline{f}}:A^p_\om\to\overline{A^q_\om}$ is bounded. By Theorem \[Theorem:D-hat\] each $k\in A^{s'}_\om$ can be factorized as $k=g\overline{h}$, where $g\in A^p_\om$ and $\overline{h}\in A^{q'}_\om$ with $\|g\|_{A^p_\om}\|\overline{h}\|_{A^{q'}_\om}\asymp\|k\|_{A^{s'}_\om}$. Therefore implies $|\langle f,k\rangle_{A^2_\om}|\lesssim\|k\|_{A^{s'}_\om}$ for all $k\in A^{s'}_\om$. The duality $(A^s_\om)^*\simeq A^{s'}_\om$ now yields $f\in A^s_\om$. Dominating sets =============== In this section we prove our results on dominating sets. We begin with Theorems \[theo:dominating-set-for-f\] and \[theo:dominating-set-for-f-2\] concerning individual functions. [Theorem \[theo:dominating-set-for-f\]]{} By [@PRS2015Berezin Lemma 8], there exists $r=r(\om)\in(0,1)$ such that $|B^\om_z(\z)|\asymp B^\om_z(z)$ for all $\z\in\Delta(z,r)$ and $z\in{\mathbb{D}}$. Further, $\|B^\om_z\|_{A^2_\om}^2\asymp({\widehat{\om}}(z)(1-|z|))^{-1}$ by [@PR2016TwoWeight Theorem 1], and ${\widehat{\om}}$ is essentially constant in each hyperbolically bounded set by Lemma \[Lemma:weights-in-D-hat\](ii). By using these facts and the subharmonicity of $|f|^q$ we deduce $$|f(z)|^q\lesssim\int_{\Delta(z,r)}|f(\z)|^q\frac{|B^\om_z(\z)|^2}{\|B^\om_z\|_{A^2_\om}^2}\frac{{\widehat{\om}}(\z)}{1-|\z|}\,dA(\z), \quad z \in {\mathbb{D}},$$ for all $f\in{\mathcal{H}}({\mathbb{D}})$. Therefore  is proved once we have shown that there exists a constant $C=C(q,r,\om)>0$ such that $$\label{eq:carleson-application} \int_{\Delta(z,r)}|f(\z)|^q|B^\om_z(\z)|^2\frac{{\widehat{\om}}(\z)}{1-|\z|}\,dA(\z) \le C\int_{\mathbb{D}}|f(\z)|^q|B^\om_z(\z)|^2\om(\z)\,dA(\z), \quad z \in {\mathbb{D}},$$ for all $f \in A^q_\om$. This looks like a consequence of a Carleson embedding theorem, but these theorems do not seem efficiently applicable as such since $|f|^q|B^\om_z|^2$ can not be written as a power of the modulus of a single analytic function because both $f$ and $B^\om_z$ may have zeros. However, a careful inspection of the proof of [@PelSum14 Theorem 3.3] shows that the argument, and in particular the proof of [@PelSum14 Lemma 3.2], carries over if any positive power of the modulus of the function involved is subharmonic. Now that $|f|^q|B^\om_z|^2$ has this local property because both $f$ and $B^\om_z$ are analytic, we deduce if the measure $\mu_z$ defined by $d\mu_z(\z) = \frac{{\widehat{\om}}(\z)}{1-|\z|}\chi_{\Delta(z,r)}(\z)\,dA(\z)$ is a $1$-Carleson measure for $A^1_\om$ with $\|M_\om(\mu_z)\|_{L^\infty}$ uniformly bounded in $z$, that is, if $\mu_z(S(a)) \lesssim \om(S(a))$ for all $a \in {\mathbb{D}}\setminus\{0\}$ and $z \in {\mathbb{D}}$. But clearly, $\Delta(z,r)\subset\{se^{i\t}:\r\le s\le\r+x(1-\r),\,\,|\t-\vp|\le x(1-\r)\}$ for some $x=x(r)\in(0,1)$, $\r=\r(z,r)\in(0,1)$ and $\vp=\vp(z)$. If now $|a|\ge \r+x(1-\r)$ there is nothing to prove, while for otherwise $$\begin{split} \int_{S(a)\cap\Delta(z,r)}\frac{{\widehat{\om}}(\z)}{1-|\z|}\,dA(\z) &\le(1-|a|)\int_{\max\{|a|,\r\}}^{\r+x(1-\r)}\frac{{\widehat{\om}}(s)}{1-s}\,ds\\ &\le(1-|a|)\widehat{\om}(a)\frac{\r+x(1-\r)-\max\{|a|,\r\}}{1-\r-x(1-\r)}\\ &\le(1-|a|)\widehat{\om}(a)\frac{x}{1-x}. \end{split}$$ It follows that $\mu_z(S(a)) \lesssim \om(S(a))$ for all $a \in {\mathbb{D}}\setminus\{0\}$ and $z \in {\mathbb{D}}$, and thus is proved. Let now $f\in A^p_\om$. Then  implies $$|f(z)|^q\chi_E(z) \le \e \int_{\mathbb{D}}|f(\z)|^q K^\om_z(\z)\om(\z)\,dA(\z),\quad z\in{\mathbb{D}}.$$ Raising this to power $p/q$ and integrating with respect to $\om\,dA$ now gives $$\int_E |f(z)|^p\om(z)\,dA(z) \le \e^{p/q} \int_{\mathbb{D}}\left(\int_{\mathbb{D}}|f(\z)|^q K^\om_z(\z)\om(\z)\,dA(\z)\right)^{p/q}\om(z)\,dA(z)=\e^{p/q}I(f).$$ Therefore  will be proved if we can show that $I(f)$ is dominated by a constant times $\|f\|_{A^p_\om}^p$. If $\om\in{\mathcal{D}}$, there is an easy to way to deduce this because the maximal Bergman projection $P^+_\om:L^p_\om\to L^p_\om$ is bounded for each $1<p<\infty$ by the proof of [@PR2016TwoWeight Theorem 3]. To see how  is obtained from this fact, denote $s = p/q$ and let $1 < x < s$ and $g \in L^s_\om$. Hölder’s inequality then gives $$\begin{split} \int_{\mathbb{D}}&\left|\int_{\mathbb{D}}g(\xi) K^\om_z(\z)\om(\z)\,dA(\z)\right|^s\om(z)\,dA(z) \\ &\le \int_{\mathbb{D}}\left(\int_{\mathbb{D}}|g(\z)|^x |B_z^\om(\z)| \om(\z)\,dA(\z)\right)^{s/x} \left(\int_{\mathbb{D}}|B_z^\om(\z)|^{\frac{2x-1}{x-1}}\om(\z)\,dA(\z)\right)^{s\frac{x-1}x} \frac{\om(z)}{\|B_z^\om\|^{2s}_{A^2_\om}}\,dA(z) \\ &\leq \int_{\mathbb{D}}\left(\int_{\mathbb{D}}|g(\z)|^x |B_z^\om(\z)| \om(\z)\,dA(\z)\right)^{s/x}\om(z)\,dA(z)\, \sup_{z\in{\mathbb{D}}} \frac{\|B_z^\om\|^{s\frac{2x-1}x}_{A^y_\om}}{\|B_z^\om\|^{2s}_{A^2_\om}}, \end{split}$$ where $y = \frac{2x-1}{x-1}$. Since $s/x > 1$ and $|g|^x \in L^{s/x}_\om$, the maximal Bergman projection $P^+_\om:L^\frac{s}{x}_\om\to L^\frac{s}{x}_\om$ is bounded, and hence $$\int_{\mathbb{D}}\left(\int_{\mathbb{D}}|g(\xi)|^x |B_z^\om(\xi)| \om(\xi)\,dA(\xi)\right)^{s/x}\om(z)\,dA(z) \lesssim \|g\|^s_{L^s_\om}.$$ On the other hand, [@PR2016TwoWeight Theorem 1] gives $$\frac{\|B_z^\om\|^{s\frac{2x-1}x}_{A^y_\om}}{\|B_z^\om\|^{2s}_{A^2_\om}} \asymp \frac{{\widehat{\om}}(z)^s(1-|z|)^s}{\left({\widehat{\om}}(z)^{y-1}(1-|z|)^{y-1}\right)^{s\frac{2x-1}{xy}}} = 1,\quad |z| \to 1^-,$$ because $$(y-1)s\frac{2x-1}{xy} = \frac{2x-1-(x-1)}{x-1}s\frac{2x-1}{x\frac{2x-1}{x-1}} = \frac{x}{x-1}s\frac{x-1}{x} = s.$$ This proves  for $\om\in{\mathcal{D}}\subsetneq{\widehat{\mathcal{D}}}$. The proof for the whole class ${\widehat{\mathcal{D}}}$ is more laborious but it also relies strongly on the kernel estimates used in the above reasoning. We argue as follows. Let $k(\z)=\widehat{\om}(\z)^\e$, where $\e<1-\frac{q}{p}$. Since $\|B^\om_z\|_{A^2_\om}^2\asymp(\widehat{\om}(z)(1-|z|))^{-1}$, Hölder’s inequality and Fubini’s theorem yield $$\begin{split} I(f) &\asymp\int_{\mathbb{D}}\left(\int_{\mathbb{D}}|f(\z)|^q|B^\om_z(\z)|k(\z)|B^\om_z(\z)|k(\z)^{-1}\om(\z)\,dA(\z)\right)^{p/q}(\widehat{\om}(z)(1-|z|))^{\frac{p}{q}}\om(z)\,dA(z)\\ &\le\int_{\mathbb{D}}\left(\int_{\mathbb{D}}|f(\z)|^p|B^\om_z(\z)|^\frac{p}{q}k(\z)^\frac{p}{q}\om(\z)\,dA(\z)\right)\\ &\quad\cdot\left(\int_{\mathbb{D}}|B^\om_z(u)|^{\frac{p}{p-q}}k(u)^{-\frac{p}{p-q}}\om(u)\,dA(u)\right)^{\frac{p-q}{q}}(\widehat{\om}(z)(1-|z|))^{\frac{p}{q}}\om(z)\,dA(z)\\ &=\int_{\mathbb{D}}|f(\z)|^pk(\z)^\frac{p}{q}\om(\z)\left(\int_{\mathbb{D}}|B^\om_z(\z)|^\frac{p}{q}(\widehat{\om}(z)(1-|z|))^{\frac{p}{q}}\om(z)\right.\\ &\quad\cdot\left.\left(\int_{\mathbb{D}}|B^\om_z(u)|^{\frac{p}{p-q}}k(u)^{-\frac{p}{p-q}}\om(u)\,dA(u)\right)^{\frac{p-q}{q}}\,dA(z)\right)\,dA(\z), \end{split}$$ where $$\begin{split} \int_{\mathbb{D}}|B^\om_z(u)|^{\frac{p}{p-q}}k(u)^{-\frac{p}{p-q}}\om(u)\,dA(u) &\lesssim\int_0^{|z|}\frac{\int_t^1\widehat{\om}(s)^{-\frac{\e p}{p-q}}\om(s)\,ds}{\widehat{\om}(t)^{\frac{p}{p-q}}(1-t)^\frac{p}{p-q}}\,dt\\ &\asymp\int_0^{|z|}\frac{dt}{\widehat{\om}(t)^{\frac{p}{p-q}-1+\frac{\e p}{p-q}}(1-t)^\frac{p}{p-q}}\\ &\asymp\frac{1}{\widehat{\om}(z)^{\frac{p}{p-q}-1+\frac{\e p}{p-q}}(1-|z|)^{\frac{p}{p-q}-1}} \end{split}$$ by [@PR2016TwoWeight Theorem 1]. This together with another application of  [@PR2016TwoWeight Theorem 1] gives $$\begin{split} I(f) &\lesssim\int_{\mathbb{D}}|f(\z)|^p\widehat{\om}(\z)^\frac{\e p}{q}\om(\z)\left(\int_{\mathbb{D}}|B^\om_z(\z)|^\frac{p}{q}\widehat{\om}(z)^{\frac{p}{q}-1-\frac{\e p}{q}}\om(z)(1-|z|)^{\frac{p-q}{q}}\,dA(z)\right)dA(\z)\\ &\lesssim\int_{\mathbb{D}}|f(\z)|^p\widehat{\om}(\z)^\frac{\e p}{q}\om(\z)\left(\int_0^{|\z|} \frac{\int_t^1\widehat{\om}(s)^{\frac{p}{q}-1-\frac{\e p}{q}}\om(s)(1-s)^{\frac{p-q}{q}}\,ds}{\widehat{\om}(t)^\frac{p}{q}(1-t)^\frac{p}{q}}\,dt\right)dA(\z). \end{split}$$ The integral from $0$ to $|\z|$ equals to $I_1+I_2$, where $$\begin{split} I_1&=\left(\int_{|\z|}^1\widehat{\om}(s)^{\frac{p}{q}-1-\frac{\e p}{q}}\om(s)(1-s)^{\frac{p-q}{q}}\,ds\right) \int_0^{|\z|} \frac{dt}{\widehat{\om}(t)^\frac{p}{q}(1-t)^\frac{p}{q}}\\ &\lesssim\left((1-|\z|)^\frac{p-q}{q}\widehat{\om}(\z)^{\frac{p}{q}-\frac{\e p}{q}}\right) \frac{1}{\widehat{\om}(\z)^\frac{p}{q}(1-|\z|)^{\frac{p}{q}-1}}=\widehat{\om}(\z)^{-\frac{\e p}{q}} \end{split}$$ and, by Fubini’s theorem, $$\begin{split} I_2&=\int_0^{|\z|} \frac{\int_{t}^{|\z|}\widehat{\om}(s)^{\frac{p}{q}-1-\frac{\e p}{q}}\om(s)(1-s)^{\frac{p-q}{q}}\,ds}{\widehat{\om}(t)^\frac{p}{q}(1-t)^\frac{p}{q}}\,dt\\ &=\int_0^{|\z|}\widehat{\om}(s)^{\frac{p}{q}-1-\frac{\e p}{q}}\om(s)(1-s)^{\frac{p-q}{q}} \left(\int_{0}^{s}\frac{dt}{\widehat{\om}(t)^\frac{p}{q}(1-t)^\frac{p}{q}}\right)ds\\ &\asymp\int_0^{|\z|}\widehat{\om}(s)^{-1-\frac{\e p}{q}}\om(s)\,ds\asymp\widehat{\om}(\z)^{-\frac{\e p}{q}}. \end{split}$$ We deduce $I(f)\lesssim\|f\|_{A^p_\om}^p$, and the theorem is proved. [Theorem \[theo:dominating-set-for-f-2\]]{} By the definition of the set $E$ and Fubini’s theorem, $$\begin{split} \int_E |f(z)|^p\om(z)\,dA(z) &\le \e \int_E \left(\frac1{\nu(E(z))}\int_{E(z)}|f(\zeta)|^p\nu(\zeta)\,dA(\zeta)\right)\om(z)\,dA(z)\\ &=\e \int_{\mathbb{D}}|f(\zeta)|^p\left(\int_E \frac{\chi_{E(z)}(\zeta)}{\nu(E(z))} \om(z)\,dA(z)\right)\nu(\zeta)\,dA(\zeta). \end{split}$$ Therefore to complete the proof, it suffices to show that the measure $\mu$ defined by $$d\mu(\z) =\left(\int_E \frac{\chi_{E(z)}(\zeta)}{\nu(E(z))} \om(z)\,dA(z)\right)\nu(\zeta)\,dA(\zeta),\quad \z\in{\mathbb{D}},$$ is a $p$-Carleson measure for $A^p_\om$. To see this, let $a\in{\mathbb{D}}\setminus\{0\}$. Then Fubini’s theorem gives $$\begin{split} \mu(S(a)) &=\int_{S(a)}\left(\int_E \frac{\chi_{E(z)}(\zeta)}{\nu(E(z))} \om(z)\,dA(z)\right)\nu(\zeta)\,dA(\zeta) =\int_E\frac{\nu(S(a) \cap E(z))}{\nu(E(z))} \om(z)\,dA(z)\\ &\le\int_{\{z:S(a)\cap \Delta(z,r)\ne\emptyset\}}\om(z)\,dA(z). \end{split}$$ Now that there exist $C=C(r)>0$ and $R=R(r,C)\in(0,1)$ such that $\{z:S(a)\cap \Delta(z,r)\ne\emptyset\}\subset S(b)$ for some $b=b(a)\in{\mathbb{D}}\setminus\{0\}$ with $\arg b=\arg a$ and $1-|b|=C(1-|a|)$ and for all $|a|\ge R$, we deduce $$\begin{split} \mu(S(a)) &\le\om(S(b)) \le C(1-|a|) \int_{1-C(1-|a|)}^1\om(s)\,ds\\ &\lesssim C(1-|a|)\left(\frac{C(1-|a|)}{1-|a|}\right)^\b\widehat{\om}(a)\asymp \om(S(a)), \quad |a|\ge R, \end{split}$$ by the hypothesis $\om\in{\widehat{\mathcal{D}}}$ and Lemma \[Lemma:weights-in-D-hat\]. Consequently, $\mu$ is a $p$-Carleson measure for $A^p_\om$ by [@PR2015Embedding Theorem 1], and the lemma is proved. We next proceed towards the proof of Theorem \[theo:dominating-sets-non-radial\]. \[Lemma:test-functions-non-radial\] Let $\om$ be a weight on ${\mathbb{D}}$ such that $\om(S(a))>0$ for all $a\in{\mathbb{D}}\setminus\{0\}$. Then the following statements are equivalent: - $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$; - there exist $\b=\b(\om)>0$ and $C=C(\om)\ge1$ such that $$\frac{\om(S(a))}{(1-|a|)^\b}\le C\frac{\om(S(a'))}{(1-|a'|)^\b},\quad 0<|a|\le|a'|<1,\quad\arg a=\arg a';$$ - for some (equivalently for each) $K>0$ there exists $C=C(\om,K)>0$ such that $$\om(S(a))\le C\om\left(S\left(\frac{K+|a|}{K+1}e^{i\arg a}\right)\right),\quad a\in{\mathbb{D}}\setminus\{0\};$$ - there exist $\eta=\eta(\om)>0$ and $C=C(\eta,\om)>0$ such that $$\displaystyle \int_{\mathbb{D}}\frac{\om(z)}{|1-\overline{a}z|^\eta}\,dA(z)\le C\frac{\om(S(a))}{(1-|a|)^\eta},\quad a\in{\mathbb{D}}\setminus\{0\}.$$ Assume (i), and let $r_n=1-2^{-n}$ for $n\in{\mathbb{N}}\cup\{0\}$. Let $0<|a|\le|a'|<1$ with $\arg a=\arg a'$, and take $k,m\in{\mathbb{N}}\cup\{0\}$ such that $r_k\le|a|<r_{k+1}$ and $r_m\le|a'|<r_{m+1}$. Then $$\begin{split} \om(S(a)) &\le\om\left(S\left(r_k\frac{a}{|a|}\right)\right) \le C\om\left(S\left(r_{k+1}\frac{a}{|a|}\right)\right) \le\cdots\le C^{m-k+1}\om\left(S\left(|a'|\frac{a}{|a|}\right)\right)\\ &=C^{2}2^{(m-k-1)\log_2 C}\om\left(S\left(a'\right)\right) \le C^2\left(\frac{1-|a|}{1-|a'|}\right)^{\log_2C}\om\left(S\left(a'\right)\right), \end{split}$$ and thus (ii) is proved. Assume (ii) and let $K>0$. For each $a\in{\mathbb{D}}\setminus\{0\}$ take $a'=\frac{K+|a|}{K+1}e^{i\arg a}$. Then (ii) implies $$\begin{split} \frac{\om(S(a))}{(1-|a|)^\b}\le C(K+1)^\beta\frac{\om(S(a'))}{(1-|a|)^\b},\quad a\in{\mathbb{D}}\setminus\{0\}, \end{split}$$ and thus it follows that (iii) is satisfied for each $K>0$. Assume now (iii) for some $K>0$, and let $a_n=(1-(K+1)^n(1-|a|))e^{i\arg a}$ for $n=0,\ldots,N=\max\{k\in{\mathbb{N}}\cup\{0\}:(K+1)^k(1-|a|)<1\}$. Let $\eta>0$ to be fixed later. If $z\in S(a_n)\setminus S(a_{n-1})$, $n \geq 2$, then $$|1-\overline{a}z|\ge|a-z|\gtrsim|a-a_{n-1}|=(1-|a|)((K+1)^{n-1}-1)\asymp(1-|a|)(K+1)^n.$$ Moreover, since $\frac{K+|a_n|}{K+1}=|a_{n-1}|$, the hypothesis (iii) yields $$\begin{split} \int_{{\mathbb{D}}\setminus S(a_1)}\frac{\om(z)}{|1-\overline{a}z|^\eta}\,dA(z) &=\sum_{n=2}^N\int_{S(a_n)\setminus S(a_{n-1})}\frac{\om(z)}{|1-\overline{a}z|^\eta}\,dA(z) +\int_{{\mathbb{D}}\setminus S(a_N)}\frac{\om(z)}{|1-\overline{a}z|^\eta}\,dA(z)\\ &\lesssim\frac1{(1-|a|)^\eta}\sum_{n=2}^N\frac{\om(S(a_n))}{(K+1)^{n\eta}}+1 \le\frac{\om(S(a_0))}{(1-|a|)^\eta}\sum_{n=2}^\infty\left(\frac{C}{(K+1)^{\eta}}\right)^{n}+1. \end{split}$$ In addition, clearly $$\int_{S(a_1)}\frac{\om(z)}{|1-\overline{a}z|^\eta}\,dA(z) \lesssim \frac{\om(S(a_1))}{(1-|a_1|)^\eta}.$$ By repeating the method used in the first part of the proof, it is easy to see that (iii) in fact implies (ii). Therefore (iv) follows by choosing $\eta>\log_{K+1}C$ sufficiently large so that $\om(S(a))(1-|a|)^{-\eta}$ is essentially increasing. Finally, assume (iv) and denote $a^\star=\frac{1+|a|}{2}e^{i\arg a}$ for $a\in{\mathbb{D}}\setminus\{0\}$. Then $$\frac{\om(S(a))}{(1-|a^\star|)^\eta} \asymp\int_{S(a)}\frac{\om(z)}{|1-\overline{a^\star}z|^\eta}\,dA(z) \leq\int_{{\mathbb{D}}}\frac{\om(z)}{|1-\overline{a^\star}z|^\eta}\,dA(z) \lesssim\frac{\om(S(a^\star))}{(1-|a^\star|)^\eta},\quad a\in{\mathbb{D}}\setminus\{0\},$$ and it follows that $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$. [Theorem \[theo:dominating-sets-non-radial\]]{} Let $G$ be a dominating set for $A^p_\om$. It suffices to show  for points $a \in {\mathbb{D}}$ close to the boundary. Let $\beta = \beta(\om) > 0$ and $\eta = \eta(\om) > 0$ be those in Lemma \[Lemma:test-functions-non-radial\]. For each $a\in{\mathbb{D}}\setminus\{0\}$, define $$\widetilde{f}_a(z) = \left(\frac{1-|a|}{1-\overline{a}z}\right)^{m/p} \frac1{\om(S(a))^{1/p}}, \quad z\in{\mathbb{D}},$$ where $m = m(\om)> \max\{\beta,\eta\}$. Then $\big\|\widetilde{f}_a\big\|_{A^p_\om}\asymp1$ for all $a\in{\mathbb{D}}\setminus\{0\}$ by Lemma \[Lemma:test-functions-non-radial\]. Set $f_a=\widetilde{f}_a/\big\|\widetilde{f}_a\big\|_{A^p_\om}$ for all $a\in{\mathbb{D}}\setminus\{0\}$. For given $a$ close to the boundary, consider the points $a_n$ defined in the proof of Lemma \[Lemma:test-functions-non-radial\]. Then the proof of (iii) implies (iv) with $K=1$ in the said lemma gives $$\begin{split} \int_{{\mathbb{D}}\setminus S(a_n)}|\widetilde{f}_a(z)|^p\om(z)\,dA(z) &=\frac{(1-|a|)^m}{\om(S(a))}\int_{{\mathbb{D}}\setminus S(a_n)}\frac{\om(z)}{|1-\overline{a}z|^m}\,dA(z)\\ &\lesssim\sum_{j=n+1}^\infty\left(\frac{C}{2^m}\right)^j\asymp\frac{1}{2^{mn}},\quad n\in{\mathbb{N}}, \end{split}$$ for all $|a|$ sufficiently close to the boundary and $m$ sufficiently large. It follows that $$\int_{{\mathbb{D}}\setminus S(a_n)}|f_a(z)|^p\om(z)\,dA(z)\lesssim\frac1{2^{mn}},\quad n\in{\mathbb{N}}.$$ Since $G$ is a dominating set for $A^p_\om$ by the hypothesis, there exists $\d=\d(G)>0$ such that $$\label{5} \d\|f\|^p_{A^p_\om}\le\int_G |f(z)|^p\om(z)\,dA(z), \quad f \in A^p_\om.$$ Fix $n\in{\mathbb{N}}$ sufficiently large such that $$\int_{{\mathbb{D}}\setminus S(a_n)} |f_a(z)|^p\om(z)\,dA(z) < \frac\delta{2}.$$ Since $f_a$ has norm one in $A^p_\om$, this together with yields $$\begin{split} \int_{G\cap S(a_n)} |f_a(z)|^p\om(z)\,dA(z) &= \int_{G} |f_a(z)|^p\om(z)\,dA(z) - \int_{G\setminus S(a_n)} |f_a(z)|^p\om(z)\,dA(z) \\ &\geq \delta - \int_{{\mathbb{D}}\setminus S(a_n)} |f_a(z)|^p\om(z)\,dA(z) > \d - \frac\d{2} = \frac\d{2}, \end{split}$$ and $$\int_{G\cap S(a_n)} |f_a(z)|^p\om(z)\,dA(z) \asymp\int_{G\cap S(a_n)} |\widetilde{f}_a(z)|^p\om(z)\,dA(z) \le\frac{\om(G\cap S(a_n))}{\om(S(a))}.$$ Since $\om\in{\widehat{\mathcal{D}}}({\mathbb{D}})$, we have $\om(S(a))=\om(S(a_0))\ge C^{-n}\om(S(a_n))$, and now that $n$ is fixed, we deduce $\om(S(a_n))\lesssim\om(G\cap S(a_n))$ for all $a\in{\mathbb{D}}$ sufficiently close to the boundary. The claim follows from this estimate. Sampling measures ================= A positive Borel measure $\mu$ on ${\mathbb{D}}$ is a $q$-Carleson measure for $A^p_\om$ if $A^p_\om$ is continuously embedded into $L^q_\mu$. In order to prove Theorem \[theo:sampling-suff\] we need the following lemma which is a generalization of [@L1985 Theorem 2.3]. It readily follows from the proof that if the hypothesis on $\nu$ is replaced by $\nu(\Delta(a,r))\lesssim\om(\Delta(a,r))$, then on the left in both occasions $S(\z)$ must be replaced by $\Delta(\z,r)$. Luecking [@L1985/2 Lemma 3.10] showed this for $\nu=\om\in C_\infty$ under the hypothesis $\mu(\Delta(a,r))\lesssim\om(\Delta(a,r))$. Recall that, as discussed after Theorem \[Theorem:D-hat\], this last requirement characterizes $p$-Carleson measures for $A^p_\om$ if $\om$ satisfies the the Bekollé-Bonami condition by [@Constantin2 Theorem 3.1]. \[lem:difference-estimate\] Let $0 < p < \infty$, $0<r<\frac{R}2\le\frac14$ and $\om$ a weight. Let $\mu$ and $\nu$ be positive Borel measures on ${\mathbb{D}}$ such that $d\widetilde{\mu}(z) = \frac{\mu(\Delta(z,R))}{(1-|z|)^2}\,dA(z)$ is a $p$-Carleson measure for $A^p_\om$ and $\nu(S(a))\lesssim\om(S(a))$ for all $a\in{\mathbb{D}}\setminus\{0\}$. Then there exists a constant $C = C(p,R,\om,\mu,\nu) > 0$ such that $$\int_{\mathbb{D}}\left(\int_{S(\z)}|f(z)-f(\zeta)|^p\,d\nu(z)\right)\frac{d\mu(\zeta)}{\om(S(\zeta))} \le r^pC \|f\|^p_{A^p_\om}, \quad f \in A^p_\om.$$ In particular, if $\om\in{\widehat{\mathcal{D}}}$ and $\mu$ is a $p$-Carleson measure for $A^p_\om$, then the statement is valid. First, it is easy to show that there exists a constant $C = C(p,R) > 0$ such that $$\left|\frac{f(z) - f(0)}{z}\right|^p \le C\int_{D(0,R)} |f(w)|^p\,dA(w)$$ for all $|z| < R/2$ and $f \in \mathcal{H}({\mathbb{D}})$. Thus, if $|z| < r < R/2$, then $$|f(z) - f(0)|^p \le r^p C \int_{D(0,R)} |f(w)|^p\,dA(w).$$ An application of this to $f\circ\vp_\zeta$ and a change of variable on the right yield $$|f(z) - f(\zeta)|^p \le r^p C \int_{\Delta(\zeta,R)} |f(w)|^p\frac{(1-|\zeta|^2)^2}{|1-\overline{\zeta}w|^4}\,dA(w),\quad z\in\Delta(\z,r).$$ By multiplying this by $\chi_{S(\z)}(z)/\om(S(\zeta))$, integrating with respect to $\nu$ in the variable $z$ and using the hypothesis on $\nu$ now give $$\begin{split} \int_{S(\z)}|f(z)-f(\zeta)|^p\frac{d\nu(z)}{\om(S(\zeta))} &\le\frac{4Cr^p}{\om(S(\zeta))} \int_{S(\z)}\left(\int_{\Delta(\zeta,R)}|f(w)|^p\frac{dA(w)}{(1-|w|)^{2}}\right)d\nu(z)\\ &\lesssim r^p \int_{\Delta(\zeta,R)} |f(w)|^p\frac{dA(w)}{(1-|w|)^{2}}. \end{split}$$ Now an integration with respect to $\mu$ in the variable $\zeta$ and Fubini’s theorem yield $$\begin{split} \int_{\mathbb{D}}\left(\int_{S(\z)}|f(z)-f(\zeta)|^p\,d\nu(z)\right)\frac{d\mu(\zeta)}{\om(S(\zeta))} &\lesssim r^p\int_{\mathbb{D}}|f(w)|^p\frac{\mu(\Delta(w,R))}{(1-|w|)^2}\,dA(w)=r^p\|f\|_{L^p_{\widetilde{\mu}}}^p. \end{split}$$ Therefore the first statement in the lemma follows by the hypothesis on $\widetilde\mu$. It remains to show that if $\om\in{\widehat{\mathcal{D}}}$, then $\widetilde{\mu}$ is a $p$-Carleson measure for $A^p_\om$ whenever $\mu$ is. To see this, use first Fubini’s theorem to deduce $$\begin{split} \widetilde{\mu}(S(a)) &= \int_{S(a)} \frac{\mu(\Delta(z,R))}{(1-|z|)^2}\,dA(z) =\int_{\{\z:S(a)\cap\Delta(\z,R)\ne\emptyset\}}\left(\int_{S(a)\cap\Delta(\z,R)}\frac{dA(z)}{(1-|z|)^2}\right)d\mu(\z)\\ &\le\int_{S(b)}\left(\int_{\Delta(\z,R)}\frac{dA(z)}{(1-|z|)^2}\right)d\mu(\z)\asymp\mu(S(b)), \end{split}$$ where $b=b(a,R)\in {\mathbb{D}}$ is such that $\arg b=\arg a$ and $1-|b|\asymp1-|a|$ for all $a\in{\mathbb{D}}$. Now that $\mu$ is a $p$-Carleson measure for $A^p_\om$ by the hypothesis, [@PR2015Embedding Theorem 1] together with Lemma \[Lemma:weights-in-D-hat\] yields $\widetilde{\mu}(S(a))\lesssim\om(S(b))\asymp\om(S(a))$ for all $a\in{\mathbb{D}}$. Consequently, $\widetilde{\mu}$ is indeed a $p$-Carleson measure for $A^p_\om$, and the proof is complete. [Theorem \[theo:sampling-suff\]]{} By Lemma \[lem:difference-estimate\], with $d\nu = \om\,dA$ and $\Delta(\z,r)$ in place of $S(\z)$, and [@L1985/2 Lemma 3.10] there exists a constant $C=C(p,R,\om,\mu)>0$ such that $$\label{eq:proof-theo-sampling-suff} \int_{\mathbb{D}}\left(\int_{\Delta(z,r)}|f(z)-f(\zeta)|^p\frac{d\mu(\zeta)}{\om(\Delta(\zeta,r))}\right)\om(z)\,dA(z) \leq r^pC \|f\|^p_{A^p_\om}, \quad f \in A^p_\om,$$ where $0<r<R/2\le1/4$. Let first $p \geq 1$. Then raising this to power $1/p$ and using Minkowski’s inequality on the left yields $$\begin{split} &\Bigg(\int_{\mathbb{D}}|f(z)|^p\left(\int_{\Delta(z,r)}\frac{d\mu(\z)}{\om(\Delta(\zeta,r))}\right)\om(z)\,dA(z)\Bigg)^{1/p}\\ &- \left(\int_{\mathbb{D}}\left(\int_{\Delta(z,r)}|f(\zeta)|^p\frac{d\mu(\zeta)}{\om(\Delta(\zeta,r))}\right)\om(z)\,dA(z)\right)^{1/p} \le rC^{1/p} \|f\|_{A^p_\om}, \end{split}$$ where, by Lemma \[Lemma:weights-in-D-hat\] and [@L1985/2 Lemma 3.4], the fact that $r$ is bounded away from 1 and the definition of $G$, $$\int_{\Delta(z,r)} \frac{d\mu(\zeta)}{\om(\Delta(\zeta,r))} \ge C_1 k_r(z) \ge C_1\e \|M_\om(\mu)\|_{L^\infty}\chi_G(z),\quad z\in{\mathbb{D}},$$ for some constant $C_1=C_1(\om)>0$, and thus $$\int_{\mathbb{D}}|f(z)|^p\left(\int_{\Delta(z,r)}\frac{d\mu(\z)}{\om(\Delta(\z,r))}\right)\om(z)\,dA(z) \ge C_1\e\|M_\om(\mu)\|_{L^\infty} \int_G |f(z)|^p\om(z)\,dA(z).$$ Further, by Fubini’s theorem, $$\begin{split} \int_{\mathbb{D}}\left(\int_{\Delta(z,r)}|f(\zeta)|^p\frac{d\mu(\zeta)}{\om(\Delta(\z,r))}\right)\om(z)\,dA(z) &=\int_{\mathbb{D}}|f(\zeta)|^p\,d\mu(\zeta), \end{split}$$ and hence $$\left(C_1\e\|M_\om(\mu)\|_{L^\infty} \int_G |f(z)|^p\om(z)\,dA(z)\right)^{1/p} - \left(\int_{\mathbb{D}}|f(\zeta)|^p\,d\mu(\zeta)\right)^{1/p} \le rC^{1/p} \|f\|_{A^p_\om}.$$ Since by the hypothesis $G$ is dominating set, there exists a constant $\alpha>0$ such that $\int_G |f(z)|^p\om(z)\,dA(z)\ge\alpha\|f\|^p_{A^p_\om}$ for all $f \in A^p_\om$. Consequently, choosing $r$ such that $r^pC < C_1\e\alpha\|M_\om(\mu)\|_{L^\infty}$ yields $$\label{eq:proof-theo-sampling-suff-2} \|f\|_{A^p_\om} \le\frac{1}{(C_1\e\alpha\|M_\om(\mu)\|_{L^\infty})^{1/p}-rC^{1/p}} \left(\int_{\mathbb{D}}|f(\zeta)|^p\,d\mu(\zeta)\right)^{1/p},$$ and thus the proof is complete when $1 \leq p < \infty$. If $p < 1$, then one can simply apply the inequality $|x - y|^p \geq |x|^p - |y|^p$ to the left hand side of  to obtain $$\begin{split} &\int_{\mathbb{D}}|f(z)|^p\left(\int_{\Delta(z,r)}\frac{d\mu(\z)}{\om(\Delta(\z,r))}\right)\om(z)\,dA(z)\\ &- \int_{\mathbb{D}}\left(\int_{\Delta(z,r)}|f(\zeta)|^p\frac{d\mu(\zeta)}{\om(\Delta(\z,r))}\right)\om(z)\,dA(z) \le r^p C \|f\|^p_{A^p_\om}. \end{split}$$ Then the estimates used in the case $p \geq 1$ yield $$C_1\e\|M_\om(\mu)\|_{L^\infty} \int_G |f(z)|^p\om(z)\,dA(z) -\int_{\mathbb{D}}|f(\zeta)|^p\,d\mu(\zeta) \le r^pC \|f\|^p_{A^p_\om},$$ from which and estimate similar to  follows as above. [Theorem \[theo:w-conv-meas\]]{} The first statement can be established by following the proof of [@L2000 Theorem 1]. Namely, since $\mu_n(S(a)\cap D(0,r))\le\|M_\om(\mu_n)\|_{L^\infty}\om({\mathbb{D}})$ for each $a\in{\mathbb{D}}$, the hypothesis $\sup_n\|M_\om(\mu_n)\|_{L^\infty}<\infty$ implies that for each $r\in(0,1)$, the sequence $\left(\mu_n|_{D(0,r)}\right)_{n\in{\mathbb{N}}}$ is bounded in $C_0(D(0,r))^*$. Hence there is a subsequence that converges in the weak$^*$-topology by the Banach-Alaoglu theorem. By diagonalization we may extract a subsequence $(\mu_{n_j})$ such that, for each $r \in (0,1)$, $(\mu_{n_j}|_{D(0,r)})$ converges to a measure $\mu_r$ supported in $\overline{D(0,r)}$. Since clearly $\mu_s = \mu_r$ on $D(0,r)$ for all $r<s$, we may define a measure $\mu$ as the limit $\lim_{r\to1^-} \mu_r$. If now $h\in C_c({\mathbb{D}})$, then it’s support is in some $D(0,r)$, and thus $$\int_{\mathbb{D}}h(z)\,d\mu_{n_j}(z) \to \int_{\mathbb{D}}h(z)\,d\mu(z),\quad j \to \infty.$$ To prove , let $f \in A_\om^p$. For $h\in C_c({\mathbb{D}})$ satisfying $h(z) \leq 1$ for all $z \in {\mathbb{D}}$ we have $$\liminf_{n\to\infty} \int_{\mathbb{D}}|f(z)|^p\,d\mu_n(z) \ge\lim_{n\to\infty} \int_{\mathbb{D}}h(z)|f(z)|^p\,d\mu_n(z) =\int_{\mathbb{D}}h(z)|f(z)|^p\,d\mu(z)$$ by Fatou’s lemma, and hence $$\label{1} \liminf_{n\to\infty} \int_{\mathbb{D}}|f(z)|^p\,d\mu_n(z) \geq \int_{\mathbb{D}}|f(z)|^p\,d\mu(z).$$ For the converse inequality, let $\e > 0$ and take $r=r(f)\in (0,1)$ such that $$\int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,\om(z)\,dA(z) < \e.$$ Let $h\in C_c({\mathbb{D}})$ such that $h(z) \leq 1$ for all $z \in {\mathbb{D}}$ and $h\equiv1$ on $\overline{D(0,r)}$. Then $$\label{eq:weak-conv-pro} \int_{\mathbb{D}}|f(z)|^p\,d\mu_n(z) \le\int_{\mathbb{D}}h(z)|f(z)|^p\,d\mu_n(z) +\int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,d\mu_n(z).$$ Let us first handle the right most integral. For a moment, let $\nu$ be a $p$-Carleson measure for $A_\om^p$ and denote $d\nu_r = \chi_{{\mathbb{D}}\setminus\overline{D(0,r)}}\,d\nu$. Then, by [@PR2015Embedding Theorem 9], $$\begin{split} \int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,d\nu(z) &= \int_{\mathbb{D}}\left|f(z)\chi_{{\mathbb{D}}\setminus\overline{D(0,r)}}(z)\right|^p\,d\nu_r(z) \\ &\leq \int_{\mathbb{D}}N\left(f\chi_{{\mathbb{D}}\setminus\overline{D(0,r)}}\right)^p(z)\,d\nu_r(z) \\ &\lesssim \int_{\mathbb{D}}N\left(f\chi_{{\mathbb{D}}\setminus\overline{D(0,r)}}\right)^p(z) M_\om(\nu_r)(z)\,\om(z)\,dA(z) \\ &\leq \|M_\om(\nu_r)\|_{L^\infty} \int_{{\mathbb{D}}\setminus\overline{D(0,r)}} N(f)^p(z)\,\om(z)\,dA(z) \\ &\asymp \|M_\om(\nu_r)\|_{L^\infty} \int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,\om(z)\,dA(z),\quad f \in A_\om^p, \end{split}$$ where the last inequality follows from the fact that the classical non-tangential maximal function is a bounded operator from $H^p$ to $L^p$ of the boundary [@Garnett1981 Theorem 3.1 on p. 57]. Therefore there exists a constant $C=C(p)>0$ such that $$\int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,d\mu_n(z) \leq C \sup_{n} \|M_\om(\mu_n)\|_{L^\infty} \int_{{\mathbb{D}}\setminus\overline{D(0,r)}} |f(z)|^p\,\om(z)\,dA(z) \leq C\Lambda\e.$$ Thus, by taking the limit superior of  we obtain $$\begin{split} \limsup_{n\to\infty} \int_{\mathbb{D}}|f(z)|^p\,d\mu_n(z) &\le\int_{\mathbb{D}}h(z)|f(z)|^p\,d\mu(z) + C\Lambda\e \le\int_{\mathbb{D}}|f(z)|^p\,d\mu(z) + C\Lambda\e, \end{split}$$ and since $\e > 0$ was arbitrary, by combining this with we deduce . Recall that if $\nu$ is a $p$-Carleson measure for $A^p_\om$, then $\|Id\|^p_{A^p_\om \to L^p_\nu}\asymp \|M_\om(\nu)\|_{L^\infty}$ by [@PelRatSie2015 Theorem 3], and therefore $\sup_n\|Id\|^p_{A^p_\om \to L^p_{\mu_n}}=\Lambda_1<\infty$. It follows that $$\int_{\mathbb{D}}|f(z)|^p\,d\mu_n(z) \le\|Id\|_{A^p_\om \to L^p_{\mu_n}}^p\|f\|^p_{A^p_\om} \le\Lambda_1\|f\|^p_{A^p_\om}, \quad f \in A^p_\om,$$ for all $n \in {\mathbb{N}}$. By the identity just proved, we may pass to the limit to obtain $$\int_{\mathbb{D}}|f(z)|^p\,d\mu(z)\le\Lambda_1\|f\|^p_{A^p_\om}, \quad f \in A^p_\om.$$ Since this is also true for any subsequence of $(\mu_n)$, we may replace $\Lambda_1$ with $\liminf_{n \to \infty} \|Id\|^p_{A^p_\om \to L^p_{\mu_n}}$. Thus $\mu$ is a $p$-Carleson measure for $A^p_\om$ with $\|Id\|_{A^p_\om \to L^p_\mu}\le\liminf_{n\to\infty}\|Id\|_{A^p_\om\to L^p_{\mu_n}}$ as claimed. In the case of sampling measures, the lower inequality follows in a manner similar to above. Details of this step are omitted. [99]{} A. Aleman, A class of integral operators on spaces of analytic functions, Topics in complex analysis and operator theory, 3–30, Univ. Málaga, Málaga, 2007. A. Aleman and J. A. Cima, An integral operator on $H^p$ and Hardy’s inequality, J. Anal. Math. 85 (2001), 157–176. D. Bekollé, Inégalités á poids pour le projecteur de Bergman dans la boule unité de ${\mathbb{C}}^n$ \[Weighted inequalities for the Bergman projection in the unit ball of ${\mathbb{C}}^n$\], Studia Math. 71 (1981/82), no. 3, 305–323. D. Bekollé and A. Bonami, Inégalités á poids pour le noyau de Bergman, (French) C. R. Acad. Sci. Paris Sr. A–B 286 (1978), no. 18, 775–778. O. Constantin, Discretizations of integral operators and atomic decompositions in vector-valued weighted Bergman spaces, Integral Equations Operator Theory 59 (2007), 523–554. O. Constantin, Carleson embeddings and some classes of operators on weighted Bergman spaces, J. Math. Anal. Appl. 365 (2010), 668–682. C. C. Cowen, and B. D. MacCluer, Composition Operators on Spaces of Analytic Functions, Studies in Advanced Mathematics, CRC Press, Boca Raton, FL 1995. J. Duoandikoetxea, F. J. Martin-Reyes and S. Ombrosi, On the $A_\infty$ conditions for general bases, Math. Z. 282 (2016), no. 3, 955–972. P. Duren, Theory of $H^p$ Spaces, Academic Press, New York-London 1970. J. Garnett, Bounded analytic functions, Academic Press, New York, 1981. H. Hedenmalm, A factoring theorem for a weighted Bergman space, Algebra i Analiz 4 (1992), no. 1, 167–176; translation in St. Petersburg Math. J. 4 (1993), no. 1, 163–174. C. Horowitz, Zeros of functions in the Bergman spaces, Duke Math. J. 41 (1974), 693–710. C. Horowitz, Factorization theorems for functions in the Bergman spaces, Duke Math. J. 44 (1977), no. 1, 201–213. C. Horowitz, Some conditions on Bergman space zero sets, J. Anal. Math. 62 (1994), 323–348. C. Horowitz, Zero sets and radial zero sets in function spaces, J. Anal. Math. 65 (1995), 145–159. C. Horowitz and Y. Schnaps, Factorization of functions in weighted Bergman spaces, Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4) 29 (2000), no. 4, 891–903. R. A. Kerman and A. Torchinsky, Integral inequalities with weights for the Hardy maximal function, Studia Math. 71(3) (1981/82), 277–284. B. Korenblum, An extension of Nevanlinna theory, Acta Math. 135 (1975), 265–283. D. Luecking, Inequalities in Bergman spaces, III. J. Math. 25 (1981), 1–11. D. Luecking, Closed ranged restriction operators on weighted Bergman spaces, Pacific J. Math. 110 (1984), no. 1, 145–160. Pacific. D. Luecking, Forward and Reverse Carleson Inequalities for Functions in Bergman Spaces and Their Derivatives, Amer. J. Math. 107 (1985), 85–111. D. Luecking, Representation and Duality in Weighted Spaces of Analytic Functions, Indiana Univ. Math. J. 34 (1985), no. 2, 319–336. D. Luecking, Zero Sequences for Bergman Spaces, Complex Variables 30 (1996), 345–362. D. Luecking, Sampling measures for Bergman spaces on the unit disk, Math. Ann. 316 (2000), 659–679. J. Pau, and R. Zhao, Weak factorization and Hankel forms for weighted Bergman spaces on the unit ball, Math. Ann. 363 (2015), no. 1–2, 363–383. J. Pau, R. Zhao and K. Zhu, Weighted BMO and Hankel operators between Bergman spaces, Indiana Univ. Math. J. 65 (2016), no. 5, 1639–1673. J. A.  Peláez, Small weighted Bergman spaces, Proceedings of the summer school in complex and harmonic analysis, and related topics, (2016). J. Peláez and J. Rättyä, Weighted Bergman spaces induced by rapidly increasing weights, Mem. Amer. Math. Soc. 227 (2014). J. Peláez and J. Rättyä, Embedding Theorems for Bergman Spaces via Harmonic Analysis, Math. Ann. 362 (2015), 205–239. J. A. Peláez and J. Rättyä, Two weight inequality for Bergman projection, J. Math. Pures Appl. (9) 105 (2016), no 1, 102–130. J. A. Peláez and J. Rättyä, Trace class criteria for Toeplitz and composition operators on small Bergman spaces, Adv. Math. 293 (2016), 606–643. J. A. Peláez and J. Rättyä, Weighted Bergman projections on $L^\infty$, preprint. J. Peláez, J. Rättyä and K. Sierra, Berezin transform and Toeplitz operators on weighted Bergman spaces induced by regular weights, J. Geom. Anal. in press. J. A.  Peláez, J. Rättyä and K. Sierra, Embedding Bergman spaces into tent spaces, Math. Z. 281 (2015), no. 3-4, 1215–1237. V. V. Peller, Hankel operators and their applications, Springer Monographs in Mathematics, Springer-Verlag, New York, 2003. S. Pott and M. C. Reguera, Sharp Bekollé estimate for the Bergman projection, J. Funct. Anal. 265 (2013), 3233–3244. K. Seip, Interpolation and sampling in small Bergman spaces, Collect. Math. 64 (2013), 61–72. K. Seip, On a theorem of Korenblum, Ark. Mat. 32 (1994), no. 1, 237–243. K. Seip, On Korenblum’s density condition for the zero sequences of $A^{-\a}$, J. Anal. Math. 67 (1995), 307–322. J. H. Shapiro, Composition Operators and Classical Function Theory, Universitext: Tracts in Mathematics. Springer-Verlag, New York, 1993. K. Zhu, Operator theory in function spaces. Monographs and Textbooks in Pure and Applied Mathematics, 139. Marcel Dekker, Inc., New York, 1990. [^1]: This research was supported in part by Ministerio de Economía y Competitivivad, Spain, project MTM2014-52865-P; by Academy of Finland project no. 268009, and by Faculty of Science and Forestry of University of Eastern Finland.

ArXiv

--- abstract: 'We will discuss the spectrum of the eta mesons making use of the Nambu-Jona-Lasinio (NJL) model supplemented with a model of confinement. We will go on to discuss the properties of mesons at finite temperature and the phenomenon of deconfinement. We will then discuss some excited states of the quark-gluon plasma calculated in lattice QCD models.These resonances are thought to be created in heavy-ion collisions.We consider the role these states play in leading to a hydrodynamic description of the plasma at early stages of its formation.' author: - 'C. M. Shakin' - Huangsheng Wang - Qing Sun - Hu Li - Xiangdong Li title: 'EXCITATIONS OF THE QUARK- GLUON PLASMA' --- Introduction ============ In this presentation we will discuss the properties of QCD as one moves upward along the temperature axis from the point at $T=0$, where the matter density is zero. Note that for experiments at RHIC the associated chemical potential is small. (430,360){width=".63\textwidth"} **Fig. 1** Schematic phase diagram of nuclear matter (Taken from F. Karsch, Lect. Notes Phys. **583**, 209 (2002) \[hep-lat/0106019\].) Properties of the $\eta$ mesons calculated with the NJL interaction and a model of confinement ============================================================================================== *C. M. Shakin and Huangsheng Wang, Physical Review D, **65**, 094003 (2002)*\ \ Recent work has shown that the singlet-octet mixing angles of the $\eta(547)$ and $\eta^\prime(958)$ are different. That may be demonstrated either in extended chiral perturbation theory or by analysis of a large body of experimental data. The conclusion is that the $\eta(547)$ is almost entirely of octet character, while the $\eta^\prime(958)$ is mainly of singlet character with about 10% octet component. It is possible to calculate the mixing angles and decay constants in our generalized Nambu–-Jona-Lasinio (NJL) model, which includes a covariant model of confinement. Our model is able to give a good account of the mass values of the $\eta(547)$, $\eta^\prime(958)$, $\eta(1295)$, and $\eta(1440)$ mesons. (We also provide predictions for the mass values of a large number of radially excited states.) It is well known that the $U_A(1)$ symmetry is broken, so that we only have eight pseudo Goldstone bosons, rather than the nine we would have otherwise. In the NJL model that feature may be introduced by including the ’t Hooft interaction in the Lagrangian. That interaction reduces the energy of the octet state somewhat and significantly increases the energy of the singlet state, making it possible to fit the mass values of the $\eta(547)$ and $\eta^\prime(958)$ in the NJL model when the ’t Hooft interaction is included. In this work, we derive the equations of a covariant random phase approximation that may be used to study the nonet of pseudoscalar mesons. We demonstrate that a consistent treatment of the ’t Hooft interaction leads to excellent results for the singlet-octet mixing angles. (The values obtained for the singlet and octet decay constants are also quite satisfactory.) It may be seen that the difference between the up (or down) constituent quark mass and the strange quark mass induces singlet-octet mixing that is too large. However, the ’t Hooft interaction contains singlet-octet coupling that enters into the theory with a sign opposite to that of the term arising from the difference of the quark mass values, leading to quite satisfactory results. In this work we present the wave function amplitudes for a number of states of the eta mesons. (The inclusion of pseudoscalar axial-vector coupling is important for our analysis and results in the need to specify eight wave function amplitudes for each state of the eta mesons.) We present the values of the various constants that parameterize our generalized NJL model and which give satisfactory values of the eta meson masses, decay constants, and mixing angles. It is found that the calculated mass values for the $\eta(1295)$ and $\eta(1440)$ are quite insensitive to variation of the parameters of the model whose values have largely been fixed in our earlier studies of other light mesons. (570,600)  **Fig. 2** (600,500)  **Fig. 3** (600,500)  **Fig. 4** Experimental and calculated spectra of the $\eta$ mesons. Columns 1 and 3 show the results for the confinement potential only. Data Set I Set II Set III Set IV ---------------------------------------- ----------------------- --------------- --------------- --------------- -------------------- $m_\eta(547)$ \[MeV\] $\cdots$ 538 536 527 555 $m_{\eta^\prime}(958)$ \[MeV\] $\cdots$ 911 942 963 949 $m_\eta(1295)$ \[MeV\] $\cdots$ 1319 1318 1317 1319 $m_\eta(1440)$ \[MeV\] $\cdots$ 1414 1416 1419 1411 $\tilde f_\eta^{(8)}$ \[MeV\] $\cdots$ 177.2 178.6 180.9 163 $\tilde f_\eta^{(0)}$ \[MeV\] $\cdots$ 27.59 24.51 18.95 52.8 $\tilde f_{\eta^\prime}^{(8)}$ \[MeV\] $\cdots$ -84.26 -84.64 -80.97 -105 $\tilde f_{\eta^\prime}^{(0)}$ \[MeV\] $\cdots$ 159.2 157.3 156.0 150 $F_8$ \[MeV\] $(1.32\pm0.06)f_\pi)$ $=174\pm8$ MeV 179.3 180.3 181.9 170 $F_0$ \[MeV\] $(1.37\pm0.07)f_\pi)$ $=181\pm9$ MeV 180.3 178.2 174.2 190 $\theta_\eta$ $(-5.7\pm2.7)^\circ$ $-8.81^\circ$ $-7.82^\circ$ $-6.26^\circ$ $-16.1^\circ$ $\theta_{\eta^\prime}$ $(-24.6\pm2.3)^\circ$ $-28.0^\circ$ $-28.0^\circ$ $-26.4^\circ$ $-38.2^\circ$ $\theta_0$ $(-7.0\pm2.7)^\circ$ $-9.83^\circ$ $-8.76^\circ$ $-6.94^\circ$ $-19.4^\circ$ $\theta_8$ $(-21.5\pm2.4)^\circ$ $-25.4^\circ$ $-25.4^\circ$ $-24.1^\circ$ $-32.8^\circ$ $\theta_0-\theta_8$ $16.4^\circ$ $15.6^\circ$ $16.6^\circ$ $17.2^\circ$ $13.4^\circ$ $\hat F_0$ \[MeV\] $(1.21\pm0.07)f_\pi)$ $=160\pm9$ MeV 161 159 157 158 $\hat F_8$ \[MeV\] $=188\pm11$ MeV 196 198 198 194 $G_D$ \[\] $\cdots$ -180 -200 -220 -161.6($G_{08}=0$) Chiral Symmetry Restoration and Deconfinement of Light Mesons at Finite Temperature =================================================================================== *Hu Li and C. M. Shakin , hep-ph/0209136*\ \ Confinement model: V\^C(r) = r (-\_0r), V\^C(k-k\^)=-8, V\^C(k-k\^)=-8, V\^C(r,T)=r,(T)=, V\_[max]{}(T)=. (200, 350)(0, -300) **Fig. 5** A comparison of quenched (open symbols) and unquenched (filled symbols) results for the interquark potential at finite temperature. The dotted line is the zero temperature quenched potential. Here, the symbols for $T=0.80T_c$ \[open triangle\], $T=0.88T_c$ \[open circle\], $T=0.94T_c$ \[open square\], represent the quenched results. The results with dynamical fermions are given at $T=0.68T_c$ \[solid downward-pointing triangle\], $T=0.80T_c$ \[solid upward-pointing triangle\], $T=0.88T_c$ \[solid circle\], and $T=0.94T_c$ \[solid square\]. (300,200)  **Fig. 6** The potential $V^C(r, T)$ is shown for $T/T_c=0$ \[solid line\], $T/T_c=0.4$ \[dotted line\], $T/T_c=0.6$ \[dashed line\], $T/T_c=0.8$ \[dash-dot line\], $T/T_c=0.9$ \[short dashes\], $T/T_c=1.0$ \[dash-dot-dot line\]. Here, $V^C(r,T)=\kappa r\exp[-\mu(T)r]$, with $\mu(T)=0.01\mbox{GeV}/[1-0.7(T/T_c)^2]$. (300, 270)  **Fig. 7** Temperature dependent constituent mass values, $m_u(T)$ and $m_s(T)$, calculated using the equation below. are shown. Here $m_u^0=0.0055$ GeV, $m_s^0=0.120$ GeV, and $G(T)=5.691[1-0.17(T/T_c)]$, if we use Klevansky’s notation.\ $m(T)=m^0+2G_S(T)N_C\frac{m(T)}{\pi^2}\int_0^\Lambda dp\frac{p^2}{E_p}\tanh(\frac 1 2\beta E_p)$. (280, 170)  **Fig. 8** Values of $m_u(T)$ are shown. The dashed curve is calculated with $m^0=5.50\,\mbox{MeV}$. Here, $G(T)=G\,[\,1-0.17\,(T/T_c)\,]$, with $G=5.691\,\mbox{GeV}^{-2}$ and $T_c=0.150\,\mbox{GeV}$. The solid curve is calculated with the same value of $G(T)$ and $T_c$, but with $m^0=0$. From the solid curve, we see that chiral symmetry is restored at $T=0.136\,\mbox{GeV}$ when $m^0=0$. (440, 300)  **Fig. 9** The mass values of the pionic states calculated in this work with $G_\pi(T)=13.49[1-0.17\,T/T_c]$ GeV, $G_V(T)=11.46[1-0.17\,T/T_c]$ GeV. The value of the pion mass is 0.223 GeV at $T/T_c=0.90$, where $m_u(T)=0.102$ GeV and $m_s(T)=0.449$ GeV. The pion is bound up to $T/T_c=0.94$, but is absent beyond that value. Calculation of Hadronic Current Correlation Functions at Finite Temperature =========================================================================== *Bing He, Hu Li, C. M. Shakin, and Qing Sun, Physical Review C **67**, 065203 (2003)*\ \ (600, 120)  **Fig. 10** The upper figure represents the basic polarization diagram of the NJL model in which the lines represent a constituent quark and a constituent antiquark. The lower figure shows a confinement vertex \[filled triangular region\] used in our earlier work. For the present work we neglect confinement for $T\geq1.2\,T_c$, with $T_c=150$ MeV.\ \ For ease of reference, we present a discussion of our calculation of hadronic current correlators. The procedure we adopt is based upon the real-time finite-temperature formalism, in which the imaginary part of the polarization function may be calculated. Then, the real part of the function is obtained using a dispersion relation. The result we need for this work has been already given in the work of Kobes and Semenoff. (The quark momentum is $k$ and the antiquark momentum is $k-P$. We will adopt that notation in this section for ease of reference to the results presented in the work of Kobes and Semenoff.) We write the imaginary part of the scalar polarization function as \[e319.1\]J\_S(*P*\^2, T)=12(2N\_c)\_S(*P*\^0)ke\^[-k\^2/\^2]{}()\ {(1-n\_1(k)-n\_2(k)) (*P*\^0-E\_1(k)-E\_2(k))\ -(n\_1(k)-n\_2(k)) (*P*\^0+E\_1(k)-E\_2(k))\ -(n\_2(k)-n\_1(k)) (*P*\^0-E\_1(k)+E\_2(k))\ -(1-n\_1(k)-n\_2(k)) (*P*\^0+E\_1(k)+E\_2(k))}.Here, $E_1(k)=[\,\vec k\,{}^2+m_1^2(T)\,]^{1/2}$. We have included a Gaussian regulator, $\exp[\,-\vec k\,{}^2/\alpha^2\,]$, with $\alpha=0.605$ GeV, which is the same as that used in most of our applications of the NJL model in the calculation of meson properties. We also note that n\_1(k)=1[e\^[E\_1(k)]{}+1]{},and n\_2(k)=1[e\^[E\_2(k)]{}+1]{}.For the calculation of the imaginary part of the polarization function, we may put $\ksq=m_1^2(T)$ and $(k-P)^2=m_2^2(T)$, since in that calculation the quark and antiquark are on-mass-shell. In Eq.(\[e319.1\]) the factor $\beta_S$ arises from a trace involving Dirac matrices, such that \_S&=&-\[(k+m\_1)(k-P+m\_2)\]\ &=&2P\^2-2(m\_1+m\_2)\^2,where $m_1$ and $m_2$ depend upon temperature. In the frame where $\vec P=0$, and in the case $m_1=m_2$, we have $\beta_S=2P_0^2(1-{4m^2}/{P_0^2})$. For the scalar case, with $m_1=m_2$, we find \[e319.6\] J\_S(P\^2, T)=(1-)\^[3/2]{} e\^[-k\^2/\^2]{}\[1-2n\_1(k)\],where \[e319.7\] k\^2=4-m\^2(T). For pseudoscalar mesons, we replace $\beta_S$ by \_P&=&-\[i\_5(k+m\_1)i\_5(k-P+m\_2)\]\ &=&2P\^2-2(m\_1-m\_2)\^2,which for $m_1=m_2$ is $\beta_P=2P_0^2$ in the frame where $\vec P=0$. We find, for the $\pi$ mesons, J\_P(P\^2,T)=(1-)\^[1/2]{} e\^[-k\^2/\^2]{}\[1-2n\_1(k)\],where $ \vec k\,{}^2={P_0^2}/4-m_u^2(T)$, as above. Thus, we see that, relative to the scalar case, the phase space factor has an exponent of 1/2 corresponding to a *s*-wave amplitude. For the scalars, the exponent of the phase-space factor is 3/2. For a study of vector mesons we consider \_\^V=\[\_(k+m\_1)\_(k-P+m\_2)\],and calculate g\^\_\^V=4\[P\^2-m\_1\^2-m\_2\^2+4m\_1m\_2\],which, in the equal-mass case, is equal to $4P_0^2+8m^2(T)$, when $m_1=m_2$ and $\vec P=0$. This result will be needed when we calculate the correlator of vector currents in the next section. Note that, for the elevated temperatures considered in this work, $m_u(T)=m_d(T)$ is quite small, so that $4P_0^2+8m_u^2(T)$ can be approximated by $4P_0^2$, when we consider the vector current correlation functions. In that case, we have J\_V(P\^2,T) J\_P(P\^2,T).At this point it is useful to define functions that do not contain that Gaussian regulator: \_P(P\^2,T)=(1-)\^[1/2]{}\[1-2n\_1(k)\],and \_V(P\^2,T)=(1-)\^[1/2]{}\[1-2n\_1(k)\],We need to use a twice-subtracted dispersion relation to obtain $\mbox{Re}\,\tilde{J}_P(P^2,T)$, or $\mbox{Re}\,\tilde{J}_V(P^2,T)$. For example, \[e319.16\]\_P(P\^2,T)=\_P(0,T)+ \[\_P(P\_0\^2,T)-\_P(0,T)\]+\ \_[4m\^2(T)]{}\^[\^[2]{}]{} ds,where $\tilde{\Lambda}^{2}$ can be quite large, since the integral over the imaginary part of the polarization function is now convergent. We may introduce $\tilde{J}_P(P^2,T)$ and $\tilde{J}_V(P^2,T)$ as complex functions, since we now have both the real and imaginary parts of these functions. We note that the construction of either $\mbox{Re}\,J_P(P^2,T)$, or $\mbox{Re}\,J_V(P^2,T)$, by means of a dispersion relation does not require a subtraction. We use these functions to define the complex functions $J_P(P^2,T)$ and $J_V(P^2,T)$. In order to make use of Eq.(\[e319.16\]), we need to specify $\tilde{J}_P(0)$ and $\tilde{J}_P(P_0^2)$. We found it useful to take $P_0^2=-1.0$ 2 and to put $\tilde{J}_P(0)=J_P(0)$ and $\tilde{J}_P(P_0^2)=J_P(P_0^2)$. The quantities $\tilde{J}_V(0)$ and $\tilde{J}_V(P_0^2)$ are determined in an analogous function. This procedure in which we fix the behavior of a function such as $\mbox{Re}\tilde{J}_V(P^2)$ or $\mbox{Re}\tilde{J}_V(P^2)$ is quite analogous to the procedure used in our earlier work. In that work we made use of dispersion relations to construct a continuous vector-isovector current correlation function which had the correct perturbative behavior for large $P^2\rightarrow-\infty$ and also described that low-energy resonance present in the correlator due to the excitation of the $\rho$ meson. In our earlier work the NJL model was shown to provide a quite satisfactory description of the low-energy resonant behavior of the vector-isovector correlation function. We now consider the calculation of temperature-dependent hadronic current correlation functions. The general form of the correlator is a transform of a time-ordered product of currents, iC(P\^2, T)=d\^4xe\^[iPx]{}&lt;&gt;,where the double bracket is a reminder that we are considering the finite temperature case. For the study of pseudoscalar states, we may consider currents of the form $j_{P,i}(x)=\tilde{q}(x)i\gamma_5\lambda^iq(x)$, where, in the case of the $\pi$ mesons, $i=1,2$ and $3$. For the study of scalar-isoscalar mesons, we introduce $j_{S,i}(x)=\tilde{q}(x)\lambda^i q(x)$, where $i=0$ for the flavor-singlet current and $i=8$ for the flavor-octet current. In the case of the pseudoscalar-isovector mesons, the correlator may be expressed in terms of the basic vacuum polarization function of the NJL model, $J_P(P^2, T)$. Thus, \[e319.18\] C\_P(P\^2, T)=\_P(P\^2, T),where $G_P(T)$ is the coupling constant appropriate for our study of $\pi$ mesons. We have found $G_P(T)=13.49$ by fitting the pion mass in a calculation made at $T=0$, with $m_u = m_d =0.364$ GeV. The result given in Eq.(\[e319.18\]) is only expected to be useful for small $P^2$, since the Gaussian regulator strongly modifies the large $P^2$ behavior. Therefore, we suggest that the following form is useful, if we are to consider the larger values of $P^2$. \[e319.19\] =.(As usual, we put $\vec{P}=0$.) This form has two important features. At large $P_0^2$, ${\mbox{Im}\,C_{P}(P_0, T)}/{P_0^2}$ is a constant, since ${\mbox{Im}\,\tilde{J}_{P}(P_0^2, T)}$ is proportional to $P_0^2$. Further, the denominator of Eq.(\[e319.19\]) goes to 1 for large $P_0^2$. On the other hand, at small $P_0^2$, the denominator is capable of describing resonant enhancement of the correlation function. As we will see, the results obtained when Eq.(\[e319.19\]) is used appear quite satisfactory. For a study of the vector-isovector correlators, we introduce conserved vector currents $j_{\mu, i}(x)=\tilde{q}(x)\gamma_{\mu}\lambda_i q(x)$ with i=1, 2 and 3. In this case we define \[e319.20\]J\_V\^(P\^2, T)=(g\^-)J\_V(P\^2, T)and \[e319.21\] C\_V\^(P\^2, T)=(g\^-)C\_V(P\^2, T),taking into account the fact that the current $j_{\mu,\,i}(x)$ is conserved. (Note that Eqs. (\[e319.20\]) and (\[e319.21\]) are valid for zero temperature. However, we still use that form at finite temperature for convenience.) We may then use the fact that J\_V(P\^2,T) = 13g\_J\_V\^(P\^2,T)and J\_V(P\^2,T)&=& 23(1-)\^[1/2]{}e\^[-k\^2/\^2]{}\[1-2n\_1(k)\]\ && J\_P(P\^2,T).(See Eq.(\[e319.7\]) for the specification of $k=|\vec k|$.) We then have C\_V(P\^2,T)=\_V(P\^2,T)1[1-G\_V(T)J\_V(P\^2,T)]{},where we have introduced \_V(P\^2,T)&=& 23(1-)\^[1/2]{}\[1-2n\_1(k)\]\ && \_P(P\^2,T). In the literature, $\omega$ is used instead of $P_0$. We may define the spectral functions \_V(, T)=C\_V(, T),and \_P(, T)=C\_P(, T), Since different conventions are used in the literature, we may use the notation $\overline{\sigma}_P(\omega, T)$ and $\overline{\sigma}_V(\omega, T)$ for the spectral functions given there. We have the following relations: \_P(, T)=\_P(, T),and =\_V(, T),where the factor 3/4 arises because there is a division by 4 in the literatures, while we have divided by 3. Calculation of the Momentum Dependence of Hadronic Current Correlation Functions at Finite Temperature ====================================================================================================== *Xiangdong Li, Hu Li, C. M. Shakin, Qing Sun and Huangsheng Wang, nucl-th/0405081*\ \ We have calculated spectral functions associated with hadronic current correlation functions for vector currents at finite temperature. We made use of a model with chiral symmetry, temperature-dependent coupling constants and temperature-dependent momentum cutoff parameters. Our model has two parameters which are used to fix the magnitude and position of the large peak seen in the spectral functions. In our earlier work, good fits were obtained for the spectral functions that were extracted from lattice data by means of the maximum entropy method (MEM). In the present work we extend our calculations and provide values for the three-momentum dependence of the vector correlation function at $T=1.5\,T_c$. These results are used to obtain the correlation function in coordinate space, which is usually parametrized in terms of a screening mass. Our results for the three-momentum dependence of the spectral functions are similar to those found in a recent lattice QCD calculation for charmonium \[S. Datta, F. Karsch, P. Petreczky and I. Wetzorke, hep-lat/0312037\]. For a limited range we find the exponential behavior in coordinate space that is usually obtained for the spectral function for $T>T_c$ and which allows for the definition of a screening mass. (280, 250)  **Fig. 11** The spectral functions $\sigma/\omega^2$ for pseudoscalar states obtained by MEM are shown. The solid line is for $T/T_c=1.5$ and the dashed line is for $T/T_c=3.0$. The second peak is a lattice artifact. (280, 250)  **Fig. 12** The spectral functions $\sigma/\omega^2$ for vector states obtained by MEM are shown. The second peak is a lattice artifact. (280, 200)  **Fig. 13** The imaginary part of the correlator $\sigma(\omega)/\omega^2$ is shown for various values of $|\vec P|$ as a function of $\omega^{2}$. Starting with the topmost curve the values of $|\vec P|$ in GeV units are 0.10, 0.30, 0.50, 0.70, 0.90, 1.10, 1.30, 1.50, 1.70, 1.90 and 2.10. Here we have used $G_S=1.2$ GeV$^{-2}$ and $k_{max}=1.22$ GeV. (280, 260)  **Fig. 14** The correlation function $C(z)$ is shown. The dotted line represents a fit using an exponential function.\ Here, $ C(z)=\frac12\int_{-\infty}^\infty\,dP_ze^{iP_zz}\int_0^\infty\,d\omega\frac{\sigma(\omega, 0, 0, P_z)}\omega\,. $ We may also use the form $ C(z)=\frac14\int_{-\infty}^\infty\,dP_ze^{iP_zz}\int_0^\infty\,dP^2\,\frac{\sigma(P^2, 0, 0, P_z)}{P^2}\,. $ Quark propagation in the quark-gluon plasma =========================================== *Xiangdong Li, Hu Li, C. M. Shakin and Qing Sun, Physical Review C, **69**, 065201 (2004)*\ \ It has recently been suggested that the quark-gluon plasma formed in heavy-ion collisions behaves as a nearly ideal fluid. That behavior may be understood if the quark and antiquark mean-free- paths are very small in the system, leading to a “sticky molasses" description of the plasma, as advocated by the Stony Brook group. This behavior may be traced to the fact that there are relatively low-energy $q\overline{q}$ resonance states in the plasma leading to very large scattering lengths for the quarks. These resonances have been found in lattice simulation of QCD using the maximum entropy method (MEM). We have used a chiral quark model, which provides a simple representation of effects due to instanton dynamics, to study the resonances obtained using the MEM scheme. In the present work we use our model to study the optical potential of a quark in the quark-gluon plasma and calculate the quark mean-free-path. Our results represent a specific example of the dynamics of the plasma as described by the Stony Brook group. (280,250)  **Fig. 15** Values of $t(P^2, p_2)$ are shown for various values of the quark momentum $|\vec{p}_2|$. Starting with the uppermost curve, the $|\vec{p}_2|$ values in GeV units are 0.01, 0.03, 0.05, 0.07, 0.09, 0.11, 0.13, 0.15, 0.17, 0.19, 0.21, 0.23, 0.25, 0.27, 0.29 and 0.31. (For large $P^2$, we have $t(P^2, p_2)\simeq(1/\pi P^2)G$.) Here $P^2=(p_1+p_2)^2$, where $p_1$ is the antiquark momentum. (280, 240)  **Fig. 16** Values of $n(p_1)$ are shown for $\mu=1.1$GeV (dotted curve), $\mu=1.3$GeV (dashed curve) and $\mu=1.5$GeV (solid curve). Here $T=1.5\,T_c$ with $T_c=270$MeV. (280,260)  **Fig. 17** The imaginary part of the quark optical potential is shown for $\mu=1.1$GeV (dotted curve), $\mu=1.3$GeV (dashed curve) and $\mu=1.5$GeV (solid curve). (We recall that the nucleon-nucleus imaginary optical potential is about 0.01GeV in magnitude.) Calculation of Screening Masses in a Chiral Quark Model ======================================================= *Xiangdong Li, Hu Li, C. M. Shakin and Qing Sun, nucl-th/0405035*\ \ We consider a simple model for the coordinate-space vacuum polarization function which is often parametrized in terms of a screening mass. We discuss the circumstances in which the value $m_{sc}=\pi T$ is obtained for the screening mass. In the model considered here, that result is obtained when the momenta in the relevant vacuum polarization integral are small with respect to the first Matsubara frequency. In order to present our results in the simplest form, we consider only the scalar interaction proportional to $(\overline{q}q)^{2}$. We also extend the definition of $\sigma(\omega,T)$ to include a dependence upon the total moment of the quark and antiquark appearing in the polarization integral. Thus we consider the imaginary part of the correlator, $\sigma(\omega,\overrightarrow{P})$. Since we place $\overrightarrow P$ along the *z*-axis this quantity may be written as $\sigma(\omega, 0, 0, P_z)$. In this work we will present our result for the coordinate-dependent correlator $C(z)$ which is proportional to the correlator, C(z)=12\_[-]{}\^dP\_ze\^[iP\_zz]{}\_0\^d. We may also use the form C(z)=14\_[-]{}\^dP\_ze\^[iP\_zz]{}\_0\^dP\^2. We have made a study of the screening mass in a simple model in order to understand the origin of exponential behavior for the correlator. We consider the Matsubara formalism and note that the quark propagator may be written, with $\beta=1/T$, S\_(,\_n)=.For bosons the vacuum polarization function is given as, (,p\^0)=\_n 1[+\^2+M\^2]{}. We modify the last equation to refer to fermions. In this case the Matsubara frequencies are \_n=and we have (,p\^0)=k,if we keep only the first term in the sum, where $\omega_0=\pi/\beta$. As a next step we drop $p^0$, so that we have (,0)=k.We then take $\overrightarrow{p}$ along the $z$ axis and write $\Pi(p_z)=\Pi(\overrightarrow{p},0)$. We define C(z)=dp\_ze\^[ip\_zz]{}(p\_z).In our calculation we replace $g^2/2\beta$ by unity and use a sharp cutoff so that $|\overrightarrow{k}|<k_{max}$. (280,250)  **Fig. 18** The function $C(z)$ is shown for a sharp cutoff of $k_{max}=0.1$ GeV. The dotted line represents an exponential fit to the curve using $m_{sc}=1.23$ GeV. (We recall that $\pi T$ is equal to 1.27 GeV.) (280,250)  **Fig. 19** The function $C(z)$ of is shown for a sharp cutoff of $k_{max}=0.4$ GeV. The dotted line represents an exponential fit to the curve using $m_{sc}=0.961$ GeV. (We recall that $\pi T$ is equal to 1.27 GeV.)

ArXiv

--- abstract: 'We prove that the Weyl algebra over $\mathbb{C}$ cannot be a fixed ring of any domain under a nontrivial action of a finite group by algebra automorphisms, thus settling a 30-year old problem. In fact, we prove the following much more general result. Let $X$ be a smooth affine variety over $\mathbb{C}$, let $D(X)$ denote the ring of algebraic differential operators on $X,$ and let $\Gamma$ be a finite group. If $D(X)$ is isomorphic to the ring of $\Gamma$-invariants of a $\mathbb{C}$-domain $R$ on which $\Gamma$ acts faithfully by $\mathbb{C}$-algebra automorphisms, then $R$ is isomorphic to the ring of differential operators on a $\Gamma$-Galois covering of $X.$' address: 'University of Toledo, Department of Mathematics & Statistics, Toledo, OH 43606, USA' author: - Akaki Tikaradze title: The Weyl algebra as a fixed ring --- *Dedicated to my parents Natalia and Soso Tikaradze with love and admiration* Introduction ============ Throughout given a variety $X,$ by $\mathcal{O}(X)$ we will denote the algebra of its global functions. Also, given a ring $R$, its center will be denoted by $Z(R).$ Let $S$ be a commutative Noetherian ring. Given a smooth affine variety $X$ over $S,$ we will denote by $D(X)$ its ring of differential operators over $S,$ as usual. Recall that $D(X)$ is generated over $\mathcal{O}(X)$ by the Lie algebra of its $S$-derivations $T^1_X=Der_S(\mathcal{O}(X), \mathcal{O}(X)),$ with the following relations $$[\tau, f]=\tau (f), \tau\tau_1-\tau_1\tau=[\tau_, \tau_1], \quad f\in \mathcal{O}(X), \tau, \tau_1\in T_X.$$ Also, by $W_n(S)$ we will denote the $n$-th Weyl algebra over $S.$ Whether the Weyl algebra over $\mathbb{C}$ can be a fixed point ring of a $\mathbb{C}$-domain under a nontrivial action of a finite group has been an open problem in ring theory for some time now. In this regard it was proved by Smith [@S] that if $R$ is a $\mathbb{C}$-domain and $\Gamma\subset Aut_{\mathbb{C}}(R)$ is a finite solvable group of $\mathbb{C}$-algebra automorphisms, such that $R^{\Gamma}=W_1(\mathbb{C}),$ then $R=R^{\Gamma}.$ This result was generalized by Canning and Holland [@CH] to the ring of differential operators of an affine smooth algebraic curve over $\mathbb{C}.$ Namely, they showed that if $R$ is a $\mathbb{C}$-domain and $\Gamma$ is a finite solvable group of $\mathbb{C}$-automorphisms of $R$ such that $R^{\Gamma}\cong D(X),$ where $X$ is a smooth affine curve over $\mathbb{C},$ then there exists a smooth affine curve $Y$ over $\mathbb{C}$, such that $Y$ is a $\Gamma$-Galois covering of $X$ and $R\cong D(Y).$ The proofs are based on the description of the Picard group of invertible bimodules on $D(X).$ No such result is available for high dimensional $X.$ Moreover these proofs do not work for non-solvable $\Gamma.$ On the other hand Alev and Polo [@AP] showed that if $\Gamma$ is an arbitrary finite group such that if both $R, R^{\Gamma}$ are isomorphic to the $n$-th Weyl algebra $W_n(\mathbb{C})$, then again $R=R^{\Gamma}.$ Similarly they proved that the enveloping algebra of a semi-simple Lie algebra over $\mathbb{C}$ cannot occur as a fixed point ring of a nontrivial finite group action on an enveloping algebra of a semi-simple Lie algebra. Recall that given an affine variety $X$ over an algebraically closed field $\bold{k}$, its $\Gamma$-Galois covering is an affine (possibly disconnected) variety $Y$ together with a $\Gamma$-equivariant étale covering $f:Y\to X,$ such that $\Gamma$ acts simply transitively on the fibers of closed points of $X.$ The main results of this paper represents the strongest possible generalization of the above results. \[Main\] Let $X$ be a smooth affine variety over $\mathbb{C}.$ Let $R$ be a $\mathbb{C}$-domain, let $\Gamma\subset Aut_{\mathbb{C}}(R)$ be a finite group of automorphisms of $R.$ If $R^{\Gamma}=D(X)$, then there exists a $\Gamma$-Galois covering $Y\to X$ such that $R\cong D(Y)$ as $\Gamma$-algebras. As an immediate corollary we get that if $X$ is a smooth affine variety over $\mathbb{C}$ such that $X(\mathbb{C})$ is simply connected, then $D(X)$, in particular the Weyl algebras $W_n(\mathbb{C})$, cannot be a fixed point ring of any $\mathbb{C}$-domain under a nontrivial finite group action. To state our next result we will recall some terminology. Let $S$ be a commutative ring. Recall that an $S$-algebra $R$ is a Galois covering of an algebra $B$ with Galois group $\Gamma$ if $\Gamma$ acts on $R$ via $S$-algebra automorphisms, $B=R^{\Gamma}$ and the skew group ring $R\rtimes \Gamma$ is isomorphic to $End_{B}(R);$ moreover $R$ is a finitely generated projective generator as a $B$-module. We will say that a $\Gamma$-Galois covering $R$ of an algebra $B$ is a trivial covering if $R\cong \Pi_{g\in \Gamma} B[e_g]$ where $e_g$ are primitive orthogonal central idempotents, $\sum_{g\in \Gamma} e_g=1$ and $\Gamma$ acts on $e_g$ by (left) multiplication on indices. Let $\mathfrak{g}$ be a semi-simple Lie algebra over $\mathbb{C}.$ Let $U\mathfrak{g}$ be its enveloping algebra. We will say that a central character $\chi:Z(U\mathfrak{g})\to \mathbb{C}$ is very generic if its values on the standard generators of $Z(U\mathfrak{g})$ are algebraically independent over $\mathbb{Z}.$ Denote by $U_{\chi}\mathfrak{g}$ the central quotient of $U\mathfrak{g}$ corresponding to $\chi.$ \[g\] Let $\mathfrak{g}$ be a semi-simple Lie algebra over $\mathbb{C}.$ Then for a very generic central character $\chi$ of $U\mathfrak{g}$, there is no integral domain $R$ with a nontrivial finite group of automorphisms $\Gamma\subset Aut_{\mathbb{C}}R$ so that $R^{\Gamma}=U_{\chi}\mathfrak{g}.$ Notice that it is necessary to require for the central character $\chi:Z(U\mathfrak{g})\to \mathbb{C}$ to be rather generic for the above result to hold, see [@S1]. If $\Gamma$ is a finite subgroup of algebra automorphisms of $R$, and $H$ is a normal subgroup of $\Gamma,$ then as $R^{\Gamma}=(R^{H})^{\Gamma/H}$, proofs of above theorems easily reduce to the case of a simple group $\Gamma.$ Thus we will be assuming from now on that $\Gamma$ is a simple group. The proof is based on the reduction $\mod p$ technique for a large enough prime $p.$ Preliminary results =================== At first we will need to recall a fundamental observation due to Bezrukavnikov, Mirkovic and Rumynin \[[@BMR], Theorem 2.2.3\] which asserts that given a smooth variety $X$ (which for us will always be affine) over an algebraically closed field $\bold{k}$ of characteristic $p>0$, then its ring of (crystalline) differential operators $D(X)$ is an Azumaya algebra over the Frobenius twist of the cotangent bundle $T^{*}(X).$ Namely, given $ \theta\in T_X=Der_{\bold{k}}(\mathcal{O}(X), \mathcal{O}(X)),$ denote by ${\theta}^{[p]}\in T_X$ the $p$-th power of derivation $\theta.$ Then the center of $D(X)$ is generated by $$f^p, \theta^p-\theta^{[p]},\quad f\in \mathcal{O}(X),\quad \theta\in T_X.$$ We will start by establishing an easy commutative version of Theorem \[Main\]. \[Poisson\] Let $X$ be a smooth affine variety over an algebraically closed field $\bold{k}.$ Let $f:Y\to T^{*}(X)$ be a Galois covering with Galois group $\Gamma.$ Assume that $p=char(\bold{k})$ does not divide $|\Gamma|.$ Then there exists an affine variety $Y'$ equipped with a Galois covering $f':Y'\to X$ with Galois group $\Gamma,$ such that $Y\cong T^{*}(Y')$ interchanging $f,T^{*}(f')$. Recall that the cotangent bundle $T^*(X)$ is a conic variety, which amounts to the fact that there is the natural contracting action of the multiplicative group $\mathbb{G}_m$ corresponding to the grading on $\mathcal{O}(T^*(X)).$ We will construct a contracting action of $\mathbb{G}_m$ on $Y$ that is compatible with the action on $T^{*}(X).$ Let $x\in Y$ be a closed point, let $Z=\overline{\mathbb{G}_mf(x)}\cong \mathbb{A}_{\bold{k}}^1.$ Thus $f^{-1}(Z)\to Z$ is a Galois covering with Galois group $\Gamma.$ Since $p$ does not divide $|\Gamma|,$ then $f^{-1}(Z)\to Z$ must be a $|\Gamma|$-fold trivial covering, hence $f^{-1}(Z)=\bigsqcup \mathbb{A}_{\bold{k}}^1.$ Let $Z_x$ be the connected component of $f^{-1}(Z)$ containing $x$. Hence $Z_x\cong \mathbb{A}^1_{\bold{k}}.$ Thus we may equip $Z_x$ with the unique $G_m$-action such that $f:Z_x\to Z$ commutes with it. Varying $x\in Y,$ we get an action of $\mathbb{G}_m$ on $Y$ commuting with $f.$ It follows that this action is a contraction of $X$ on $f^{-1}(X)$ and has no negative eigenvalues. Hence $\mathcal{O}(Y)$ is a nonnegatively graded algebra $$\mathcal{O}(Y)=\bigoplus_{i\geq 0}B_, \quad {\text{Spec}}(B_0)=f^{-1}(X).$$ Put $Y'=f^{-1}(X).$ Next we claim that $T^{*}(Y')\cong Y.$ Indeed, put $B'=B_0[T_X].$ Then ${\text{Spec}}(B')\to T^{*}(X)$ is a $\Gamma$-Galois covering, hence $Y={\text{Spec}}(B')$ and ${\text{Spec}}(B')\cong T^{*}(Y').$ Next we will recall the Howlett-Isacs theorem [@HI] which will be used crucially in Lemma \[twisted\]. Let $G$ be a finite group, $F$ an algebraically closed field whose characteristic does not divide $|G|.$ Let $\rho\in Z^2(G, F^*)$ be a two-cocycle, and let $F_{\rho}[G]$ denote the corresponding $\rho$-twisted group algebra: $$F_{\rho}[G]=\oplus_{g\in G}Fe_g,\quad e_{g_1}\cdot e_{g_2}=\rho(g_1,g_2)e_{g_1g_2}, \quad g_1, g_2\in G.$$ Then the Howlett-Isaacs theorem asserts that if $G$ is a simple group, then for any 2-cocycle $\rho$, $F_{\rho}[G]$ is not a simple $F$-algebra. [^1] Now we have the following key \[twisted\] Let $Z$ be a commutative domain over an algebraically closed field $\bold{k}$ of characteristic $p$, and let $A$ be a finite algebra over $Z,$ such that $Z(A)=Z$ and $A$ has no $Z$-torsion. Let $\Gamma\to Aut_Z(A)$ be a finite simple group of central automorphisms of $A,$ such that $p$ does not divide $|\Gamma|$. Then $A\rtimes \Gamma$ is not an Azumaya algebra over $Z.$ Proof by contradiction. Throughout given a finite dimensional algebra $S,$ its Jacobson radical will be denoted by $J(S).$ We will make a base change from $Z$ to an algebraically closed field $F$ containing $Z$ and put $A_F=A\otimes_ZF.$ We conclude that $Z(A_F)=F$ and $J(A_F)=0$ since $J( A_F\rtimes\Gamma)=0$. So $A_F$ is a central simple algebra over $F$, therefore $A_F$ is the matrix algebra $M_n(F)$ for some $n$. Hence, we have that $$M_n(F)\rtimes\Gamma\cong M_{m}(F).$$ Let $\rho\in Z^2(\Gamma, F^{*})$ be a 2-ccocycle corresponding to the projective representation $\Gamma\to Aut(M_{n}(F)).$ Thus $$F_{\rho}[\Gamma]\otimes_F M_{n}(F)\cong M_{m}(F),$$ where $F_{\rho}[\Gamma]$ is the twisted group algebra of $\Gamma$ corresponding to cocycle $\rho.$ Therefore $F_{\rho}[\Gamma]$ must be a central simple algebra over $F$, which is a contradiction by the Howlett-Isaacs theorem. We will make use of the following lemma which is an immediate corollary of a result by Brown and Gordon \[[@BG] Theorem 4.2\]. At first let us recall their definition of Poisson orders. Let $A$ be a Poisson algebra over $\mathbb{C}.$ Then a Poisson order over $A$ is an algebra $B$ containing $A$ as a central subalgebra such that it is finitely generated as a $A$-module, and is equipped with a $\mathbb{C}$-linear map $A\to Der_{\mathbb{C}}(B, B)$ satisfying the Leibnitz identity, such that it restricts to the Poisson bracket on $A$. \[BGord\] Let $ A$ be an affine Poisson $\mathbb{C}$-algebra such that the Poisson bracket on $A$ induces a structure of a smooth symplectic variety on ${\text{Spec}}(A).$ Let $B$ be a Poisson order over $A.$ Then for all $t\in{\text{Spec}}(A),$ algebras $B/tB$ are mutually isomorphic. In particular, the restriction map ${\text{Spec}}(B)\to {\text{Spec}}(A)$ is an étale covering. We will need the following result which follows immediately from [@M]. \[Susan\] Let $A$ be a simple Noetherian domain over $\mathbb{C}$ such that $Z(A)=\mathbb{C}.$ Let $R$ be a $\mathbb{C}$-domain equipped with a faithful $\Gamma$-action, such that $R^{\Gamma}=A.$ Then $R$ is a $\Gamma$-Galois covering of $A.$ It follows that $R$ is a simple domain: If $I$ is a nonzero proper two-sided ideal of $R$, then $ J=\bigcap_{g\in \Gamma} g(I)$ is a $\Gamma$-invariant nonzero two-sided ideal in $R.$ Then $J^{\Gamma}\neq 0$ since $J$ is not nilpotent as follows from \[[@M], Theorem 1.7\]. Hence $J\cap R \neq 0$ is a proper nonzero two-sided ideal in $R$, a contradiction. It follows that $Z(R)=\mathbb{C}$ and $\Gamma$ acts on $R$ by outer automorphisms: If $g\in \Gamma$ such that $g(x)=axa^{-1}$ for some $a\in R,$ then $a^{|\Gamma|}\in Z(R)$, but since $Z(R)=\mathbb{C},$ we get that $a^{|\Gamma|}\in \mathbb{C}$. So $a$ is a root of unity and $g$=Id. Hence by \[[@M], Theorem 2.5\], $R$ is a $\Gamma$-Galois covering of $A.$ Finally we have the following crucial \[Azu\] Let $Z$ be an affine commutative domain over an algebraically closed field $\bold{k}$ of characteristic $p.$ let $A$ be an Azumaya algebra over $Z$, and let $R$ be a $\Gamma$-Galois covering of $A.$ Assume that $p$ does not divide $|\Gamma|.$ Then ${\text{Spec}}(Z(R))\to {\text{Spec}}(Z)$ is a $\Gamma$-Galois covering and $R=A\otimes_ZZ(R).$ Let us put $Z(R)=Z_1.$ Remark that $Z=Z_1^{\Gamma}.$ Indeed, since $ R\rtimes\Gamma$ is an Azumaya algebra over $Z$, then $Z(R\rtimes\Gamma)=Z,$ in particular $R$ commutes with $Z,$ therefore $Z\subset Z_1$. Hence $Z_1^{\Gamma}=Z.$ Since $R$ is a $\Gamma$-Galois covering of $A$, it follows that $R$ is a projective left, right $A$-modules of rank $|\Gamma|.$ Since $Z\subset Z_1,$ it follows that $R$ is a module over $ A\otimes_{Z}A^{op}$. As $A$ is an Azumaya algebra over $Z$, it follows that $R\cong A\otimes_{Z}B,$ where $B$ is the centralizer of $A$ in $R.$ Thus we conclude that $B\rtimes\Gamma$ is an Azumaya algebra over $Z$ of rank $|\Gamma|^2.$ Also $Z(B)=Z_1.$ It follows from Lemma\[twisted\] that the action of $\Gamma$ on $Z_1$ must be faithful. Also it is clear that $Z_1$ has no nilpotent elements. Let $ \eta\in {\text{Spec}}(Z)$ be the generic point. Then $Z(B_{\eta})=(Z_1)_{\eta}$ and $B_{\eta}$ is a $|\Gamma|$-dimensional semi-simple $Z_{\eta}$-algebra. On the other hand, since $\Gamma$ acts faithfully on $(Z_1)_{\eta}$ and $(Z_1)^{\Gamma}_{\eta}=Z_{\eta},$ it follows that $(Z_1)^{\Gamma}_{\eta}$ is $|\Gamma|$-dimensional over $(Z)_{\eta}.$ Therefore $(Z_1)_{\eta}=B_{\eta}.$ Hence $B=Z_1,$ since $B$ has no $Z$-torsion. Now we claim that ${\text{Spec}}(Z_1)\to {\text{Spec}}(Z)$ is an étale covering with Galois group $\Gamma.$ Indeed, for any $\chi\in {\text{Spec}}(Z)$, we have that $(Z_1)_{\chi}\rtimes\Gamma$ is a matrix algebra. Therefore $J((Z_1)_{\chi})=0$. Since $\dim (Z_1)_{\chi}=|\Gamma|,$ it follows that $$(Z_1)_{\chi}=\bold{k}\times\cdots\times\bold{k}.$$ Hence ${\text{Spec}}(Z_1)\to {\text{Spec}}(Z)$ is an étale covering with Galois group $\Gamma,$ as desired. We will also need the following result on lifting of $p'$-order Galois coverings from characteristic $p$ to characteristic 0 [^2]. It will only be used in the proof of Theorem \[g\], not for Theorem \[Main\]. \[lifting\] Let $X$ be a smooth geometrically connected affine variety over $S$, where $S\subset \mathbb{C}$ is a finitely generated ring. Suppose that a finite group $\Gamma$ appears as a quotient of the étale fundamental group of $X_{\bold{k}}$ for all large enough primes $p>>0$, where $X_{\bold{k}}=X\times_{{\text{Spec}}(S)}{\text{Spec}}(\bold{k})$ is a $\mod p$ reduction of $X$ by a base change $S\to \bold{k}$ to an algebraically closed field of characteristic $p.$ Then $\Gamma$ is a quotient of the fundamental group of $X(\mathbb{C}).$ This follows from general results about étale coverings tamely ramified across normal crossing divisors. Indeed, let $\overline{X}$ be a good compactification of $X_{\mathbb{C}}$, thus it is a smooth projective variety such that $\overline{X}\setminus X_{\mathbb{C}}$ is a normal crossings divisor. By enlarging $S$ we may assume that $\overline{X}$ is defines over $S$ and $\overline{X}\setminus X$ is a normal crossings divisor over $S$. For $p>>0$, for a base change $S\to \bold{k},$ $\overline{X_{\bold{k}}}\setminus X_{\bold{k}}$ is a divisor with normal crossings. Let us choose an embedding $S\to V$, where $V$ is a complete discrete valuation ring with the residue field $\bold{k}$-an algebraically closed field of characteristic $p.$ For example we may take $V$ to be the ring of Witt vectors over $\bold{k},$ where $\bold{k}$ is large enough algebraically closed field of characteristic $p.$ Then it follows from \[[@LO], Corollary A.12\] that $X_V$ has a connected Galois covering with the Galois group $\Gamma,$ Applying a base change $V\to \mathbb{C}$ we obtain the desired $\Gamma$-Galois covering of $X_{\mathbb{C}}.$ The proof of the main result for rings of differential operators ================================================================ We start by the following Lemma. It shows that to prove Theorem \[Main\], it will suffice to prove that $R$ as a bimodule over $D(X)$ is supported on the diagonal of $X\times X.$ \[K\] Let $X$ be a smooth affine variety over $\mathbb{C},$ and let $R$ be a $\mathbb{C}$-domain. Assume that a finite simple group $\Gamma$ acts faithfully on $R$ so that $R^{\Gamma}=D(X).$ Moreover assume that for any $f\in \mathcal{O}(X), {\text{ad}}(f)=[f, -]$ acts locally nilpotently on $R.$ Then $R=D(Y),$ where $Y$ is a $\Gamma$-Galois covering of $X.$ We may view $R$ as a left module over $D(X\times X)$. Assumptions above imply that $R$ is supported on the diagonal $\Delta(X)\subset X\times X.$ Let $R'$ be the centralizer of $\mathcal{O}(X)$ in $R.$ Clearly $R'$ is preserved by ${\text{ad}}(y)$ for all $y\in Der_{\mathbb{C}}(\mathcal{O}(X),\mathcal{O}(X))=T_X.$ Thus $R'\otimes_{\mathcal{O}(X)}D(X)$ is naturally a $D(X)$-bimodule equipped with the bimodule map $R'\otimes_{\mathcal{O}(X)}D(X)\to R.$ Now recall that Kashiwara’s theorem \[[@GM], Theorem 4.9.1\] establishes an equivalence between the category of $D(X\times X)$-modules supported on the diagonal and the category of $D(X)$-modules, given by direct and inverse image functors corresponding to the diagonal embedding $\Delta:X\to X\times X.$ Applying this to $R$, we get that $R\cong R'\otimes_{\mathcal{O}(X)}D(X).$ Next we will introduce an ascending filtration on $R$ as follows. Let $$R_n=\lbrace x\in R,\quad {\text{ad}}(f)^{n+1}(x)=0,\forall f\in \mathcal{O}(X)\rbrace, n\geq 0.$$ Then $$R_0=R', \quad R_1=R'T_X=T_XR', \quad R_nR_m\subset R_{n+m}.$$ This filtration restricts on $D(X)$ to the usual filtration by the order of differential operators. It follows that ${\text{gr}}R=R'\otimes_{\mathcal{O}(X)}\mathcal{O}(T^*(X))$ is a finitely generated module over $\mathcal{O}(T^{*}(X))={\text{gr}}(D(X)),$ moreover ${\text{gr}}R$ is a Poisson order over $\mathcal{O}(T^{*}(X)).$ Now lemma \[BGord\] implies that for all $\eta\in T^{*}(X)$, algebras $({\text{gr}}R)_{\eta}$ are isomorphic to each other. Therefore, for all $\eta\in X,$ algebras $R'_{\eta}$ are mutually isomorphic. Put $Y={\text{Spec}}(Z(R')).$ We claim that $Y\to X$ is finite étale map and $R'$ is an Azumaya algebra over $Y.$ Indeed, since for any $y\in T_X$ $${\text{ad}}(y)(\mathcal{O}(Y))\subset \mathcal{O}(Y),$$ then $R''=Z(R')\otimes_{\mathcal{O}(X)}D(X)$ is a $\Gamma$-invariant subalgebra of $R$, and $R''^{\Gamma}=D(X).$ Thus applying the above argument to $R''$ instead of $R$ we get that ${\text{Spec}}(Z(R'))\to X$ is a finite étale map. Then we have a homomorphism of algebras $\theta:R''=\mathcal{O}(Y)\otimes_{\mathcal{O}(X)}D(X)\to D(Y)$ given by mapping $y\in T_X$ to $ {\text{ad}}(y)\in Der_{\mathbb{C}}(\mathcal{O}(Y), \mathcal{O}(Y))=T_{Y}.$ As $T_Y=T_X\otimes_{\mathcal{O}(X)}\mathcal{O}(Y),$ it follows that $\theta$ is an isomorphism: $R''=D(Y)$ and as a bimodule over $D(Y), R$ is supported on the diagonal of $Y\times Y$. Hence in the above argument we may replace $D(X)$ by $D(Y)$ to conclude that $R'$ is an Azumaya agebra over $Y$. Notice that up to now we have not used the action of $\Gamma$ on $R.$ Since $D(X)$ is a simple ring, it follows from lemma \[Susan\] that $R$ is a $\Gamma$-Galois covering of $D(X).$ Hence $R$ is a projective $D(X)$-module of rank $|\Gamma|$. Therefore $R'$ is also a projective $\mathcal{O}(X)$-module of rank $|\Gamma|.$ If the action of $\Gamma$ on $Z(R')$ is faithful then $Y\to X$ is a $\Gamma$-Galois covering, hence $R''=R$ and we are done. Thus we may assume that $Z(R')=\mathcal{O}(X).$ Then $R'^{\Gamma}=\mathcal{O}(X)$ and $\Gamma\subset Aut_{\mathcal{O}(X)}R'.$ Since $R\rtimes\Gamma$ is Morita equivalent to $D(X),$ It follows that $R\rtimes\Gamma'$ is Morita equivalent to $\mathcal{O}(X)$, which is a contradiction by Lemma \[twisted\]. It follows from the above proof that if $X$ is a smooth algebraic curve and $R$ is a $\mathbb{C}$-domain containing $D(X)$, such that it is finite as both left and right $D(X)$-module, and $R$ is supported on the diagonal of $X\times X$ as a $D(X)$-bimodule, then $R\cong D(Y)$, where $Y\to X$ is a finite étale map. Indeed, this is because the Azumaya algebra in the proof must be split as the Brauer group over curves is trivial over algebraically closed fields. As in the proof above, since $D(X)$ is a simple ring, it follows from \[Susan\] that $R$ is a $\Gamma$-Galois covering of $D(X).$ There exists finitely generated $\mathbb{Z}$- algebra $S\subset \mathbb{C}, \frac{1}{|\Gamma|}\in S,$ such that $X=X'_\mathbb{C},$ where $X'$ is a smooth affine variety over $S,$ and an $S$-subalgebra $R'$ of $ R$ such that $R=R'\otimes_{S}\mathbb{C},$ $\Gamma$ acts on $R'$ by $S$-automorphisms and $R'$ is a $\Gamma$-Galois covering of $D(X').$ Therefore for all large primes $p>>0$ there exists a homomorphism $\rho:S\to \bold{k}$, where $\bold{k}$ is an algebraically closed field of characteristic $p$, such that $R_{\bold{k}}=R'\otimes_S\bold{k}$ is a $\Gamma$-Galois covering of $D(X_{\bold{k}}),$ where $X_{\bold{k}}=X'\times _{{\text{Spec}}(S)}{\text{Spec}}(\bold{k}).$ For simplicity we will denote by $Z_0, Z_1$ the centers of $D(X_{\bold{k}}), R_{\bold{k}}$ respectively. Recall that $D(X_{\bold{k}})$ is an Azumaya algebra over $Z_0$, and ${\text{Spec}}(Z_0)=T^{*}(X_{\bold{k}}).$ It follows from Proposition \[Azu\] that $\Gamma$ acts faithfully on $Z_1, Z_0=Z_1^{\Gamma}$ and ${\text{Spec}}(Z_1)\to {\text{Spec}}(Z_0)$ is a $\Gamma$-Galois covering. Applying Lemma\[Poisson\] to the $\Gamma$-Galois covering ${\text{Spec}}(Z_1)\to {\text{Spec}}(Z_0)$, we conclude that there exists an affine $\bold{k}$-variety $Y_{\bold{k}}$ and an étale covering $Y_{\bold{k}}\to X_{\bold{k}}$ with the Galois group $\Gamma,$ such that ${\text{Spec}}(Z_1)\cong T^{*}(Y_{\bold{k}})$ as $\Gamma$-varieties. Since $$D(Y_{\bold{k}})\cong D(X_{\bold{k}})\otimes_{Z_0}Z(D(Y_{\bold{k}})),$$ and $$Z(D(Y_{\bold{k}}))\cong \mathcal{O}(T^{*}(Y_{\bold{k}})),$$ it follows that $D(Y_{\bold{k}})\cong R_{\bold{k}}$ as $\Gamma$-algebras. Next we will show that if $f\in\mathcal{O}(X'),$ then ${\text{ad}}(f)$ is locally nilpotent on $R'.$ In view of Lemma \[K\], this will yield the theorem. Let $x\in R'.$ Let $f_1,\cdots, f_m\in D(X')$ be such that $$x^m+\sum_{i=0}^{m-1} f_ix^{m-i}=0 .$$ Let $N=Max_i(\deg(f_i))$, here $\deg$ means the degree of a differential operator in $D(X')$. We claim that ${\text{ad}}(f)^{N+1}(x)=0$. We will show this by proving that ${\text{ad}}(\bar{f})^m(\bar{x})=0$ in $R_{\bold{k}}$ for all large enough $p,$ where $\bar{y}$ denotes the image of an element $y\in R'$ under the base change map $R'\to R_{\bold{k}}.$ Since $R_{\bold{k}}\cong D(Y_{\bold{k}}),$ we will equip $R_{\bold{k}}$ with the filtration corresponding to the degree filtration of differential operators in $D(Y_{\bold{k}})$. Hence ${\text{gr}}R_{\bold{k}}=\mathcal{O}(T^{*}(Y_{\bold{k}})).$ Clearly this filtration restricts on $D(X_{\bold{k}})$ to the filtration corresponding to the order of differential operators on $X_{\bold{k}}.$ Thus ${\text{gr}}(\bar{f_i})\leq N, 1\leq i\leq m.$ Now it follows that ${\text{gr}}R_{\bold{k}}$ has no nilpotent elements and is a torsion free ${\text{gr}}D(X_{\bold{k}})=\mathcal{O}(T^{*}(X_{\bold{k}}))$-module. Thus we have that $${\text{gr}}(\bar{x})^m=-\sum_{i=1}^m {\text{gr}}(\bar{f_i}){\text{gr}}(\bar{x})^{m-i}.$$ This implies that ${\text{gr}}(\bar{x})\leq N.$ Hence for any $g\in \mathcal{O}(X_{\bold{k}}),$ we have ${\text{ad}}(g)^m(\bar{x})=0.$ In particular, ${\text{ad}}(\bar{f})^m(\bar{x})=0.$ Therefore, ${\text{ad}}(f)^m(x)=0$ and we are done. The proof for enveloping algebras ================================= Before proving Theorem \[g\] we will need to recall some standard facts about enveloping algebras of semi-simple Lie algebras in characteristic $p>0.$ Let $G$ be a semi-simple, simply connected algebraic group of rank $n$ over $\mathbb{C}$, let $\mathfrak{g}$ be its corresponding Lie algebra. Let $p>>0$ be a very large prime, and let $G_{\bf{k}}, \mathfrak{g}_{\bold{k}}$ denote reductions of $G, \mathfrak{g}$ to an algebraically closed field $\bold{k}$ of characteristic $p.$ We will be identifying $\mathfrak{g}_{\bold{k}}$ with $\mathfrak{g}_{\bold{k}}^{*}$ via a nondegenerate $G_{\bold{k}}$-invariant bilinear form on $\mathfrak{g}_{\bold{k}}$, as usual. Let $f_1,\cdots, f_n$ be homogeneous generators of $\bold{k}[\mathfrak{g}^*]^G=\bold{k}[f_1,\cdots, f_n].$ Let $\tilde{f_1},\cdots, \tilde{f_n}$ be the corresponding generators of $(U\mathfrak{g}_{\bold{k}})^{G_{\bold{k}}}=\bold{k}[\tilde{f_1},\cdots, \tilde{f_n}].$ Given $\chi \in {\text{Spec}}(\bold{k}[\mathfrak{g}^*_{\bold{k}}]^{G_{\bold{k}}}),$ we will denote by $\mathcal{N}_{\chi}$ the preimage of $\chi$ under the map $F=(f_1,\cdots, f_n):\mathfrak{g}_{\bold{k}}\to \bold{k}^n.$ Of course for the origin $\chi=0, \mathcal{N}_0$ is the nilpotent cone of $\mathfrak{g}_{\bold{k}}.$ While for a generic $\chi, \mathcal{N}_{\chi}$ is the conjugacy class of a regular semi-simple element of $\mathfrak{g}_{\bold{k}}.$ Given $\chi\in {\text{Spec}}(\bold{k}[\mathfrak{g_{\bold{k}}}^*]^{G_{\bold{k}}}) $, we will say that $\chi$ is very generic if its values on $f_1,\cdots,f_n$ are algebraically independent over $F_p.$ Similarly, given $\chi\in {\text{Spec}}(U\mathfrak{g})^{G_{\bold{k}}}$ , we will say that it is very generic if its values on $\tilde{f_1},\cdots,\tilde{f_n}$ are algebraically independent over $F_p.$ Let $\chi:U\mathfrak{g}_{\bold{k}}^{G_{\bf{k}}}\to \bold{k}$ be a character, and let $U_{\chi}\mathfrak{g}_{\bold{k}}$ be the corresponding central reduction of $U\mathfrak{g}_{\bold{k}}.$ Then it is known that $Z(U_{\chi}\mathfrak{g}_{\bold{k}})\cong \mathcal{O}(\mathcal{N}_{\tilde{\chi}})$ for the corresponding $\tilde{\chi}.$ Moreover for a very generic $\chi$ character $\tilde{\chi}$ is also very generic, and $ U_{\chi}(\mathfrak{g}_{\bold{k}})$ is an Azumaya algebra. Indeed, Let $V$ be a simple $U_{\chi}(\mathfrak{g}_{\bold{k}})$-module. Then the $p$-character of $V$ viewed as a simple $\mathfrak{U}\mathfrak{g}$-module is regular. Hence $\dim V=p^d$, where $d$ is half the dimension of the nilpotent cone of $\mathfrak{g}_{\bold{k}}.$ Therefore all simple $U_{\chi}(\mathfrak{g}_{\bold{k}})$-modules have the same dimension. So $U_{\chi}(\mathfrak{g}_{\bold{k}})$ is an Azumaya algebra by \[[@BG1], Proposition 3.1\]. As varieties $\mathcal{N}_{\chi}$ are simply connected for very generic characters over $\mathbb{C},$ Lemma \[lifting\] yields the following \[Congugacy\] Let $G, \mathfrak{g}$ be as above and $p>>0.$ Let $\chi:\bold{k}[\mathfrak{g_{\bold{k}}}]^G_{\bold{k}}\to \bold{k}$ be a very generic character. Then the étale fundamental group of $\mathcal{N}_{\chi}$ has no $p'$ finite group quotients. We start exactly as in the beginning of the proof of Theorem \[Main\]. Since ${U}_{\chi}\mathfrak{g}$ is a simple $\mathbb{C}$-algebra, it follows from Lemma \[Susan\] that $R$ is a Galois covering of ${U}_{\chi}\mathfrak{g}.$ Let $S$ be a finitely generated subring of $\mathbb{C}$ over which $U_{\chi}\mathfrak{g}$ is defined, and let $R'$ be a $S$-subring of $R$ such that $R=R'\otimes_S\mathbb{C}$ and $R'$ is $\Gamma$-Galois covering of ${U}_{\chi}(\mathfrak{g}).$ Let $S\to \bold{k}$ be a base change to algebraically closed field of characteristic $p>>0.$ Put $R_{\bold{k}}=R'\otimes\bold{k}.$ Since $U_{\chi}(\mathfrak{g}_{\bold{k}})$ is an Azumaya algebra over $\mathcal{O}(\mathcal{N}_{\bar{\chi}})$ it follows from Proposition \[Azu\] that ${\text{Spec}}(Z(R_{\bold{k}}))\to \mathcal{N}_{\bar{\chi}}$ is a $\Gamma$-Galois covering. But by Lemma \[Congugacy\], $\mathcal{N}_{\bar{\chi}}$ has no nontrivial $\Gamma$-Galois covering. Therefore $$Z(R_{\bold{k}})\cong \Pi_{i=1}^{|\Gamma|} \mathcal{O}(\mathcal{N}_{\bar{\chi}})e_i,$$ where $e_i$ are orthogonal central idempotents $\sum_{i=1}^{|\Gamma|} e_i=1.$ Hence by Lemma \[Azu\], it follows that $$R_{\bold{k}}\cong \Pi_{i=1}^{|\Gamma|}{U}_{\chi}(\mathfrak{g}_{\bold{k}})e_i.$$ In particular $R_{\bold{k}}$ has no nonzero nilpotent elements. Next we claim that $R'$ is a Harish-Chandra bimodule over $\mathfrak{g}:$ the adjoint action of $\mathfrak{g}$ on $R'$ is locally finite. Indeed, let $x\in R'$. It will suffice to show that the nilradical of a Borel subalgebra $\mathfrak{n}\subset \mathfrak{g}$ acts nilpotently on $x.$ Using the above isomorphism $R_{\bold{k}}\cong \Pi_{i=1}^nU_{\chi}(\mathfrak{g}_{\bold{k}})e_i$ we will introduce a PBW filtration on $R_{\bold{k}}$ which will restrict to the usual PBW filtration on $U_{\chi}(\mathfrak{g}_{\bold{k}}).$ Thus, $${\text{gr}}R_{\bold{k}}\cong \Pi_{i=1}^n{\text{gr}}(U_{\chi}(\mathfrak{g}_{\bold{k}}))e_i,\quad \deg(e_i)=0.$$ Just as in the proof of Theorem \[Main\], since $x$ is integral over $U_{\chi}\mathfrak{g},$ let $$x^m+\sum f_ix^{m-i}=0,\quad f_i\in U_{\chi}\mathfrak{g}.$$ Put $N=Max_i(\deg f_i)$, here $\deg f_i$ is understood as the filtration degree according to the PBW filtration on $U_{\chi}(\mathfrak{g}_{\bold{k}}).$ Then it follows that $\deg (\bar{x})\leq N$ for all $p>>0,$ where $\bar{x}$ is the image of $x$ under the map $R\to R_{\chi}=\Pi_iU_{\chi}(\mathfrak{g}_{\bold{k}})e_i.$ Hence $$x=\sum x_ie_i, \quad x_i\in U_{\chi}(\mathfrak{g})_{\bold{k}},\quad\deg(x_i)\leq N.$$ Let $l$ be such that ${\text{ad}}(\mathfrak{n})^l(\mathfrak{g})=0$. Then ${\text{ad}}(\mathfrak{n})^{lN}(x_i)=0,$ so ${\text{ad}}(\mathfrak{n})^{lN}(\bar{x})=0$. Hence ${\text{ad}}(\mathfrak{n})^{lN}(x)=0.$ Therefore the adjoint action of $\mathfrak{g}$ on $R'$ is locally finite as desired. Now it follows that for any semi-simple $h\in \mathfrak{g}, R$ is $\mathbb{Z}$-graded according to eigenvalues of ${\text{ad}}(h)$ on $R.$ Let $h_1,\cdots, h_m\in \mathfrak{g}$ be semi-simple element such that they generate $\mathfrak{g}$ as a Lie algebra. By enlarging $S$ if necessary we may assume that $R'$ is $\mathbb{Z}^m$-graded algebra according to eigenvalues of ${\text{ad}}(h_i), 1\leq i\leq m.$ Hence the degree 0 component $R'$ is the centralizer $\mathfrak{g}$ in $R'.$ Thus $(R_{\bold{k}})$ is also $\mathbb{Z}^m$-graded. Since $e_i^2=e_i$ and $R_{\bold{k}}$ has no nilpotent elements, this forces $e_i$ to have degree 0. Denote the centralizer of $\mathfrak{g}$ in $R'$ (respectively in $R$) by $R'_0$ (respectively $R_0$). Now we claim that $R_0$ is commutative. Indeed, since $(R'_0)_{\bf{k}}$ is in the centralizer of $\mathfrak{g}$ in $R_{\bf{k}}$, which is $Z(U_{\chi}(\mathfrak{g}_{\bold{k}}))e_i,$ hence commutative. So $(R'_0)_{\bf{k}}$ is commutative for all $p>>0$, hence so is $R'_0$ and $R_0.$ We also have that $e_i\in (R'_0)_{\bf{k}}, 1\leq i\leq n.$ In particular $\mathbb{C}\neq R_0.$ Thus $R_0$ is a commutative $\mathbb{C}$-domain, equipped with a $\mathbb{C}$-action of $\Gamma$, such that $\mathbb{C}=R_0^{\Gamma}\neq R_0$. This is a contradiction. **Acknowledgement:** I am extremely grateful to A. Eshmatov for many useful discussion, particularly for asking me the question whether the Weyl algebra can be Morita equivalent to a nontrivial skew group ring, answering which led to this paper. I would also like to thank the referee for many helpful suggestions. [qowq]{} J. Alev, P. Polo, [*A rigidity theorem for finite group actions on enveloping algebras of semisimple Lie algebras*]{}, Adv. Math. 111 (1995), no. 2, 208–226. K. Brown, K. Goodearl, [*Homological Aspects of Noetherian PI Hopf Algebras and Irreducible Modules of Maximal Dimension*]{}, J. Algebra (1997), 198, 240–265. K. Brown, I. Gordon, [*Poisson orders, symplectic reflection algebras and representation theory*]{}, J. Reine Angew. Math. 559 (2003), 193–216. R. Bezrukavnikov, I. Mirkovic, D. Rumynin, [*Localization of modules for a semisimple Lie algebra in prime characteristic*]{}, Annals of Mathematics, 167 (2008), 945–991. S. Gelfand, Y. Manin, [*Homological Algebra*]{}, Springer. R. Howlett, M. Isaacs, [*On groups of central type*]{}, Math. Z. 179 (1982), no. 4, 555–569. M. Lieblich, M. Olsson, [*Generators and relations for the etale fundamental group*]{}, Pure and Applied Math. Quarterly, (2010) vol. 6 no. 1 209–243. S.P. Smith, [*Can the Weyl algebra be a fixed ring?*]{}, Proc. Amer. Math. Soc., 107 (1989) 587–589. S.P. Smith, [*Overrings of primitive factor rings of $U(sl(2,C))$*]{}, J. Pure Appl. Algebra, 63 (1990) 207–218. R. Cannings, M. Holland, [*Étale covers, bimodules and differential operators*]{}, Math. Z. 216 (1994), no. 2, 179–194. S. Montgomery, [*Fixed rings of finite automorphism groups of associative rings*]{}, (1980) Lecture Notes in Math. [^1]: We thank Professor Pham Tiep for telling us about the Howlett-Isaacs theorem. [^2]: I am grateful to D.Harbater, especially P.Achinger and A.Javanpeykar for the proof.

ArXiv

Introduction ============ Consider a set of $K$ elements. These elements are repeatedly compared with one another in pairs. For two elements $i$ and $j$ of this set, [@Bradley1952] suggested the following model$$\Pr(i\text{ beats }j)=\frac{\lambda_{i}}{\lambda_{i}+\lambda_{j}} \label{eq:BTmodel}$$ where $\lambda_{l}>0$ is a parameter associated to element $l\in\left\{ 1,2,\ldots,K\right\} $ that represents its skill rating and we denote $\lambda:=\left\{ \lambda_{i}\right\} _{i=1}^{K}$. This model has found numerous applications. As mentioned in [@Hunter2004], as early as 1976, a published bibliography on paired comparisons includes several hundred entries [@Davidson1976]. For example, it has been adopted by the World Chess Federation and the European Go Federation to rank players and it is a standard approach to build multiclass classifiers based on the output of binary classifiers [@Hastie1998]. Various extensions have been proposed to handle home advantage [@Agresti1990], draws [@Rao1967], multiple [@Plackett1975; @Luce1959] and team comparisons [@Huang2006]. In particular, the popular extension to multiple comparisons, named the Plackett-Luce model [@Plackett1975; @Luce1959], defines a prior distribution over permutations and has been used for ranking of multiple individuals and for choice models [@Luce1977]. The monographs of [@David1988] and @Diaconis1988 [Chap. 9] provide detailed discussions on the statistical foundations of these models. For the basic Bradley-Terry model (\[eq:BTmodel\]), it is possible to find the maximum likelihood (ML) estimate of the skill ratings $\lambda$ using a simple iterative procedure [@Zermelo1929; @Hunter2004]. [@Lange2000] established that this procedure is a specific case of the general class of algorithms referred to as MM algorithms. Generally speaking, MM algorithms use surrogate minimizing functions of the log-likelihood to define an iterative procedure converging to a local maximum. EM algorithms are thus just a special case of MM algorithms. An excellent survey of the MM approach and its applications can be found in [@Lange2000]. [@Hunter2004] further derived MM algorithms for generalized Bradley-Terry models and established sufficient conditions under which these algorithms are guaranteed to converge towards the ML estimate. Recently several authors have proposed to perform Bayesian inference for (generalized) Bradley-Terry models [@Adams2005; @Gormley2009; @Gorur2006; @Guiver2009]. The resulting posterior density is typically not tractable and needs to be approximated. An Expectation-Propagation method is developed in [@Guiver2009]; this yields an approximation of the posterior which can be computed quickly and might be suitable for very large scale applications. However, it relies on a functional approximation of the posterior and the convergence properties of this algorithm are not well-understood. M-H algorithms have been proposed in [@Adams2005; @Gormley2009; @Gorur2006]. [@Gormley2009] suggested a carefully designed proposal distribution, though it can perform poorly in some scenarios as demonstrated in section \[sec:experiments\]. Our contribution here is three-fold. First, we show that by introducing suitable sets of latent variables, the MM algorithms proposed by [@Hunter2004] for the basic Bradley-Terry model and its generalizations to take into account home advantage, ties and multiple comparisons can be reinterpreted as standard EM algorithms. We believe that this non-trivial reinterpretation is potentially fruitful for statisticians who usually like thinking in terms of latent variables. Note that the latent variables introduced here differ from the ones introduced in the standard Thurstonian interpretation of the Bradley-Terry model [@Diaconis1988 Chap. 9] and lead to more efficient algorithms as discussed in section \[sec:bt\]. Second, using similar ideas, we propose original EM algorithms for some recent generalizations of the Bradley-Terry model including group comparisons and random graphs. Third, based on the sets of latent variables introduced to derive these EM algorithms, we propose Gibbs samplers to perform Bayesian inference in this important class of models. To the best of our knowledge, no Gibbs sampler has ever been proposed in this context. These algorithms have the great advantage of allowing us to bypass the design of proposal distributions for M-H updates and we demonstrate experimentally that they perform very well. The rest of this paper is organized as follows. In section \[sec:bt\], we consider the basic Bradley-Terry model (\[eq:BTmodel\]). Based on the introduction of a suitable set of latent variables, we present an EM reinterpretation of the MM algorithm presented by [@Hunter2004] for Maximum a Posteriori (MAP) parameter estimation and an original data augmentation algorithm to sample from the posterior. In section \[sec:genbt\], various standard extensions of the Bradley-Terry model allowing for home advantage, ties and competition between teams are described. EM algorithms and original Gibbs sampling schemes are proposed. The Plackett-Luce model [@Plackett1975; @Luce1959], a very popular generalization of the Bradley-Terry model for multiple comparisons, is presented in section \[sec:plackettluce\]. Algorithms applicable to further extensions of the Bradley-Terry model to choice models, random graphs and classification are presented in section \[sec:discussion\]. In section \[sec:experiments\], these algorithms are applied to the NASCAR 2002 dataset and to chess competition data. Bradley-Terry model\[sec:bt\] ============================= Suppose we have observed a number of statistically independent pairwise comparisons among $K$ individuals. We denote by $D$ the associated data. Let also $w_{ij}$ denote the number of comparisons where $i$ beats $j$ and $n_{ij}=w_{ij}+w_{ji}$ the total number of comparisons between $i$ and $j$. Based on the Bradley-Terry model (\[eq:BTmodel\]), the log-likelihood function is given by$$\begin{aligned} \ell(\lambda) & =\sum_{1\leq i\neq j\leq K}\text{ }w_{ij}\log\lambda _{i}-w_{ij}\log(\lambda_{i}+\lambda_{j})\\ & =\sum_{1\leq i\neq j\leq K}\text{ }w_{ij}\log\lambda_{i}-\sum_{1\leq i<j\leq K}\text{ }n_{ij}\log(\lambda_{i}+\lambda_{j})\end{aligned}$$ where the notation $1\leq i\neq j\leq K$ is an abusive notation to denote the set $\left\{ \left( i,j\right) \in\left\{ 1,...,K\right\} ^{2}\right .$ $\left .\text{ such that }i\neq j\right\} $ and $1\leq i<j\leq K$ stands for $\left\{ \left( i,j\right) \in\left\{ 1,...,K\right\} ^{2}\text{ such that }i<j\right\} $. We seek to introduce latent variables which are such that the resulting complete log-likelihood admits a simple form. It is well-known that the Bradley-Terry model enjoys the following Thurstonian interpretation [@Diaconis1988 Chap. 9]: for each pair $1\leq i<j\leq K$ and for each associated pair comparison $k=1,\ldots,n_{ij}$, let $Y_{ki}\sim\mathcal{E}(\lambda_{i})$ and $Y_{kj}\sim\mathcal{E}(\lambda_{j})$ where $\mathcal{E}\left( \varsigma\right) $ is the exponential distribution of rate parameter $\varsigma$ then $$\Pr(Y_{ki}<Y_{kj})=\frac{\lambda_{i}}{\lambda_{i}+\lambda_{j}}.$$ These latent variables can be interpreted as arrival times and the individual with the lowest arrival time wins. These latent variables would allow us to define EM and data augmentation algorithms. However, if we only introduce for each pair $i,j$ the sum of the lowest arrival times $$Z_{ij}=\sum_{k=1}^{n_{ij}}\min(Y_{kj},Y_{ki})\sim\mathcal{G(}n_{ij},\lambda_{i}+\lambda_{j})$$ instead of $\left\{ Y_{ki},Y_{kj}\right\} $ for $k=1,\ldots,n_{ij}$, then the resulting complete log-likelihood remains simple. Here $\mathcal{G}\left( \alpha,\beta\right) $ denote the Gamma distribution of shape $\alpha$ and inverse scale $\beta$. As the fraction of missing information is reduced, this leads to faster rates of convergence for the resulting EM and data augmentation algorithms [@Liu2001 Chap. 6]. We will essentially proceed similarly to introduce latent variables for generalized Bradley-Terry models. To summarize, for $1\leq i<j\leq K$ such that $n_{ij}>0$, we introduce the latent variables $Z=\left\{ Z_{ij}\right\} $ which are such that $$p\left( \left. z\right\vert D,\lambda\right) ={\displaystyle\prod\limits_{1\leq i<j\leq K|n_{ij}>0}} \mathcal{G}(z_{ij};n_{ij},\lambda_{i}+\lambda_{j})\label{eq:conditionallatent}$$ The resulting complete data log-likelihood is given by$$\ell(\lambda,z)=\sum_{1\leq i\neq j\leq K|w_{ij}>0}w_{ij}\log\lambda_{i}-\sum_{1\leq i<j\leq K|n_{ij}>0}(\lambda_{i}+\lambda_{j})z_{ij}+(n_{ij}-1)\log z_{ij}-\log\Gamma(n_{ij})\label{eq:completelikelihood}$$ where $\Gamma$ is the Gamma function. If we assign additionally a prior to $\lambda$ such that $$p\left( \lambda\right) ={\displaystyle\prod\limits_{i=1}^{K}} \mathcal{G}(\lambda_{i};a,b) \label{eq:prior}$$ as in [@Gormley2009; @Guiver2009] then we can maximize the resulting log-posterior using the EM algorithm which proceeds as follows at iteration $t$:$$\lambda^{(t)}=\underset{\lambda}{\arg\max}\text{ }Q(\lambda,\lambda^{\left( t-1\right) }). \label{eq:maxEM}$$ We have $$\begin{aligned} Q(\lambda,\lambda^{\ast}) & =\mathbb{E}_{\left. Z\right\vert D,\lambda ^{\ast}}\left[ \ell(\lambda,Z)\right] +\log\text{ }p\left( \lambda\right) \label{eq:QBT}\\ & \equiv\sum_{i=1}^{K}\left( a-1+w_{i}\right) \log\lambda_{i}-b\lambda _{i}-\sum_{1\leq i<j\leq K}(\lambda_{i}+\lambda_{j})\frac{n_{ij}}{\lambda _{i}^{\ast}+\lambda_{j}^{\ast}}\nonumber\end{aligned}$$ with $\equiv$ meaning equal up to terms independent of the first argument of the $Q$ function and $w_{i}=\sum_{j=1,j\neq i}^{K}w_{ij}$ is the number of wins of element $i$. Using (\[eq:maxEM\]), it follows that$$\lambda_{i}^{(t)}=\frac{a-1+w_{i}}{b+\sum_{j\neq i}\frac{n_{ij}}{\lambda _{i}^{(t-1)}+\lambda_{j}^{(t-1)}}}. \label{eq:BTEMupdate}$$ For $a=1$ and $b=0$, the MAP and ML estimates coincide. In this case (\[eq:QBT\]) is exactly the minorizing function of the MM algorithm proposed in [@Hunter2004 Eq. (10)] and thus the MM algorithm is given by (\[eq:BTEMupdate\]). Based on the same latent variables, we present a simple data augmentation algorithm for sampling from the posterior distribution $p\left( \left. \lambda,z\right\vert D\right) \propto p\left( \lambda\right) \exp\left( \ell(\lambda,z)\right) $. By construction, we can update $Z$ conditional upon $\lambda$ using (\[eq:conditionallatent\]) and the conditional for $\lambda$ given $Z$ can be expressed easily so that the data augmentation sampler at iteration $t$ proceeds as follows: - For $1\leq i<j\leq K$ s.t. $n_{ij}>0$, sample $$Z_{ij}^{(t)}|D,\lambda^{(t-1)}\sim\mathcal{G}(n_{ij},\lambda_{i}^{(t-1)}+\lambda_{j}^{(t-1)}).$$ - For $i=1,\ldots,K,$ sample $$\lambda_{i}^{(t)}|D,Z^{(t)}\sim\mathcal{G(}a+w_{i},b+\sum_{i<j|n_{ij}>0}Z_{ij}^{(t)}+\sum_{i>j|n_{ij}>0}Z_{ji}^{(t)}).$$ Generalized Bradley-Terry models\[sec:genbt\] ============================================= Home advantage -------------- Consider now that the pairwise comparisons are modeled using the Bradley-Terry model with home-field advantage" [@Agresti1990] where$$\Pr(i\text{ beats }j)=\left\{ \begin{array} [c]{ll}\frac{\theta\lambda_{i}}{\theta\lambda_{i}+\lambda_{j}} & \text{if }i\text{ is home,}\\ \frac{\lambda_{i}}{\lambda_{i}+\theta\lambda_{j}} & \text{if }j\text{ is home.}\end{array} \right.$$ The parameter $\theta$, $\theta>0$, measures the strength of the home-field advantage ($\theta>1$) or disadvantage ($\theta<1$). Let $a_{ij}$ be the number of times that $i$ is at home and beats $j$ and $b_{ij}$ is the number of times that $i$ is at home and loses to $j$. The log-likelihood of the skill ratings $\lambda$ and $\theta$ is given by$$\ell(\lambda,\theta)=c\log\theta+\sum_{i=1}^{K}w_{i}\log\lambda_{i}-\sum_{1\leq i\neq j\leq K}n_{ij}\log(\theta\lambda_{i}+\lambda_{j})$$ where $n_{ij}=a_{ij}+b_{ij}$ is the number of times $i$ plays at home against $j$, $c=\sum_{1\leq i\neq j\leq K}a_{ij}$ is the total number of home-field wins and $w_{i}$ is the total number of wins of element $i$. For $1\leq i\neq j\leq K$ such that $n_{ij}>0$, let us introduce the latent variables $Z=\left\{ Z_{ij}\right\} $ which are such that $$p\left( \left. z\right\vert D,\lambda\right) ={\displaystyle\prod\limits_{1\leq i\neq j\leq K|n_{ij}>0}} \mathcal{G}(z_{ij};n_{ij},\theta\lambda_{i}+\lambda_{j}).$$ The associated complete data log-likelihood is given by $$\ell(\lambda,\theta,z)=c\log\theta+\sum_{i=1}^{K}w_{i}\log\lambda_{i}-\sum_{1\leq i\neq j\leq K|n_{ij}>0}(\theta\lambda_{i}+\lambda_{j})z_{ij}+\log\Gamma\left( n_{ij}\right) .$$ Using independent priors for $\lambda$ and $\theta$, i.e. $p\left( \lambda,\theta\right) =p\left( \lambda\right) p\left( \theta\right) $, where $p\left( \lambda\right) $ is defined as (\[eq:prior\]) and $$\theta\sim\mathcal{G}(a_{\theta},b_{\theta}), \label{eq:priortheta}$$ then we can maximize the resulting posterior using the EM algorithm $$\left( \lambda^{(t)},\theta^{\left( t\right) }\right) =\text{ }\underset{\left( \lambda,\theta\right) }{\arg\max}\text{ }Q(\left( \lambda,\theta\right) ,\left( \lambda^{\left( t-1\right) },\theta^{\left( t-1\right) }\right) )$$ where we have $$\begin{aligned} Q(\left( \lambda,\theta\right) ,\left( \lambda^{\ast},\theta^{\ast}\right) ) & =\mathbb{E}_{Z|D,\lambda^{\ast},\theta^{\ast}}\left[ \ell (\lambda,\theta,Z)\right] +\log\text{ }p\left( \lambda,\theta\right) \\ & \equiv\left( a_{\theta}-1+c\right) \log\theta-b_{\theta}\theta+\sum _{i=1}^{K}\left( a-1+w_{i}\right) \log\lambda_{i}-b\sum_{i=1}^{K}\lambda _{i}-\sum_{1\leq i\neq j\leq K}n_{ij}\frac{\theta\lambda_{i}+\lambda_{j}}{\theta^{\ast}\lambda_{i}^{\ast}+\lambda_{j}^{\ast}}.\end{aligned}$$ We obtain $$\begin{aligned} \lambda_{i}^{(t)} & =\frac{a-1+w_{i}}{b+\sum_{1\leq i\neq j\leq K}\text{ }\left\{ \frac{\theta^{(t-1)}n_{ij}}{\theta^{(t-1)}\lambda_{i}^{(t-1)}+\lambda_{j}^{(t-1)}}+\frac{n_{ji}}{\theta^{(t-1)}\lambda_{j}^{(t-1)}+\lambda_{i}^{(t-1)}}\right\} }\text{ for }i=1,\ldots,K,\\ \theta^{(t)} & =\frac{a_{\theta}-1+c}{b_{\theta}+\sum_{1\leq i\neq j\leq K}\frac{n_{ij}\text{ }\lambda_{i}^{(t)}}{\theta^{(t-1)}\lambda_{i}^{(t-1)}+\lambda_{j}^{(t-1)}}}.\end{aligned}$$ For $a=a_{\theta}=1$ and $b=b_{\theta}=0$, i.e. if we use flat priors, this EM algorithm is similar to the MM algorithm proposed in [Hunter2004]{}. Using the same latent variables, we can sample from the posterior distribution of $\left( \lambda,\theta,Z\right) $ using the Gibbs sampler which updates iteratively $Z,$ $\lambda$ and $\theta$ as follows at iteration $t$: - For $1\leq i\neq j\leq K$ s.t. $n_{ij}>0$, sample $$Z_{ij}^{(t)}|D,\lambda^{(t-1)},\theta^{(t-1)}\sim\mathcal{G}\left( n_{ij},\theta^{(t-1)}\lambda_{i}^{(t-1)}+\lambda_{j}^{(t-1)}\right) .$$ - For $i=1,\ldots,K,$ sample $$\lambda_{i}^{(t)}|D,\theta^{(t-1)},Z^{(t)}\sim\mathcal{G}\left( a+w_{i},b+\theta^{\left( t-1\right) }\sum_{j\neq i|n_{ij}>0}Z_{ij}^{(t)}+\sum_{j\neq i|n_{ij}>0}Z_{ji}^{\left( t\right) }\right) .$$ - Sample $$\theta^{\left( t\right) }|D,\lambda^{(t)},Z^{(t)}\sim\mathcal{G}\left( a_{\theta}+c,b_{\theta}+\sum_{i=1}^{K}\lambda_{i}^{\left( t\right) }\sum_{j\neq i|n_{ij}>0}Z_{ij}^{\left( t\right) }\right) .$$ Model with ties\[sec:ties\] --------------------------- If we now want to allow for ties in pairwise comparisons, we can use the following model proposed by [@Rao1967] $$\begin{aligned} \Pr(i\text{ beats }j) & =\frac{\lambda_{i}}{\lambda_{i}+\theta\lambda_{j}},\\ \Pr(i\text{ ties }j) & =\frac{(\theta^{2}-1)\lambda_{i}\lambda_{j}}{(\lambda_{i}+\theta\lambda_{j})(\theta\lambda_{i}+\lambda_{j})}$$ where $\theta>1$. The log-likelihood function for $\left( \lambda ,\theta\right) $ is given by$$\begin{aligned} \ell(\lambda,\theta) & =\sum_{1\leq i\neq j\leq K}w_{ij}\log\frac {\lambda_{i}}{\lambda_{i}+\theta\lambda_{j}}+\frac{t_{ij}}{2}\log\frac {(\theta^{2}-1)\lambda_{i}\lambda_{j}}{(\theta\lambda_{i}+\lambda_{j})(\lambda_{i}+\theta\lambda_{j})}\\ & =\sum_{1\leq i\neq j\leq K}s_{ij}\log\frac{\lambda_{i}}{\lambda_{i}+\theta\lambda_{j}}+\frac{t_{ij}}{2}\log(\theta^{2}-1)\end{aligned}$$ where $t_{ij}=t_{ji}$ is the number of ties between $i$ and $j$ and $s_{ij}=w_{ij}+t_{ij}$. For $1\leq i\neq j\leq K$ such that $s_{ij}>0$, let us introduce the latent variables $Z=\left\{ Z_{ij}\right\} $ which are such that $$p\left( \left. z\right\vert D,\lambda\right) ={\displaystyle\prod\limits_{1\leq i\neq j\leq K|s_{ij}>0}} \mathcal{G}(z_{ij};s_{ij},\lambda_{i}+\theta\lambda_{j})$$ which yields the following complete log-likelihood$$\ell(\lambda,\theta,z)=T\log(\theta^{2}-1)+\sum_{1\leq i\neq j\leq K|s_{ij}>0}s_{ij}\log\lambda_{i}-(\lambda_{i}+\theta\lambda_{j})z_{ij}+(s_{ij}-1)\log z_{ij}-\log\Gamma(s_{ij})$$ where $T=\frac{1}{2}\sum_{1\leq i\neq j\leq K}t_{ij}$ is the total number of ties. If we adopt for $\theta$ a flat improper prior on $\left[ 1,\infty\right) $ and we select $p\left( \lambda\right) $ as (\[eq:prior\]) then we obtain$$\begin{aligned} Q(\left( \lambda,\theta\right) ,\left( \lambda^{\ast},\theta^{\ast}\right) ) & =\mathbb{E}_{Z|D,\lambda^{\ast},\theta^{\ast}}\left[ \ell (\lambda,\theta,Z)\right] +\log\text{ }p\left( \lambda,\theta\right) \\ & \equiv T\log(\theta^{2}-1)+\sum_{1\leq i\neq j\leq K}s_{ij}\left( \log\lambda_{i}-\frac{\lambda_{i}+\theta\lambda_{j}}{\lambda_{i}^{\ast}+\theta^{\ast}\lambda_{j}^{\ast}}\right) +\sum_{i=1}^{K}\left( a-1\right) \log\lambda_{i}-b\lambda_{i}$$ and we recover once again the minorizing function in [Hunter2004]{} for $a=1$ and $b=0$. This can be maximized using the following procedure - For $i=1,\ldots,K,$ set $$\lambda_{i}^{(t)}=\left( a-1+\sum_{j\neq i}s_{ij}\right) \left[ b+\sum_{j\neq i}\frac{s_{ij}}{\lambda_{i}^{(t-1)}+\theta^{(t-1)}\lambda _{j}^{(t-1)}}+\frac{\theta^{(t-1)}s_{ji}}{\theta^{(t-1)}\lambda_{i}^{(t-1)}+\lambda_{j}^{(t-1)}}\right] ^{-1}\text{.}$$ - Set $$\theta^{(t)}=\frac{1}{2c^{(t)}}+\sqrt{1+\frac{1}{4c^{(t)\text{ }2}}}$$ where $$c^{(t)}=\sum_{1\leq i\neq j\leq K}\frac{s_{ij}\lambda_{j}^{(t)}}{\lambda _{i}^{(t-1)}+\theta^{(t-1)}\lambda_{j}^{(t-1)}}\text{.}$$ Using the same latent variables, we can sample from the posterior distribution of $\left( \lambda,\theta,Z\right) $ using the following Gibbs sampler which updates iteratively $Z,$ $\lambda$ and $\theta$ as follows at iteration $t$: - For $1\leq i\neq j\leq K$ s.t. $s_{ij}>0$, sample $$Z_{ij}^{\left( t\right) }|D,\lambda^{(t-1)},\theta^{(t-1)}\sim \mathcal{G}\left( s_{ij},\lambda_{i}^{\left( t-1\right) }+\theta ^{(t-1)}\lambda_{j}^{\left( t-1\right) }\right) .$$ - For $i=1,\ldots,K$, sample$$\lambda_{i}^{(t)}|D,\theta^{(t-1)},Z^{(t)}\sim\mathcal{G}\left( a+\sum_{j\neq i}s_{ij},b+\sum_{j\neq i|s_{ij}>0}Z_{ij}^{\left( t\right) }+\theta ^{(t-1)}\sum_{j\neq i|s_{ij}>0}Z_{ji}^{\left( t\right) }\right) .$$ - Sample $$\theta^{\left( t\right) }|D,\lambda^{(t)},Z^{(t)}\sim p(\theta |D,\lambda^{(t)},Z^{\left( t\right) })$$ where$$p(\theta|D,Z,\lambda)\propto(\theta^{2}-1)^{T}\exp\left( -\sum_{1\leq i\neq j\leq K|s_{ij}>0}Z_{ij}\text{ \ }\theta\right) 1_{\theta>1}. \label{eq:conditionalnonstandardtheta}$$ It is possible to sample from (\[eq:conditionalnonstandardtheta\]) exactly. By performing a change of variable $\overline{\theta}=\theta-1$, we obtain $$p(\overline{\theta}|D,Z,\lambda)\propto(\overline{\theta}^{2}+2\overline {\theta})^{T}\exp\left( -\sum_{1\leq i\neq j\leq K|s_{ij}>0}Z_{ij}\text{ \ }\overline{\theta}\right)$$ which is a mixture of Gamma distributions. Group comparisons ----------------- Consider now that we have $n$ pairwise comparisons betweem teams. For each comparison $i=1,\ldots,n$, let $T_{i}^{+}\subset\{1,\ldots,K\}$ be the winning team and $T_{i}^{-}\subset\{1,\ldots,K\}$ the losing team where $T_{i}^{+}\cap T_{i}^{-}=\varnothing$ and $T_{i}=T_{i}^{+}\cup T_{i}^{-}$. Recently [@Huang2006] have proposed the following model$$\Pr(T_{i}^{+}\text{ beats }T_{i}^{-})=\frac{\sum_{j\in T_{i}^{+}}\lambda_{j}}{\sum_{j\in T_{i}}\lambda_{j}}\text{.}$$ The log-likelihood function for $\lambda$ is thus given by $$\ell(\lambda)=\sum_{i=1}^{n}\log\left( \sum_{j\in T_{i}^{+}}\lambda _{j}\right) -\log\left( \sum_{j\in T_{i}}\lambda_{j}\right) \text{.}$$ For $i=1,...,n$ we introduce the latent variables $Z=\left\{ Z_{i}\right\} $ and $C=\left\{ C_{i}\right\} $ such that $$p\left( \left. z,c\right\vert D,\lambda\right) =p\left( \left. z\right\vert D,\lambda\right) P\left( \left. c\right\vert D,\lambda\right)$$ with$$\begin{aligned} p\left( \left. z\right\vert D,\lambda\right) & ={\displaystyle\prod\limits_{i=1}^{n}} \mathcal{E}\left( z_{i};\sum_{j\in T_{i}}\lambda_{j}\right) ,\\ \Pr\left( \left. c\right\vert D,\lambda\right) & ={\displaystyle\prod\limits_{i=1}^{n}} \frac{\lambda_{c_{i}}}{\sum_{j\in T_{i}^{+}}\lambda_{j}}\text{ with }c_{i}\in T_{i}^{+}$$ where $\mathcal{E}\left( x;\alpha\right) $ is the exponential density of argument $x$ and parameter $\alpha$. It follows that the complete log-likelihood is given by$$\ell(\lambda,z,c)=\sum_{i=1}^{n}\log\lambda_{c_{i}}-\left( \sum_{j\in T_{i}}\lambda_{j}\right) z_{i}.$$ The $Q$ function associated to the EM algorithm is given by $$\begin{aligned} Q(\lambda,\lambda^{\ast}) & =\mathbb{E}_{Z,C|D,\lambda^{\ast}}\left[ \ell(\lambda,Z,C)\right] +\log\text{ }p\left( \lambda\right) \\ & \equiv\sum_{i=1}^{n}\sum_{j\in T_{i}^{+}}\log\lambda_{j}\frac{\lambda _{j}^{\ast}}{\sum_{k\in T_{i}^{+}}\lambda_{k}^{\ast}}-\frac{\sum_{j\in T_{i}}\lambda_{j}}{\sum_{j\in T_{i}}\lambda_{j}^{\ast}}+\sum_{k=1}^{K}\left( a-1\right) \log\lambda_{k}-b\lambda_{k}\\ & \equiv\sum_{k=1}^{K}\left( a-1+\lambda_{k}^{\ast}\sum_{i=1}^{n}\frac{\alpha_{ik}}{\sum_{j\in T_{i}^{+}}\lambda_{j}^{\ast}}\right) \log\lambda_{k}-\lambda_{k}\left( b+\sum_{i=1}^{n}\frac{\gamma_{ik}}{\sum_{j\in T_{i}}\lambda_{j}^{\ast}}\right)\end{aligned}$$ where $\alpha_{ik}=1$ if $k\in T_{i}^{+}$ and $0$ otherwise and $\gamma _{ik}=1$ if $k\in T_{i}$ and $0$ otherwise. It follows that the EM update is given by$$\lambda_{k}^{\left( t\right) }=\frac{a-1+\lambda_{k}^{\ast}\sum_{i=1}^{n}\frac{\alpha_{ik}}{\sum_{j\in T_{i}^{+}}\lambda_{j}^{\left( t-1\right) }}}{b+\sum_{i=1}^{n}\frac{\gamma_{ik}}{\sum_{j\in T_{i}}\lambda_{j}^{\left( t-1\right) }}}.$$ Using the same latent variables, we obtain a data augmentation sampler to sample from $p\left( \left. \lambda,z,c\right\vert D\right) $ by iteratively sampling $\left( Z,C\right) $ and $\lambda.$ This proceeds as follows at iteration $t$: - For $i=1,...,n$, sample $$\begin{tabular} [c]{l}$Z_{i}^{\left( t\right) }|D,\lambda^{(t-1)}\sim\mathcal{E}\left( \sum_{j\in T_{i}}\lambda_{j}^{\left( t-1\right) }\right) ,$\\ $\Pr\left( C_{i}^{\left( t\right) }=k|D,\lambda^{(t-1)}\right) =\frac{\lambda_{k}^{\left( t-1\right) }}{\sum_{j\in T_{i}^{+}}\lambda _{j}^{\left( t-1\right) }},$ $k\in T_{i}^{+}.$\end{tabular} \ $$ - For $k=1,\ldots,K$, sample$$\lambda_{k}^{(t)}|D,Z^{(t)},C^{(t)}\sim\mathcal{G}\left( a+\sum_{i=1}^{n}\delta_{k,C_{i}^{\left( t\right) }},b+\sum_{i=1}^{n}\gamma_{ik}Z_{i}^{\left( t\right) }\right)$$ where $\delta_{u,v}=1$ if $u=v$ and $0$ otherwise. Multiple comparisons\[sec:plackettluce\] ======================================== We now consider the popular Plackett-Luce model [@Luce1959; @Plackett1975] which is an extension of the Bradley-Terry model to comparisons involving more than two elements. Assume that $p_{i}\leq K$ individuals are ranked for comparison $i$ where $i=1,...,n$. We write $\rho_{i}=(\rho_{i1},\ldots ,\rho_{ip_{i}})$ where $\rho_{i1}$ is the first individual, $\rho_{i2}$, the second, etc. The Plackett-Luce model assumes$$\Pr(\rho_{i}|\lambda)=\prod_{j=1}^{p_{i}}\frac{\lambda_{\rho_{ij}}}{\sum _{k=j}^{p_{i}}\lambda_{\rho_{ik}}}=\prod_{j=1}^{p_{i}-1}\frac{\lambda _{\rho_{ij}}}{\sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}}}. \label{eq:plackettluce}$$ For $i=1,\ldots,n$ and $j=1,\ldots,p_{i}-1$, we introduce the following latent variables $Z=\left\{ Z_{ij}\right\} $ $$p\left( \left. z\right\vert D,\lambda\right) =\prod_{i=1}^{n}\prod _{j=1}^{p_{i}-1}\mathcal{E(}z_{ij};\sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}})$$ which leads to the complete log-likelihood $$\ell(\lambda,z)=\sum_{i=1}^{n}\sum_{j=1}^{p_{i}-1}\text{ }\log\lambda _{\rho_{ij}}-\left( \sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}}\right) z_{ij}.$$ The $Q$ function associated to the EM algorithm is given by $$\begin{aligned} Q(\lambda,\lambda^{\ast}) & =\mathbb{E}_{Z|D,\lambda^{\ast}}\left[ \ell(\lambda,Z)\right] +\log\text{ }p\left( \lambda\right) \\ & \equiv\sum_{i=1}^{n}\sum_{j=1}^{p_{i}-1}\log\lambda_{\rho_{ij}}-\frac {\sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}}}{\sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}}^{\ast}}+\sum_{k=1}^{K}\text{ }\left( a-1\right) \log\lambda_{k}-b\lambda_{k}$$ which is once again equivalent to the majorizing function in [@Hunter2004 pp. 398] for $a=1,$ $b=0$. It follows that the EM algorithm is given at iteration $t$ by $$\lambda_{k}^{(t)}=(a-1+w_{k})\left[ b+\sum_{i=1}^{n}\left( \sum_{j=1}^{p_{i}-1}\frac{\delta_{ijk}}{\sum_{k=j}^{p_i}\lambda_{\rho_{ik}}^{(t-1)}}\right) \right] ^{-1}$$ where $w_{k}$ is the number of rankings where the $k^{\text{th}}$ individual is not in the last ranking position and $\delta_{ijk}$ is defined as$$\delta_{ijk}=\left\{ \begin{array} [c]{ll}1 & \text{if }k\in\{\rho_{ij},\ldots,\rho_{ip_{j}}\}\\ 0 & \text{otherwise}\end{array} \right.$$ i.e. $\delta_{ijk}$ is the indicator of the event that individual $k$ receives a rank no better than $j$ in the $i^{\text{th}}$ ranking. To sample from $p\left( \left. \lambda,z\right\vert D\right) $, we can use the following data augmentation sampler. At iteration $t$, this proceeds as follows: - For $i=1,...,n$, for $j=1,\ldots,p_{i}-1$, sample$$Z_{ij}^{\left( t\right) }|D,\lambda^{\left( t-1\right) }\sim \mathcal{E}(\sum_{k=j}^{p_{i}}\lambda_{\rho_{ik}}^{\left( t-1\right) }).$$ - For $k=1,\ldots,K$, sample $$\lambda_{k}^{\left( t\right) }|D,Z^{\left( t\right) }\sim\mathcal{G}(a+w_{k},b+\sum_{i=1}^{n}\sum_{j=1}^{p_{i}-1}\delta_{ijk}Z_{ij}^{\left( t\right) }).$$ Using exactly the same augmentation, EM and Gibbs samplers can be defined for further extensions of these models such as mixtures of Plackett-Luce models [@Gormley2008a]. Discussion\[sec:discussion\] ============================ Identifiability --------------- Consider the basic Bradley-Terry model and its extensions to group comparisons and multiple comparisons. Let us define $$\Lambda=\sum_{i=1}^{K}\lambda_{i}\text{ \ and }\pi_{i}=\frac{\lambda_{i}}{\Lambda}$$ and write $\pi:=\left\{ \pi_{i}\right\} _{i=1}^{K}$. The likelihood is invariant to a rescaling of the vector $\lambda$ so the parameter $\Lambda$ is not likelihood-identifiable and $$p(\pi,\Lambda|D)=p(\left. \pi\right\vert D)p(\Lambda).$$ From (\[eq:prior\]), it follows that $\pi\sim\mathcal{D}(a,\ldots,a)$ where $\mathcal{D}$ is the Dirichlet distribution and $\Lambda\sim\mathcal{G}(Ka,b)$, hence $$\Lambda^{MAP}=\frac{aK-1}{b}.$$ To improve the mixing of the MCMC algorithms in this context, an additional sampling step can be added where we normalize the current parameter estimate $\lambda^{\left( t\right) }$ and then rescale them randomly using a prior draw for $\Lambda$. - For $i=1,\ldots,K$, set $\lambda_{i}^{\ast(t)}=\frac{\lambda_{i}^{(t)}}{\sum_{j=1}^{K}\lambda_{j}^{(t)}}\Lambda^{(t)}$ where $\Lambda^{(t)}\sim\mathcal{G}(Ka,b)$. This step can drastically improve the mixing of the Markov chain. However, if we are only interested in the normalized values $\pi$ of $\lambda$ then this additional step is useless. As an alternative, it is also possible to consider an EM algorithm for the basic Bradley-Terry model which does not require the introduction of a scale parameter. Assume $\pi\sim\mathcal{D}(a,\ldots,a)$ and let us introduce latent variables $M_{ij}$, $C_{ij}=(C_{ij1},\ldots,C_{ijM_{ij}})$ for $1\leq i\neq j\leq K$ such that $n_{ij}>0$ $$\begin{aligned} M_{ij} & \sim NB(n_{ij},\pi_{i}+\pi_{j}),\\ \Pr\left( C_{ijk}=l\right) & =\frac{\pi_{l}}{\sum_{n\neq i,j}\pi_{n}}\text{ for }k=1,\ldots M_{ij}\text{ and }l\neq i,j\end{aligned}$$ where $NB(r,p)$ is the negative binomial distribution. The complete log-likelihood is given by $$\ell(\pi,m,c)=\sum_{i=1}^{K}\sum_{j=1,j\neq i}^{K}w_{ij}\log\pi_{i}+\sum _{i=1}^{K}\sum_{j=1,j\neq i}^{K}\sum_{k\neq i,j}\left[ \log\binom {n_{ij}+m_{ij}-1}{n_{ij}-1}+r_{ijk}\log\pi_{k}\right]$$ where $r_{ijk}$ is the number of $c_{ijl},l=1,\ldots,m_{ij}$ that take value $k$. Omitting the terms independent of $\pi$, the $Q$ function is given by$$\begin{aligned} Q(\pi,\pi^{\ast}) & =\mathbb{E}_{M|D,\pi^{\ast}}\left[ \mathbb{E}_{C|D,M,\pi^{\ast}}\left[ \ell(\pi,M,C)\right] \right] +\log\text{ }p\left( \pi\right) \\ & \equiv\mathbb{E}_{M|D,\pi^{\ast}}\left[ \sum_{i=1}^{K}\sum_{j=1,j\neq i}^{K}w_{ij}\log\pi_{i}+\sum_{i=1}^{K}\sum_{j=1,j\neq i}^{K}\sum_{k\neq i,j}\log\binom{n_{ij}+M_{ij}-1}{n_{ij}-1}+M_{ij}\frac{\pi_{k}^{\ast}}{\sum_{l\neq i,j}\pi_{l}^{\ast}}\log\pi_{k}\right] \\ &+\log\text{ }p\left( \pi\right) \\ & \equiv\sum_{i=1}^{K}\sum_{j=1,j\neq i}^{K}w_{ij}\log\pi_{i}+\sum_{i=1}^{K}\sum_{j=1,j\neq i}^{K}\sum_{k\neq i,j}n_{ij}\frac{\pi_{k}^{\ast}}{\pi _{i}^{\ast}+\pi_{j}^{\ast}}\log\pi_{k}+\left( a-1\right) \sum_{i=1}^{K}\log\pi_{k}+C\\ & \equiv\sum_{k=1}^{K}(w_{k}+\pi_{k}^{\ast}\sum_{i=1,i\neq k}^{K}\sum_{j=1,j\neq i,k}^{K}\frac{n_{ij}}{\pi_{i}^{\ast}+\pi_{j}^{\ast}}\log \pi_{k})+\left( a-1\right) \sum_{i=1}^{K}\log\pi_{k}+C\end{aligned}$$ where $C$ is a term independent of $\pi$. It follows that the EM update is given by$$\pi_{k}^{(t)}\propto a-1+w_{k}+\pi_{k}^{(t-1)}\sum_{i=1,i\neq k}^{K}\sum_{j=1,j\neq i,k}^{K}\frac{n_{ij}}{\pi_{i}^{(t-1)}+\pi_{j}^{(t-1)}}$$ with $\sum_{k=1}^{K}\pi_{k}^{\left( t\right) }=1.$ Although the above EM algorithm does not rely on unidentifiable scaling parameters, it suffers from a slow convergence rate. When $\pi_{k}$ takes small values, $\sum_{i\neq k}\sum_{j\neq k}\frac{n_{ij}}{\pi_{i}+\pi_{j}}$ is large and it slows down the convergence of the EM algorithm. The same augmentation can be used to define a Gibbs sampler, but the same slow mixing issues arise for the Markov chain. Hyperparameter estimation ------------------------- The prior (\[eq:prior\]) is specified by the hyperparameters $a$ and $b$. However, the inverse scale parameter $b$ is not likelihood identifiable so there is no point assigning a prior to it. However it might be interesting to set a prior $p(a)$ on $a$ and estimate it from the data. Given $\lambda$, we have $$p\left( \left. a\right\vert \lambda\right) \propto p\left( a\right) \underset{l_{1}\left( a\right) }{\underbrace{\left( b^{K}\prod_{i=1}^{K}\lambda_{i}\right) ^{a}\text{\ }}}\text{ }\underset{l_{2}(a)}{\underbrace{\Gamma^{-K}(a)}}.$$ It is possible to sample from this density using auxiliary variables $U_{1},U_{2}$ defined on $\left( 0,\infty\right) $ as described in [@Damien1999]. We introduce $$p\left( \left. a,u_{1},u_{2}\right\vert \lambda\right) \propto p\left( a\right) \mathbb{I}\left\{ u_{1}<l_{1}\left( a\right) \right\} \mathbb{I}\left\{ u_{2}<l_{2}\left( a\right) \right\} .$$ A Gibbs sampler can now be implemented to sample from $p\left( \left. a,u_{1},u_{2}\right\vert \lambda\right) $. We can directly sample from the full conditionals of $U_{1}$ and $U_{2}$ $$\left. U_{1}\right\vert \lambda\sim\mathcal{U}\left( 0,l_{1}\left( a\right) \right) ,\text{ \ }\left. U_{2}\right\vert \lambda\sim \mathcal{U}\left( 0,l_{2}\left( a\right) \right)$$ where $\mathcal{U}\left( \alpha,\beta\right) $ is the uniform distribution on $\left( \alpha,\beta\right) $. The full conditional of $a$ given $u_{1},u_{2}$ is given by $$p\left( \left. a\right\vert \lambda,u_{1},u_{2}\right) \propto p\left( a\right) \mathbb{I}_{A_{1}\cap A_{2}}\left( a\right)$$ where $$A_{i}=\left\{ a:l_{i}\left( a\right) >u_{i}\right\} .$$ Alternatively we can update $a$ using a M-H random walk on $\log(a)$. We can propose $a^{\star}=\exp(\sigma_{a}^{2}z)a$ where $z\sim\mathcal{N}(0,1)$ and $a^{\star}$ is accepted with probability $$\min\left\{ 1,\frac{p(a^{\star})}{p(a)}\left( \frac{\Gamma(a)}{\Gamma(a^{\star})}\right) ^{K}\left( b^{K}\prod_{i=1}^{K}\lambda _{i}\right) ^{a^{\star}-a}\right\} .$$ Further extensions ------------------ ### Random graphs with a given degree sequence A model closely related to Bradley-Terry has been proposed for undirected random graphs with $K$ vertices [@Holland1981; @Chatterjee2010; @Park2004]. In this model, the degree sequence $(d_{1},\ldots,d_{K})$ of a given graph, where $d_{i}$ is the degree of node $i$, is supposed to capture all the information in the graph. It can be formalized by saying that the degree sequence is a sufficient statistic for a probability distribution on graphs [@Chatterjee2010]. In this model an edge is inserted between vertices $i$ and $j$ for $1\leq i<j\leq K$ with probability$$\frac{\lambda_{i}\lambda_{j}}{1+\lambda_{i}\lambda_{j}}$$ where $\lambda_{k}>0$ for $k\in\{1,\ldots,K\}$. Let $r_{ij}=1$ if there is an edge between $i$ and $j$ and $0$ otherwise. Given the observations $D=\left\{ r_{ij}\right\} _{1\leq i<j\leq K}$, the log-likelihood function for $\lambda$ is given by $$\ell(\lambda)=\sum_{1\leq i<j\leq K}r_{ij}\log\left( \lambda_{i}\lambda _{j}\right) -\log\left( 1+\lambda_{i}\lambda_{j}\right) .$$ We introduce the following latent variables $Z=\left\{ Z_{ij}\right\} _{1\leq i<j\leq K}$ such that$$p\left( \left. z\right\vert D,\lambda\right) ={\displaystyle\prod\limits_{1\leq i<j\leq K}} \mathcal{E}(z_{ij};\lambda_{i}+\frac{1}{\lambda_{j}})\text{.}$$ The complete log-likelihood is given by$$\ell(\lambda,z)=\sum_{1\leq i<j\leq K}r_{ij}\log\lambda_{i}-(1-r_{ij})\log\lambda_{j}-(\lambda_{i}+\frac{1}{\lambda_{j}})z_{ij}$$ The $Q$ function associated to the EM algorithm is given by$$\begin{aligned} Q(\lambda,\lambda^{\ast}) & =\mathbb{E}_{Z|D,\lambda^{\ast}}\left[ \ell(\lambda,Z)\right] +\log\text{ }p\left( \lambda\right) \\ & \equiv\sum_{i=1}^{K}\log\lambda_{i}\left[ \left( a-1\right) +\sum _{j>i}r_{ij}-\sum_{j<i}(1-r_{ij})\right] -\lambda_{i}\left( b+\sum _{j>i}\frac{1}{\lambda_{i}^{\ast}+\frac{1}{\lambda_{j}^{\ast}}}\right) -\frac{1}{\lambda_{i}}\sum_{j<i}\frac{1}{\lambda_{j}^{\ast}+\frac{1}{\lambda_{i}^{\ast}}}.\end{aligned}$$ Solving $\partial Q(\lambda,\lambda^{\ast})/\partial\lambda_{i}=0$ requires solving a quadratic equation. For sake of brevity, we do not present these details here. Once again, we can define a data augmentation sampler to sample from $p\left( \left. \lambda,z\right\vert D\right) $ by iteratively sampling $Z$ and $\lambda.$ This proceeds as follows at iteration $t$: - For $1\leq i<j\leq K$, sample$$\left. Z_{ij}^{\left( t\right) }\right\vert D,\lambda^{(t-1)}\sim\mathcal{E(}\lambda_{i}^{\left( t-1\right) }+\frac{1}{\lambda _{j}^{\left( t-1\right) }}).$$ - For $i=1,...,K$, sample$$\lambda_{i}^{\left( t\right) }|D,Z^{\left( t\right) }\sim\mathcal{GIG}\left( 2(\sum_{j>i}Z_{ij}^{\left( t\right) }+b),2\sum_{j<i}Z_{ij}^{\left( t\right) },a+\sum_{j>i}r_{ij}-\sum_{j<i}(1-r_{ij})\right) .$$ Here $\mathcal{GIG}\left( \alpha,\beta,\gamma\right) $ denotes the generalized inverse Gaussian distribution (see e.g. [@Barndorff-Nielsen2001]) whose density for an argument $x$ is proportional to $$x^{\gamma-1}\exp\left\{ -\left( \alpha x+\beta/x\right) /2\right\} .$$ Algorithms to sample exactly from this distribution are available. ### Choice models Other extensions of the Bradley-Terry model are the choice models introduced by [@Restle1961] and [@Tversky1972; @Tversky1972a] in psychology; see also [@Wickelmaier2004; @Gorur2006]. In these models, we are given a set of $n$ elements. To each element $i$ is associated a set of $K$ features represented by a binary vector $f_{i}\in\{0,1\}^{K}$. The probability that element $i$ is chosen over element $j$ is given by$$\pi_{ij}=\frac{\sum_{k=1}^{K}\lambda_{k}f_{ik}(1-f_{jk})}{\sum_{k=1}^{K}\lambda_{k}f_{ik}(1-f_{jk})+\sum_{k=1}^{K}\lambda_{k}f_{jk}(1-f_{ik})}$$ where $\lambda_{k}>0$ is a weight representing the importance of feature $k$. The term $\sum_{k=1}^{K}\lambda_{k}f_{ik}(1-f_{jk})$ corresponds to the sum of the weights of features possessed by object $i$ but not object $j$. EM and Gibbs algorithms can be derived by following the same construction as with group comparisons. ### Classification model Let consider the following original model for categorical data analysis$$\begin{aligned} \Pr(Y_{i}=k) & =\frac{\sum_{j=1}^{p}\exp(X_{ij})\lambda_{kj}}{\sum_{l=1}^{K}\sum_{j=1}^{p}\exp(X_{ij})\lambda_{lj}}\nonumber\\ & =\frac{\exp(X_{i})^{T}\lambda_{k}}{\sum_{l=1}^{K}\exp(X_{i})^{T}\lambda _{l}}\label{eq:classif}$$ where $X_{i}\in\mathbb{R}^{p}$ is a vector of covariates and $\lambda_{k}\in\mathbb{R}_{+}^{p}$ for $k=1,\ldots,K$. This model could be used as an alternative to the multinomial logit model [@Agresti1990]. By introducing latent variables $Z_{i}\sim\mathcal{E}\left( \sum_{l=1}^{K}\exp(X_{i})^{T}\lambda_{l}\right) $, we can define EM and Gibbs algorithms resp. to maximize the posterior distribution of the parameters $\lambda_{k}$ and sample from it when the prior is given by (\[eq:prior\]). Experimental results\[sec:experiments\] ======================================= In all the above models, the parameter $b$ is just a scaling parameter on $\lambda_{k}$. As the likelihood is invariant to a rescaling of the $\lambda_{k}$, this parameter does not have any influence on inference. Hence to ensure that the MAP estimate satisfies $\sum_{k=1}^{K}\widehat{\lambda}_{k}=1$, we set $b=Ka-1$ henceforth as explained in section \[sec:discussion\]. We demonstrate our algorithms on one synthetic and two real-world data sets. Synthetic Data -------------- We first study the Plackett-Luce model, comparing experimentally the mixing properties of the Gibbs sampler relative to a slightly modified version of the M-H algorithm proposed by [@Gormley2009]. In this latter paper, the authors propose to update the skill parameters simultaneously using the following proposal distribution[^1] at iteration $t$ $$\text{for }i=1,\ldots,K\text{, }\lambda_{i}^{\star}\sim\mathcal{G}\left( a+w_{k},b+\sum_{i=1}^{n}\left( \sum_{j=1}^{p_{i}-1}\frac{\delta_{ijk}}{\sum_{k=j}^{p}\lambda_{\rho_{ik}}^{(t-1)}}\right) \right)$$ We simulated $500$ dataset of $n$ rankings of $K=4$ individuals, for various values of $n$ with $a=5$. For each dataset, 10,000 iterations of the Gibbs sampler presented in section \[sec:plackettluce\] were run. The sample lag-1 autocorrelation was then computed for the four skill parameters. For a given sample size $n$, the mean value over skill parameters and simulated data is reported on Figure \[fig:ACFsimu\] together with 90% confidence bounds. The algorithm of [@Gormley2009] performs reasonably well when the sample size is large, which is the case for the voting data they considered, but poorly for small sample sizes. ![Sample lag-1 autocorrelation as a function of the sample size $n$ for the Gibbs sampler and a modified version of the M-H algorithm of Gormley and Murphy (2009).[]{data-label="fig:ACFsimu"}](ACFsimu){width="8cm"} Nascar 2002 dataset ------------------- NASCAR is the primary sanctioning body for stock car auto racing in the United States. Each race involves 43 drivers. During the 2002 season, 87 different drivers participated in 36 races. Some drivers participated in all of the races while others participated in only one. We propose to apply the Plackett-Luce model with gamma prior on the parameters. The NASCAR dataset[^2] has been studied by [@Hunter2004], who noted that the MLE cannot be found for the original data set as four drivers placed last in each race they entered, and therefore had to be removed. This does not need to be done if we follow a Bayesian approach. We focus here on predicting the outcome of the next race based on the previous ones, starting from race 5; i.e. we predict the results of race 6 based on the MAP estimates obtained with the first 5 races, then the results of race 7 based on the MAP estimated obtained with the first 6 races, etc. For each race, we compute the test log-likelihood using the MAP estimates. The mean value and 90% confidence bounds are represented in Figure \[fig:nascar2002\] w.r.t. the value of $a$. The EM algorithm was initialized using $(\lambda_{1}^{\left( 0\right) },\ldots,\lambda_{83}^{\left( 0\right) })=(\frac{1}{83},\ldots,\frac{1}{83})$. ![Test log-likelihood on the Nascar 2002 dataset. From race 5 to 35, we compute the log-likelihood of the next race based on the MAP estimates. Mean and 90% interval of the log-likelihood is represented w.r.t. to the parameter $a$. The straight line represents the test log-likelihood obtained with a uniform prior.[]{data-label="fig:nascar2002"}](nascar2002pred){width="8cm"} The Gibbs sampler was also applied to the same dataset. The skill parameters were initialized at the same value, and the parameter $a$ was assigned a flat improper prior and sampled as described in section \[sec:discussion\]. We ran 50,000 iterations with 2,000 burn-in. As detailed in Section \[sec:discussion\], only the normalized weights $\pi_{i}$ are likelihood identifiable. Skill ratings are usually represented on the real line, and we use the following one-to-one mapping $\beta_{i}=\log\pi_{i}-\log1/83$. The marginal posterior densities of the reparameterized skill ratings for the first four drivers according to their average place are reported in Figure \[fig:postnascar\]. The Bayesian approach can effectively capture the uncertainty in the skill ratings of the drivers. ML and MMSE (minimum mean squared error) estimates together with standard deviations are reported in Table \[tab:nascar\] for the first ten and last ten drivers according to average place. \[c\][cccccc]{}& & Average & MLE & MMSE & Standard\ Driver & Races & place & estimate & estimate & Deviation\ P. Jones & 1 & 4.00 & 2.74 & 0.11 & 0.48\ S. Pruett & 1 & 6.00 & 2.21 & 0.10 & 0.48\ M. Martin & 36 & 12.17 & 0.67 & 0.79 & 0.17\ T. Stewart & 36 & 12.61 & 0.42 & 0.60 & 0.17\ R. Wallace & 36 & 13.17 & 0.65 & 0.78 & 0.17\ J. Johnson & 36 & 13.50 & 0.53 & 0.68 & 0.17\ S. Marlin & 29 & 13.86 & 0.33 & 0.49 & 0.19\ M. Bliss & 1 & 14.00 & 0.82 & 0.04 & 0.48\ J. Gordon & 36 & 14.06 & 0.33 & 0.53 & 0.17\ K. Busch & 36 & 14.06 & 0.24 & 0.46 & 0.17\ ... & & & & &\ C. Long & 2 & 40.50 & -1.73 & -0.67 & 0.46\ C. Fittipaldi & 1 & 41.00 & -1.85 & -0.51 & 0.50\ H. Fukuyama & 2 & 41.00 & -2.17 & -0.81 & 0.50\ J. Small & 1 & 41.00 & -1.94 & -0.60 & 0.51\ M. Shepherd & 5 & 41.20 & -1.86 & -1.05 & 0.39\ K. Shelmerdine & 2 & 41.50 & -1.73 & -0.72 & 0.46\ A. Cameron & 1 & 42.00 & -1.41 & -0.44 & 0.49\ D. Marcis & 1 & 42.00 & -1.38 & -0.43 & 0.49\ D. Trickle & 3 & 42.00 & -1.72 & -0.87 & 0.42\ J. Varde & 1 & 42.00 & -1.55 & -0.48 & 0.50\ ![Marginal posterior distribution for the modified skill ratings $\beta_{i}$ of the first 4 drivers according to their average place. P. Jones and S. Pruett only participated in 1 race each, while M. Martin and T. Stewart participated in 36 races.[]{data-label="fig:postnascar"}](margpost){width="8cm"} Chess data ---------- Rating the skills of chess players is of major practical interest. It allows organizers of a tournament to avoid having strong players playing against each other at early stages, or to restrict the tournament to players with skills above a given threshold. The international chess federation adopted the so-called Elo system which is based on the Bradley-Terry model [@Elo1978]. For historical considerations about the rating system in chess, the reader should refer to [@Glickman1995]. We consider here game-by-game chess results over 100 months, consisting of 65,053 matches between 8631 players[^3]. The outcome of each game is either win (+1), tie (+0.5) or loss (0). We estimate the parameters of the Bradley-Terry model with ties presented in section \[sec:ties\] on the first 95 months and then predict the outcome of the games of the last 5 months. The hyperparameters for the tie parameter $\theta$ are set to $a_{\theta}=1,$ $b_{\theta}=0$. Given the large sample size, it is not possible to sample from Eq. as the number of elements in the mixture is very large. We therefore use a M-H step with a normal random walk proposal of standard deviation $0.1$. The mean squared error is reported for predictions based on MAP estimates and full Bayesian predictive based on the Gibbs sampler outcomes, for different values of the hyperparameter $a$. EM and Gibbs samplers were initialized at $(\lambda_{1}^{\left( 0\right) },\ldots,\lambda_{8631}^{\left( 0\right) })=(\frac{1}{8631},\ldots,\frac {1}{8631})$ and $\theta^{\left( 0\right) }=1,5$. The Gibbs samplers were run with 10,000 iterations and 1,000 burn-in iterations. The results are reported in Figure \[fig:chesstest\]. The results demonstrate the benefit of penalizing the skill rating parameters and the improvement brought up by a full Bayesian analysis. In Figure \[fig:ACF\] we also report the autocorrelation function associated to the parameter $\theta$ and the skill parameters with largest mean values. The Markov chain displays good mixing properties. ![Test mean square error on the chess dataset for different values of the parameter $a$. Based on an history of 95 months, we predict the outcome of the games of the last 5 months. []{data-label="fig:chesstest"}](chesstest){width="8cm"} Conclusion ========== The Bradley-Terry model and its generalizations arise in numerous applications. We have shown here that most of the MM algorithms proposed in [@Hunter2004] can be reinterpreted as special cases of EM algorithms. Additionally we have proposed original EM algorithms for some recent generalizations of the Bradley-Terry models. Finally we have shown how the latent variables introduced to derive these EM algorithms lead straightforwardly to Gibbs sampling algorithms. These elegant MCMC algorithms mix experimentally well and outperform a recently proposed M-H algorithm. [Acknowledgment]{}. The authors are grateful to Persi Diaconis for helpful discussions and pointers to references on the Plackett-Luce and random graph models and to Luke Bornn for helpful comments. [^1]: The authors actually use a normal approximation of the gamma distribution, and work with normalized data. To obtain similar algorithms, we consider unnormalized data. [^2]: The data can be downloaded from [http://www.stat.psu.edu/ dhunter/code/btmatlab/]{} [^3]: Chess data can be downloaded from http://kaggle.com/chess

ArXiv

--- abstract: 'We present a detailed study of the spin-torque diode effect in CoFeB/MgO/CoFe/NiFe magnetic tunnel junctions. From the evolution of the resonance frequency with magnetic field at different angles, we clearly identify the free-layer mode and find an excellent agreement with simulations by taking into account several terms for magnetic anisotropy. Moreover, we demonstrate the large contribution of the out-of-plane torque in our junctions with asymmetric electrodes compared to the in-plane torque. Consequently, we provide a way to enhance the sensitivity of these devices for the detection of microwave frequency.' author: - 'R. Matsumoto' - 'A. Chanthbouala' - 'J. Grollier' - 'V. Cros' - 'A. Fert' - 'K. Nishimura' - 'Y. Nagamine' - 'H. Maehara' - 'K. Tsunekawa' - 'A. Fukushima' - 'S. Yuasa' title: 'Spin-Torque Diode Measurements of MgO-Based Magnetic Tunnel Junctions with Asymmetric Electrodes' --- Spin-transfer torque (STT) in MgO-based magnetic tunnel junctions (MTJs)[@Parkin1; @Yuasa1] is under development for device applications such as STT random access memories (STT-RAM)[@Hosomi; @Yuasa2], domain-wall-motion MRAM,[@FukamiNEC] racetrack memory[@Parkin2] and spintronic memristors.[@SeagateIEEE; @Andre] Recently, spin-torque diodes[@Tulapurkar] have attracted much attention because their sensitivity for the detection of microwave frequency may exceed that of semiconductor diodes.[@WangJAP; @Ishibashi] In the spin diode effect, an applied rf current to the MTJ exerts an oscillating spin torque on the magnetization of the free layer, leading to excitation of the ferromagnetic resonance (FMR) mode. The dynamics of the free layer cause oscillations of the tunnel magnetoresistance (TMR). As a result, the oscillating resistance partially rectifies the rf current and dc voltage is obtained (*V*$_{\rm diode}$). The spin diode effect depends on the relative amplitudes (*a$_{J}$* and *b$_{J}$*) of the classical in-plane torque (T$_{\rm IP}$) and the out-of-plane field-like torque (T$_{\rm OOP}$).[@Sankey; @Kubota] Here, the torques are expressed as T$_{\rm IP}$ = -$\gamma($*a$_{J}$*/*M*$_{\rm s}$)**m**$\times$(**m**$\times$**M$_{\rm ref}$**) and T$_{\rm OOP}$ = -$\gamma$*b$_{J}$***m**$\times$**M$_{\rm ref}$** with **m** (*M*$_{\rm s}$) being the magnetization vector (the saturation magnetization) of the free layer, **M$_{\rm ref}$** the magnetization vector of the reference layer, and $\gamma$ the gyromagnetic ratio. In spin diode spectra (*V*$_{\rm diode}$ as a function of frequency (*f*)), the contribution of T$_{\rm IP}$ (resp. T$_{\rm OOP}$) results in a peak with a Lorentzian component (resp. an anti-Lorentzian component); $$\begin{aligned} \mathbf{\textit{V}_{\rm diode}} & = & \frac{A \left(f_1^2- f^2\right) + B f^2}{\left(f_1^2- f^2\right)^2 + \left(\Delta f \right)^2}, \label{Vdiode} \\ A &\propto& \gamma^2 H_{\rm d} \frac{\partial b_J}{\partial I} {\rm TMR \ sin^{2}}\theta, \label{A} \\ B &\propto& \gamma \Delta \frac{\partial a_J}{\partial I} {\rm TMR \ sin^{2}}\theta,\label{B} \end{aligned}$$ where *A* and *B* are the amplitudes of the anti-Lorentzian component and Lorentzian component, *f*$_{\rm 1}$ is the resonance frequency, $\Delta$ is the peak linewidth, *H*$_{\rm d}$ is the out-of-plane demagnetization field, and $\theta$ is the relative angle between the free layer and the reference layer. Experimentally evaluated T$_{\rm OOP}$ was reported to reach over 25$\%$ of T$_{\rm IP}$ in conventional CoFeB/MgO/CoFeB MTJs with symmetric electrodes.[@Sankey; @Kubota] However, in these MTJs, the dc bias dependence of T$_{\rm OOP}$ is quadratic and symmetric with respect to the polarity of bias, leading to *A* = 0 at zero dc bias voltage. For MTJs with asymmetric electrodes, on the other hand, the bias dependence of T$_{\rm OOP}$ is expected to be asymmetric and linear at low bias,[@Xiao; @Oh] leading to larger *A* at zero dc bias voltage. In this study, we perform spin diode measurements of MgO-based MTJs with asymmetric electrodes. We also measure its dependence on magnitude and angle of the in-plane external magnetic field (*H*$_{\rm ext}$) to identify the free-layer excitation modes. Thin films of MTJs are deposited with a magnetron sputtering system (Canon ANELVA C-7100). The stacking structure (see Fig. 1(a)) is IrMn(7)/Co$_{70}$Fe$_{30}$(2.5)/Ru(0.9)/Co$_{60}$Fe$_{20}$B$_{20}$(3)/MgO tunnel barrier(1.1)/Co$_{70}$Fe$_{30}$(1)/Ni$_{83}$Fe$_{17}$(4)/capping layers (thickness in nm) deposited on thermally-oxidized Si substrate/buffer layers. Annealing treatment in a high-vacuum furnace at 330 $^{\circ}$C for 2 hours is then made under a 1 T magnetic field. These MTJ films are micro-processed into nano-pillars with an elliptic junction area of 70 ${\times}$ 270 nm$^{2}$. All the measurements presented here were done at room temperature. We first present in Fig. 1(b) the resistance versus magnetic field (*R(H)*) curves obtained for the in-plane *H*$_{\rm ext}$ applied along 0$^{\circ}$ and 90$^{\circ}$ (definition of the angle for *H*$_{\rm ext}$ ($\phi$$_{\rm H}$) is schematically shown in an inset of Fig. 1(b)). The TMR ratio is 67.7$\%$ and the resistance is 162 $\Omega$ in the parallel magnetic state. The angular spin diode measurements presented in the following are performed with in-plane *H*$_{\rm ext}$ at various $\phi$$_{\rm H}$. The injected microwave power is kept constant at -15 dBm, and no dc current is applied. ![(a) Sketch of the MgO-based magnetic tunnel junction (MTJ) stack. (b) Resistance versus magnetic field curves obtained for the in-plane magnetic field applied along 0$^{\circ}$(dotted line) and 90$^{\circ}$ (solid line). Inset in (b) is schematic explaining angle ($\phi_{\rm H}$) of in-plane magnetic field (*H*$_{\rm ext}$) and magnetization direction of reference layer (*M*$_{\rm ref}$). (c) Spin-torque diode signal (*V*$_{\rm diode}$) versus frequency measured under *H*$_{\rm ext}$ = +20 mT with [*$\phi_{\rm H}$*]{} = 0$^{\circ}$ with fitted curve.[]{data-label="f1"}](fig1.eps){width="7.5"} As an example of spin diode measurements, in Fig. 1(c) we show the spectrum obtained at *H*$_{\rm ext}$ = +20 mT with [*$\phi_{\rm H}$*]{} = 0$^{\circ}$. In the MTJs of this study, the spin diode spectrum typically has two peaks that we label Mode 1 and Mode 2. As expected from Eq. (\[A\]), in the case of the MTJs with asymmetric electrodes, the peaks have significant anti-Lorentzian component. The peaks are fitted well by Eq. (\[Vdiode\]) (see Fig. 1(c)). First, we focus on the property of Mode 1. The resonance frequency of Mode 1 versus *H*$_{\rm ext}$ (*f*$_{\rm 1}$(*H*)) at different angles is shown in Figs. 2(a), (c), and (e). For higher ${|}$*H*$_{\rm ext}$${|}$ ${>}$ 500 Oe, *f*$_{\rm 1}$(*H*) increases with increasing *H*$_{\rm ext}$ following Kittel formula.[@Kittel] Fitting *f*$_{\rm 1}$(*H*) for $\phi$$_{\rm H}$ = 0$^{\circ}$ with the Kittel formula gives *H*$_{\rm d}$ = 1.1 T. For small *H*$_{\rm ext}$ ${<}$ 500 Oe, *f*$_{\rm 1}$(*H*) exhibits an asymmetric behavior with respect to the polarity of *H*$_{\rm ext}$, especially for *H*$_{\rm ext}$ with $\phi$$_{\rm H}$ = 0$^{\circ}$ (Fig. 2(a)). This characteristic cannot be explained by macro-spin model taking into account only in-plane shape-anisotropy field (*H*$_{\rm S.A.}$). The in-plane crystalline-anisotropy field (*H*$_{\rm C.A.}$) due to Co$_{70}$Fe$_{30}$ in the free layer can be responsible for this asymmetric *f*$_{\rm 1}$(*H*). The CoFe layer is amorphous in the as-grown state but it crystallizes in a cubic-crystal texture by post-annealing[@Yuasa2] then leading to cubic anisotropy. It should be noted that its in-plane crystalline orientation can be arbitrary. ![Resonance frequency of free-layer mode versus external magnetic field (*H*$_{\rm ext}$) with various angles (f) of (a)-(b) 0$^{\circ}$, (c)-(d) 30$^{\circ}$, and (e)-(f) 50$^{\circ}$°. \[(a), (c), (e)\] Experimental results. Insets in (a), (b) and (c) are schematic explaining angle ($\phi$$_{\rm H}$) of in-plane magnetic field (*H*$_{\rm ext}$) and magnetization direction of reference layer (*M*$_{\rm ref}$). \[(b), (d), (f)\] Simulation results with consistent parameters: in-plane shape-anisotropy field (*H*$_{\rm S.A.}$) of 20 mT, in-plane crystalline-anisotropy field (*H*$_{\rm C.A.}$) of 10 mT, angle of crystalline-anisotropy field in CoFe layer ($\beta$) of 63$^{\circ}$, and *H*$_{\rm d}$ = 1.1 T.[]{data-label="f2"}](fig2.eps){width="7.5"} To understand this asymmetric *f*$_{\rm 1}$(*H*), we analytically simulate the resonance frequency versus *H*$_{\rm ext}$ of the free layer (*f*$_{\rm free}$(*H*)) with macro-spin model taking into account not only *H*$_{\rm S.A.}$ but also fourfold *H*$_{\rm C.A.}$[@Goryunov] due to Co$_{70}$Fe$_{30}$ having an arbitrary crystalline orientation. First, the equilibrium angle of free-layer magnetization ($\phi$) under *H*$_{\rm ext}$ with $\phi$$_{\rm H}$ is given by Eq. (\[Eq.Cond\]); $$\label{Eq.Cond} \textit{H}_{\rm ext}{\rm sin}(\phi-\phi_{H})=\frac{1}{2}H_{\rm S.A.}{\rm sin}2\phi - \frac{1}{4}H_{\rm C.A.}{\rm cos}4(\phi-\beta).$$ Here, $\beta$ is the angle of crystalline easy axis in CoFe layer, defined with respect to the minor axis of the patterned ellipse. Then, *f*$_{\rm free}$(*H*) is given by Eq.(\[freq\]); $$\begin{aligned} \label{freq} \textit{f} & = & \frac{\gamma}{2\pi} \left[ H_{\rm d} \{H_{\rm ext} {\rm cos}(\phi-\phi_{H})\right. \nonumber\\ & & \left. -H_{\rm S.A.}{\rm cos}2\phi + H_{\rm C.A.}{\rm cos}4(\phi-\beta) \}\right]^\frac{1}{2}.\end{aligned}$$ The simulation results are shown in Figs. 2 (b), (d), and (f). Here, we fix all parameters except for *H*$_{\rm ext}$ and $\phi$$_{\rm H}$ : *H*$_{\rm S.A.}$ = 20 mT, *H*$_{\rm C.A.}$ = 10 mT, $\beta$ = 63$^{\circ}$, and *H*$_{\rm d}$ = 1.1 T. The qualitative correspondence between the analytically simulated *f*$_{\rm free}$(*H*) and experimental *f*$_{\rm 1}$(*H*) indicates that Mode 1 corresponds to the free-layer mode, and that the contribution of *H*$_{\rm C.A.}$ is not negligible. We also measure *f*$_{\rm 1}$(*H*) of other samples on the same lot. Although each sample exhibits the asymmetric *f*$_{\rm 1}$(*H*) with different characteristic at $\phi$$_{\rm H}$ = 0$^{\circ}$, each *f*$_{\rm 1}$(*H*) can be qualitatively explained by Eqs. (3) and (4) with various $\beta$ (0$^{\circ}$ - 90$^{\circ}$), similar values of *H*$_{\rm S.A.}$ (20 - 40 mT) and *H*$_{\rm C.A.}$ (15 - 20 mT), and the same *H*$_{\rm d}$ = 1.1 T. This result also supports the measurable contribution of *H*$_{\rm C.A.}$ To further obtain quantitative agreement between the simulated *f*$_{\rm free}$(*H*) and experimental *f*$_{\rm 1}$(*H*) especially in the low-field range, we need to take into account the coupling with Mode 2. However, the origin of Mode 2 is still under debate while such higher-order modes are often observed in MgO-based MTJs. Depending on the authors, the origin is attributed to edge mode, higher-order spin wave mode, or mode of CoFe/Ru/CoFeB synthetic antiferromagnetic layers.[@WangJAP; @Helmer] To discuss the origin of Mode 2 is beyond the scope of this letter. ![Magnetic field dependence of (a) fitting parameter *A* and (b) fitting parameter *B* of Eq. (\[Vdiode\]). Insets in (a) are schematic explaining angle ($\phi_{\rm H}$) of in-plane magnetic field (*H*$_{\rm ext}$) and magnetization directions of free layer (*m*$_{\rm free}$) and reference layer (*M*$_{\rm ref}$).[]{data-label="f3"}](fig3.eps){width="7.5"} The fitting parameters *A* (anti-Lorentzian component) and *B* (Lorentzian component) of Eq. (2) versus *H*$_{\rm ext}$ with $\phi$$_{\rm H}$ = 0$^{\circ}$ that maximizes the spin diode effect are shown in Figs. 3 (a) and (b), respectively. First, the amplitude of the parameter *A* is one order of magnitude larger than that of parameter *B*. Indeed, the parameter *A* is proportional to large $\gamma$*H*$_{\rm d}$ whereas the parameter *B* is proportional to small $\Delta$ $\approx$ $\gamma$$\alpha$*H*$_{\rm d}$, where $\alpha$ is the Gilbert damping.[@Kubota; @Sankey; @WangPRB]. The linear bias dependence of T$_{\rm OOP}$ in the MTJs with asymmetric electrodes might open a way to enhance the diode sensitivity (*V*$_{\rm diode}$ divided by injected rf power). At *H*$_{\rm ext}$ = 20 - 40 mT, the relative angle between the free layer and the reference layer ($\theta$) reaches its maximum. Then, it leads to maximum value of the parameter *A* and *B* for the chosen $\phi$$_{\rm H}$ = 0$^{\circ}$ because T$_{\rm IP}$ and T$_{\rm OOP}$ are proportional to sin $\theta$ (see Eqs. (\[A\]) and (\[B\])). We also check the diode sensitivity with a similar sample on the same lot. Here, we carefully measure the bias dc dependence of resistance and diode effect,[@Andre] and calibrate the impedance mismatch of the sample using the measured bias dependence of the resistance and backgrounds of the spin diode spectra.[@WangPRB] At *H*$_{\rm ext}$ = 42 mT, we obtain a diode sensitivity of 100 mV/mW. This sensitivity is competitive compared to those of past studies although the TMR ratio in our study is about half.[@WangJAP; @Ishibashi] For higher ${|}$*H*$_{\rm ext}$${|}$ ${>}$ 500 Oe, parameter *A* changes its sign depending on the polarity of *H*$_{\rm ext}$ while the sign of parameter *B* is independent of the polarity of *H*$_{\rm ext}$. This result agrees with the vectorial expression of the spin torques because the vector product in T$_{\rm OOP}$ changes polarity depending on the polarity of *H*$_{\rm ext}$ as schematically shown in insets of Fig. 3, while the vector product in T$_{\rm IP}$ does not. Above-mentioned characteristics of *H*$_{\rm ext}$ dependence of parameters *A* and *B* also support that Mode 1 corresponds to the free-layer mode. In summary, spin diode measurements are performed in CoFeB/MgO/CoFe/NiFe MTJs having asymmetric electrodes without dc bias current. Their spin diode spectra exhibit peaks with strong anti-Lorentzian components originating from T$_{\rm OOP}$ in the asymmetric MTJs. Because of the significant contribution of T$_{\rm OOP}$, under an in-plane magnetic field of 42 mT, we obtain as large diode sensitivity as 100 mV/mW (after impedance matching correction). We also perform the spin diode measurements under various *H*$_{\rm ext}$ with $\phi$$_{\rm H}$. By comparing experimental *f*$_{\rm 1}$(*H*) with the simulation results of *f*$_{\rm free}$(*H*), we can identify the free-layer excitation modes where the contribution of *H*$_{\rm C.A.}$ is observed to be measurable as well as *H*$_{\rm S.A.}$ This work was supported by the European Research Council (ERC Stg 2010 No. 259068) and JSPS Postdoctoral Fellowships for Research Abroad. [20]{} S. S. P. Parkin, C. Kaiser, A. Panchula, P.M. Rice, B. Hughes, M. Samant, S.-H. Yang: Nat. Mat. **3** (2004) 862. S. Yuasa, T. Nagahama, A. Fukushima, Y. Suzuki, K. Ando, Nat. Mater. **3** (2004) 868. M. Hosomi, H. Yamagishi, T. Yamamoto, K. Bessho, Y. Higo, K. Yamane, H. Yamada, M. Shoji, H. Hachino, C. Fukumoto, H. Nagao, and H. Kano: IEDM Tech. Dig., 2005, p. 459. S. Yuasa, D. D. Djayaprawira: J. Phys. D: Appl. Phys. **40** (2007) R337. S. Fukami, T. Suzuki, K. Nagahara, N. Ohshima, Y. Ozaki, S. Saito, R. Nebashi, N. Sakimura, H. Honjo, K. Mori, C. Igarashi, S. Miura, N. Ishiwata, and T. Sugibayashi, Dig. Tech. Pap. - Symp. VLSI Technol., 2009, p. 230. S. S. P.Parkin, M. Hayashi, and L. Thomas: Science **320** (2008) 190. X. Wang, Y. Chen, H. Xi, H. Li, and D. Dimitrov: IEEE Electron Device Letters, **30** (2009) 294. A. Chanthbouala, R. Matsumoto, J. Grollier, V. Cros, A. Anane, A. Fert, A. V. Khvalkovskiy, K. A. Zvezdin, K. Nishimura, Y. Nagamine, H. Maehara, K. Tsunekawa, A. Fukushima, and S. Yuasa: arXiv:1102.2106v1. A. A. Tulapurkar, Y. Suzuki, A. Fukushima, H. Kubota, H. Maehara, K. Tsunekawa, D. D. Djaraprawira, N. Watanabe, and S. Yuasa: Nature **438** (2005) 339. C. Wang, Y.-T. Cui, J. Z. Sun, J. A. Katine, R. A. Buhrman, and D. C. Ralph: J. Appl. Phys. **106** (2009) 053905. S. Ishibashi, T. Seki, T. Nozaki, H. Kubota, S. Yakata, A. Fukushima, S. Yuasa, H. Maehara, K. Tsunekawa, D. D. Djayaprawira, and Y. Suzuki: Appl. Phys. Express **3** (2010) 073001. H. Kubota, A. Fukushima, K. Yakushiji, T. Nagahama, S. Yuasa, K. Ando, H. Maehara, Y. Nagamine, K. Tsunekawa, D.D. Djayaprawira, N. Watanabe, Y. Suzuki: Nat. Phys. **4** (2008) 37. J. C. Sankey, Y.-T. Cui, J.Z. Sun, J.C. Slonczewski, R.A. Buhrman, D.C. Ralph: Nat. Phys. **4** (2008) 67. C. Wang, Y.-T. Cui, J. Z. Sun, J. A. Katine, R. A. Buhrman, and D. C. Ralph: Phys. Rev. B **79** (2009) 224416. J. Xiao, G. E. W. Bauer, and A. Brataas: Phys. Rev. B **77** (2008) 224419. S. -C. Oh, S. -Y. Park, A. Manchon, M. Chshiev, J. -H. Han, H. -W. Lee, J. -E. Lee, K. -T. Nam, Y. Jo, Y.-C. Kong, B. Dieny, and K.-J. Lee: Nat. Phys. **5** (2009) 898. A. Helmer, S. Cornelissen, T. Devolder, J.-V. Kim, W. van Roy, L. Lagae, and C. Chappert: Phys. Rev. B **81** (2010) 094416. C. Kittel, *Introduction to Solid State Physics*, 7th ed. (Wiley, New York, 1996), p. 505. Yu. V. Goryunov, N. N. Garif’yanov, G. G. Khaliullin, I. A. Garifullin, L. R. Tagirov, F. Schreiber, Th. Mühge, and H. Zabel: Phys. Rev. B **52** (1995) 13450.

ArXiv

--- author: - | Jinming Duan, Ghalib Bello, Jo Schlemper, Wenjia Bai, Timothy J W Dawes, Carlo Biffi,\ Antonio de Marvao, Georgia Doumou, Declan P O’Regan, Daniel Rueckert, [^1] bibliography: - 'ref.bib' title: 'Automatic 3D bi-ventricular segmentation of cardiac images by a shape-refined multi-task deep learning approach' --- [^1]: JD, JS, WB, CB and DR are with the Biomedical Image Analysis Group, Imperial College London, London, UK. JD, GB, TJWD, A de M, GD and DPO’R are with MRC London Institute of Medical Sciences, Imperial College London, London, UK. TJWD is with the National Heart and Lung Institute, Imperial College London, London, UK.

ArXiv

--- abstract: 'We consider the possibility of looking for CP-mixing effects in two-Higgs doublet models (and particularly in the MSSM) by studying the lineshape of the CP-even ($H$) and CP-odd ($A$) neutral scalars. In most cases $H$ and $A$ come quite degenerate in mass, and their $s$-channel production would lead to nearly overlapping resonances. CP-violating effects may connect these two Higgs bosons, giving origin to one-loop particle mixing, which, due to their mass proximity, can be resonantly enhanced. The corresponding transition amplitude contains then CP-even and CP-odd components; besides the signal of intereference between both amplitudes, leading to a CP-odd asymmetry, we propose to look for the mixing probability itself, a quantity which, although CP-even, can originate only from a CP-odd amplitude. We show that, in general, the effect of such a mixing probability cannot be mimicked by (or be re-absorbed into) a simple redefinition of the $H$ and $A$ masses in the context of a CP-conserving model. Specifically, the effects of the CP-mixing are such that, either the mass-splitting of the $H$ and $A$ bosons cannot be accounted for in the absence of CP-mixing, and/or the detailed energy dependence of the produced lineshape is clearly different from the one obtained by redefining the masses, but not allowing any mixing. This analysis suggests that the detailed study of the lineshape of this Higgs system may provide valuable information on the CP nature of the underlying theory.' author: - 'J. Bernabéu$^a$, D. Binosi$^{b}$ and J. Papavassiliou$^a$' title: 'CP violation through particle mixing and the $H$-$A$ lineshape' --- FTUV-0603-30\ ECT\*-06-03\ Introduction ============ Despite the impressive success of the Standard Model (SM) in describing a plethora of high-precision experimental data [@PDB], a number of important theoretical issues related to its structure remain largely unexplored. Perhaps the most prominent among them is the nature of the very mechanism which endows the elementary particles with their observed masses. The best established way for introducing masses at tree-level, without compromising renormalizability, is the Higgs mechanism [@Higgs:1964ia; @Englert:1964et; @Guralnik:1964eu], where the crucial ingredient is the coupling of all would-be massive particles to a complex scalar field with a non-vanishing vacuum expectation value. The most characteristic physical remnant of this procedure is a massive scalar particle in the spectrum of the theory, the Higgs boson. In addition to the minimal SM, the Higgs mechanism is employed in many popular new physics scenaria, most notably in the Minimal Supersymmetric Standard Model (MSSM) and its variants [@Nilles:1983ge], leading to yet richer scalar sectors. However, to date, neither the Higgs boson of the minimal SM, nor the additional scalars predicted by its supersymmetric extensions have been observed, and their discovery and subsequent detailed study is considered as one of the main priorities for the upcoming collider experiments. Several of the aforementioned scalars, and especially the “standard” Higgs boson, are expected to be discovered at the LHC [@LHC]. In addition, in the past few years a significant amount of technical and theoretical research has been invested in the evaluation of the feasibility and physics potential of muon colliders [@Cline:1994tk; @Marciano:1998ue]. Such machines will have the particularly appealing feature of variable centre-of-mass energy, which will allow the resonant enhancement of $s$-channel interactions in order to produce Higgs bosons copiously. Thus, muon colliders are expected to operate as Higgs factories, in the energy range of up to .5 TeV. Given the additional features of small energy spread and precise energy determination, these machines offer a unique possibility for studying the line-shape of the SM Higgs boson, and determining its mass and width with an accuracy of about .1 MeV and .5 MeV, respectively. In addition, if supersymmetric particles will be discovered at the LHC, then the muon colliders would play the role of a precision machine for studying their main properties, such as masses, widths, and couplings, and delineating their line-shapes [@Barger:1996jm; @Binoth:1997ev; @Krawczyk:1997be; @Grzadkowski:2000fg; @Blochinger:2002hj]. In the two-Higgs doublet models in general [@Lee:1973iz; @Branco:1985aq; @Liu:1987ng; @Weinberg:1990me; @Wu:1994ja], and in most SUSY scenarios in particular [@Nilles:1983ge], the extended scalar sector contains five physical fields: a couple of charged Higgs bosons ($H^\pm$), a CP-odd scalar $A$, and two CP-even scalars $h$ (the lightest, which is to be identified with the SM Higgs) and $H$ (the heaviest). The detailed study of the $H$-$A$ system is particularly interesting because in most [*beyond the SM*]{} scenarios (such as SUSY) the masses of these two particles are almost degenerate, and therefore their resonant production is expected to give rise to nearly overlapping resonances [@Barger:1996jm]. Specifically, the tree-level mass eigenvalues of the Higgs mass matrices are given by $$\begin{aligned} m^2_{H^\pm}&=&m^2_A+M^2_W, \nonumber \\ m^2_{h,H}&=&\frac12\left[M^2_Z+m^2_A\mp\sqrt{(M^2_Z+m^2_A)^2-4m^2_AM^2_Z\cos^22\beta}\right],\end{aligned}$$ and therefore one will have the tree-level bounds $M_Z\in(m_h,M_H)$ and $M_W<m_{H^\pm}$. Furthermore, in the decoupling limit $M_A\gg M_Z$, the neutral sector (tree-level) masses satisfy the relations  [@declim] $$\begin{aligned} m^2_h&\approx&M^2_Z\cos^22\beta,\nonumber \\ m^2_H&\approx& m^2_A + M_Z^2 \sin^22\beta, \label{deg}\end{aligned}$$ which, for $\tan\beta\ge2$ (and thus $\cos^22\beta\approx1$), imply the degeneracies $m_h\approx M_Z$ and $m_H\approx m_A$. In general, the inclussion of radiative corrections is known to modify significantly the above tree-level relations, mainly due to the large Yukawa coupling of the top-quark. Such corrections are particularly important for the lower bound on the mass of the lightest Higgs boson $h$: for the largest bulk of the parameter space, $h$ is in fact heavier than the $Z$ boson, and can be as heavy as 130 GeV [@Mh]. It is important to emphasize, however, that the inclusion of radiative corrections does [*not*]{} lift the mass degeneracy in the $H$-$A$ system under consideration, especially in the parameter space region where $m_A>2M_Z$ and $\tan\beta\ge2$  [@KKRW]. In particular, as we will see in detail below, quantum effects do shift the values of $m_H$ and $m_A$, but by amounts that are numerically comparable; as a result, the small mass splitting between $H$ and $A$ remains still valid (but is not described quantitatively by Eq. (\[deg\])). In addition, as was shown by Pilaftsis [@Pilaftsis:1997dr], the near degeneracy of the $H$-$A$ system may give rise to a resonant enhancement of CP violation due to particle mixing. Specifically, if the CP symmetry is exact, the $H$ cannot mix with $A$, at any given order. However, in the presence of a CP-violating interactions [@Bernabeu:1987gr; @Pilaftsis:1992st; @Korner:1992zk; @Bernabeu:1993up; @Ilakovac:1993pt; @Bernabeu:1995ph; @Tommasini:1995ii; @Pilaftsis:1998dd], in addition to a variety of interesting implications [@Pilaftsis:1998pe], the $H$ can mix with the $A$ already at one-loop level, giving rise to a non-vanishing mixing self-energy $\Pi_{HA}(s)$. Such mixing, in turn, can be measured through the study of appropriate CP-odd observables, such as left-right asymmetries. As has been explained in [@Pilaftsis:1997dr], in general the CP-violating amplitude is particularly enhanced near resonace, if the two mixing particles are nearly degenerate, a condition which is naturally fullfilled in the $H$-$A$ system. Furthermore, the mixing between $H$ and $A$ has an additional profound effect for the two masses: the near degeneracy of the two particles is lifted, and the pole masses move further apart [@Carena:2001fw]; as a result, the originally overlapping resonances of the CP-invariant theory tend to be separated. (300,340)(90,-170) ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*The mass spectrum of the $H-A$ system in the MSSM with and without CP-breaking effects. Upper panels: the variation of $m_A$ and $m_H$ as a function of $\tan\beta$, for different values of $\phi$ and $|\mu|$; Lower panels: the variation of $\delta m=m_H-m_A$ as a function of $\phi$, for different values of $\tan\beta$ and $|\mu|$ (spikes appearing in the lower panels are code artifacts and should be ignored).*[]{data-label="spectrum"}](m_vs_tb-1-1.eps "fig:"){width="7.5cm"} ![*The mass spectrum of the $H-A$ system in the MSSM with and without CP-breaking effects. Upper panels: the variation of $m_A$ and $m_H$ as a function of $\tan\beta$, for different values of $\phi$ and $|\mu|$; Lower panels: the variation of $\delta m=m_H-m_A$ as a function of $\phi$, for different values of $\tan\beta$ and $|\mu|$ (spikes appearing in the lower panels are code artifacts and should be ignored).*[]{data-label="spectrum"}](m_vs_tb-1.5-1.eps "fig:"){width="7.5cm"} ![*The mass spectrum of the $H-A$ system in the MSSM with and without CP-breaking effects. Upper panels: the variation of $m_A$ and $m_H$ as a function of $\tan\beta$, for different values of $\phi$ and $|\mu|$; Lower panels: the variation of $\delta m=m_H-m_A$ as a function of $\phi$, for different values of $\tan\beta$ and $|\mu|$ (spikes appearing in the lower panels are code artifacts and should be ignored).*[]{data-label="spectrum"}](dm_vs_phi-1-1.eps "fig:"){width="7.4cm"} ![*The mass spectrum of the $H-A$ system in the MSSM with and without CP-breaking effects. Upper panels: the variation of $m_A$ and $m_H$ as a function of $\tan\beta$, for different values of $\phi$ and $|\mu|$; Lower panels: the variation of $\delta m=m_H-m_A$ as a function of $\phi$, for different values of $\tan\beta$ and $|\mu|$ (spikes appearing in the lower panels are code artifacts and should be ignored).*[]{data-label="spectrum"}](dm_vs_phi-1.5-1.eps "fig:"){width="7.4cm"} ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- (-305,155) (-10.7,5.15)(-10.1,5.15) (-280,145) (-10.7,4.9)(-10.1,4.9) (-10.41,4.9)(-10.39,4.9) (-280,137.5) (-10.7,4.65)(-10.1,4.65) (-280,130) (-415,127) (-415,60) (-465,85) (-350,15) (-86,58) (-3.0,1.75)(-2.4,1.75) (-63,48) (-3.0,1.5)(-2.4,1.5) (-2.71,1.5)(-2.69,1.5) (-63,40.5) (-3.0,1.25)(-2.4,1.25) (-63,33) (-180,147) (-180,80) (-230,85) (-115,15) (-287.5,-135) (-385,-25) (-405,-50) (-405,-65) (-405,-79) (-407,-89) (-407,-103) (-465,-100) (-347,-168) (-60,-135) (-165,-25) (-202,-35) (-202,-55) (-202,-70) (-202,-89) (-202,-100) (-230,-100) (-112,-168) Using the code `CPsuperH` [@Lee:2003nt], one can study the $H-A$ system mass spectrum in the CP-invariant ($\phi=0$) and CP-breaking ($\phi\ne0$) limits of the MSSM. For our purposes we have assumed (motivated by SUGRA models) universal squarks soft masses $M_0$ (with $M_0=.5$ TeV), trilinear couplings $A$ (with $\vert A\vert=1$ TeV) and CP breaking phases $\phi_A=\phi$, with the MSSM $\mu$ parameter real and varying between 1.0 and 1.5 TeV. Finally we set $m_{H^\pm}=.4571$ TeV which sets the tree-level mass of the CP-odd Higgs at the value of .450 TeV. In the first two panels of Fig. \[spectrum\], we plot the neutral CP-even and CP-odd scalar masses as a function of $\tan\beta$ for three values of the CP-breaking phase and two different values of $\mu$ ($\vert\mu\vert=1$ TeV in the left panel, while $\vert\mu\vert=1.5$ TeV in the right one). When the CP-breaking phases are absent (black continuous curves), the $H$-$A$ resonance can be clearly seen at $\tan\beta=5$ and $\tan\beta=4$ respectively. The degeneracy is lifted when the phases are switched on (blue dashed-dotted and red dahsed curves, corresponding to $\phi=30^\circ$ and $\phi=90^\circ$ degrees, respectively), their effect becoming rapidly smaller for larger $\tan\beta$ values: from $\tan\beta\gsim 20$ they cannot be disentangled from radiative effects calculated in the CP-invariant limit of the theory. A more in depth view of how the CP-breaking phases affect the $H$-$A$ mass spectrum can be achieved by studying the mass difference $\delta m= m_H-m_A$ versus $\phi$ for different values of $\tan\beta$, as it is shown in the two lower panels of Fig.\[spectrum\] (as before, $\vert\mu\vert=1$ TeV in the left panel, while $\vert\mu\vert=1.5$ TeV in the right one). In both cases we see that when $\tan\beta\gsim20$ (orange dotted lines) the difference between the CP-invariant and CP-breaking case is barely visibile. There are some differences, however, depending on the value of the $\mu$ parameter. For $|\mu|=1$ TeV, the mass difference shows a maximum around $\phi=100^\circ$; moreover in the CP-invariant limit the mass splitting is bigger for lower values of $\tan\beta$. When $|\mu|=1.5$ TeV, instead, $\delta m$ has no maximum and in the CP-invariant limit the splitting is bigger for larger $\tan\beta$ values. The above analysis suggests that, when studying the lineshape of the $H$-$A$ system, one may envisage two, physically very different, scenarios. In the first one, the CP symmetry is exact, with the position of resonances determined by Eq.(\[deg\]) plus its radiative corrections; the relative position between the two resonance will then specify $\tan\beta$. In the second scenario, CP is violated, resulting in mixing at one-loop level between CP-even and CP-odd states wich translates into a non-vanishing off-shell transition amplitude $\Pi_{HA}(s)$. Then, for the same mass splitting $\delta m$ as in the previous case, one may not reach the same conclusion for the value of $\tan\beta$, because one could have started out with the two masses almost degenerate, corresponding to a different value for $\tan\beta$ (lower or higher depending on the value of the $\mu$ parameter), and the observed separation between $m_{H}$ and $m_{A}$ may be due to the lifting of the degeneracy produced by the presence of the aforementioned $\Pi_{HA}(s)$. To address these questions in detail, we then turn to the MSSM, in which the one-loop mixing between the CP-even and CP-odd Higgs will be due to third-family squarks circulating in the loop of Fig.\[feyndiag\], and study the $H$-$A$ lineshape, which will be derived explicitly at one-loop level for general values of $\Pi_{HH}(s)$, $\Pi_{AA}(s)$, and $\Pi_{HA}(s)$. Our main results may be then summarized as follows. In the cases for which the CP-breaking phases are sizeable (we will set $\phi=90^\circ$ for this case), one can always clearly distinguish between the CP-invariant and CP-beaking scenarios, either because the mass splitting in the latter case is just too big for being due to CP-invariant radiative corrections only, or because of the different energy dependence of the cross section. For smaller values of the CP-breaking phases (say $\phi=30^\circ$), one can still distinguish between the CP-invariant and CP-breaking case only when $\tan\beta$ is small; when $\tan\beta\gsim10$ one cannot tell the two cases apart simply by studying the lineshape. We therefore conclude that the experimental determination of the Higgs line-shape, in conjunction with CP-odd asymmetries and other suitable observables, may provide valuable information for settling the issue regarding CP-mixing effects in two-Higgs models. The line-shape in the presence of H-A mixing ============================================ In this section we will concentrate on the calculation of the line-shape for the resonant process $\mu^+\mu^-\to A^*,\ H^*\to f\bar f$ in the presence of the CP-violating one-loop mixing diagrams of Fig.1 (a) between the CP-even ($H$) and the CP-odd ($A$) Higgses. ![(a) [*The one-loop $AH$ mixing responsible for the CP-violating effects. Particles circulating in the loop may vary according to the model (in our case they are either squarks or heavy Majorana neutrinos)*]{}. (b) *The cross section for $\sigma^{ij}_{kl}$ the resonant process $\mu^+\mu^-\to A^*,\ H^*\to f\bar f$. In the resonant region and at the one-loop level the only diagram contributing is the one shown. Notice that ${i,j,k,l}$ may be either $A$ or $H$ (with $A$-$H$ mixing due to the one-loop diagram (a)); however due to the chirality difference between $A$ and $H$, the only non-zero combinations are obtained when $i=l$ and $j=k$.*[]{data-label="feyndiag"}](feyndiag.eps){width="16cm"} The amplitude for such process can be written as $$T=T^s+T^t=V_i^P\widehat\Delta_{ij}V_j^D+T^t,$$ where $i,j=A,H$, $V_i^P$ ($V_j^D$) is the production (decay) amplitude of the process $\mu^+\mu^-\to i$ ($j\to\bar f f$), $\widehat\Delta_{ij}$ is the propagator matrix in the presence of mixing, and $T^t$ represents collectively all the $t$-channel and box diagrams. The propagator matrix $\widehat\Delta_{ij}(s)$ appearing in the equation above is constructed from self-energies resummed in the pinch technique framework [@Papavassiliou:1995fq]; they are gauge-independent, display only physical thresholds, have the correct unitarity, analiticity, and renormalization group properties, and satisfy the equivalence theorem. In particular, one has $$\begin{aligned} \widehat\Delta_{AA}(s) &=& [s-M_H^{2} + \widehat\Pi_{HH} (s)] \widehat\Delta^{-1}(s), \nonumber\\ \widehat\Delta_{HH}(s) &=& [s-M_A^{2} + \widehat\Pi_{AA} (s)] \widehat\Delta^{-1}(s), \nonumber\\ \widehat\Delta_{HA}(s) &=& \widehat\Delta_{AH}(s) = - \widehat\Pi_{AH}(s) \widehat\Delta^{-1}(s),\end{aligned}$$ with $\widehat\Delta(s)$ the determinat of the propagator matrix, i.e., $$\widehat\Delta(s)=[s-M_H^{2} + \widehat\Pi_{HH} (s)][s-M_A^{2} + \widehat\Pi_{AA} (s)] - \widehat\Pi_{AH}(s) \widehat\Pi_{HA}(s).$$ In the resonant region on which we concentrate here, $M_A\approx M_H$, the $t$-channel and box diagrams are subdominant, with $T^t\ll T^s$, and can be safely ignored; moreover, at the one-loop level, the production and decay amplitudes $V^P_i$ and $V^D_j$ coincide with the corresponding tree-level (model dependent) vertices $\Gamma^{(0)}_{\mu^+\mu^-i}=\chi_{\mu i}$ and $\Gamma^{(0)}_{jf\bar f}=\chi_{jf}$. Finally, in the resonant region, the dispersive (real) part of the self-energies are very small, so that one can neglect them and concentrate only on their absorptive (imaginary) parts. The cross sections leading to the lineshape we look for, can then be calculated by just squaring the single diagram shown in Fig. \[feyndiag\] (b); helicity mismatches limit the possible combinations to the factors $$\begin{aligned} D_{AA}^{AA}(s) &=& \left[(s-M_H^{2})^{2} + (\mathrm{Im}\;\!\widehat\Pi_{HH} (s))^{2}\right] D^{-1}(s), \nonumber\\ D_{HH}^{HH}(s) &=& \left[(s-M_A^{2})^{2} + (\mathrm{Im}\;\!\widehat\Pi_{AA} (s))^{2}\right] D^{-1}(s)\, , \nonumber\\ D_{HA}^{AH}(s) &=& D_{AH}^{HA}= \widehat\Pi_{AH}^{2}(s) D^{-1}(s),\end{aligned}$$ with $$\begin{aligned} D(s) &=& \left[(s-M_A^{2})(s-M_H^{2}) - \mathrm{Im}\;\!\widehat\Pi_{AA} (s) \mathrm{Im}\;\!\widehat\Pi_{HH} (s) - \widehat\Pi_{AH}^{2}(s)\right]^{2} \nonumber\\ &+& \left[(s-M_A^{2})\mathrm{Im}\;\!\widehat\Pi_{HH}(s) + (s-M_H^{2})\mathrm{Im}\;\!\widehat\Pi_{AA}(s)\right]^{2}.\end{aligned}$$ In the formulas above, the standard thresholds are given by $$\begin{aligned} \mathrm{Im}\;\!\widehat\Pi_{HH}^{(f\bar{f})}(s) &=& \frac{\alpha_\mathrm{w} N^f_c}{8} \chi_{Hf}^2 \frac{m^2_f}{M^2_W} s\ \left(1-\frac{4m^2_f}{s}\right)^{3/2} \theta (s-4m^2_f ),\nonumber \\ \mathrm{Im}\;\!\widehat\Pi_{AA}^{(f\bar{f})}(s) &=& \frac{\alpha_\mathrm{w} N^f_c}{8} \chi_{Af}^2 \, \frac{m^2_f}{M^2_W}\, s\, \left(1-\frac{4m^2_f}{s}\right)^{1/2} \theta (s-4m^2_f ),\nonumber \\ \mathrm{Im}\;\!\widehat\Pi_{HH}^{(VV)}(s) &=& \frac{n_V\alpha_\mathrm{w}}{32} \chi_{HV}^2 \frac{M^4_H}{M^2_W} \left(1-\frac{4M^2_V}{s}\right)^{1/2}\nonumber\\ &\times& \left[ 1+4\frac{M^2_V}{M^2_H}-4\frac{M^2_V}{M^4_H}(2s-3M^2_V) \right] \theta (s-4M^2_V ), \label{thresh}\end{aligned}$$ where, $\alpha_\mathrm{w}=g^2/4\pi$, $n_V=2$ (respectively 1) for $V\equiv W$ (respectively $V\equiv Z$), and $N^f_c=1$ for leptons and 3 for quarks. Restoring the couplings, summing over final states and averaging over initial polarizations, one is then left with the cross-sections $$\sigma(s)=\sum_{i,j}\sigma^{ij}_{ji}(s) = \frac{\pi \alpha_\mathrm{w}^2 m_{\mu}^2 m_{b}^2}{16 M_W^4}s \sum_{i,j}\chi_{i\mu}^{2} \chi_{jb}^{2} D_{ji}^{ij}(s), \label{cs}$$ where we have specialized to the case in which the final fermions are bottom quarks. Numerical results ================= In this section we present numerical results for the cross section of Eq.(\[cs\]), in order to see how CP violating effects through the one-loop mixing of Fig.\[feyndiag\] (a) affects the Higgs line-shape. As anticipated in the introduction, the MSSM the Higgs sector is composed by two Higgs doublets $(A_1,A_2)$ and $(H_1,H_2)$. After performing an orthogonal rotation by the angles $\beta$ (with $\tan\beta=v_2/v_1$ and $v_i$ the VEVs of the Higgs doublets) and $\alpha$ (with $\tan\alpha\propto\tan\beta$), the doublets end up in the physical basis $(G^0,A)$ and $(h,H)$, with masses $m_A$, $m_h$ and $m_H$ ($G_0$ turns out to be the true would-be Goldstone boson, absorbed by the longitudinal part of the $Z$ boson). Here we use the same conventions of [@Pilaftsis:1998dd], to which the reader is referred for details concerning the model \[and the model dependent factor of, for example, Eq.(\[thresh\])\]. It turns out that the $AH$ mixing is dominated by the large CP-violating Yukawa couplings to the top and bottom squarks, for which one has (see again [@Pilaftsis:1998dd], where all the definitions of the quantities appearing in the formula below can be found) $$\begin{aligned} \Pi_{H_iA}(s) &=& \frac{1}{16\pi^2}\, \sum_{q=t,b}N^q_c\,\mathrm{Im}(h^q_1) \Bigg\{\frac{r_i s_{2q}}{s_\beta v}\Delta M^2_{\tilde{q}} \Big[B_0(s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2}) -B_0 (0,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2})\Big] \nonumber\\ &+& \mathrm{Re}(h^q_i)\frac{s^2_{2q}}{s_\beta} \Big[B_0 (s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_1}) + B_0 (s,M^2_{\tilde{q}_2},M^2_{\tilde{q}_2}) - 2B_0 (s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2})\Big] \nonumber\\ &+& \frac{s_{2q}}{s_\beta}\Big\{(c^2_q g^{L,q}_i + s^2_q g^{R,q}_i )\Big[B_0 (s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2}) - B_0 (s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_1})\Big]\nonumber\\ &+& (s^2_q g^{L,q}_i + c^2_q g^{R,q}_i ) \Big[B_0 (s,M^2_{\tilde{q}_2},M^2_{\tilde{q}_2}) - B_0 (s,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2})\Big]\Big\}\Bigg\}, \label{AHsquarks}\end{aligned}$$ with $s=q^2$, and $B_0(s,m_1,m_2)$ the standard Passarino-Veltman function defined as $$\begin{aligned} B_0(s,m_1,m_2)&=& C_{\mbox{\scriptsize{UV}}}+2-\ln(m_1m_2)+\frac1s\left[(m_2^2-m_1^2)\ln\left(\frac{m_1}{m_2}\right)\right.\nonumber \\ &+&\left.\sqrt{\lambda(s,m^2_1,m^2_2)}\cosh^{-1}\left(\frac{m_1^2+m_2^2-s}{2m_1m_2}\right)\right],\nonumber \\ B_0(0,m_1,m_2)&=&C_{\mbox{\scriptsize{UV}}}+1-\ln(m_1m_2)+\frac{m_1^2+m_2^2}{m_1^2-m_2^2}\ln\left(\frac{m_2}{m_1}\right),\end{aligned}$$ where $C_{\mbox{\scriptsize{UV}}}=1/\epsilon-\gamma_E+\ln(4\pi\mu^2)$ is the UV cutoff of dimensional regularization, and $\lambda(x,y,z)=(x-y-z)^2-4yz$. As emphasized in [@Pilaftsis:1998dd], the tadpole contribution $B_0 (0,M^2_{\tilde{q}_1},M^2_{\tilde{q}_2})$ is crucial for accomplishing the UV-finiteness of the $H_iA$ self-energies in question. Notice also that the $\Pi_{H_iA}(s)$ given by (\[AHsquarks\]) are real, because, due to the heaviness of the squarks appearing in the loops, no particle thresholds can open. The CP violating self-energy transition $HA$ is then obtained through an orthogonal rotation of the above result, giving $$\Pi_{HA}(s)=-\sin\alpha\ \Pi_{H_1A}(s)+\cos\alpha\ \Pi_{H_2A}(s). \label{MSSMAH}$$ Notice that $hA$ mixing is also possible, but $m_h\ll m_A,\ m_H$ and therefore its contribution is negligible, since one is far from the resonant enhancement region. The kinematic regime analized in this section is that where $m_A>2M_Z$ and $\tan\beta\geq2$; then since $\tan\beta\approx\tan\alpha$, Eq.(\[MSSMAH\]) reduces to $\Pi_{HA}(s)\approx-\Pi_{H_1A}(s)$. Also, the rotation angle of the scalar top and bottom quarks are equal to $\pi/4$, [*i.e.*]{} $s_{2q}\approx1$. As already mentioned before we assume universal squarks soft masses ($\tilde{M_Q}=\tilde{M_t}=\tilde{M_b}=\ M_0=.5$ TeV) and trilinear couplings ($A_t=A_b=A$ with $|A|=1$). Finally $m_{H^\pm}$ will be fixed at the value of $.4571$ TeV. (300,340)(90,-170) ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ![*The MSSM Higgs lineshape for different values of $|\mu|$, $\tan\beta$ and $\phi$. The black continuous curves correspond to the CP-invariant limit of the theory ($\phi=0^\circ$), while the red dashed one correspond to the case in which CP-breaking phases have been switched on (with $\phi=90^\circ$). Finally, the dashed-dotted blue curve, when present, correspond to the CP-invariant limit of the theory re-calculated for a different value of $\tan\beta$ to accommodate the mass-splitting observed in the CP-breaking case.*[]{data-label="cross-sec90"}](cs-1-1-5-90.eps "fig:"){width="7.5cm"} ![*The MSSM Higgs lineshape for different values of $|\mu|$, $\tan\beta$ and $\phi$. The black continuous curves correspond to the CP-invariant limit of the theory ($\phi=0^\circ$), while the red dashed one correspond to the case in which CP-breaking phases have been switched on (with $\phi=90^\circ$). Finally, the dashed-dotted blue curve, when present, correspond to the CP-invariant limit of the theory re-calculated for a different value of $\tan\beta$ to accommodate the mass-splitting observed in the CP-breaking case.*[]{data-label="cross-sec90"}](cs-1.5-1-5-90.eps "fig:"){width="7.4cm"} ![*The MSSM Higgs lineshape for different values of $|\mu|$, $\tan\beta$ and $\phi$. The black continuous curves correspond to the CP-invariant limit of the theory ($\phi=0^\circ$), while the red dashed one correspond to the case in which CP-breaking phases have been switched on (with $\phi=90^\circ$). Finally, the dashed-dotted blue curve, when present, correspond to the CP-invariant limit of the theory re-calculated for a different value of $\tan\beta$ to accommodate the mass-splitting observed in the CP-breaking case.*[]{data-label="cross-sec90"}](cs-1-1-10-90.eps "fig:"){width="7.5cm"} ![*The MSSM Higgs lineshape for different values of $|\mu|$, $\tan\beta$ and $\phi$. The black continuous curves correspond to the CP-invariant limit of the theory ($\phi=0^\circ$), while the red dashed one correspond to the case in which CP-breaking phases have been switched on (with $\phi=90^\circ$). Finally, the dashed-dotted blue curve, when present, correspond to the CP-invariant limit of the theory re-calculated for a different value of $\tan\beta$ to accommodate the mass-splitting observed in the CP-breaking case.*[]{data-label="cross-sec90"}](cs-1.5-1-10-90.eps "fig:"){width="7.5cm"} ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ (-435,160) (-435,152.5) (-15.3,5.1)(-14.7,5.1) (-415,143.5) (-435,136) (-435,128.5) (-15.3,4.25)(-14.7,4.25) (-415,120) (-435,112.5) (-435,105) (-470,55) (-345,15) (-198,160) (-198,152.5) (-7.0,5.1)(-6.4,5.1) (-178,143.5) (-198,136) (-198,128.5) (-7.0,4.25)(-6.4,4.25) (-178,120) (-198,112.5) (-198,105) (-233,55) (-105,15) (-435,-17.5) (-435,-25) (-15.3,-1.15)(-14.7,-1.15) (-415,-34) (-435,-41.5) (-435,-49) (-15.3,-1.95)(-14.7,-1.95) (-415,-57.5) (-435,-65) (-435,-72.5) (-470,-120) (-345,-165) (-198,-17.5) (-198,-25) (-7.0,-1.15)(-6.4,-1.15) (-178,-34) (-198,-41.5) (-198,-49) (-1.8,-1.15)(-1.2,-1.15) (-30,-34) (-50,-41.5) (-50,-49) (-233,-120) (-105,-165) For relatively big values of the CP-breaking phase (we set $\phi=90^\circ$) our results are shown in Fig.\[cross-sec90\]. As anticipated in the analysis carried out in Section I, we can see that as $\tan\beta$ decreases (upper panels, where $\tan\beta=5$), the difference between the two line-shapes (the CP-invariant limit, black continuous curves, and the CP-breaking phase, red dashed curves) increases, the overlapping resonance get resolved and the two mass peaks are clearly separated. At this point one can determine directly from the lineshape the new pole mass of the Higgs bosons, which turn out to be always in good agreement with the masses obtained algebraically by re-diagonalizing the one-loop propagator matrix, [*i.e.*]{} by solving the characteristic equation $\Delta(s)=0$, yielding $$s_i=M^2_i-\widehat\Pi_{ii}(s_i)+\frac{\widehat\Pi_{ij}(s_i)\widehat\Pi_{ji}(s_i)}{s_i-M^2_j+\widehat\Pi_{jj}(s_i)},$$ where $i,j=A,H$ and $i\ne j$  [@Carena:2001fw]. (300,340)(90,-170) -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![*Same as before but for the CP-breaking phase which is now set to $\phi=30^\circ$. Notice that in the upper left panel, the same masses could have been also reached from a CP-invariant theory with $\tan\beta\sim25$, which however would give a much higher cross section which is not shown.*[]{data-label="cross-sec30"}](cs-1-1-5-30.eps "fig:"){width="7.5cm"} ![*Same as before but for the CP-breaking phase which is now set to $\phi=30^\circ$. Notice that in the upper left panel, the same masses could have been also reached from a CP-invariant theory with $\tan\beta\sim25$, which however would give a much higher cross section which is not shown.*[]{data-label="cross-sec30"}](cs-1.5-1-5-30.eps "fig:"){width="7.5cm"} ![*Same as before but for the CP-breaking phase which is now set to $\phi=30^\circ$. Notice that in the upper left panel, the same masses could have been also reached from a CP-invariant theory with $\tan\beta\sim25$, which however would give a much higher cross section which is not shown.*[]{data-label="cross-sec30"}](cs-1-1-10-30.eps "fig:"){width="7.5cm"} ![*Same as before but for the CP-breaking phase which is now set to $\phi=30^\circ$. Notice that in the upper left panel, the same masses could have been also reached from a CP-invariant theory with $\tan\beta\sim25$, which however would give a much higher cross section which is not shown.*[]{data-label="cross-sec30"}](cs-1.5-1-10-30.eps "fig:"){width="7.5cm"} -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- (-435,160) (-435,152.5) (-15.3,5.1)(-14.7,5.1) (-415,143.5) (-435,136) (-435,128.5) (-15.3,4.25)(-14.7,4.25) (-415,120) (-435,112.5) (-435,105) (-470,55) (-345,15) (-198,160) (-198,152.5) (-7.0,5.1)(-6.4,5.1) (-178,143.5) (-198,136) (-198,128.5) (-7.0,4.25)(-6.4,4.25) (-178,120) (-198,112.5) (-198,105) (-233,55) (-105,15) (-435,-17.5) (-435,-25) (-15.3,-1.15)(-14.7,-1.15) (-415,-34) (-435,-41.5) (-435,-49) (-15.3,-1.95)(-14.7,-1.95) (-415,-57.5) (-435,-65) (-435,-72.5) (-470,-120) (-345,-165) (-198,-17.5) (-198,-25) (-7.0,-1.15)(-6.4,-1.15) (-178,-34) (-198,-41.5) (-198,-49) (-7.0,-1.95)(-6.4,-1.95) (-178,-57.5) (-198,-65) (-198,-72.5) (-233,-120) (-105,-165) In particular, notice that when $\vert\mu\vert=1.5$ TeV(upper-right panel) the mass splitting $\delta m\sim11$ GeV and cannot be explained in terms of CP-invariant radiative corrections only, no matter how big $\tan\beta$ is chosen. In the remaining panels, $\delta m$ allows for smaller (upper-left panel, $\tan\beta\sim2.5$) or larger (lower panels, left $\tan\beta\sim20$ and right $\tan\beta\sim30$) values of $\tan\beta$ which are consequently plotted in th blue dashed-dotted curves. Also in these cases one can distinguish the CP-invariant and CP-breaking phase of the theory due to the very different energy behavior of the cross section; also notice that too lower values of $\tan\beta$ predicts, in the parameter range chosen here, a too low mass for the lightest Higgs, and therefore would be in any case excluded ([*e.g.*]{} when $\tan\beta\sim2.5$, $m_h=94.8$ GeV). For smaller values of the CP-breaking phase $\phi$ we get a somewhat different situation. In Fig.\[cross-sec30\], the same set of plots as before is shown for $\phi=30^\circ$: for small values of $\tan\beta$ (upper panels) it is still possible to discriminate the phase of the theory, while for larger values (lower panels) the effects of $A-H$ mixing vanishes much more rapidly than in the previous case. Conclusions =========== The CP-even and CP-odd neutral Higgs scalars, $H$ and $A$, respectively, together with their charged companions and the “standard” Higgs boson, constitute the extended scalar sector of generic two-Higgs doublet models, and appear naturally in supersymmetric extensions of the SM. Therefore it is expected that their discovery and subsequenet study of their fundamental physical properties should be of central importance in the next decades. In addition to the LHC, where their primary discovery might take place, further detailed studies of their characteristics have been proposed in recent years, most notably in the context of muon-[@Cline:1994tk; @Marciano:1998ue] and photon colliders [@Ginzburg:2001ph; @Choi:2004kq]. Of particular interest is the possibility of encountering CP violation in the Higgs sector of two-Higgs doublet models, induced by the one-loop mixing of $H$ and $A$. The latter is produced due to CP-violating (tree-level) interactions of each one of these two Higgs bosons with particles such as scalar quarks or Majorana neutrinos, to name a few, which will circulate in the (otherwise vanishing) $H$-$A$ transition amplitude (Fig.1). A most appealing feature of the $H$-$A$ system is that, due to their relative small mass-difference, possible CP-mixing effects undergo resonant enhancement [@Pilaftsis:1997dr]. Such effect may be observed in appropriately constructed left-right asymmetries, as proposed in [@Pilaftsis:1997dr]. Since these observables are odd under CP, a non-vanishing experimental result would constitute an unequivocal signal of CP violation. In this paper we have explored the possibility of detecting the presence of CP-mixing between $H$ and $A$ through the detailed study of the cross-section of the $s$-channel process $\mu^+\mu^-\to A^*,H^*\to f\bar f$ as function of the center-of-mass energy. Although the lineshape is a CP-even quantity, there are certain characteristics that signal the presence of a CP-violating mixing. In the context of the photon colliders, Ref.[@Choi:2004kq] also discusses both CP-even and CP-odd observables in terms of photon polarizations. As has been explained in a series of papers  [@Barger:1996jm; @Binoth:1997ev; @Krawczyk:1997be; @Grzadkowski:2000fg; @Blochinger:2002hj], the high-resolution scanning of the lineshape of this process is expected to reveal two relatively closely spaced, or even superimposed, resonances, corresponding to the nearly degenerate $H$ and $A$. As was recognized in [@Carena:2001fw], the presence of CP-mixing between $H$ and $A$ modifies the position of the pole mass, and tends to push the two resonances further apart. Of course, this fact by itself could not serve as a signal for the presence of CP-mixing; one could simply envisage the situation where all fundamental interactions are CP-conserving and there is no CP-mixing, and the masses of $H$ and $A$ assume simply from the beginning their shifted values. The result of our analysis is very positive and represents, in our opinion, a significant step in the search of the CP-properties of two scalar-doublet models: [*either the mass-splitting of the H and A Higgs boson, in a CP-mixing scenario, cannot be accounted for in absence of CP-mixing, and/or the detailed energy dependence of the line-shape allows to discriminate between both scenarios*]{}. Even though in our analysis we have considered only the MSSM due to its phenomenological relevance, we however expect that our analysis will hold in general for two Higgs doublet models, independently from the specific type of mechanism which will give rise to the CP-breaking loops of Fig.\[feyndiag\] (a). For example similar results are expected in the case of a two Higgs doublet model in which the particles circulating in the loop are three generations of heavy Majorana neutrinos $N_i$ ($i=1,2,3$) for which the mixing amplitude will be given by [@Pilaftsis:1997dr] $$\begin{aligned} & &\Pi_{AH}(s)=\ -\frac{\alpha_w \,s}{4\pi}\chi_{Au}\chi_{Hu}\sum\limits_{j>i}^{3} \mathrm{Im}(C^2_{ij}) \sqrt{\lambda_i\lambda_j}\left[ B_0(s/M^2_W,\lambda_i,\lambda_j)+\, 2B_1(s/M^2_W,\lambda_i,\lambda_j)\right],\nonumber \\ & & B_1(s,m_i,m_j) = \frac{m_j^2 - m_i^2}{2s}B_0(s, m_i, m_j) - B_0(0,m_i, m_j) - \frac12B_0(s, m_i, m_j),\end{aligned}$$ with $\lambda_i = m^2_i/M^2_W$, and $\mathrm{Im}(C^2_{ij})$ playing the role of the MSSM CP-breaking phase $\phi$. Thus, the combined study of CP-odd observables, such as left-right asymmetries, together with the $H$-$A$ lineshape, may furnish a powerful panoply for attacking the isssue of CP violation in two-Higss doublet models. [*Acknowledgments:*]{} This research has been supported by the Spanish MEC and European FEDER, under the Grant FPA 2005-01678. Useful correspondence with A. Pilaftsis and J. S. Lee is gratefully acknowledged. We thank Peter Zerwas for bringing reference [@Choi:2004kq] to our attention. JaxoDraw [@Binosi:2003yf] has been used. [99]{} See, for example, S. Eidelman [*et al.*]{} \[Particle Data Group\], Phys. Lett. B [**592**]{} (2004); G. Altarelli and M. W. Grunewald, Phys. Rept.  [**403-404**]{}, 189 (2004); \[arXiv:hep-ph/0404165\]. A. Sirlin, J. Phys. G [**29**]{}, 213 (2003) \[arXiv:hep-ph/0209079\]. P. W. Higgs, Phys. Lett.  [**12**]{}, 132 (1964). F. Englert and R. Brout, Phys. Rev. Lett.  [**13**]{}, 321 (1964). G. S. Guralnik, C. R. Hagen and T. W. B. Kibble, Phys. Rev. Lett.  [**13**]{}, 585 (1964). For reviews, see, H. P. Nilles, Phys. Rept.  [**110**]{} (1984) 1; H. E. Haber and G. L. Kane, Phys. Rept.  [**117**]{} (1985) 75; A. B. Lahanas and D. V. Nanopoulos, Phys. Rept.  [**145**]{}, 1 (1987); J. F. Gunion, H. E. Haber, G. L. Kane and S. Dawson, “The Higgs Hunter’s Guide,” (Addison-Wesley, Reading, MA, 1990). See, for example, J. Haller \[ATLAS Collaboration\], arXiv:hep-ex/0512042; J. Tanaka, Nucl. Phys. Proc. Suppl.  [**144**]{} (2005) 341; R. Kinnunen, S. Lehti, A. Nikitenko and P. Salmi, J. Phys. G [**31**]{}, 71 (2005); M. Schumacher, arXiv:hep-ph/0410112; U. Ellwanger, J. F. Gunion, C. Hugonie and S. Moretti, arXiv:hep-ph/0401228. D. B. Cline, AIP Conf. Proc.  [**352**]{}, 70 (1996). W. J. Marciano, AIP Conf. Proc.  [**441**]{}, 347 (1998). V. D. Barger, M. S. Berger, J. F. Gunion and T. Han, Phys. Rept.  [**286**]{}, 1 (1997) \[arXiv:hep-ph/9602415\]. T. Binoth and A. Ghinculov, Phys. Rev. D [**56**]{}, 3147 (1997) \[arXiv:hep-ph/9704299\]. M. Krawczyk, arXiv:hep-ph/9803484. B. Grzadkowski, arXiv:hep-ph/0005170. C. Blochinger [*et al.*]{}, arXiv:hep-ph/0202199. T. D. Lee, Phys. Rev. D [**8**]{}, 1226 (1973). G. C. Branco and M. N. Rebelo, Phys. Lett. B [**160**]{}, 117 (1985). J. Liu and L. Wolfenstein, Nucl. Phys. B [**289**]{}, 1 (1987). S. Weinberg, Phys. Rev. D [**42**]{}, 860 (1990). Y. L. Wu and L. Wolfenstein, Phys. Rev. Lett.  [**73**]{}, 1762 (1994) See, for instance, the last item of [@Nilles:1983ge], and references therein, and H. E. Haber, arXiv:hep-ph/9707213. S.P. Li and M. Sher, Phys.  Lett. [**B140**]{}, 339 (1984); Y. Okada, M. Yamaguchi and T. Yanagida, Phys. Lett. [ **B262**]{}, 54 (1991); H.E. Haber and R. Hempfling, Phys.  Rev. Lett.  [**66**]{}, 1815 (1991); J. Ellis, G. Ridolfi and F. Zwirner, Phys. Lett. [**B257**]{}, 83 (1991); R. Barbieri, M. Frigeni and F. Caravaglios, Phys. Lett.  [**B258**]{}, 167 (1991); D.M. Pierce, J.A. Bagger, K.T. Matchev and R.-J. Zhang, Nucl.  Phys. [**B491**]{}, 3 (1997). See, [*e.g.*]{}, G.L. Kane, C. Kolda, L. Roszkowksi and J.D. Wells, Phys. Rev. [**D49**]{}, 6173 (1994). A. Pilaftsis, Nucl. Phys. B [**504**]{}, 61 (1997) J. Bernabeu, A. Santamaria, J. Vidal, A. Mendez and J. W. F. Valle, Phys. Lett. B [**187**]{}, 303 (1987). A. Pilaftsis, Phys. Lett. B [**285**]{}, 68 (1992). J. G. Korner, A. Pilaftsis and K. Schilcher, Phys. Rev. D [**47**]{}, 1080 (1993); Phys. Lett. B [**300**]{}, 381 (1993). J. Bernabeu, J. G. Korner, A. Pilaftsis and K. Schilcher, Phys. Rev. Lett.  [**71**]{}, 2695 (1993). A. Ilakovac, B. A. Kniehl and A. Pilaftsis, Phys. Lett. B [**317**]{}, 609 (1993) \[Phys. Lett. B [**320**]{}, 329 (1994)\]. \[arXiv:hep-ph/9308318\]. J. Bernabeu and A. Pilaftsis, Phys. Lett. B [**351**]{}, 235 (1995). D. Tommasini, G. Barenboim, J. Bernabeu and C. Jarlskog, Nucl. Phys. B [**444**]{}, 451 (1995). A. Pilaftsis, Phys. Lett. B [**435**]{}, 88 (1998) A. Pilaftsis, Phys. Rev. D [**58**]{}, 096010 (1998); A. Pilaftsis and C. E. M. Wagner, Nucl. Phys. B [**553**]{}, 3 (1999); D. A. Demir, Phys. Rev. D [**60**]{}, 055006 (1999); S. Y. Choi, M. Drees and J. S. Lee, Phys. Lett. B [**481**]{}, 57 (2000); M. Carena, J. R. Ellis, A. Pilaftsis and C. E. M. Wagner, Nucl. Phys. B [**586**]{}, 92 (2000); G. L. Kane and L. T. Wang, Phys. Lett. B [**488**]{}, 383 (2000); S. Y. Choi, K. Hagiwara and J. S. Lee, Phys. Rev. D [**64**]{}, 032004 (2001); T. Ibrahim and P. Nath, Phys. Rev. D [**63**]{}, 035009 (2001). M. Carena, J. R. Ellis, A. Pilaftsis and C. E. Wagner, Nucl. Phys. B [**625**]{}, 345 (2002). J. S. Lee, A. Pilaftsis, M. Carena, S. Y. Choi, M. Drees, J. R. Ellis and C. E. M. Wagner, Comput. Phys. Commun.  [**156**]{}, 283 (2004) J. Papavassiliou and A. Pilaftsis, Phys. Rev. Lett.  [**75**]{}, 3060 (1995); Phys. Rev. D [**53**]{}, 2128 (1996); Phys. Rev. D [**54**]{}, 5315 (1996); D. Binosi and J. Papavassiliou, Phys. Rev. D [**66**]{}, 076010 (2002); D. Binosi, J. Phys. G [**30**]{}, 1021 (2004). I. F. Ginzburg, M. Krawczyk and P. Osland, arXiv:hep-ph/0101331; Nucl. Instrum. Meth. A [**472**]{}, 149 (2001); arXiv:hep-ph/0101208; W. Khater and P. Osland, Nucl. Phys. B [**661**]{}, 209 (2003); W. Bernreuther, A. Brandenburg and M. Flesch, arXiv:hep-ph/9812387; P. Niezurawski, A. F. Zarnecki and M. Krawczyk, JHEP [**0502**]{}, 041 (2005); M. Krawczyk, arXiv:hep-ph/0512371. S. Y. Choi, J. Kalinowski, Y. Liao and P. M. Zerwas, Eur. Phys. J. C [**40**]{}, 555 (2005). D. Binosi and L. Theussl, Comput. Phys. Commun.  [**161**]{}, 76 (2004).

ArXiv

--- abstract: 'In this paper some new cases of Knaster’s problem on continuous maps from spheres are established. In particular, we consider an almost orbit of a $p$-torus $X$ on the sphere, a continuous map $f$ from the sphere to the real line or real plane, and show that $X$ can be rotated so that $f$ becomes constant on $X$.' address: - ' Roman Karasev, Dept. of Mathematics, Moscow Institute of Physics and Technology, Institutskiy per. 9, Dolgoprudny, Russia 141700' - ' Alexey Volovikov, Department of Mathematics, University of Texas at Brownsville, 80 Fort Brown, Brownsville, TX, 78520, USA' - ' Alexey Volovikov, Department of Higher Mathematics, Moscow State Institute of Radio-Engineering, Electronics and Automation (Technical University), Pr. Vernadskogo 78, Moscow 117454, Russia' author: - 'R.N. Karasev' - 'A.Yu. Volovikov' title: 'Knaster’s problem for almost $(Z_p)^k$-orbits' --- [^1] [^2] Introduction ============ In [@kna1947] the following conjecture (Knaster’s problem) was formulated. \[knaster\] Let $S^{d-1}$ be a unit sphere in $\mathbb R^d$. Suppose we are given $m = d-k+1$ points $x_1, \ldots, x_m\in S^{d-1}$ and a continuous map $f: S^{d-1}\to \mathbb R^k$. Then there exists a rotation $\rho\in SO(d)$ such that $$f(\rho(x_1)) = f(\rho(x_2)) = \dots = f(\rho(x_m)).$$ In papers [@kasha2003; @hiri2005] it was shown that for certain sets $\{x_1, \ldots, x_m\}\subset S^{d-1}$ Knaster’s conjecture fails, such counterexamples exist for every $k>2$, for $k=2$ and $d\ge 5$, for $k=1$ and $d\ge 67$. Still it is possible to prove Knaster’s conjecture in some particular cases of sets. In [@vol1992] the set of points was some orbit of the action of a $p$-torus $G=(Z_p)^k$ on $\mathbb R[G]$ for $k=1$ and on $\mathbb R[G]\oplus\mathbb R$ for $k=2$. Here we prove some similar results, the set of points being a $(Z_p)^k$-orbit minus one point. The group algebra $\mathbb R[G]$ is supposed to have left $G$-action, unless otherwise stated. Considered as a $G$-representation, $\mathbb R[G]$ may have a $G$-invariant inner product. In fact, the space of invariant inner products has the dimension equal to the number of distinct irreducible $G$-representations in $\mathbb R[G]$ (for a commutative $G$), for a $p$-torus $G=(Z_p)^k$ the dimension of this space is $\frac{p^k+1}{2}$ for odd $p$, and $p^k$ for $p=2$. Denote $I[G]\subset\mathbb R[G]$ the $G$-invariant subspace in $\mathbb R[G]$ consisting of $$\sum_{g\in G} \alpha_g g,\quad\text{with}\quad \sum_{g\in G} \alpha_g = 0.$$ Note that its orthogonal complement (w.r.t. any $G$-invariant inner product) is the one-dimensional space with trivial $G$-action. In the sequel we consider a $p$-torus $G=(Z_p)^k$ and denote $q=p^k$. \[knaster1\] Let $S^{q-2}$ be the unit sphere of $I[G]$ w.r.t. some $G$-invariant inner product, denoted by $(\cdot, \cdot)$. Then conjecture \[knaster\] holds for $k=1$, the rotations w.r.t. $(\cdot, \cdot)$, and the set $Gx\setminus\{x\}$, where $x\in S^{q-2}$ is any point. \[knaster2\] Let $S^{q-1}$ be the unit sphere of $\mathbb R[G]$ w.r.t. some $G$-invariant inner product $(\cdot, \cdot)$. Then conjecture \[knaster\] holds for $k=2$, $q$ odd, the rotations w.r.t. $(\cdot, \cdot)$, and the set $Gx\setminus\{x\}$, where $x\in S^{q-1}$ is any point. In fact, the last theorem may be formulated a little stronger. For example, Theorem \[knaster2d\] (see below) gives the following statement. Let $x\in S^{q-1}$ be as in the theorem, and let $f_1, f_2 : S^{q-1}\to \mathbb R$ be two continuous functions. Then for some rotation $\rho$ and two constants $c_1, c_2$ $$\begin{aligned} &\forall g\in G\ &f_1(\rho(gx)) = c_1\\ &\forall g\in G,\ g\not=e\ &f_2(\rho(gx))=c_2.\end{aligned}$$ Equivariant cohomology of $G$-spaces {#eq-cohomology} ==================================== We consider topological spaces with continuous action of a finite group $G$ and continuous maps between such spaces that commute with the action of $G$. We call them $G$-spaces and $G$-maps. Let us consider the group $G=(Z_p)^k$ and list the results (mostly from [@vol2005]) that we need in this paper. The cohomology is taken with coefficients in $Z_p$, in the notation we omit the coefficients. Consider the algebra of $G$-equivariant (in the sense of Borel) cohomology of the point $A_G = H_G^*({\mathrm{pt}}) = H^*(BG)$. For any $G$-space $X$ the natural map $X\to{\mathrm{pt}}$ induces the natural map of cohomology $\pi_X^* : A_G\to H_G^*(X)$. For a group $G=(Z_p)^k$ the algebra $A_G$ (see [@hsiang1975]) has the following structure. For odd $p$, it has $2k$ multiplicative generators $v_i,u_i$ with dimensions $\dim v_i = 1$ and $\dim u_i = 2$ and relations $$v_i^2 = 0,\quad\beta{v_i} = u_i.$$ Here we denote $\beta(x)$ the Bockstein homomorphism. For a group $G=(Z_2)^k$ the algebra $A_G$ is the algebra of polynomials of $k$ one-dimensional generators $v_i$. The powerful tool of studying $G$-spaces is the following spectral sequence (see [@hsiang1975; @mcc2001]). \[specseqeq\] There exists a spectral sequence with $E_2$-term $$E_2^{x, y} = H^x(BG, \mathcal H^y(X, Z_p)),$$ that converges to the graded module, associated with the filtration of $H_G^*(X, Z_p)$. The system of coefficients $\mathcal H^y(X, Z_p)$ is obtained from the cohomology $H^y(X, Z_p)$ by the action of $G = \pi_1(BG)$. The differentials of this spectral sequence are homomorphisms of $H^*(BG, Z_p)$-modules. For every term $E_r(X)$ of this spectral sequence there is a natural map $\pi^*_r : A_G\to E_r(X)$. Denote the kernel of the map $\pi^*_r$ by $\operatorname{Ind}^r_G X$. The ideal-valued index of a $G$-space was introduced in [@fh1988], the above filtered version was introduced in [@vol2000]. Remind the properties of $\operatorname{Ind}^r_G X$, that are obvious by the definition. We omit the subscript $G$ when a single group is considered. - If there is a $G$-map $f:X\to Y$, then $\operatorname{Ind}^r X\supseteq \operatorname{Ind}^r Y$. - $\operatorname{Ind}^{r+1} X$ may differ from $\operatorname{Ind}^r X$ only in dimensions $\ge r$. - $\bigcup_r \operatorname{Ind}^r X = \operatorname{Ind}X = \ker \pi_X^* : A_G\to H_G^*(X)$. The first property in this list is very useful to prove nonexistence of $G$-maps. Following [@vol2005] we define a numeric invariant of this ideal filtering $\operatorname{Ind}_G^r X$. Put $$i_G(X) = \max \{r : \operatorname{Ind}_G^r X = 0\}.$$ It is easy to see that $i_G(X)\ge 1$ for any $G$-space $X$, $i_G(X)\ge 2$ for a connected $G$-space $X$, and $i_G(X)$ may be equal to $+\infty$. Moreover, for a $G$-space $X$ without fixed points, $G$-homotopy equivalent to a finite $G$-$CW$-complex, $i_G(X)\le \dim X + 1$. From the definition of $\operatorname{Ind}_G^r X$ it follows that if there exists a $G$-map $f:X\to Y$, then $i_G(X)\le i_G(Y)$ (the monotonicity property). The definition of $i_G(X)$ can be further extended. Define the index of a cohomology class $\alpha\in A_G$ on a $G$-space $X$ $$i_G(\alpha, X) = \max\{r: \alpha\not\in \operatorname{Ind}_G^r X\}.$$ It may equal $+\infty$ if $\alpha\not\in \operatorname{Ind}_G X$ It is clear from the definition that either $i_G(\alpha, X)=+\infty$, or $i_G(\alpha, X)\le \dim \alpha$ and $i_G(\alpha, X)\le \dim X + 1$ (for a finite $G$-$CW$-complex). Moreover, for any $G$-map $f : X\to Y$ we have the monotonicity property $$i_G(\alpha, X)\le i_G(\alpha, Y).$$ Reformulations ============== We reformulate Theorems \[knaster1\] and \[knaster2\] in a more general way. \[knaster1d\] Let $S^{q-2}$ be the unit sphere of $I[G]$ w.r.t. some $G$-invariant inner product, and let $f:S^{q-2}\to\mathbb R$ be some continuous function. Consider $x\in S^{q-2}$, the vector $v=\sum_{g\in G} g\in\mathbb R[G]$ and some other vector $w\in\mathbb R[G]$, non-collinear to $v$. Then for some rotation $\rho\in SO(q-1)$ the vector $\sum_{g\in G}f(\rho(gx))g\in\mathbb R[G]$ is in the linear span of $v$ and $w$. Theorem \[knaster1\] follows from this theorem in the following way. Put $w=e\in\mathbb R[G]$. Then by Theorem \[knaster1d\] there exists a rotation $\rho$ such that for some $\alpha,\beta\in\mathbb R$ $$\forall g\in G,\ g\not=e,\ f(\rho(gx)) = \alpha,\quad f(\rho(x)) = \alpha +\beta.$$ That is exactly the statement of Theorem \[knaster1\]. \[knaster2d\] Let $S^{q-1}$ be the unit sphere of $\mathbb R[G]$ w.r.t. some $G$-invariant inner product, and let $f:S^{q-1}\to\mathbb R^2$ be some continuous map with coordinates $f_1, f_2$. Let $q$ be odd. Consider $x\in S^{q-1}$, the vectors $v=\sum_{g\in G}g\oplus 0\in\mathbb R[G]\oplus\mathbb R[G]$, $u=0\oplus\sum_{g\in G} g\in\mathbb R[G]\oplus\mathbb R[G]$ and some other vector $w\in\mathbb R[G]\oplus\mathbb R[G]$, non-coplanar to $v, u$. Then for some rotation $\rho\in SO(q)$ the vector $$\sum_{g\in G}f_1(\rho(gx))g\oplus\sum_{g\in G}f_2(\rho(gx))g\in\mathbb R[G]\oplus\mathbb R[G]$$ is in the linear span of $v, u, w$. Again, Theorem \[knaster2\] (and its stronger version in the remark after Theorem \[knaster2\]) follows from this theorem by taking a vector $w=e\oplus 0$, similar to the previous remark. Proof of Theorem \[knaster1d\] in the case of odd $q$ {#knaster1d-odd} ===================================================== In this section $q=p^k$, $p$ is an odd prime, $G=(Z_p)^k$. Define for any $\rho\in SO(q-1)$ $$\phi(\rho) = \sum_{g\in G} f(\rho(g(x)))g\in\mathbb R[G].$$ For any $h\in G$ we have $$\phi(\rho\circ h^{-1}) = \sum_{g\in G} f(\rho(h^{-1}g(x)))g = \sum_{g\in G} f(\rho(h^{-1}g(x)))hh^{-1}g=\sum_{g\in G} f(\rho(g(x)))hg.$$ Thus the map $\phi : SO(q-1)\to \mathbb R[G]$ is a $G$-map for the left action of $G$ on $SO(q-1)$ by right multiplications by $g^{-1}\in G$, and for the standard left action of $G$ on $\mathbb R[G]$. Denote for any $g\in G$ by $L_g=(v, gw)\subset\mathbb R[G]$ the $2$-dimensional subspaces. Assume the contrary: that is the image of $\phi$ does not intersect $\bigcup_{g\in G} L_g$. So $\phi$ maps $SO(q-1)$ to the space $Y=\mathbb R[G]\setminus \bigcup_{g\in G} L_g$. The natural projection $\pi : Y\to\mathbb R[G]/(v)=V$ gives a homotopy equivalence between $Y$ and $V\setminus \bigcup_{g\in G} \mathbb R\pi(gw)$, the latter space is homotopically $q-2$-dimensional sphere without several points, hence it is a wedge of $q-3$-dimensional spheres. $G$ acts on $Y$ without fixed points, so $i_G(Y)\le q-2$. In [@vol1992] it was shown that $i_G(SO(q-1)) = q-1$ w.r.t. the considered $G$-action. Here we give a short explanation. In the spectral sequence of Theorem \[specseqeq\] all multiplicative generators of $H^*(SO(q-1), Z_p)$ are transgressive, because they are pullbacks of the transgressive generators of $H^*(SO(q-1), Z_p)$ in the spectral sequence of the fiber bundle $\pi_{SO(q-1)} : ESO(q-1)\to BSO(q-1)$. So the first nonzero $\operatorname{Ind}_G^r SO(q-1)$ corresponds to the first nonzero characteristic class of the $G$-representation $I[G]$ in the cohomology ring $A_G$. It was shown in [@vol1992] that this is the Euler class of $I[G]$ of dimension $q-1$. So we have a contradiction with the monotonicity of $i_G(X)$. Proof of Theorem \[knaster2d\] {#knaster2d-proof} ============================== Similar to the previous proof, we consider the $G$-map $\phi: SO(q)\to \mathbb R[G]\oplus\mathbb R[G]$, given by the formula $$\phi(\rho) = \sum_{g\in G} f_1(\rho(g(x)))g\oplus \sum_{g\in G} f_2(\rho(g(x)))g\in\mathbb R[G]\oplus\mathbb R[G].$$ Take the composition $\psi = \pi\cdot \phi$ with the projection $\pi:\mathbb R[G]\oplus\mathbb R[G]\to I[G]\oplus I[G] = V$. Assume the contrary: that is the map $\phi$ does not intersect the linear span of $u$ and $v$ in $\mathbb R[G]\oplus\mathbb R[G]$ and $\psi$ does not intersect the linear span of $gw$ for any $g\in G$ in $V$, it means that the image of $\psi$ is in the space $Y = V\setminus \bigcup_{g\in G} \mathbb R\pi(gw)$. Let $e\in A_G$ be the Euler class of $V$. From the spectral sequence of Theorem \[specseqeq\] it is obvious that $i_G(e, V\setminus\{0\}) = 2q-2$, because the spectral sequence for $V\setminus\{0\}$ has the only nontrivial differential that kills the Euler class $e$. Since $Y\subset V\setminus\{0\}$, then $i_G(e, Y) < +\infty$. Similar to the previous proof, the space $Y$ is homotopically a wedge of $2q-4$-dimensional spheres, so $i_G(e, Y) \le \dim Y + 1 = 2q-3$. In [@vol1992] it was shown that $i_G(e, SO(q)) = 2q-2$, because $e$ is in the image of the transgression in the spectral sequence and $e$ is not contained in the ideal of $A_G$, generated by the characteristic classes of $SO(q)$ of lesser dimension. So we again have a contradiction with the monotonicity of $i_G(e, X)$. Proof of Theorem \[knaster1d\] in the case of even $q$ ====================================================== In this section $q=2^k$, $G=(Z_2)^k$. We use the notation from the odd case in Section \[knaster1d-odd\]. Note that the case $q=2$ is trivial, and if $q\ge 4$ then $G$ acts on $I[G]$ by transforms with positive determinant, so the group $SO(q-1)$ can be considered as the configuration space. Assume the contrary: the image $\phi(SO(q-1))$ is in $Y=\mathbb R[G]\setminus \bigcup_{g\in G} L_g$. Denote the Stiefel-Whitney classes of $I[G]$ in $A_G$ by $w_k$. We need the following lemma, stated in [@vol1992], based on results from [@dic1911; @mui1975]. \[euler-nz\] The only nonzero Stiefel-Whitney classes of $I[G]$ are $w_{q-2^l}\in A_G$ $(l=0,\ldots, k)$, the classes $w_{q-2^l}$ $(l=0,\ldots, k-1)$ are algebraically independent and form a regular sequence, hence $w_{q-1}$ is nonzero and not contained in the ideal of $A_G$, generated by $w_k$ with $k<q-1$. Similar to the proof of Theorem \[knaster2d\] in Section \[knaster2d-proof\], we find that $i_G(w_{q-1}, Y)\le \dim Y + 1 = q-2$. Now we apply the spectral sequence of Theorem \[specseqeq\] to the $G$-space $SO(q-1)$. The action of $G$ on $SO(q-1)$ is the restriction of action of $SO(q-1)$ on itself, the latter group being connected, hence $G$ acts trivially on $H^*(SO(q-1), Z_2)$. The results of [@bor1953] imply that the differentials in this spectral sequence are generated by transgressions that send the primitive (in terms of [@bor1953]) elements of $H^*(SO(q-1), Z_2)$ to the Stiefel-Whitney classes $w_k$ (see Proposition 23.1 in [@bor1953]). Thus Lemma \[euler-nz\] implies that $i_G(w_{q-1}, SO(q-1)) = q-1$, and the existence of the $G$-map $\phi$ contradicts the monotonicity of $i_G(w_{q-1}, X)$. [99]{} A. Borel. Sur la cohomology des espaces fibrés principaux et des espace homogènes de groupes de Lie compact. // Ann. Math., 57, 1953, 115–207. L.E. Dickson. A fundamental system of invariants of the general modular linear group with a solution of the form problem. // Trans. Amer. Math. Soc., 12(1), 1911, 75–98. E. Fadell, S. Husseini. An ideal valued cohomological index theory with applications to Borsuk-Ulam and Bourgain-Yang theorems. // Ergod. Th. & Dynam. Sys. 8, 1988, 73–85. A. Hinrichs, C. Richter. The Knaster problem: More counterexamples. // Israel Journal of Mathematics, 145(1), 2005, 311–324. Wu Yi Hsiang. Cohomology theory of topological transformation groups. Springer Verlag, 1975. B.S. Kashin, S.J. Szarek. The Knaster problem and the geometry of high-dimensional cubes. // Comptes Rendus Mathematique, 336(11), 2003, 931–936. B. Knaster. Problem 4. // Colloq. Math., 30, 1947, 30–31. J. McCleary. A user’s guide to spectral sequences. Cambridge University Press, 2001. H. Mùi. Modular invariant theory and cohomology algebras of symmetric groups. // J. Fac. Sci. U. Tokyo, 22, 1975, 319–369. A.Yu. Volovikov. A Bourgin-Yang-type theorem for $Z\sp n\sb p$-action (In Russian). // Mat. Sbornik, 183(2), 1992, 115–144; translation in Russian Acad. Sci. Sb. Math., 76(2), 1993, 361–387. A.Yu. Volovikov. On the index of $G$-spaces (In Russian). // Mat. Sbornik, 191(9), 2000, 3–22; translation in Sbornik Math., 191(9), 2000, 1259–1277. A.Yu. Volovikov. Coincidence points of maps of $\mathbb Z_p^n$-spaces. // Izvestiya: Mathematics, 69(5), 2005, 913-–962. [^1]: The research of R.N. Karasev was supported by the President’s of Russian Federation grant MK-1005.2008.1, and partially supported by the Dynasty Foundation. [^2]: The research of A.Yu. Volovikov was partially supported by the Russian Foundation for Basic Research grant No. 08-01-00663

ArXiv

--- author: - | \ [Department of Probability Theory and Statistics, Eötvös Loránd University, Budapest, Hungary]{}\ [Faculty of Informatics, University of Debrecen, Debrecen, Hungary]{} --- Abstract {#abstract .unnumbered} ======== Control charts have traditionally been used in industrial statistics, but are constantly seeing new areas of application, especially in the age of Industry 4.0. This paper introduces a new method, which is suitable for applications in the healthcare sector, especially for monitoring a health-characteristic of a patient. We adapt a Markov chain-based approach and develop a method in which not only the shift size (i.e. the degradation of the patient’s health) can be random, but the effect of the repair (i.e. treatment) and time between samplings (i.e. visits) too. This means that we do not use many often-present assumptions which are usually not applicable for medical treatments. The average cost of the protocol, which is determined by the time between samplings and the control limit, can be estimated using the stationary distribution of the Markov chain. Furthermore, we incorporate the standard deviation of the cost into the optimisation procedure, which is often very important from a process control viewpoint. The sensitivity of the optimal parameters and the resulting average cost and cost standard deviation on different parameter values is investigated. We demonstrate the usefulness of the approach for real-life data of patients treated in Hungary: namely the monitoring of cholesterol level of patients with cardiovascular event risk. The results showed that the optimal parameters from our approach can be somewhat different from the original medical parameters. 0.5cm KEYWORDS: control chart; cost-effectiveness; Markov-chain; healthcare Introduction ============ Statistical process control, and with it control charts enjoy a wide range of use today, and have seen great developments, extensions and generalisations since the original design of Shewhart.[@Montgomery] This proliferation of methods and uses can at part be attributed to the fact that information is becoming available in ever increasing quantities and in more and more areas. Even though control charts have originally been designed with statistical criteria in mind, the development of cost-efficient control charts also began early. Duncan in his influential work from 1956 developed the basis for a cycle based cost-effective control chart framework which is still very popular today and is implemented in statistical software packages such as in ****.[@Duncan; @Mortarino; @Zhu] Cost-efficient solutions are often the focus in industrial and other settings besides statistical optimality. Statistical process control, including control charts can be found in very different environments in recent literature. For example in mechanical engineering, e.g. Zhou et al.[@Zhou], where the authors develop a $T^2$ bootstrap control chart, based on recurrence plots. Another example is the work of Sales et al.[@Sales] at the field of chemical engineering, in their work they use multivariate tools for monitoring soybean oil transesterification. It is not a surprise that the healthcare sector has seen an increased use of control chart too. Uses within this sector range from quality assurance to administrative data analysis and patient-level monitoring, and many more.[@Thor; @Suman] In one example control charts were used as part of system engineering methods applied to healthcare delivery.[@Padula] Other examples include quality monitoring in thyroid surgery[@Duclos], monitoring quality of hospital care with administrative data[@Coory] and chronic respiratory patient control[@Correia]. Cost-efficient control charts have not been widely used in healtcare settings, but there are some examples which deal with cost monitoring and management. In one work $\overline{X}$ and $R$ charts were used to assess the effect of a physician educational program on the hospital resource consumption[@Johnson]. Another article is about a hospital which used control charts for monitoring day shift, night shift and daily total staffing expense variance, among other variables[@Shaha]. Yet another paper documents a case study about primary care practice performance, where control charts were used to monitor costs associated with provider productivity and practice efficiency. Further costs monitored were net patient revenue per relative value unit, and provider and non‐provider cost as a percent of net revenue[@Stewart]. Even though these studies used control charts for cost monitoring or optimisation purposes, they did not deal with the same problems as this paper, as our method focuses on cost-optimisation by finding the optimal parameters of the control chart setup. The aim of this study is to present a cost-efficient control chart framework which was specifically designed with use on healthcare data in mind. Specifically, we would like to use control charts for the purposes of analysing and controlling a healthcare characteristic of a patient over time, such as the blood glucose level. Traditionally, the minimal monitoring and process cost is achieved by finding the optimal parameters, namely the sample size, time between samplings and critical values[@Montgomery]. If one desires to find the optimal parameters for a cost-optimal control chart for a healthcare process, then the proper modelling of the process is critical, since if the model deviates greatly from reality, then the resulting parameters and costs may not be appropriate. Of course this presents several problems as certain typical assumptions in control chart theory will not hold for these kind of processes. Namely the assumption of fixed and known shift size, which is problematic because healthcare processes can often be drifting in nature and can produce different shift sizes. Another assumption is the perfect repair, which in this case means that the patient can always be fully healed, which is obviously often impossible. Lastly, the time between shifts is usually set to be constant, but in a healthcare setting with human error and non-compliance involved, this also needs to be modified. Furthermore, since these processes can have undesirable erratic behaviour, the cost standard deviation also needs to be taken into account besides the expected cost. The control chart design which is presented here for dealing with these problems is the Markov chain-based framework. This framework was developed by Zempléni et al. in 2004 and was successfully applied on data collected from a Portuguese pulp plant[@Zempleni]. Their article provided suitable basis for generalisations. Cost-efficient Markov chain-based control charts build upon Duncan’s cycle based model, as these also partition the time into different segments for cost calculation purposes. The previous paper from 2004[@Zempleni] has already introduced the random shift size element, in this paper we expand upon that idea and develop further generalisations to create cost-efficient Markov chain-based control charts for healthcare data. The article is organized in the following way: Section 2 starts with basic definitions and notions needed to understand the more complicated generalisations. Subsection 2.2 discusses the mathematics behind the aforementioned new approaches in the Markov chain-based framework. Subsection 2.3 deals with the problems introduced by discretisation - which is needed for the construction of the transition matrix - and also explains the calculation of the stationary distribution, and defines the cost function. Analysis of results and examples will be provided with the help of program implementations written in ****. Section 3 shows an example of use for the new control chart methods in healthcare settings. This example involves the monitoring of low-density lipoprotein levels of patients at risk of cardiovascular events. Section 4 concludes the theoretical and applied results. Methods ======= The methods and **** implementation in this paper largely depend on the works of Zempléni et al. The first part of this section gives a brief introduction to the methods used in their paper.[@Zempleni] The Markov-chain-based Framework -------------------------------- Consider a process which is monitored by a control chart. We will show methods for processes which can only shift to one (positive) direction, monitored by a simple $X$-chart, with sample size $n=1$. This aims to model the monitoring of a characteristic where the shift to only one direction is interesting or possible, and the monitoring is focused at one patient at a time. Several assumptions are made for the base model. The process distribution is normal with known parameters $\mu_0$ and $\sigma$. We will denote its cumulative distribution function by $\phi$. The shift intensity $1/s$ is constant and known and the shift size $\delta^*$ is fixed. It is also assumed that the process does not repair itself, but when repair is carried out by intervention, it is perfect. The repair is treated as an instantaneous event. All costs related to repairing should be included in the repair cost, for example if a major repair entails higher cost, then this should also be reflected in the calculation. The time between shifts is assumed to be exponentially distributed. The above assumptions ensure that the probabilities of future transitions are only dependent on the current state. This is the so-called Markov property, and the model under consideration is called a Markov chain. The states of this Markov chain are defined at the sampling times and the type of the state depends on the measured value and the actual (unobservable) background process, namely whether there was a shift from the target value in the parameter. This way four basic state types are defined: - No shift - no alarm: in-control (INC) - Shift - no alarm: out-of-control (OOC) - No shift - alarm: false alarm (FA) - Shift - alarm: true alarm (TA) A graphical representation of this process can be seen on Figure 1.  The transition matrix of this Markov chain can be seen below. $\begin{matrix} {\scalebox{1}{\mbox{\ensuremath{\displaystyle \begin{bmatrix} { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix} (1-F(h))\phi(k)\end{matrix}}}^{\text{In-control}}}$} (1-F(h))\phi(k)} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}F(h))\phi(k-\delta^*)\end{matrix}}}^{\text{Out-of-control}}}$}F(h))\phi(k-\delta^*)} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}(1-F(h))(1-\phi(k))\end{matrix}}}^{\text{False alarm}}}$}(1-F(h))(1-\phi(k))} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}F(h)(1-\phi(k-\delta^*))\end{matrix}}}^{\text{True alarm}}}$}F(h)(1-\phi(k-\delta^*))} \\ 0 & \phi(k-\delta^*) & 0 & 1-\phi(k-\delta^*) \\ (1-F(h))\phi(k) & F(h))\phi(k-\delta^*) & (1-F(h))(1-\phi(k)) & F(h)(1-\phi(k-\delta^*)) \\ (1-F(h))\phi(k) & F(h))\phi(k-\delta^*) & (1-F(h))(1-\phi(k)) & F(h)(1-\phi(k-\delta^*)) \end{bmatrix}}}}} \end{matrix}$ Here $F()$ is the cumulative distribution function of the exponential distribution with expectation $1/s$, where $s$ is the expected number of shifts in a unit time interval. $h$ is the time between consecutive observations and $\phi$ is the cumulative distribution function of the process. $k$ is the control limit and $\delta^*$ is the size of the shift. After calculating the stationary distribution of this Markov chain, one can define a cost function for the average total cost: $$E(C) =\frac{c_s + p_3c_f + p_4c_r}{h} + p_2c_o + p_4c_oB$$ where $p_i$ is the probability of state $i$, in the stationary distribution. $c_s$, $c_f$, $c_o$ and $c_r$ are the cost of sampling, false alarm, out-of-control operation and repair respectively. $B$ is the fraction of time between consecutive observations where the shift occurred and remained undetected[@Mortarino]: $$B=\frac{hse^{hs} - e^{hs} + 1}{hs(e^{hs} - 1)} \nonumber$$ An economically optimal chart, based on the time between samplings and the control limit can be constructed by minimising the cost function. Generalisation -------------- The previous simple framework can be used for more general designs. Zempléni et al.[@Zempleni] used this method to set up economically optimal control charts where shifts are assumed to have a random size and only the distribution of the shift size is known. This means that the shift size is no longer assumed to be fixed and known, which is important in modeling various processes not just in healthcare, but in industrial or engineering settings too. This requires expanding the two shifted states to account for different shift sizes. Let $\tau_i$ denote the random shift times on the real line and let $\rho_i$ be the shift size at time $\tau_i$. Let the probability mass function of the number of shifts after time $t$ from the start be denoted by $\nu_t$. $\nu_t$ is a discrete distribution with support over $\mathbb{N}^0$. Assume that $\rho_i$ follows a continuous distribution, which has a cumulative distribution function with support over $(0, \infty)$, and that the shifts are independent from each other and from $\tau_i$. If the previous conditions are met, the resulting random process has monotone increasing trajectories between samplings, which are step functions. The cumulative distribution function of the process values for a given time $t$ from the start can be written the following way: $$Z_t(x)= \begin{dcases} 0 &{\text{if }} x < 0, \\ \nu_t(0) + \sum_{k=1}^\infty \nu_t(k) \Psi_k(x) &{\text{if }} x \geq 0 \end{dcases}$$ where $\Psi_k()$ is the cumulative distribution function of the sum of $k$ independent, identically distributed $\rho_i$ shift sizes. The $x=0$ case means there is no shift. The probability of zero shift size is just the probability of no shift occurring, which is $\nu_t(0)$. Let us assume now that the shift times form a homogeneous Poisson process, and the shift size is exponentially distributed, independently of previous events. The choice of the exponential distribution was motivated by the tractability of its convolution powers as a gamma distribution. The cumulative distribution function of the shift size by (2) is then: $$Q_t(x)= \begin{dcases} 0 &{\text{if }} x < 0, \\ n_t(0) + \sum_{k=1}^\infty n_t(k) Y_k(x) &{\text{if }} x \geq 0 \end{dcases}$$ where $n_t$ is the probability mass function of the Poisson distribution, with parameter $ts$ - the expected number of shifts per unit time multiplied by the time elapsed - representing the number of shifts in the time interval $(0;t)$. $Y_k$ - the shift size cumulative distribution function, for $k$ shift events - is a special case of the gamma distribution, the Erlang distribution $E(k,\frac{1}{\delta})$, which is just the sum of $k$ independent exponential variates each with mean $\delta$. The framework can be generalised even further by not assuming perfect repair after a true alarm signal. This means that the treatment will not have perfect results on the health of the patient, or - in industrial or engineering settings - that the machines cannot be fully repaired to their original condition. In this case, the imperfectly repaired states act as out-of-control states. It is assumed that the repair cannot worsen the state of the process, but an imperfectly repaired process will still cost the same as an equally shifted out-of-control process, thus repaired and out-of-control states do not need distinction during the cost calculation. Different approaches can be considered for modeling the repair size distribution. The one applied here uses a random variable to determine the distance from the target value after repair. A natural choice for the distribution of this random variable could be the beta distribution since it has support over $[0,1]$ - the portion of the distance compared to the current one after repair. Also, the flexible shape of its density function can model many different repair processes. Because of these considerations we will assume that the proportion of the remaining distance from $\mu_0$ after repair - $R$ - is a $Beta(\alpha,\beta)$ random variable, with known parameters. Yet another generalisation is the random sampling time. In certain environments, the occurrence of the sampling at the prescribed time is not always guaranteed. For example in healthcare, the patient or employee compliance can have significant effect on the monitoring, thus it is important to take this into account during the modeling too. Here it is modeled in a way, that the sampling is not guaranteed to take place - e.g. the patient may not show up for control visit. This means that the sampling can only occur according to the sampling intervals, for example at every $n$th days, but is not a guaranteed event. One can use different approaches when modeling the sampling probability, here we consider two cases. The first one assumes that too frequent samplings will decrease compliance. This assumption is intended to simulate the situation in which too frequent samplings may cause increased difficulty for the staff or the patient - leading to decreased compliance. The probability of a successful sampling as a function of the prescribed time between samplings is modeled using a logistic function: $$\begin{aligned} T^*_h=\frac{1}{1+e^{-q(h-z)}} \nonumber \end{aligned}$$ where $q>0$ is the steepness of the curve, $z \in \mathbb{R}$ is the value of the sigmoid’s midpoint and $h$ is the time between samplings. In the other approach it is assumed that too frequent samplings will decrease compliance and increased distance from the target value will increase compliance. This assumption means that a heavily deteriorated process or health state will ensure a higher compliance. The probability of a successful sampling as a function of the prescribed time between samplings and the distance from the target value is modeled using a modified beta distribution function: $$\begin{aligned} T^*_h(w) &= P\Bigg(W_h<\frac{w+\zeta^*}{V+2\zeta^*}\Bigg) \nonumber \end{aligned}$$ where $W_h$ is a beta distributed $Beta($a$/h,$b$)$ random variable, $w$ is the distance from the target value, $h$ is the time between samplings, $V$ is the maximum distance from $\mu_0$ taken into account - the distance where we expect maximal possible compliance. The shifts in the values of $w$ and $V$ are needed to acquire probabilities strictly between $0$ and $1$, since deterministic behaviour is undesirable even in extreme cases. These shifts are parametrised by the $\zeta^*>0$ value, which should typically be a small number. It is important to note, that these are just two ways of modeling the sampling probability. Other approaches and distributions may be more appropriate depending on the process of interest. The shift size distribution, the repair size distribution and the sampling probability, together with the control chart let us model and monitor the behaviour of the process. The resulting process is monotone increasing between samplings and has a downward “jump” at alarm states - as the repair is assumed to be instantaneous. Usually a wide range of different cost types are associated with the processes and their monitoring, these include the costs associated with operation outside the target value. Since the operator of the control chart only receives information about the process at the time of the sampling, the proper estimation of the process behaviour between samplings is critical. Previously, at the perfect repair and non-stackable, fixed shift size model, this task was reduced to estimating the time of shift in case of a true alarm or out-of-control state, since otherwise the process stayed at the previous value between samplings. The estimation of the process behavior with random shift sizes and random repair is more difficult. The expected out-of-control operation cost can be written as the expectation of a function of the distance from the target value. At (2) the shift size distribution was defined for a given time $t$, but this time we are interested in calculating the expected cost for a whole interval. We propose the following calculation method for the above problem: Let $H_j$ be a random process whose trajectories are monotone increasing step functions defined by $\tau_i$ shift times and $\rho_i$ shift sizes as in (2), with starting value $j \geq 0$. The expected value of a function of $H_j$ over an $\epsilon$ long interval can be written as: $$E_\epsilon(f(H_j)) = \frac{\bigintsss_{t_0}^{t_0+\epsilon} \bigintsss_{0}^{\infty} 1-Z_t(f^{-1}(x-j)) dx dt}{\epsilon}$$ where $Z_t()$ is the shift size cumulative distribution function given at (2), $f()$ is an invertible, monotonic increasing function over the real line and $t_0$ is the start point of the interval. Let us observe that the inner integral in the numerator - $\int_{0}^{\infty} 1-Z_t(f^{-1}(x-j)) dx$ - is just the expected value of a function of the shift size at time $t$, since we know that if $X$ is a non-negative random variable, then $E(X)=\int_0^\infty (1-F(x))dx$, where $F()$ is the cumulative distribution function of $X$. In other words, the inner integral in the numerator is the expected value the process $f(H_j)$, $t$ time after the start. Furthermore, observe that this expected value is a continuous function of $t$. We are looking for the expectation of $\int_{0}^{\infty} 1-Z_t(f^{-1}(x-j)) dx$ over $[t_0,t_0+\epsilon]$, which is (4). For practical purposes we can apply the previous, general proposition to our model of Poisson-gamma mixture shift size distribution, thus we assume that the shift size distribution is of the form of (3). The connection between the distance from the target value and the resulting cost is often assumed not to be linear: often a Taguchi-type loss function is used - the loss is assumed to be proportional to the squared distance, see e.g. the book of Deming[@Deming]. Applying this to the above proposition means $f(x)=x^2$. Since we are interested in the behaviour of the process between samplings, $t_0=0$ and $\epsilon=h$, thus: $$\begin{aligned} E_h(H_j^2) &= \frac{\bigintsss_{0}^{h} \Big[e^{-ts}j^2 + \Big(\sum_{k=1}^\infty \frac{(ts)^k e^{-ts}}{k!} \cdot \bigintsss_{0}^{\infty} (x+j)^2 \frac {(1/\delta)^{k}x^{k-1}e^{-x/\delta}}{(k-1)!} dx \Big) \Big] dt}{h} \nonumber \\ &= \frac{\bigintsss_{0}^{h} e^{-ts}j^2 + \sum_{k=1}^\infty \frac{(ts)^k e^{-ts}}{k!}\big(k\delta^2+(k\delta + j)^2\big) dt}{h} = \frac{\bigintsss_{0}^{h} 2 \delta^2 ts + (\delta ts + j)^2dt}{h} \nonumber \\ &= hs\delta \Bigg(\delta + \frac{hs\delta}{3} + j\Bigg) + j^2 \end{aligned}$$ where first we have used the law of total expectation - the condition being the number of shifts within the interval. If there is no shift, then the distance is not increased between samplings, this case is included by the $e^{-ts}j^2$ term before the inner integral. Note that the inner integral is just $E(X+j)^2$ for a gamma - namely an Erlang($k,\frac{1}{\delta}$) - distributed random variable. When calculating the sum, we used the known formulas for $E(N^2)$, $E(N)$ and the Poisson distribution itself - where $N$ is a Poisson($ts$) distributed random variable. Implementation -------------- #### Discretisation For cost calculation purposes we would like to find a discrete stationary distribution which approximates the distribution of the monitored characteristic at the time of samplings. This requires the discretisation of the above defined functions, which in turn will allow us to construct a discrete time Markov chain with discrete state space. A vector of probabilities is needed to represent the shift size probability mass function $q_t()$ during a sampling interval: $$q_t(i)= \begin{dcases} n_t(0) &{\text{if }}i=0, \\ \sum_{k=1}^\infty n_t(k) \big(Y_k(i\Delta) - Y_k((i-1)\Delta)\big) &{\text{if }} i \in \mathbb{N}^{+} \nonumber \end{dcases}$$ where $\Delta$ stands for the length of an interval, one unit after discretisation. For $i=0$ the function is just the probability of no shift occurring. The discretised version of the repair size distribution can be written the following way: $$R(l,m) = P\Bigg(\frac{m}{l+1/2}\le R <\frac{m+1}{l+1/2}\Bigg) \nonumber$$ where $l$ is the number of discretised distances closer to $\mu_0$ than the current one - including $\mu_0$. $m$ is the index for the repair size segment we are interested in, with $m=0$ meaning perfect repair ($m\leq l$). The repair is assumed to move the expected value towards the target value by a random percentage, governed by $R()$. Even though discretisation is required for practical use of the framework, in reality the repair size distribution is continuous. To reflect this continuity in the background, the probability of perfect repair is set to be $0$ when the repair is random. $l$ is set to be $0$ when there is no repair, meaning $R(0,m)\equiv1$. The $1/2$ terms are necessary for correcting the overestimation of the distances from the target value, introduced by the discretisation: in reality, the distance can fall anywhere within the discretised interval, without correction the maximum of the possible values would be taken into account, which is an overestimation of the actual shift size. After correction the midpoint of the interval is used, which can still be somewhat biased, but in practical use with fine disretisation this effect is negligible. When the sampling probability depends on the time between samplings only, the model is unchanged, since both the time between samplings and the sampling probability can be continuous. However, when the probability also depends on the shift size, discretisation is required here as well: $$T_h(v) = P\Bigg(W_h<\frac{v+\zeta}{V_d+\zeta}-\frac{1}{2(V_d+\zeta)}\Bigg) \nonumber$$ now $v$ is the state distance from the target value in discretised units, and $V_d$ is the number of considered intervals - discretised shift sizes. The $\frac{1}{2(V_d+\zeta)}$ term is necessary for correcting the overestimation of the distances from the target value. The denominators contain simply $V_d+\zeta$ instead of $V_d+2\zeta$, because $v+\zeta$ is already strictly smaller than $V_d+\zeta$, since the smallest discretised state is 0 and thus the greatest is $V_d-1$. This ensures that the probability can never reach 1. Example curves for the successful sampling probability can be seen in Figure 2. It shows that longer time between samplings and greater distances from the target value increase the probability of successful sampling.  #### Transition Matrix and Stationary Distribution The transition probabilities can be written using the $\phi()$ process distribution, the $q_t()$ shift size distribution, the $R()$ repair size distribution and the $T_h()$ sampling probability. For the ease of notation let us define the $S()$ and $S'()$ functions, the first for target states without alarm and the latter for target states with alarm: $$\begin{aligned} S(g,v,l)& = \big[T_h(v) \phi(k-\Delta'(v)) + (1-T_h(v))\big] \smashoperator[lr]{\sum_{m\in\mathbb{N}^{0}, m\leq l, m\leq v}^{}} q_h(g-m)R(l,m) \nonumber \\ S'(g,v,l)& = T_h(v) \big[1-\phi\big(k-\Delta'(v)\big)\big] \smashoperator[lr]{\sum_{m\in\mathbb{N}^{0}, m\leq l, m\leq v}^{}} q_h(g-m)R(l,m) \nonumber \end{aligned}$$ where $\Delta'(v)$ is a function defined as $$\Delta'(v)= \begin{dcases} 0 &{\text{if }}v=0, \\ i\Delta-\frac{\Delta}{2} &{\text{if }} v = 1,\hdots,V_d-1 \nonumber \end{dcases}$$ $k-v\Delta$ would simply be the critical value minus the size of the shift in consideration. The $-\frac{\Delta}{2}$ term is added, because without it, the shift size would be overestimated. The total number of discretised disctances is $V_d$, thus the number of non-zero distances is $V_d-1$. $g$ is the total shift size possible in discretised units when moving to a given state, $v$ is the distance, measured as the number of discretised units from the target value, $l$ is the index of the actual partition and $m$ is the index for the repair size segment we are interested in with $m=0$ meaning perfect repair. Note, that as before, $l$ is set to be $0$ when there is no repair, meaning $R(0,m)\equiv1$. $S()$ and $S'()$ are functions which combine the three distributions into the probability we are interested in. The possible values of $m$ are restricted this way, because the parameter of the $q_h()$ function must be non-negative. A more intuitive explanation is, that we assumed positive shifts and a negative $g-m$ would imply a negative shift. It can be seen that the probability of a state without an alarm - the $S'()$ function - is increased with the probability of unsuccessful sampling - the $1- T_h(v)$ term. Of course it is still possible to not receive an alarm even though the sampling successfully occurred. Using the $S()$ and $S'()$ functions one can construct the transition matrix: $$\Pi= \begin{matrix} {\scalebox{0.9}{\mbox{\ensuremath{\displaystyle \begin{bmatrix} { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}S(0,0,0)\end{matrix}}}^{\text{In-control}}}$}S(0,0,0)} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}S(1,1,0) & S(2,2,0) & \dots\end{matrix}}}^{\text{Out-of-control}}}$}S(1,1,0) & S(2,2,0) & \dots} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}S'(0,0,0)\end{matrix}}}^{\text{False alarm}}}$}S'(0,0,0)} & { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}S'(1,1,0) & S'(2,2,0) & \dots\end{matrix}}}^{\text{True alarm}}}$}S'(1,1,0) & S'(2,2,0) & \dots}\\ 0 & S(0,1,0) & S(1,2,0) & \dots & 0 & S'(0,1,0) & S'(1,2,0) & \dots \\ 0 & 0 & S(0,2,0) & \dots & 0 & 0 & S'(0,2,0) & \dots \\ \vdots & \vdots & \vdots & & \vdots & \vdots & \vdots & \\ S(0,0,0) & S(1,1,0) & S(2,2,0) & \dots & S'(0,0,0) & S'(1,1,0) & S'(2,2,0) & \dots \\ 0 & S(1,1,1) & S(2,2,1) & \dots & 0 & S'(1,1,1) & S'(2,2,1) & \dots \\ 0 & S(1,1,2) & S(2,2,2) & \dots & 0 & S'(1,1,2) & S'(2,2,2) & \dots \\ \vdots & \vdots & \vdots & & \vdots & \vdots & \vdots & \ddots \end{bmatrix} }}}} {\scalebox{0.66}{\mbox{\ensuremath{\displaystyle \begin{aligned} &\left.\begin{matrix} \\[0.1em] \end{matrix} \right\}}}} \\ &\left.\begin{matrix} \\ \\[1.8em] \\ \end{matrix}\right\} \\ &\left.\begin{matrix} \\[0.1em] \end{matrix}\right\} \\ &\left.\begin{matrix} \\ \\[1.7em] \\ \end{matrix}\right\} \\ \end{aligned} \nonumber } \end{matrix}$$ The size of the matrix is $2V_d\times2V_d$ since every shift size has two states: one with and one without alarm. The first $V_d$ columns are states without alarm, the second $V_d$ are states with alarm. One can observe, that once the process leaves the healthy state it will never return. This is due to the nature of the imperfect repair we have discussed. The transition matrix above defines a Markov chain with a discrete, finite state space with one transient, inessential class (in-control and false alarm states) and one positive recurrent class (out-of-control and true alarm states). The starting distribution is assumed to be a deterministic distribution concentrated on the in-control state, which is to say that the process is assumed to always start from the target value. In finite Markov chains, the process leaves such a starting transient class with probability one. The problem of finding the stationary distribution of the Markov chain is thus reduced to finding a stationary distribution within the recurrent classes of the chain. Since there is a single positive recurrent class which is also aperiodic, we can apply the Perron–Frobenius theorem to find the stationary distribution[@Meyer]: Let $A$ be an $n \times n$, irreducible matrix with non-negative elements, $x\in \mathbb{R}^n$ and $$\lambda_0 = \lambda_0(A) = sup \{ \lambda : \exists x > 0 : Ax \geq \lambda x \}.$$ Then the following statements hold: - $\lambda_0$ is an eigenvalue of $A$ with algebraic multiplicity of one, and its corresponding eigenvector $x_0$ has strictly positive elements. - The absolute value of all other eigenvalues is less than or equals $\lambda_0$. - If $A$ is also aperiodic, then the absolute value of all other eigenvalues is less than $\lambda_0$. One can apply this theorem to find the stationary distribution of $\Pi$. If we consider now $\Pi$ without the inessential class - let us denote it with $\Pi'$ - then $\lambda_0(\Pi') = \lambda_0(\Pi'^T) = 1$. Moreover, the stationary distribution - which is the left eigenvector of $\Pi'$, normalised to sum to one - is unique and exists with strictly positive elements. Finding the stationary distribution is then reduced to solving the following equation: $\Pi'^T f_0 = f_0$, where $f_0$ is the left eigenvector of $\Pi'$. This amounts to solving $2V_d-2$ equations - the number of states minus the in-control and false alarm states - for the same number of variables, so the task is easily accomplishable. The stationary distribution is then: $$P = \frac{f_0}{\sum _{i{\mathop {=}}1}^{2V_d-2}f_{0_i}}. \nonumber$$ #### Cost Function Using the stationary distribution, the expected cost can be calculated: $$E(C) = c_s\frac{1}{h}(T' \cdot P) + \frac{\sum_{i=1}^{V_d-1} \big(c_{rb} + c_{rs}\Delta'^2(i)\big) P_{r_i}}{h} + c_o (A^2 \cdot P) \nonumber$$ This cost function incorporates similar terms as previously $(1)$. The first term deals with the sampling cost: $T' = \{ T_h(1), T_h(2),\ldots,T_h(V_d-1), T_h(1), T_h(2),\ldots,T_h(V_d-1)\}$ is the vector of successful sampling probabilities repeated in a way to fit the length and order of the stationary distribution. This first term uses the expected time between samplings, $\frac{1}{h}(T' \cdot P)$, instead of simply $h$, which would just be the minimal possible time between samplings. The second term deals with the repair costs and true alarm probabilities. $P_{r_i}$ is the true alarm probability for shift size $i$. The repair cost is partitioned into a base and shift-proportionate part: $c_{rb}$ and $c_{rs}$. The true alarm probability is used, since it is assumed that repair occurs only if there is an alarm. The last term is the average cost due to operation while the process is shifted. The connection between the distance from the target value and the resulting cost is assumed to be proportional to the squared distance. This is modeled using the $A^2$ vector, which contains the weighted averages of the expected squared distances from the target value between samplings: $$A^2_i = \sum_{j=1}^{i} E_h\big(H_{\Delta'(j)}^2\big) M_{ij} \nonumber$$ where $E_h\big(H_{\Delta'(j)}^2\big)$ is calculated using (5). $j$ indicates one of the possible starting distances immediately after the sampling, and $i$ indicates the state - shift - of the process at the current sampling. $M_{ij}$ is the probability that $\Delta'(j)$ will be the starting distance after the sampling, given that the current state is $i$. These probabilities can be written in a matrix form: $$\begin{matrix} {\scalebox{0.9}{\mbox{\ensuremath{\displaystyle \begin{bmatrix} { \makebox[0pt][l]{$\smash{\overbrace{\phantom{ \begin{matrix}\ \ \ \ \ 0 \ \ \ \ \ \ & \ \ \ \ 1 \ \ \ \ \ & \ \ \ \ 0 \ \ \ \ \ & \ \ \ \ \ 0 \ \ \ \ \ \ & \ \ \dots \ \ \ \end{matrix}}}^{\text{Distance from the target value starting from 0}}}$}\ \ \ \ \ 0 \ \ \ \ \ \ & \ \ \ \ 1 \ \ \ \ \ & \ \ \ \ 0 \ \ \ \ \ & \ \ \ \ \ 0 \ \ \ \ \ \ & \ \ \dots \ \ \ } \\ 0 & 0 & 1 & 0 & \dots \\ 0 & 0 & 0 & 1 & \dots \\ \vdots & \vdots & \vdots & \vdots & \\ 0 & 1 & 0 & 0 & \dots \\ 0 & R(1,0) & R(1,1) & 0 & \dots \\ 0 & R(2,0) & R(2,1) & R(2,2) & \dots \\ \vdots & \vdots & \vdots & \vdots & \end{bmatrix} }}}} {\scalebox{0.63}{\mbox{\ensuremath{\displaystyle \begin{aligned} &\left.\begin{matrix} \\ \\ \\ \\[1.5em] \\ \end{matrix}\right\}}}} \text{Out-of-control} \\ &\left.\begin{matrix} \\ \\ \\ \\[1.5em] \\ \end{matrix}\right\} \text{True alarm} \\ \end{aligned} \nonumber } \end{matrix}$$ It can be seen, that when the process is out-of-control without alarm, the distance is not changed. The probabilities for the alarm states are calculated using the $R()$ repair size distribution. So far only the expected cost was considered during the optimisation. In certain fields of application the reduction of the cost standard deviation can be just as or even more important than the minimisation of the expected cost. Motivated by this, let us consider now the weighted average of the cost expectation and the cost standard deviation: $$G = pE(C) + (1-p)\sigma(C) \nonumber$$ Now $G$ is the value to be minimised and $p$ is the weight of the expected cost ($0\le p\le 1$). The cost standard deviation can easily be calculated by modifying the cost function formula. All of the previous models can be used without any significant change, one simply changes the value to be minimised from $E(C)$ to $G$. Comparison of Different Scenarios --------------------------------- Implementation of the methods was done using the **** programming language. Supplying all the necessary parameters, one can calculate the $G$ value of the process for one time unit. It is also possible to minimise the $G$ value by finding the optimal time between samplings and control limit. All the other parameters are assumed to be known. The optimization step can be carried out using different tools, the results presented here were obtained with the built-in `optim()` **** function: box-constrained optimization using PORT routines[@Gay], namely the Broyden-Fletcher-Goldfarb-Shanno (L-BFGS-B) algorithm. The optimisation procedure can be divided into three steps. First, the transition matrix needs to be constructed from the given parameters. After this, the stationary distribution of the Markov chain is computed. In the third step, the $G$ value is calculated using the stationary distribution and the cost function. The optimisation algorithm then checks the resulting $G$ value and iterates the previous steps with different time between sampling and/or control limit parameters until it finds the optimal ones. #### Dependence on Parameters The testing was done using a moderate overall compliance level. This is required because low compliance can result in extreme behaviour in the optimisation such as taking the maximum or minimum allowed parameter value. An example for this is when the sampling probability depends on both the time between samplings and the state of the process: if the compliance level is also relatively low, then given certain parameter setups the optimal time between samplings will tend to zero. This will essentially create a self-reporting system as the increased distance from the target value will eventually increase the compliance and the sampling will take place. In a healthcare environment this would mean that the patient is told to come back for a control visit as soon as possible, but the patient will show up only when the symptoms are severe enough. This kind of process behaviour is undesirable in many cases, for example when the time between samplings cannot be set arbitrarily. The results obtained are shown in Figure 3.  One may observe the weak dependence of the critical value on the out-of-control cost. The time between samplings should be decreased with the increase of the out-of-control cost, and the average cost and the cost standard deviation increase with the out-of-control cost, as expected. The effect of the expected shift size on the critical value depends on the process standard deviation, as increased shift size results in markedly increased critical values only in the case of $\sigma=1$. Higher expected shift size entails less time between samplings, and increased expected cost and cost standard deviation. One can assess, that if the cost standard deviation is taken into account during the optimisation procedure ($p=0.9$), then lower critical value should be used with increased time between samplings. This is logical, because the increased time between sampling will lead to less frequent interventions, thus a less erratic process. Of course, at the same time we do not want to increase the expected cost, so the critical value is lowered. The cost standard deviation is decreased, as expected. What is interesting to note is that the expected costs have barely increased compared to the $p=1$ setup. This is important, because it shows that by changing the parameters appropriately, the cost standard deviation can be lowered - sometimes substantially - while the expected cost is only mildly increased. Several scenarios entail relatively large cost standard deviations, this is party due to the Taguchi-type loss funcion used during the calculations. The process standard deviation $\sigma$ has noticeable effect on the critical value only: lower critical values should be used for lower standard deviations, as expected. #### Sensitising Rules The effect of sensitising rules[@Montgomery] was investigated using simulation, since the implementation of theoretical calculations would have resulted in a hugely inflated transition matrix which poses a serious obstacle in both programming a running times. Optimal parameters were calculated for $p=0.9$, $\sigma=1$, $s=0.2$, $\delta=2$, $\alpha=1$, $\beta=3$, a $=0.01$, b $=1$, $c_s=1$, $c_{rb}=10$, $c_{rs}=10$, $c_{rs}=20$. The resulting optimal parameters were $h=0.38$ and $k=1.14$. This parameter setup entailed an expected cost of $E(C) = 37.75$ and cost standard deviation of $\sigma(C) = 150.33$. The probability of alarm states together was $\sum_iP_{r_i}=0.201$. Simulations were run for 50000 sampling intervals which equals 19000 unit time. Simulations from the first 100 sampling intervals were discarded as it was regarded as a burn-in stage. First, we present the baseline simulation results - the ones without additional rules. Overall, the simulation achieved an acceptable level of convergence to the theoretical stationary distribution. The empirical expected cost was $\overline C = 36.51$. The proportion of alarm states was 0.192. The calculation of the empirical standard deviation was carried out by taking into account the data of only every 30th sampling interval to deal with the autocorrelation of neighboring values. The empirical standard deviation using this method was $s^* = 199.37$. It is important to note that the empirical results can be somewhat inaccurate, depending on the particular simulation results and the parameters used. This is due to the large variance and slow convergence of the standard deviation. Nonetheless, for this particular scenario the theoretical and empirical results were quite close to each other, thus we will compare the effect of sensitising rules to this baseline simulation. The first rule we investigated was the one which produces an alarm in case of three consecutive points outside the $\frac{2}{3}k$ warning limit but still inside the control limit. Running the simulation with the extra rule resulted in $\overline C = 37.42$, $s^* = 171.57$ and a ratio of all alarm states together of 0.194, all of these values are within the hypothesised confidence interval. We can see no major difference in any of these values compared to the baseline. The second rule was the same as the first one, except this time two consecutive points outside the $\frac{2}{3}k$ warning limit were enough to produce an alarm signal. The results were $\overline C = 36.54$, $s^* = 190.91$ and 0.200 for the proportion of alarm states. Again, no apparent differences can be seen, but it is worth noting the proportion of alarm states is somewhat higher in this case than the at the baseline or the previous rule, and this was also seen with repeated simulation. Overall, the effect of the investigated sensitising rules seems to be minimal on the results. Further investigation is required of the the possible effects in case of other parameter setups and rules. Application =========== In the following paragraphs we show a real life healthcare example as the application of the previously introduced methods. Two approaches will be presented: one with and one without taking the standard deviation into account. The example problem is to minimise the healthcare burden generated by patients with high cardiovascular (CV) event risk. The model is built upon the relationship between the low-density lipoprotein (LDL) level and the risk of CV events, thus the LDL level is the process of interest.[@Boekholdt] Parameters were estimated from several sources. The list below gives information about the meaning of the parameter values in the healthcare setting and shows the source of the parameter estimations. [p[3cm]{}C[3.5cm]{}p[7cm]{}]{} Parameter value & Meaning & Parameter source $\mu_0$=3 mmol/l & Target value. & Set according to the European guideline for patients at risk.[@Garmendia] $\sigma$=0.1 mmol/l & Process standard deviation. & Estimated using real life data from Hungary, namely registry data through the Healthware Consulting Ltd. $\delta$=0.8/3 & Expected shift size, 0.8 increase in LDL per year on average. & Estimated with the help of a health professional. $s$=1/120 & Expected number of shifts in a day, 3 shifts per year on average. & Estimated with the help of a health professional. $\alpha = 0.027$, $\beta = 1.15$ & Parameters of the repair size beta distribution. & Estimated using an international study which included Hungary.[@Garmendia] $q = 0.1$, $z = 30$, & Parameters of the sampling probability logistic function. & Patient non-compliance in LDL controlling medicine is quite high, and this is represented through the parametrisation of the logistic function.[@Lardizabal] $c_s$=€5.78 & Sampling cost. & Estimated using the LDL testing cost and visit cost in Hungary. $c_o$=€5.30 & Shift-proportional daily out-of-control cost. & Estimated using real world data of cardiovascular event costs from Hungary $c_{rb}$=€11.50 & Base repair cost. & Estimated using the simvastatin therapy costs in Hungary $c_{rs}$=€8.63 & Shift-proportional repair cost. & Estimated using the simvastatin therapy costs in Hungary It is very difficult to give a good estimate for the type and the parameters of a distribution that properly models the non-compliance, thus the results here can at best be regarded as close approximations to a real life situation. This is not a limiting factor, as patients themselves can have vast differences in their behaviour, so evaluation of different scenarios are often required, and will also be presented here. Since high LDL levels rarely produce noticeable symptoms, the sampling probability only depends on the time between samplings, thus the sampling probability was modeled by the logistic function and not by the beta distribution.[@cholesterol] It is important to note that the proportional costs increase according to a Taguchi-type loss function, thus huge expenses can be generated if the patient’s health is highly out-of-control. #### Optimisation using only the cost expectation The optimal parameters for the case when the cost standard deviation was not taken into account were $56.57$ days and $0.143$ mmol/l for the time between samplings and the critical increase in the LDL level from the guideline value respectively. These parameters entailed an average daily cost of €$0.469$ and standard deviation of €$0.562$. This result is interesting, because the optimisation says that we should use a somewhat higher critical value than the one according to the guideline - 0 mmol/l critical increase would be the original 3 mmol/l value - but we should monitor the patients more often than usual - times of the LDL measurements are usually several months or years apart. It is important to note, that this is a strictly cost effective viewpoint which could be overwritten by a health professional. Nonetheless, the results provide a possible new approach to the therapeutic regime for controlling LDL level. Often, it is good to look at the interaction between the parameters and the resulting average cost, especially in situations where the optimal parameters cannot be used because of real life reasons.  The heat map of Figure 4 shows the average cost as the function of the different parameter values. The dot in the lightest area of the figure corresponds to the optimal cost. Any other point entails a higher average cost. It can clearly be seen that too low or high critical values will both increase the average daily cost. What is interesting - for this scenario - is that the change in the time between samplings entails relatively low change in the critical LDL increase: even if the time between control visits is changed the critical value should stay around $0.12 - 0.18$ mmol/l. #### Optimisation using cost expectation and cost standard deviation In this part, the cost standard deviation is also taken into account with $p=0.9$, thus the weight of the standard deviation in the calculation of $G$ is 0.1. The optimal parameters found by our approach were $64.76$ days and $0.129$ mmol/l for the time between samplings and critical increase in the LDL level respectively. These parameters entailed an average daily cost of €$0.477$ and standard deviation of €$0.418$. The inclusion of the cost standard deviation into the model has somewhat increased the time between control visits and decreased the critical value. The expected cost somewhat increased, while the cost standard deviation was moderately decreased. Figure 5 shows the previous heat map with non-compliance included in the model.  It can be seen that the elliptical shape of the heat map has not changed: the change in the time between control visits still does not entail great change in the critical value. #### Sensitivity Analysis As there were uncertainty about the estimation of several parameters, it is important to assess the effect of different parameter setups. The results for different out-of-control costs are plotted for both approaches. The results can be seen in Figure 6.  The critical value and time between samplings decrease and the average cost and cost standard deviation increase with higher out-of-control costs. Just as on the heat maps, one can observe here, that if the cost standard deviation is taken into account in the optimisation, then the critical value should be lowered and the time between samplings increased. One can observe, that a substantial decrease can be achieved in the cost standard deviation while the cost expectation barely changes. Uncertainty was also high around the estimation of the sampling probability. The sigmoid’s midpoint so far was $z=30$ days, meaning that the probability of sampling was $0.5$ at $h=30$ and increased with $h$. Figure 7 contains results for different $z$ values.  One can observe, that as the probability of successful sampling decreases - the value of $z$ is increased - the critical value decreases and the time between samplings increases. This can be explained by the increased uncertainty of the sampling: More time between samplings entails higher patient compliance, and when the visit occurs a low critical value is used to ensure treatment. The cost expectation and standard deviation increases with lower sampling probabilities. There are only minor differences between the stationary distributions, nonetheless it can be seen that lower sampling probability is associated with higher probabilities for greater distances from the target value. The last panel shows how the sampling probability decreases with increasing $z$ values for a fixed $h$. Conclusions =========== Cost-optimal control charts based on predominantly Duncan’s cycle model are found in a wide variety of areas. Even though the benefit of using these charts in industrial and engineering settings is unquestionable, the numerous assumptions needed about the processes makes the applicability of the traditional models problematic in certain environments. Motivated by the desire to apply cost-optimal control charts on processes with erratic behaviour and imperfect repair mechanism - such as ones found in healthcare - this paper presented a Markov chain-based framework for the generalisation of these control charts, which enabled the loosening of some of the usual assumptions. Cost-optimisation of control charts are usually carried out by finding the optimal critical value, time between samplings and sample size. Our work concentrated on the monitoring of a single element at a time - e.g. a patient - thus the methods presented here always used a sample size of $1$. Building on and expanding the work of Zempléni et al.[@Zempleni] we discussed three types of generalisations: the random shift size, the imperfect repair and the non-compliance. The random shift size means that only the distribution of the shift size and its parameters are assumed to be known. This let us monitor processes which are potentially drifting in nature. The second generalisation - the imperfect repair - assumed that the process stays out of control even after repair, but on a level closer to the target value than before. This type of repair response is often observed in treatments in healthcare, but may be found in other areas too. The third generalisation was intended to help the modeling of patient or staff non-compliance. We implemented this concept in a way that allows sampling times to be skipped by a probability governed by a distribution or function with known parameters. Since the processes modeled with the above loosened assumptions can create complicated trajectories between samplings, the mathematical description of these was also necessary. We proposed an expectation calculation method of a function of the values taken on by the process between samplings. The application of this proposition while assuming exponentially distributed shift sizes, Poisson distributed event numbers and Taguchi-type loss function yielded a compact formula for the expectation. We implemented our theoretical results in the **** programming language and investigated the effect of parameter estimation uncertainty on the optimal parameters and the resulting expected cost and cost standard deviation. We also tested the effect of involving the cost standard deviation in the optimisation procedure itself. We found that typically the critical value increases and the time between samplings decreases with the expected shift size. Also, higher expected shift sizes entail higher expected costs and cost standard deviations. It was seen that with the increase of the out-of-control cost - in most cases - the critical value stagnated, the time between samplings decreased, and the expected cost and the cost standard deviation increased. The involvement of the cost standard deviation in the optimisation procedure lowered the standard deviation while the cost expectation barely increased. We have found no evidence that sensitising rules - such as values outside the warning limit - would change the results substantially. We presented an example of real-life application involving low-density lipoprotein monitoring. The results indicated that the cost-optimal critical value is somewhat higher and the cost-optimal time between control visits is less than the ones usually used according to medical guidelines. In the era of Industry 4.0, the cost-optimal control charts presented here can be applied to a wider range of processes than the traditional ones. Nonetheless there are still areas worth investigating. One of the features still missing is the proper modeling of the repair procedure, since it was assumed to be an instantaneous event, which may not be appropriate in many situations. The mathematical elaboration of a continuous model involving e.g. time series could also be beneficial. Acknowledgements ================ The authors would like to express their gratitude towards the Healthware Consulting Ltd. (Budapest, Hungary) for their help in providing data and professional knowledge during the writing of the Application section. The work was supported by the project EFOP-3.6.2-16-2017-00015, which was supported by the European Union and co-financed by the European Social Fund. [9]{} Montgomery DC. Introduction to statistical quality control, Sixth Edition. *John Wiley & Sons Association* 2009; 51/274. Duncan AJ. The Economic Design of $X$-Charts Used to Maintain Current Control of a Process. *Journal of the American Statistical Association* 1956; 51/274: 228-242. DOI: 10.1080/ 01621459.1956.10501322 Mortarino C. Duncan’s Model for $\bar{X}$-Control Charts: Sensitivity Analysis to Input Parameters. *Quality and Reliability Engineering International* 2010; 26/1: 17-26. DOI: 10.1002/ qre.1026 Zhu W, Park C. edcc: An R Package for the Economic Design of the Control Chart. *Journal of Statistical Software* 2013; 52/9. DOI: 10.18637/jss.v052.i09 Zhou C, Zhang W. Recurrence Plot Based Damage Detection Method by Integrating $T^2$ Control Chart. *Entropy* 2015; 17/5: 2624-2641. DOI: 10.3390/e17052624 Sales RF, Vitale R, de Lima SM, Pimentel MF, Stragevitch L, Ferrer A. Multivariate statistical process control charts for batch monitoring of transesterification reactions for biodiesel production based on near-infrared spectroscopy. *Computers & Chemical Engineering* 2016; 94: 343-353. DOI: 10.1016/ j.compchemeng. 2016.08.013 Thor J, Lundberg J, Ask J, Olsson J, Carli C, Härenstam KP, Brommels M. Application of statistical process control in healthcare improvement: systematic review, *Quality and Safety in Health Care* 2007; 16/5: 387–399. DOI: 10.1136/qshc.2006.022194 Suman G, Prajapati D. Control chart applications in healthcare: a literature review. *International Journal of Metrology and Quality Engineering* 2018; 9/5. DOI: 10.1051/ijmqe/ 2018003 Padula WV, Duffy MP, Yilmaz T, Mishra MK, Integrating systems engineering practice with health-care delivery. *Health Systems* 2014; 3/3: 159–164. DOI: 10.1057/hs.2014.3 Duclos A, Touzet S, Soardo P, Colin C, Peix JL, Lifant JC. Quality monitoring in thyroid surgery using the Shewhart control chart. *British Journal of Surgery* 2009; 96/2: 171-174. DOI: 10.1002/bjs.6418 Coory M, Duckett S, Sketcher-Baker K. Using control charts to monitor quality of hospital care with administrative data. *International Journal for Quality in Health Care* 2008; 20/1: 31–39. DOI: 10.1093/intqhc/mzm060 Correia F, Neveda R, Oliveira P. Chronic respiratory patient control by multivariate charts. *International Journal of Health Care Quality Assurance* 2011; 24/8: 621–643. DOI: 10.1108/ 09526861111174198 Johnson CC, Martin M. Effectiveness of a physician education program in reducing consumption of hospital resources in elective total hip replacement. *Southern Medical Journal* 1996; 89/3: 282-289. PMID: 8604457 Shaha SH. Acuity systems and control charting. *Quality Management in Health Care* 1995; 3/3: 22-30. PMID: 10143553 Stewart LJ, Greisler D. Measuring primary care practice performance within an integrated delivery system: a case study. *Journal of Healthcare Management* 2002; 47/4: 250-261. PMID: 12221746 Zempléni A, Véber M, Duarte B, Saraiva P. Control Charts: A cost-optimization approach for processes with random shifts. *Applied Stochastic Models in Business and Industry* 2004; 20/3: 185-200. DOI: 10.1002/asmb.521 Deming WE. The new economics for industry, government, education. *MIT Press* 2018; 151. Meyer C. Matrix analysis and applied linear algebra. *SIAM* 2000; 663-669. Gay DM. Usage summary for selected optimization routines. *Computing Science Technical Report, AT&T Bell Laboratories, Murray Hill* 1990; 153: 1-21. Boekholdt SM, Arsenault BJ, Mora S, Pedersen TR, LaRosa JC, Nestel PJ, Simes RJ, Durrington P, Hitman GA, Welch KM, DeMicco DA, Zwinderman AH, Clearfield MB, Downs JR, Tonkin AM, Colhoun HM, Gotto AM, Ridker PM, Kastelein JJP. Association of LDL cholesterol, non–HDL cholesterol, and apolipoprotein B levels with risk of cardiovascular events among patients treated with statins: a meta-analysis. *Journal of the American Medical Association* 2012; 307/12: 1302-1309. DOI: 10.1001/jama.2012.366 Garmendia F, Brown AS, Reiber I, Adams PC. Attaining United States and European Guideline LDL-cholesterol Levels with Simvastatin in Patients with Coronary Heart Disease (the GOALLS Study). *Current Medical Research and Opinion* 2000; 16/3: 208-219. PMID: 11191012 Lardizabal JA, Deedwania PC. Benefits of statin therapy and compliance in high risk cardiovascular patients. *Vascular health and risk management* 2010; 6: 843–853. DOI: 10.2147/VHRM.S9474 High cholesterol: Overview. *Institute for Quality and Efficiency in Health Care* \[1 December 2018\] Bookshelf ID: NBK279318 https://www.ncbi.nlm.nih.gov/books/NBK279318/

ArXiv

--- author: - The ATLAS Collaboration bibliography: - 'sct2011.bib' title: Operation and performance of the ATLAS semiconductor tracker --- =1 Introduction ============ The SCT detector ================ Operation ========= Offline reconstruction and simulation ===================================== Monitoring and data quality assessment ====================================== Performance =========== Detector occupancy ------------------ Noise ----- Alignment stability {#sec:fsires} ------------------- Intrinsic hit efficiency ------------------------ Lorentz angle ------------- Energy loss and particle identification --------------------------------------- Measurement of $\delta$-ray production -------------------------------------- Radiation effects ================= Conclusions =========== The operation and performance of the ATLAS semiconductor tracker during 2009–2013 are described in this paper. During this period, more than 99% of detector modules were operational, and more than 99% of data collected by the ATLAS experiment had good SCT data quality. The frequency-scannning interferometry system showed the position of the detector to be stable at the micron level over long periods of time. Measurements of the increase in leakage currents with time are consistent with the radiation-damage predictions. The differences between data and simulation are typically less than 30%. This level of agreement exceeds expectations, and provides confidence in the fluence predictions. The verification of the simulations will be repeated at higher beam energies in future. Single event upsets have been identified and measured in the barrel module data, and a strategy for their mitigation implemented. The detector occupancy was found to vary linearly with the number of interactions per bunch crossing, up to the maximum of 70 interactions per crossing, where it is less than 2% in the innermost barrel layer (which has highest occupancy). The intrinsic hit efficiency of the detector was measured to be (99.74$\pm$0.04)%, and the noise occupancies of almost all chips remained below the design requirement of $5 \times 10^{-4}$. Measured values of the Lorentz angle are compatible with model predictions within at most twice the estimated uncertainties on those predictions. The measured values for sensors with $<$100$>$ crystal orientation are approximately $1^{\circ}$ lower than for those with $<$111$>$ crystal orientation, contrary to the expectation that a higher expected charge-carrier mobility in the sensors with $<$100$>$ crystal orientation should result in a higher value of the Lorentz angle. Despite the binary readout, some particle identification from energy-loss measurements is possible: the discriminating power arises from the number of time bins above threshold and cluster widths. The position of the proton energy-loss peak was found to be stable at the 5–10% level during 2010–2012. The position of this peak may become a useful tool for monitoring radiation damage in future. The production of $\delta$-rays in the silicon sensors was measured, and is found to be in good agreement with expectations.

ArXiv

--- abstract: 'We study dispersion properties of linear surface gravity waves propagating in an arbitrary direction atop a current profile of depth-varying magnitude using a piecewise linear approximation, and develop a robust numerical framework for practical calculation. The method has been much used in the past [[for the case of waves propagating along the same axis as the background current]{}]{}, and we herein extend and apply it to [[problems with an arbitrary angle between the wave propagation and current directions]{}]{}. Being valid for all wavelengths without loss of accuracy, the scheme is particularly well suited to solve problems involving [[a broad range of wave vectors, such as ship waves and Cauchy-Poisson initial value problems for example.]{}]{} We examine the group and phase velocities over different wavelength regimes and current profiles, highlighting characteristics due to the depth-variable vorticity. We show an example application to ship waves on an arbitrary current profile, and demonstrate qualitative differences in the wake patterns between concave down and concave up profiles when compared to a constant shear profile with equal depth-averaged vorticity. [[We also discuss]{}]{} the nature of [[additional]{}]{} solutions to the dispersion relation when using the piecewise-linear model. [[These are vorticity waves, drifting vortical structures which are artifacts of the piecewise model. They are absent for a smooth profile and are spurious in the present context.]{}]{}' author: - 'Benjamin K. Smeltzer' - 'Simen Å. Ellingsen' bibliography: - 'nlayer\_references.bib' title: 'Surface waves on arbitrary vertically-sheared currents' --- Introduction ============ A complete understanding of surface water wave propagation on a background current profile is of great importance in areas within oceanography, marine and coastal engineering, and naval architecture[@Pere]. The presence of an underlying current modifies the wave dispersion potentially affecting key quantities such as wave loads on structures, wave propagation near coastlines, or ship wave resistance. Furthermore, measurements of wave frequencies at known wavelengths (e.g. using high-frequency radar) can be used to infer the underlying current profile[@StewartJoy; @Graber96; @Fernandez96; @Lund15], relevant for predicting storm surges and understanding the mechanisms of climate change[@Lund15]. Many studies and models in these areas have used simple velocity profiles such as depth-uniform or linear depth dependence, largely due to mathematical tractability as analytical solutions exist only for a select few of these current profiles[@Pere]. Various approximation techniques have been developed for more realistic profiles[@StewartJoy; @KirbyChen; @Skop87; @SwanJames; @Shrira93], yet these have limited range of applicability. The goal of this work is to demonstrate an approximation method for calculating the dispersion relation on an arbitrary current profile in three dimensions valid for all wavelengths that is suitable for practical calculations by engineers. For the purposes of this work we consider [[infinitesimal]{}]{}-amplitude surface waves propagating on a background rotational current flow that is [[steady,]{}]{} incompressible[[,]{}]{} and inviscid. [[Although]{}]{} viscosity [[is neglected for the wave motion]{}]{}, viscous effects are certainly involved in [[generating the shear current itself. We are not, however, concerned with how the background current may have come about. Thus, in]{}]{} this small wave amplitude regime, we assume the wave-current interaction to be unidirectional: the current affect[[s]{}]{} the wave motion but not [[*vice versa*]{}]{}. The current profiles we consider are of depth-variable magnitude yet constant direction [[and are assumed to be homogeneous in the horizontal directions]{}]{}. The vast majority of the body of work on surface waves and shear currents [[considers wave propagation parallel or anti-parallel to the current, which we refer to as two dimensional (2D) with as single vertical and horizontal spatial axis. The generalized case consists of a horizontal plane with waves propagating at an arbitrary oblique angle to the direction of the current, referred to three dimensional (3D).]{}]{} [[It was recently shown]{}]{} how 3D solutions to the Euler equations in the presence of a linear shear current (constant vorticity) can be used to solve classical problems such as ship waves and ring waves[[[@Ell14; @Li16; @Ellcp]]{}]{}. Shear currents were found to have the potential to significantly alter the characteristics of wave propagation in inherently 3D problems, evidenced by the behavior of ship waves as well as solutions to Cauchy-Poisson initial-value problems[@Ellcp]. In the former case it was shown that the Kelvin angle (the maximum wake angle with appreciable wave energy) is a function of shear strength and orientation angle of the current relative to ship motion. For initial-value problems it was shown that the difference between phase velocity and group velocity can be very different in propagation directions where waves are assisted or inhibited by the sub-surface shear, respectively, leading to anisotropic behavior in the time evolution of an initial surface disturbance. For most realistic current profiles however, the vorticity is not constant with depth. To treat profiles with arbitrary current depth-dependence, various approximation techniques have been developed, typically involving expansions in a small parameter representing the magnitude of the current velocity relative to the phase velocity of the waves[@StewartJoy; @KirbyChen; @Skop87; @SwanJames], or the departure from a velocity potential solution[@Shrira93]. These methods have been used for many practical calculations such as inferring the background current from phase velocity measurements[@StewartJoy; @Graber96; @Fernandez96; @Lund15], yet complications occur when applying them to problems involving the entire wave-spectrum as their accuracy [[is difficult to predict *a priori* and can suffer]{}]{} in certain wavelength regimes. Many problems such as the above-mentioned ship waves and ring waves are conveniently solved in Fourier space, whereupon integration over all horizontal wave vectors is performed, and a fast dispersion calculation method giving the same approximation accuracy over the entire wave-spectrum at [[little]{}]{} extra cost is thus desired. In this work we use a method based on a piecewise linear approximation (PLA) to the background current’s velocity profile. The profile is divided into vertical layers each assumed to have constant vorticity. Within each layer, solutions to the linearized Euler equations are found, and these solutions are matched appropriately at the layer interfaces to yield the full solution over the entire domain[@DrazinReed]. This method has been extensively used in the past perhaps first by Lord Rayleigh [@rayleigh1879; @rayleigh1892], mostly with 2-3 layers. The simplest two-layer case with constant shear in the lower layer and constant current in the upper layer (no shear) was analyzed by Thompson[@thomp49], and constant shear in an upper layer on an infinite lower layer of zero current by Taylor[@taylor] to investigate the potential of a current produced by a bubble curtain used as a breakwater to stop waves. The generalized two-layer result was later given by Dalrymple[@DalrTR]. Zhang[@Zhang05] compared the PLA method to other approximation methods[@StewartJoy; @KirbyChen; @Shrira01], showing how it is able to accurately calculate dispersion properties at all wavelengths. [[With]{}]{} the implementation described herein we calculate phase velocities at the $1\%$ accuracy level or better with just $4$-$5$ layers in the entire wave vector plane, making the method calculationally cheap, conceptually simple, easily implement and hence ideal for Fourier transformation purposes. The primary difficulty in using the piecewise linear approximation involves extra solutions to the polynomial equations that are solved [[in order]{}]{} to find the phase velocity for a given wavevector[@thomp49; @Shrira01; @Zhang05]. These solutions [[are spurious in the present context, and]{}]{} have phase velocities near the velocity of the background flow at the layer interfaces[@Zhang05][[. They are artifacts introduced by]{}]{} the discontinuities in the shear, something we discuss [[further]{}]{} in Section \[sec:extrasol\]. Despite the complication of discarding the spurious solutions, the piecewise linear approximation has been much used in studying the stability of small disturbances on shear flows using a perturbation type approach[@LH98; @Cap91; @Cap92]. In some cases there is disagreement with work considering similar smooth flows, raising questions about the accuracy of the method[@Eng00; @Mor98; @Shrira01]. Zhang[@Zhang05] addressed these issues showing the convergence of the piecewise profile to be $O(\Delta z^2)$ where $\Delta z$ is the layer thickness, and studied the [[extra]{}]{} solutions. As many applications are inherently 3D (initial value problems, ship waves, radiation, refraction) [[we]{}]{} implement and demonstrate the PLA method in 3D. Further analysis of the nature of the [[extra]{}]{} solutions is given by considering a simplified two-layer fluid. We demonstrate the convergence and approximation accuracy over a range of wavelength scales as a function of the number of layers. The PLA is then further applied to [[calculating]{}]{} the directional dependence of the group and phase velocities on two profiles with non-constant vorticity. Finally, we [[solve the classical ship wave problem on an arbitrary shear profile as an example application.]{}]{} Formulation of the model {#sec:der} ======================== We consider 3D surface waves propagating on a depth-varying current $U(z)$ oriented along the horizontal $x$-axis. The velocity field can be written: $$\mathbf{v} = \left(U(z) + \hat{u}, \hat{v}, \hat{w}\right)$$ following the notation of Ellingsen[@Ell14] where hatted quantities are assumed to be small perturbations due to the waves. We assume a progressive surface wave with infinitesimal surface height $\hat{\zeta}(x,y)$ in the horizontal plane, wavevector $\mathbf{k} = (k_x,k_y)$ making an angle $\theta$ to the $x$-axis[[, pressure $P = -\rho g z + \hat{p}$,]{}]{} and frequency $\omega$ such that the velocity and pressure perturbations are expressed as: $$\left(\hat{\zeta},\hat{u}, \hat{v}, \hat{w}, \hat{p}\right) = \left(\zeta,u(z),v(z),w(z),p(z)\right)e^{i\left(\mathbf{k}\cdot\mathbf{r}-\omega t\right)},$$ where $\mathbf{r}$ is the position vector in the $xy$-plane. We will artificially divide up the fluid column into $N$ layers in the vertical direction so that the vorticity of the background flow be constant inside each layer, as shown in Fig. \[fig:geom\]. Each layer has a thickness $h_j$ and vertical coordinate within each layer $z_j = z + \sum_{l=1}^{j-1}h_l$ as shown in Fig. \[fig:geom\]. The approximate current profile in layer $j$ is $$U_j^{PL}(z_j) = U_{j-1} + S_jz_j,$$ where $U_j \equiv U (z = -\sum_{l=1}^jh_l)$ is the value of the non-linear current profile at the layer interfaces, and $S_j = (U_j-U_{j+1})/h_j$ is the layer mean vorticity. ![The geometry and definitions used in the text demonstrated for a 3-layer piecewise linear profile. Vertical coordinates $z_j$ are defined within each layer. The current is assumed to be oriented along the $x$-direction, and varies linearly within each layer described by the function $U_j^{PL}(z_j)$.[]{data-label="fig:geom"}](geometry_oneCol.eps) The linearized Euler and continuity equations inside layer $j$ are $$\begin{aligned} \left[-i\omega + ik_xU_j^{PL}(z_j)\right]u_j + S_jw_j =& -ik_xp_j/\rho;\label{eq:eulerx} \\ \left[-i\omega + ik_xU_j^{PL}(z_j)\right]v_j =& -ik_yp_j/\rho; \label{eq:eulery}\\ \left[-i\omega + ik_xU_j^{PL}(z_j)\right]w_j =& -{}p_j^\prime/\rho; \label{eq:eulerz}\\ ik_xu + ik_yv + w' =&0, \label{eq:cont}\end{aligned}$$ where the prime denotes $\partial/\partial z$. In the interior of each layer this system of equations leads to the Rayleigh equation for the vertical velocity component $w_j$ $$w_j^{\prime\prime} - k^2w_j = 0, \label{eq:rayleigh}$$ with general solution $$w_j = A_j\sinh{k(z_j+h_j)} + B_j\cosh{k(z_j+h_j)}, \label{eq:wj}$$ where $k = |\mathbf{k}|$. When $A_j$ and $B_j$ are known, the two other velocity components and the pressure can be found by inserting this solution back into Eqs. -. There are thus $2N$ unknowns for $N$ layers. Four types of boundary conditions [[for $w$ and $p$]{}]{} provide the necessary $2N$ equations; [[one ]{}]{} at the bottom[[, one at the]{}]{} free surface, [[and $2N-2$]{}]{} matching conditions at the layer interfaces. Considering finite, uniform depth, the vertical velocity component must vanish at the bottom ($z_N = -h_N$), $$B_N = 0. \label{eq:bottom}$$ The vertical velocity component $w(z)$ must be continuous everywhere leading to a kinematic boundary condition at the layer interfaces $z_j = h_j$: $w_j(-h_j) = w_{j+1}(0)$, which gives $$B_j = A_{j+1}\sinh kh_{j+1} + B_{j+1}\cosh kh_{j+1}. \qquad j\in (1,N-1). \label{eq:vmatch}$$ The second matching condition at the interfaces is the continuity of pressure (a dynamic boundary condition)[[. The pressure in each layer]{}]{} can be formulated in terms of $w_j$ and its derivative as $$-k^2\frac{p_j}{\rho} = -ik_xS_jw_j + \left[-i\omega + ik_xU_j^{PL}(z_j)\right]w'_j. \label{eq:pressure}$$ Inserting Eq.  yields $$\begin{aligned} -k^2\frac{p_j}{\rho} &= \left[-ik_xS_{j}A_{j} + \left(-i\omega + ik_xU_j^{PL}(z_j)\right)kB_{j}\right]\sinh kh_{j} \nonumber\\ &+\left[-ik_xS_{j}B_{j} + \left(-i\omega + ik_xU_j^{PL}(z_j)\right)kA_{j}\right]\cosh kh_{j}.\end{aligned}$$ Continuity of pressure requires that $p_j(z_j=-h_j) = p_{j+1}(z_{j+1}=0)$. We further insert Eq.  to eliminate coefficients $A_{j\neq 1}$ and express a combined kinematic and dynamic condition at the layer interfaces in the form $$\begin{aligned} k\sigma_1 A_1 + \left[\gamma_1 - \gamma_2 - k\sigma_1\coth kh_2\right]B_1 + k\sigma_1\left[\cosh kh_2\coth kh_2 - \sinh kh_2\right]B_2 &=0 \label{eq:pmatcha}\\ \left(k\sigma_j/\sinh kh_j\right) B_{j-1} + \left[\gamma_j - \gamma_{j+1} - k\sigma_j\left(\coth kh_j+\coth kh_{j+1}\right)\right]B_j & \nonumber\\ +k\sigma_j\left[\cosh kh_{j+1}\coth kh_{j+1} - \sinh kh_{j+1}\right]B_{j+1} &= 0, \label{eq:pmatchb}\end{aligned}$$ for $j\in (2,N-1)$, where $\gamma_j \equiv -ik_xS_j$ and $\sigma_j \equiv -i\omega + ik_xU_j$. The final condition is at the free surface $\zeta$, a combined kinematic and dynamic boundary condition, expressed here neglecting surface tension: $$A_1\left[\gamma_1\sigma_0\tanh kh_1 + \sigma_0^2k + gk^2\tanh kh_1\right] + B_1\left[\gamma_1\sigma_0 + \sigma_0^2k\tanh kh_1 + gk^2\right] = 0. \label{eq:dkc}$$ Eqs. (\[eq:bottom\]), (\[eq:pmatcha\]), (\[eq:pmatchb\]), and (\[eq:dkc\]) for $A_1$ and $B_j$ form an $(N+1)\times (N+1)$ homogeneous linear system with coefficient matrix $\mathbf{M}$. The determinant of $\mathbf{M}$ must be zero for non-trivial solutions of the vertical velocity coefficients to exist, leading to a polynomial equation for the unknown $\omega$. This equation is degree $N+1$, giving in general $N+1$ solutions for the dispersion relation $\omega(\mathbf{k})$, or the phase velocity $C(\mathbf{k}) = \omega \mathbf{k}/k^2$. [[A computationally efficient method for finding $\omega(\mathbf{k})$ scalable to many layers ($N>100$) involves formulating the linear system as a quadratic eigenvalue problem, expressing the coefficient matrix $\mathbf{M} = \mathbf{L_0} + \mathbf{L_1}\omega + \mathbf{L_2}\omega^2$. The eigenvalues $\omega$ and corresponding eigenvectors $\mathbf{x}$ of the resulting equation $\left(\mathbf{L_0} + \mathbf{L_1}\omega + \mathbf{L_2}\omega^2\right)\mathbf{x}=0$ can be found using a standard polynomial eigenvalue solver.]{}]{}  [[As an illustrative example of the $N+1$ solutions to the PLA dispersion relation, we consider a current profile of constant vorticity $U(z) = z\sqrt{g/h}$ with total depth $h$,]{}]{} divided artificially into 4 layers [[($N=4$). Fig. \[fig:roots\] shows the $N+1$ solutions for a given value of $kh$ for wave propagation along the axis of the direction of $U(z)$ ($\theta = 0$). For comparison, we show the two phase velocity solutions from the single layer ($N=1$) case as solid red lines. The phase velocity solutions from the $N=1$ case correspond]{}]{} to phase [[velocities]{}]{} of plane waves propagating in directions $\mathbf{k}$ and $-\mathbf{k}$, respectively. [[For $N=4$ there are two solutions that agree with the [[well known exact solutions]{}]{}, and three additional solutions that have phase velocities approximately equal the value of the current profile at the three layer interfaces, $U(-h/4)$, $U(-h/2)$, and $U(-3h/4)$ respectively. The vertical velocity profiles $w(z)$ for each of the phase velocity solutions is plotted in the small sub-figures in Fig. \[fig:roots\]. For two velocity profiles corresponding to the solutions from the $N=1$, $w(z)$ is peaked at the surface, while for the three [[extra solutions]{}]{}, $w(z)$ is peaked at layer interfaces $z = -h/4$, $z = -h/2$, and $z = -3h/4$. Given the constant vorticity of the fluid, the use of the PLA with $N>1$ [[cannot]{}]{} change the physical nature of the problem, [[highlighting the spuriousness (in this context) of]{}]{} the extra solutions.]{}]{} Nature of additional solutions: vorticity waves {#sec:extrasol} =============================================== The presence of sharp changes in vorticity allows the model system of Fig. \[fig:geom\] to support $N-1$ wave solutions of a different physical nature than the gravity waves at the free surface. These have been studied to some extent in the context of internal waves in the atmosphere[@holmboe62; @caulfield94], and are referred to as Rayleigh waves, from being first discussed by Lord Rayleigh[@rayleigh1879], [[as]{}]{} counter-propagating Rossby wave[[s]{}]{}[@heifetz99] or a[[s]{}]{} vorticity wave[[s]{}]{}[@carpenter11]. A very readable review of the physical mechanism involved is found in section 4 of Ref. , and we shall only here recount a few main points in order to understand their appearence in the piecewise-linear model. Let us consider the simplest possible model of a kink in a doubly infinite piecewise linear velocity profile, as shown in Fig. \[fig:kink\]. Let the basic velocity be $U(z)=U_0 + S^\pm z$ so that $S^-$ and $S^+$ are the vorticities below and above the kink, respectively[[, and let the]{}]{} fluid [[be]{}]{} uniform. Assume moreover that the interface between the regions of different vorticity is slightly perturbed from $0$ to $\hat{\zeta}(x,t)\propto \exp({\mathrm{i}}{\mathbf{k}}\cdot\mathbf{r}-{\mathrm{i}}kCt)$. The Rayleigh equation again gives simple solutions for $w$ each side of the boundary: $w^\pm = A {\mathrm{e}}^{\mp kz}$ with upper (lower) sign again denoting $z-\hat{\zeta}$ positive (negative). From the Euler equations we have, similarly to Eq. , [$$}k p^\pm/\rho = {\mathrm{i}}(C-U_0\cos\theta)(w^\pm)' + {\mathrm{i}}S^\pm \cos \theta w^\pm. {$$]{}Demanding continuity of pressure at $z=0$ (linearized dynamic boundary condition) gives [$$}\label{cdrift} C({\mathbf{k}}) = U_0\cos\theta + \frac{(S^+-S^-)\cos\theta}{2k}. {$$]{}Noticing that $C= U_0\cos\theta$ would represent a perturbation that is simply drifting passively downstream, this wave mode has a nonzero intrinsic phase velocity when $S^+\neq S^-$. The vorticity wave is not a wave in the same sense as surface gravity waves, but is better thought of as a train of vortical structures which has come about due to perturbation $\hat{\zeta}$. Where $\hat{\zeta}<0$, fluid of vorticity $S^+$ is brought into the domain of background vorticity $S^-$, and *vice versa*. The resulting train of vortical structures is instructively illustrated e.g. in figure 2 of Ref.  and figure 4 of Ref. . (The vorticity equation also has an additional term when $\theta\neq 0,\pi$, due to undulations of the vortex lines of the background flow, see Ref. ). The key point to notice in the present context is that such a wave mode can only be supported when there is a sharp change on a vertical lengthscale of a wave amplitude or less. Consequently, in a linear wave theory where wave amplitudes are infinitesimal, a smooth velocity profile will not support these modes. They are, in the present context, purely an artifact of the piecewise linear model. Note that although ‘spurious’ in the system we consider here, vorticity waves *can* be observed in other systems, such as atmospheric waves. Given the artificial nature of the $N-1$ extra solutions in the context of linear waves considered herein, a technique for identifying and discarding them is necessary. The most natural method is the consider the resulting vertical velocity profiles shown in Fig. \[fig:roots\]. The physical solutions corresponding to surface wave propagation have vertical velocity profiles peaked at the free surface, whereas in the case of the extra solutions, the velocity is peaked at a layer interface. By comparing $w(z)$ evaluated at the surface and interface heights for each of the $N+1$ solutions to the dispersion relation for a given $\mathbf{k}$, the desired surface wave solutions can be selected. A simpler[[, pragmatic]{}]{} method without considering $w(z)$ can be used for [[fast-moving wave modes which often occur for sufficiently]{}]{} small values of [[$|\mathbf{k}|$, by]{}]{} exploiting the property that extra solutions have phase velocities equal to the background flow at some depth[[, i.e., $\min[U(z)\cos\theta]\leq C\leq \max[U(z)\cos\theta]$]{}]{}. For [[a range of]{}]{} long wavelengths [[one or both of the desired phase velocities then exceed this range]{}]{}, and [[can be immediately recognized as physical]{}]{}. [[For monotonous $U(z)$ this will always be true for the shear-inhibited solution propagating along the current (It is possible in principle to construct a $U^\text{PL}(z)$ with very sharp kinks whose extra phase velocity solutions lie outside the range of $U(z)$. In keeping with the pragmatic nature of this method we may safely neglect this possibility since it does not occur for even very rough models of realistic flows.)]{}]{} Results ======= The aim of this section is twofold. First, we verify and validate the numerical scheme as well as investigate the accuracy as a function of the number of layers. Secondly, we demonstrate the utility of the model in finding the dispersion relation for oft-occurring general profiles, and apply it to the ship wave problem, highlighting as an example of the use for Fourier transformation in the horizontal plane. We highlight some of the notable wave propagation characteristics that occur due to the curvature of the velocity profile as compared to a couette flow model. Verification and Validation --------------------------- The convergence of the piecewise linear approximation when the number of layers increases has been proven in general by Zhang[@Zhang05], and we verify it for our implementation as well. We apply the piecewise linear approximation to a class of profiles where an analytical solution to the dispersion relation can be found for the special case $C(\mathbf{k})=0$, analyzed by Peregrine[@Pere]: $$U(z) = U_0\cosh\alpha^{1/2}z + U^\prime_0\alpha^{-1/2}\sinh\alpha^{1/2}z, \label{eq:uper}$$ where $U_0$ and $U^\prime_0$ are the velocity and shear values at the surface respectively, and $\alpha$ is chosen here such that $U(-h) = 0$ for bottom depth $h$. Following Peregrine[@Pere] the wave number $k_0$ satisfying $C(\mathbf{k}_0)=0$ solves the following equation: $$(k_0^2 +\alpha)^{1/2}h\coth[(k_0^2 +\alpha)^{1/2}h] = (gh/U_0^2) + (U_0^\prime h/U_0). \label{eq:perdr}$$ To compare the N-layer model to this result, the phase velocity was evaluated by first choosing streamwise wave number $k_x = -k_0$ found numerically from Eq. , with the wavevector orientation opposite that of the current ($\theta = \pi$). The exact result for a smooth profile is $C = 0$ and the approximation error is shown in Fig. \[fig:conv\] for [[a concave up profile with parameters $U_0/\sqrt{gh} = 0.45$, $U_0^\prime = 0$, $\alpha=-0.62$, and a concave down profile with $U_0\sqrt{gh} = 0.45$, $U_0^\prime/\sqrt{g/h} = 1.36$, and $\alpha=2.23$.]{}]{} In both these cases, the phase velocity tends to zero $\sim N^{-2}$ in the limit of large $N$ in agreement with the result of Zhang[@Zhang05]. ![The phase velocity $C$ as a function of the number of layers $N$ evaluated at the wavevector $k_0$ satisfying the analytical dispersion relation (Eq. ) for stationary waves. Two current profiles of the form of (Eq. ) were used with parameters $U_0/\sqrt{gh} = 0.45$, $U_0^\prime = 0$ [[$\alpha=-0.62$, ]{}]{}(squares), and $U_0\sqrt{gh} = 0.45$, $U_0^\prime/\sqrt{g/h} = $${{1.36}}$, [[$\alpha=2.23$, ]{}]{}(circles).[]{data-label="fig:conv"}](fig_conv_oneCol) When naively dividing the entire liquid column into equal layers, more layers are required to achieve a given level of accuracy for short wavelengths (large $kh$). This can be easily understood physically by noting that the influence on a regular wave from currents beneath the surface decreases expo[[n]{}]{}entially with depth, and at the 1% level dispersion properties are not influenced by currents deeper than a depth of $\lambda/2$, $\lambda$ being the wavelength. More quantitatively one could consider the equation for the first-order correction to the phase velocity presented by Stewart and Joy[@StewartJoy] in infinite depth: $$C \approx \sqrt{g/k} + 2k\int_{-\infty}^0U(z) e^{2kz} dz. \label{eq:sj}$$ Eq.  is a weighted average of the current profile, with exponentially decreasing weight in the vertical direction. For shorter wavelengths, the layers in the PLA are more coarsely spaced in the depth-range where the weighting term is large, resulting in greater approximation error. An improvement in accuracy is immediately achieved by diving the fluid into $N$ layers down to some intermediate depth determined by the desired accuracy. In our example, we divide the column into $N$ layers of equal width down to a depth equal to the smaller of $h$ and $\lambda/2$. When the column is cut off at $z=-\lambda/2$ this limits the accuracy that can be achieved: Since the weighting function $\exp(2kz)$ in the integral of Eq.  is approximately $0.002$ at depth $\lambda/2$, even deeper waters must be included for accuracies better than the $10^{-3}$ level. The procedure works well, however, since this level of accuracy is achieved with only a small number of layers, typically $4$ or $5$ at the $1\%$ level. We thus achieve a computationally cheap scheme solving the dispersion problem with a uniform level of accuracy across the wave vector plane, with little variation in computational effort. Dispersion relation ------------------- In this section we compare the dispersion relation among two different profiles defined by Eq. . In particular we examine the phase velocity and group velocity, $ {\mathbf{C}_g}= \nabla_\mathbf{k}\omega(\mathbf{k})$, as a function of propagation angle $\theta$ relative to the current direction, where $\nabla_\mathbf{k} = (\partial/\partial k_x, \partial/\partial k_y)$. In general the direction of ${\mathbf{C}_g}$ is not the same as the wavevector and phase velocity, and to simplify the analysis we consider here a scalar group velocity along the direction of $\mathbf{k}$, $C_g = {\mathbf{C}_g}\cdot\mathbf{k}/k$. For a more direct comparison, we define a shear Froude number $Fr_\text{sh} \equiv U_0/\sqrt{gh} = 0.45$, which is the same for both profiles. The concave-up profile then is prescribed $U_0^\prime = 0$, while the concave-down profile $U_0^\prime = 3U_0/h$. Three different wavelength scales relative to the depth are shown in Fig. \[fig:cg\] corresponding to shallow, intermediate, and deep-water regimes. The phase velocities $C$ and group velocities $C_g$ relative to the surface velocity display different characteristics in each case. In shallow water ($kh = 0.1$), $C_g \approx C$ as usual since the medium becomes approximately non-dispersive, yet there is a directionally ($\theta$)-dependent propagation velocity magnitude due to the current profile. For the intermediate case ($kh = 1$), the latter remains true (to a lesser extent) but there is now a difference between the velocities the magnitude of which is also directionally dependent. In the deep water regime ($kh = 10$) the wavelength becomes small and the dispersion relation is influenced only by the near-surface current profile. For the concave up profile with zero surface shear, the well-known limit $C_g \approx C/2$ for deep water waves without current is approached, with the velocities being independent of propagation direction. This occurs to a lesser extent with the profile to the right due to finite surface shear.  Example application: ship waves ------------------------------- [[The effects of a background shear flow of constant vorticity on the behavior of ship wakes has been studied recently[@Ell14; @Li16], yet many shear profiles encountered in reality have depth-variable vorticity. A model taking into account arbitrary vorticity depth-dependence may be necessary in obtaining quantitatively accurate results for ship wakes and related parameters such as wave resistance in the presence of realistic shear flows. In this section we derive the solution to the ship wave problem with a piecewise linear background flow, and analyze the qualitative features of ship wakes in the presence of two illustrative example profiles of depth-varying vorticity approximately representative of flows encountered in reality.]{}]{} The ship wave problem can be solved numerically for arbitrary shear profiles using the piecewise linear approximation in a direct generalization of the recent theory for the constant vorticity profile[@Ell14; @Li16]. Assuming a stationary wave solution in a reference frame moving with the ship, a coordinate transformation ${\boldsymbol \xi} = \mathbf{x}-\mathbf{V}t$ is introduced where $\mathbf{V}$ is the velocity of a moving prescribed pressure source $\hat{p}_\text{ext}({\boldsymbol \xi})$ representing the ship relative to the undisturbed free surface. With this coordinate transformation used in the Fourier formulation, the time derivative becomes $\partial/\partial t = -i\mathbf{k}\cdot\mathbf{V}$. Thus, $-i\omega \rightarrow -i\mathbf{k}\cdot\mathbf{V}$ in the Euler equations (Eqs. -). The reader is referred to Refs.  for more detail. The problem can be formulated in essentially the same way as described in section \[sec:der\] by formulating a system of $N+1$ equations from matching conditions and boundary conditions. The velocity and pressure conditions are the same as Eq.  and Eq. -, respectively, save for replacing $-i\omega$ with $-i\mathbf{k}\cdot\mathbf{V}$ as explained above. A radiation condition $\mathbf{k}\cdot\mathbf{V} \rightarrow \mathbf{k}\cdot\mathbf{V} + i\epsilon$, $\epsilon\rightarrow 0+$ is necessary in practice to avoid singularities in the integral over the $\mathbf{k}$-plane. The dynamic boundary condition is different from section \[sec:der\], where in the case of ship waves the pressure in Fourier space equals the prescribed pressure distribution $p_\text{ext}(\mathbf{k})$ (the Fourier transform of $\hat{p}_\text{ext}({\boldsymbol \xi})$) at the free surface. This leads to a non-zero term $p_\text{ext}(\mathbf{k})/\rho$ on the right side of Eq. , which becomes: $$\begin{aligned} A_1\left[\gamma_1\sigma_\text{ship}\tanh kh_1 + \sigma_\text{ship}^2k + gk^2\tanh kh_1\right] +& \nonumber\\ B_1\left[\gamma_1\sigma_\text{ship} + \sigma_\text{ship}^2k\tanh kh_1 + gk^2\right] &= -\frac{k^2p_{ext}(\mathbf{k})}{\rho}\sigma_\text{ship}, \label{eq:dkcship}\end{aligned}$$ where $\sigma_\text{ship} \equiv -i\mathbf{k}\cdot\mathbf{V} + ik_xU_0$. The resulting $N+1$ equation system is inhomogeneous in this case and a solution to the coefficients $A_j$ and $B_j$ determining the vertical velocity can be found directly. The free surface in Fourier space representing the ship wake component at the given value of $\mathbf{k}$ is then found from the kinematic boundary condition: $$\left(-i\mathbf{k}\cdot\mathbf{V} + ik_xU_0\right)\zeta = A_1\sinh kh_1 + B_1\cosh kh_1. \label{eq:kbc}$$ The real space pattern $\hat{\zeta}({\boldsymbol \xi})$ is found using an inverse fast Fourier transform. To demonstrate the utility of the piecewise linear method for ship waves on an arbitrary current, we consider [[two current profiles of depth-varying vorticity. The first is]{}]{} an exponential current profile defined as $U(z) = U_0\left(\exp[z/d]-1\right)$ with water depth $h = 10d$[[.]{}]{} [[The profile is similar to realistic wind-driven or river plume profiles with vorticity peaked at the surface and decaying with depth. The second profile is concave up defined by Eq.  with equal $U_0$ and $h$ to the exponential profile, $U_0^\prime=0$, and $\alpha = -0.62$. Both profiles are approximated with $N=4$ layers.]{}]{} To compare with the constant shear results in the literature, we here assign a shear Froude number $Fr_s \equiv VU_0/gh$, corresponding to the “shear Froude number” used in Refs. . In this sense we are assuming that the [[depth-variable vorticity profiles are]{}]{} being compared with a constant model with vorticity equal to the depth-averaged value. We assume a Gaussian pressure distribution with half-width equal to water depth $b = h$ as $p_\text{ext}(\mathbf{\xi}) = p_0e^{-\pi^2\xi^2/b^2}$. The results of this comparison for Froude number $Fr = V/\sqrt{bg} = 0.6$ are shown in Fig. \[fig:wake\], along with the solution on a quiescent profile. Ship velocity $\mathbf{V}$ is expressed relative to the surface current and is assumed to be along the x-axis (as is the background flow). There are clear qualitative differences between the solutions. Insight can be gained by considering the transverse wavelengths, pertaining to the waves propagating parallel to the direction of ship motion in the central wake region. The transverse waves (having the same phase velocity as the source) in the [[three]{}]{} rightmost columns are inhibited by the shear and therefore must be longer in wavelength in order to travel at the same speed as the corresponding waves in quescient waters. Furthermore, given the non-constant vorticity of the [[profiles in the two rightmost columns, the transverse wavelengths are notably changed from the constant vorticity case. For the concave down]{}]{} exponential current profile, the current strength into the fluid is stronger relative to the constant shear approximation, which further increases the transverse wavelengths. [[For the concave up profile in the rightmost column, the vorticity is zero at the surface and the current strength weaker than the constant vorticity model, resulting in shorter transverse wavelengths.]{}]{} The differences are exaggerated in the bottom row, where the shear is strong enough such that $Fr$ is supercritical [[for the exponential profile in (g)]{}]{}, disallowing transverse waves (see Ref.  for details). For [[the current profiles]{}]{} considered here, inclusion of the depth-varying vorticity of the profile is essential to obtain a realistic solution to the wake, and consequently the wave resistance, of the ship. Conclusion ========== This study has demonstrated the use of a piecewise linear approximation (PLA) to an arbitrary current profile to model [[linear]{}]{} wave propagation in 3D. The method is valid for any current magnitude and wavelength and does not rely on assumptions of weak current, near-potentiality, or small vorticity as do many other approximation techniques presented in the literature. The approximation accuracy is relatively unchanged over all wavelengths making the technique well suited for solving problems formulated in Fourier wavevector space, where integration over the full plane of wave vectors is necessary. [[The accuracy of the PLA method could be further verified through comparison to experimental data of the dispersion relation where independent measurements of the background flow are performed, such as in Lund *et al*.[@Lund15].]{}]{} Additional [[discussion]{}]{} concerning the nature and characteristics of extra spurious solutions to the PLA is given showing that the solutions [[to the dispersion relation which the PLA produces]{}]{} represent vortical structures flowing along at a velocity near that of the layer interface. [[They are artifacts of the discontinuities in vorticity introduced by the model, are spurious in the present context, and should be discarded. Details of the procedures used for quickly identifying the two physical solutions for a given ${\mathbf{k}}$ are provided.]{}]{} The importance of including non-uniform vorticity into wave propagation models is highlighted by considering the directional-dependence of the group and phase velocities for concave-up vs concave-down profiles, as well as solution to the ship wave problem. The PLA in 3D is a practical method for solving a wide variety classical wave problems such as ship waves as elaborated upon herein, ring waves, and problems of radiation and refraction, in the presence of a shear current of arbitrary depth-dependence. Further extensions of the model could include velocity profiles where the direction varies with depth, as well as non-linear waves as been done by e.g. Dalrymple[@Dalrymple74] and Swan *et al.*[@Swan01]. The quadratic eigenvalue formulation of our PLA algorithm described briefly in section \[sec:der\] was conceived and implemented by Peter Maxwell. We gratefully acknowledge his efforts, expecting to draw on this improved capability in several future applications. SÅE was partly funded by the Norwegian Research Council (FRINATEK), project 249740. BKS is funded by the Department of Energy and Process Engineering, NTNU.

ArXiv

--- abstract: 'We generate translationally invariant systems exhibiting many-body localization from All-Bands-Flat single particle lattice Hamiltonians dressed with suitable short-range many-body interactions. This phenomenon – dubbed Many-Body Flatband Localization (MBFBL) – is based on symmetries of both single particle and interaction terms in the Hamiltonian, and it holds for any interaction strength. We propose a generator of MBFBL Hamiltonians which covers both interacting bosons and fermions for arbitrary lattice dimensions, and we provide explicit examples of MBFBL models in one and two lattice dimensions. We also explicitly construct an extensive set of local integrals of motion for MBFBL models. Our results can be further generalized to long-range interactions as well as to systems lacking translational invariance.' author: - Carlo Danieli - Alexei Andreanov - Sergej Flach bibliography: - 'general.bib' - 'flatband.bib' - 'mbl.bib' title: 'Many-Body Flatband Localization' --- [*Introduction* —]{} Understanding the lack of thermalization in quantum interacting systems has been an active topic since Anderson predicted in 1958 the absence of transport in single particle lattices due to spatial disorder [@anderson1958absence]. This localization phenomenon has been extensively studied theoretically and experimentally [@kramer1993localization], with the impact of interaction between localized particles as one of the main open questions. Weak interactions were predicted to preserve the absence of transport of interacting particles [@basko2006metal; @aleiner2010finite] about fifty years after Anderson original work, leading to the phenomenon of *Many-Body Localization* (MBL). The study of MBL systems and their properties is nowadays a very active topic of research with several open issues and active fronts - for a survey of the state of the art, see [@abanin2017recent; @abanin2019colloquium]. MBL was initially predicted for interacting disordered systems emerging as an interplay of disorder and weak interactions. However it was later realized that the presence of disorder is not essential, launching the search for disorder-free MBL systems. Several possible scenarios emerged as a result: from non-ergodic behavior in networks of Josephson junctions [@pino2016nonergodic] to 1D fermionic lattices involving different species of particles [@schiulaz2015dynamics] or the presence of d.c. field [@schulz2019stark], local constraints due to gauge invariance [@karpov2020disorderfree], presence of a large number of conserved quantities [@smith2017disorder; @smith2018dynamical], quasi-periodic long-range interactions [@mondaini2017many], among others. Some proposals also explored the connection to glasses, predicting MBL in glassy systems [@laumann2014many; @baldwin2015the; @baldwin2017clustering; @mossi2017on], [*e.g.*]{} kinetically constrained models [@vanhorssen2015dynamics] and geometrically frustrated models [@zhao2020glass]. However, the validity of some of the proposals were later doubted, as it was shown that several disorder-free MBL systems rely on vastly different energy scales and finite-size constraints [@papic2015many]. In other cases instead ([*e.g.*]{} [@pino2016nonergodic]), disorder-free MBL requires high temperatures or specific strong interaction regimes, likewise the original MBL requests weak interaction regimes. In this letter, we propose a generator of disorder-free MBL systems which is free of the above-mentioned requirements (specific interaction or temperature regimes, finite-size constraints, type of many-body statistics, among others) and applies for arbitrary spatial dimensions. This generator relies on geometrical frustration of the translationally invariant single particle Hamiltonians which yields no single particle dispersion - i.e. all Bloch bands are dispersionless (or flat) - and suitably chosen many-body interactions. The resulting models exhibit non-ergodic behavior with lack of transport of particles for any interaction strength, and this phenomenon is dubbed [*Many-Body Flatband Localization*]{} (MBFBL). The study of networks with one or several flatbands (FB) is an active topic of research on its own. They were first discussed in the context of groundstate ferromagnetism [@mielke1993ferromagnetism], but were later identified in various other systems [@leykam2018artificial; @leykam2018perspective] and they have been experimentally realized in several setups, using e.g. ultra cold atoms [@taie2015coherent] and photonic lattices [@mukherjee2015observation; @vicencio2015observation; @weimann2016transport]. An important property of FB systems is the presence of compact localized states (CLS) - eigenstates with strictly finite support. These were used to systematically construct FB models [@maimaiti2017compact; @maimaiti2019universal; @maimaiti2020flatband] along with other methods [@flach2014detangling; @dias2015origami; @ramachandran2017chiral; @roentgen2018compact; @toikka2018necessary]. Their fine-tuned character makes FB systems an ideal platform to study diverse localization phenomena in the presence of onsite disorder [@leykam2013flat; @bodyfelt2014flatbands; @danieli2015flatband; @leykam2017localization], DC fields [@kolovsky2018topological], and nonlinearities [@danieli2018compact; @ramachandran2018fano], among many others. We introduce MBFBL networks formed by single particle All-Bands-Flat lattice Hamiltonians dressed with suitable short-range many-body interactions, and provide explicit examples in one and two spatial dimensions. We also discuss distinct interaction terms (including long-range interactions) in order to cover different types of particle statistics. We construct an extensive set of local integrals of motion present in MBFBL networks, and explicitly derive these integrals for some of the examples presented. We extend our generator scheme by removing the assumption of translation invariance of the lattice. [*Setup* —]{} We consider a translationally invariant many-body Hamiltonian $\hat{{\mathcal{H}}}$ on a lattice $$\begin{gathered} \label{eq:Ham1} \hat{{\mathcal{H}}} = {\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}+ {\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}\;,\qquad {\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}= \sum_k {\hat{f}}_k\;,\; \; {\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}= \sum_\kappa {\hat{g}}_\kappa\end{gathered}$$ with both single particle part ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ written as sums of local operators ${\hat{f}}_k$ and ${\hat{g}}_\kappa$. The integers $k$ and $\kappa$ label unit cells of the lattice in a direct space for two different unit cell choices $A$ and $B$. We assume that the sites from one unit cell of e.g. choice $A$ belong to [*different*]{} unit cells of choice $B$. Regardless of the choice, each unit cell contains $\nu$ lattices sites or single particle levels. The operators are expressed through creation and annihilation operators ${\hat{c}}_{k,a}^\dagger, {\hat{c}}_{k,a}$ which create or annihilate a single particle on a given lattice site $k,a$ with $1\leq a \leq \nu$. Then the local operators read $$\begin{gathered} {\hat{f}}_k = \sum_{a,b=1}^{\nu} t_{ab} {\hat{c}}_{k,a}^\dagger {\hat{c}}_{k,b} + \text{h.c.} \label{eq:fk}\end{gathered}$$ We assume the *interaction* Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ to be two-body, so that the local operators are $$\begin{gathered} {\hat{g}}_\kappa = \sum_{\alpha,\beta,\gamma,\delta=1}^\nu J_{\alpha \beta \gamma \delta} {\hat{c}}_{\kappa,\alpha}^\dagger {\hat{c}}_{\kappa,\beta} ^\dagger {\hat{c}}_{\kappa,\gamma} {\hat{c}}_{\kappa,\delta} + \text{h.c.} \label{eq:gk}\end{gathered}$$ By the above definitions both single particle and interaction Hamiltonians are [*semi-detangled*]{} (SD) as $[ {\hat{f}}_k , {\hat{f}}_{k'} ] = [ {\hat{g}}_{\kappa} , {\hat{g}}_{\kappa'}] =0$ for any $k,k',\kappa ,\kappa '$. The spectrum of the single particle eigenvalue problem with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ yields $\nu$ flatbands with each being an eigenenergy of any of the local operators $f_k$. It follows that ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ enforces full localization and absence of transport. The same is true for ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$. However, because of the different unit cell choices $A,B$, in general it follows that $[{\hat{f}}_k, {\hat{g}}_\kappa]\neq 0$ for any given $k$ and at least a pair of different values of $\kappa$ (and vice versa). Consequently, the combination of both ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ into ${\mathcal{H}}$ in general yields transporting many-body eigenstates [@vidal1998aharonov; @vidal2000interaction; @Tilleke2020nearest; @danieli2020cagingprep; @*danieli2020cagingprepII]. If $t_{ab}=t_{aa} \delta_{a,b}$ (with the Kronecker symbol $\delta_{a,b}$), the ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ is coined [*fully detangled*]{} (FD) [@flach2014detangling] since it depends on the particle number operators ${\hat{n}}={\hat{c}}^\dagger {\hat{c}}$ only, and does not move any particles from any lattice site to any other one. Together with ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ being SD, the full Hamiltonian ${\mathcal{H}}$ preserves full localization of particles, which is an example of *many-body flatband localization* (MBFBL). Likewise, if we assume that $J_{\alpha \beta \gamma \delta} = J_{\alpha \beta \alpha \beta} \delta_{\alpha,\gamma} \delta_{\beta,\delta}$ it follows that ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is FD and does not move any particles from site to site. Together with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ being SD, we again arrive at the result that the full Hamiltonian ${\mathcal{H}}$ lacks transporting eigenstates and is MBFBL. The relation between the FD/SD character of the Hamiltonians and the presence of MBFBL is summarized in Table \[tab:H0H1\]. We refer to all the other types of Hamiltonians as *non-detangled* (ND). SD FD ---- ------- ------- SD — MBFBL FD MBFBL MBFBL : Existence of MBFBL for different types of single particle Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and interaction Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ as discussed in the main text.[]{data-label="tab:H0H1"} We generate MBFBL Hamiltonians by choosing any of the FD/SD MBFBL models from Table . We then perform a unitary transformation (rotation) on each unit cell in either of the two unit cell choices $A, B$. This results in general in some complicated Hamiltonian with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ being ND and ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ being FD/SD, or vice versa - ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ being FD/SD and ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ being ND - depending on which unit cell type the transformation was applied to. Furthermore these transformations can be chosen unit cell dependent resulting in non-translationally invariant Hamiltonians. Conventional disordered MBL systems are known to possess an extensive set of local integrals of motion [@serbyn2013local; @Ros2015integrals; @abanin2017recent], though explicit derivations are complicated. These integrals are used to explain relevant properties of these systems. Local integrals of motion can be explicitly derived for MBFBL networks. With our proposed scheme and considering a SD single particle Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ in ${\mathcal{H}}$ , it follows that the expectation values of the operators ${\hat{I}}_k = \sum_{a=1}^{\nu} {\hat{n}}_{k,a}$ measure the number of particles in each local unit ${\hat{f}}_k$ of ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$. These numbers are conserved in the presence of a FD interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ (since ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ does not move particles from one to another site). It follows that each ${\hat{I}}_k$ commutes with the full Hamiltonian. The unitary transformations used to recast ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ as ND yield $N$ local integrals of motion ${\hat{I}}_k$ expressed in the new basis for the generated MBFBL lattice. The very same follows if a pair of FD single particle ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and a SD interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is picked from Table \[tab:H0H1\]. In this case, the operators ${\hat{I}}_\kappa = \sum_{\alpha=1}^{\nu} {\hat{n}}_{\kappa,\alpha} $ defined in each local unit ${\hat{g}}_\kappa$ of ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ as well lead to $N$ local integrals of motion of the MBFBL lattice after the unitary transformations have been applied. In the case of FD-FD Hamiltonians ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}},{\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$, the extensive set of local integrals of motion contains $\nu\times N$ elements, since each particle number operator ${\hat{n}}_{k,a}$ commutes with the full Hamiltonian ${\mathcal{H}}$. Most of the generated MBFBL models, while being appealing from a mathematical point of view, could be hard to implement in experiments due to the complicated structure of the interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ spanning several unit cells. Experimental feasibility instead favors fully detangled ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$, which result e.g. from Coulomb interactions between density operators in real space [@ziman1972principles]. Therefore we refine our generator scheme by choosing SD single particle ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and FD interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$, and recast ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ to a ND Hamiltonian via unitary transformations that keep ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ fully detangled. This algorithm works for any number of bands $\nu$ of ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$, in any dimension, and any type of many-body statistics. [*Results* —]{} We will now discuss concrete examples in one and two spatial dimensions. We consider the SD Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and conveniently restate it in the unit cell representation $B$ of ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$. We then apply the subsequent unitary transformations. This change of unit cell introduces hopping terms between neighboring unit cells in each local Hamiltonian ${\hat{f}}_\kappa$. Without loss of generality, we assume nonzero hoppings between nearest-neighboring unit cells only, and we adopt the conventions used in Refs. [@maimaiti2017compact; @maimaiti2019universal] for flatband networks generators. Then a possible $D=1$ Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ reads $$\begin{gathered} {\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}= \sum_\kappa {\hat{f}}_\kappa = \sum_\kappa \left[\frac{1}{2}\hat{C}_\kappa^{\dagger T} H_0\hat{C}_\kappa + \hat{C}_\kappa^{\dagger T} H_1\hat{C}_{\kappa+1} + \text{h.c.}\right] \label{eq:H0}\end{gathered}$$ where we grouped the annihilation (creation) operators ${\hat{c}}_{\kappa,a}$ (${\hat{c}}_{\kappa,a}^\dagger$) in $\nu$-dimensional vectors $\hat{C}_\kappa( \hat{C}_\kappa^\dagger)$. The matrices $H_0,H_1$ describe intra- and intercell hopping respectively, and are chosen so as to enforce the SD condition $[{\hat{f}}_\kappa,{\hat{f}}_{\kappa'}] = 0$ for all $\kappa,\kappa'$. We remark that this ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ is only one of the infinitely many realizations of a SD single particle Hamiltonian. The FD two-body interaction Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ introduced above (\[eq:Ham1\],\[eq:gk\]) is taken with the coefficients $J_{\alpha\beta\gamma\delta} = J_{\alpha\beta\alpha\beta} \delta_{\alpha,\gamma} \delta_{\beta,\delta}$ for each local component ${\hat{g}}_\kappa$: $J_{\alpha\beta\alpha\beta}=1$ for $\alpha = \beta$ and $J_{\alpha\beta\alpha\beta}=2$ for $\alpha \neq \beta$. Then ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is preserved as FD with the same coefficients $J_{\alpha\beta\gamma\delta}$ by any $2\times 2$ unitary transformation $$\begin{gathered} U_{ab}:\ \begin{cases} {\hat{c}}_{\kappa,a} = z {\hat{d}}_{\kappa,a} + w {\hat{d}}_{\kappa,b}\\ {\hat{c}}_{\kappa,b} = -w^* {\hat{d}}_{\kappa,a} + z^* {\hat{d}}_{\kappa,b}\ \end{cases} \label{eq:rot1}\end{gathered}$$ parameterized by two complex numbers $z,w$ such that $|z|^2 + |w|^2 = 1$ and any pair of sites ${\hat{c}}_{\kappa,a}, {\hat{c}}_{\kappa,b}$. The resulting Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ for $\nu=2$ bands describes a two-body interaction among the sites ${\hat{a}}_\kappa = {\hat{c}}_{\kappa,a},{\hat{b}}_\kappa = {\hat{c}}_{\kappa,b}$ $$\begin{aligned} {\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}&= \sum_\kappa\bigl[{\hat{a}}_\kappa^\dagger{\hat{a}}_\kappa^\dagger{\hat{a}}_\kappa{\hat{a}}_\kappa + {\hat{b}}_\kappa^\dagger{\hat{b}}_\kappa^\dagger{\hat{b}}_\kappa{\hat{b}}_\kappa + 2{\hat{a}}_\kappa^\dagger{\hat{a}}_\kappa{\hat{b}}_\kappa^\dagger{\hat{b}}_\kappa\bigr] \notag\\ &= \sum_\kappa \bigl[ {\hat{n}}_{a,\kappa} + {\hat{n}}_{b,\kappa} - 1\bigr] \bigl[ {\hat{n}}_{a,\kappa} + {\hat{n}}_{b,\kappa}\bigr] \label{eq:H1_nu2}\end{aligned}$$ with ${\hat{n}}_{a,\kappa}={\hat{a}}_\kappa^\dagger{\hat{a}}_\kappa$ and ${\hat{n}}_{b,\kappa} = {\hat{b}}_\kappa^\dagger{\hat{b}}_\kappa$. We refer to such interaction as an *extended Hubbard* interaction, which applies to both bosons and fermions with spin. [*1D networks* —]{} We now present two concrete examples of MBFBL networks. We first start with the simplest MBFBL network with $\nu=2$ bands. It is based on the Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ in Eq.  with $$\begin{gathered} H_0 = \begin{pmatrix} 1 & 0\\[0.3em] 0 & 0 \end{pmatrix}, \qquad H_1 = \begin{pmatrix} 0 & t\\[0.3em] 0 & 0 \end{pmatrix}, \label{eq:H0H1_nu2}\end{gathered}$$ and a free complex parameter $t$. It is straightforward to check that this Hamiltonian is SD and has all bands flat. Next we pick the extended Hubbard interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ . The structure of ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is shown in Fig. \[fig:2bands\](a) with solid lines and red shaded rods respectively. The rotation $U_{ab}$  recasts $H_0,H_1$  as $$\begin{gathered} H_0 = \begin{pmatrix} |z|^2 & -z w\\[0.3em] - z^* w^* & |w|^2 \end{pmatrix}, \quad H_1 = t \begin{pmatrix} z w^* & z^2\\[0.3em] - (w^*)^2 & -z w^* \end{pmatrix}, \label{eq:H0H1_nu2_2}\end{gathered}$$ and makes Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ ND, while ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ remains FD. The resulting MBFBL network is shown in Fig. \[fig:2bands\](b). The local integrals of motion read (after the rotation) $$\begin{aligned} {\hat{I}}_k & = {\hat{n}}_{a,\kappa-1} + |z|^2{\hat{n}}_{b,\kappa} + |w|^2{\hat{n}}_{b,\kappa+1} \notag\\ & + z^*w ({\hat{a}}_\kappa^\dagger {\hat{b}}_{\kappa+1} - {\hat{a}}_{\kappa-1}^\dagger {\hat{b}}_\kappa) + \text{h.c.} \label{eq:d1nu2_liom}\end{aligned}$$ ![(Color online) One dimensional two band MBFBL network with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ SD (a) and with the cross-stitch lattice profile (b). The black circles indicate the unit cell choice, the solid lines correspond to sites connected by ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ before (a) and after (b) the rotation, and the red shaded rods indicate the sites connected by the extended Hubbard terms  of ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$.[]{data-label="fig:2bands"}](fig1.pdf){width="0.925\columnwidth"} For three bands $\nu=3$ with operators $a_\kappa, b_\kappa, c_\kappa$ corresponding to the three sites of the unit cell, the SD Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$  has the following hopping matrices $$\begin{gathered} H_0 = \begin{pmatrix} 1 & 0 & 0 \\[0.3em] 0 & 0 & t_1 \\[0.3em] 0 & t_1^* & \mu \end{pmatrix}, \qquad H_1 = \begin{pmatrix} 0 & 0 &t_2 \\[0.3em] 0 & 0 & 0 \\[0.3em] 0 & 0 & 0 \end{pmatrix}, \label{eq:H0H1_nu3}\end{gathered}$$ with two free complex ($t_1,t_2$) and one free real ($\mu$) parameters. This network is shown in Fig. \[fig:1D\_3bands\_DC\](a) with gray solid lines. The interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ consists of the extended Hubbard interaction  between the top and the bottom sites $(a_\kappa,b_\kappa)$ of each plaquette (red shaded rods in Fig. \[fig:1D\_3bands\_DC\](a)) and an additional optional onsite Hubbard interaction for the central site $c_\kappa$. Then the rotation $U_{ab}$  is applied to the pair $(a_\kappa,b_\kappa)$ only while leaving the sites $c_\kappa$ untouched. This recasts $H_0,H_1$  into $$\begin{gathered} H_0 = \begin{pmatrix} |z|^2 & -z w & t_1 w \\[0.3em] -z^* w^* & |w|^2 & t_1 z^* \\[0.3em] t_1^*w^* & t_1^* z & \mu \end{pmatrix} \quad H_1 = t_2 \begin{pmatrix} 0 & 0 & z \\[0.3em] 0 & 0 & -w^* \\[0.3em] 0 & 0 & 0 \end{pmatrix} \label{eq:H0H1_nu3_2}\end{gathered}$$ defining a ND Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ while the interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ remains FD. The resulting diamond-shaped MBFBL network is shown in Fig. \[fig:1D\_3bands\_DC\](b). That diamond-shape profile has been realized in diverse experimental setups for flatband and compact localized state studies [@mukherjee2015observation1; @mukherjee2017observation; @mukherjee2018experimental; @xia2020observation; @Kremer2020square]. Experimentally, the selective extended Hubbard interaction involving only the top and bottom sites ${\hat{a}}_\kappa,{\hat{b}}_\kappa$ of the diamond plaquette might be achieved by reducing the distance between these sites as compared to the distance to the middle site ${\hat{c}}_\kappa$. The parameter $t_1$ could be used to adjust the hoppings. The local integrals of motion ${\hat{I}}_\kappa$ for this model are given by Eq.  plus the additional particle number operator ${\hat{n}}_{c,\kappa}$ for the central site ${\hat{c}}_\kappa$ of the lattice, since it is unaffected by the rotation. ![(Color online) One dimensional three band MBFBL network with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ SD (a) and with the diamond-shaped lattice profile (b). The black circles indicate the unit cell choice, the solid lines correspond to ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ before (a) and after (b) the rotations, and the red shaded rods indicate the extended Hubbard terms  of ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$.[]{data-label="fig:1D_3bands_DC"}](fig2.pdf){width="0.975\columnwidth"} [*2D networks* —]{} Construction of higher dimensional MBFBL networks follows a procedure similar to that of 1D systems. In the simplest setting, the single particle Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ can be taken as a straightforward extension of Eq. , where matrices $H_1$ are replaced with matrices $H_0$ and $H_1^{(1)},\dots,H_1^{(D)}$ describing the intercell hopping along different spatial directions. The matrices are chosen to ensure that ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ is SD. Now taking a suitable FD interaction ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$, Eq.  or its generalizations, and picking a unitary transformation that leaves this ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ FD, we obtain a ND Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$. The full Hamiltonian ${\mathcal{H}}$ exhibits MBFBL [^1]. A notable two dimensional lattice exhibiting MBFBL obtained by applying these rules is the *decorated Lieb* lattice [@rontgen2019quantum]. This is a five-band $\nu=5$ network, whose SD Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ is shown in Fig. \[fig:2D\_5bands\](a), with matrices $H_0, H_1^{(1)}, H_1^{(2)}$. In each unit cell, we use the extended Hubbard Hamiltonians ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$  for the two site pairs indicated by red shaded rods in Fig. \[fig:2D\_5bands\](a), and an onsite Hubbard interaction for the central site. The two rotations $U_{ab}$  applied to the highlighted pairs (leaving the central site untouched) yield a ND ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ shown in Fig. \[fig:2D\_5bands\](b), and the resulting full Hamiltonian ${\mathcal{H}}$ is MBFBL. The local integrals of motion for the decorated Lieb lattice can be easily derived and have similar but more involved expressions to those of the previous models . ![(Color online) Two dimensional five band MBFBL network with ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ SD (a) and with the decorated Lieb lattice profile (b). The black circles indicate the unit cell choice, the solid lines correspond to ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ before (a) and after (b) the rotations, and the red shaded rods indicate the extended Hubbard terms  of ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$.[]{data-label="fig:2D_5bands"}](fig3.pdf){width="0.975\columnwidth"} [*Perspectives* —]{} The proposed scheme relies on the two-body Hamiltonian ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ with onsite terms in the interaction, restricting the interacting particles to bosons or spinful fermions. However, the same construction can be implemented for spinless fermions by *e.g.* choosing local operators ${\hat{g}}_{\kappa}^{\sigma} = \sum_{\alpha,\beta,\gamma,\delta=1}^{\nu} J_{\alpha \beta \gamma \delta}^{\sigma} {\hat{c}}_{\kappa,\alpha}^\dagger {\hat{c}}_{\kappa+\sigma,\beta} ^\dagger {\hat{c}}_{\kappa,\gamma} {\hat{c}}_{\kappa+\sigma,\delta} + h.c.$ with exclusively inter-site interaction terms between unit cell $\kappa$ and unit cell $\kappa+\sigma$. In particular, ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is FD for $J_{\alpha\beta\gamma\delta}^\sigma=J \delta_{\alpha,\gamma} \delta_{\beta,\delta}$ and it is preserved as FD by the same transformation . This yields a generator of $D$-dimensional $\nu$-band MFBFL lattices for spinless fermions, with the recent work of Kuno *et al.* [@kuno2020flat] being a particular $D=1$ $\nu=2$ band example. The construction can be further extended to long-range all-to-all interaction Hamiltonians ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ by setting ${\hat{g}}_\kappa = \sum_\sigma v_\sigma {\hat{g}}_\kappa^\sigma$ and even infinite-range interactions ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}=J/N\sum_{\kappa\neq\kappa', a} {\hat{n}}_{\kappa, a} {\hat{n}}_{\kappa',a}$. The latter example is valid because the interaction is a function of the total density $\hat\rho = \sum_{\kappa,a} {\hat{n}}_{\kappa,a}$ only and is therefore invariant under the transformation . We note that it is possible to extend the generator by abandoning the translational invariance of the Hamiltonian ${\mathcal{H}}$. We can choose the hopping parameters $t_{ab}$ and the interaction matrix elements $J_{\alpha \beta \gamma \delta}$ in the starting Hamiltonians ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ and ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ respectively to be unit cell dependent. To stick with the proposed scheme where ${\ensuremath{\hat{{\mathcal{H}}}_\text{int}}}$ is FD and is preserved by unitary transformations , the unit cell dependent terms are restrained to the SD ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ only ([*e.g.*]{} onsite or hopping disorder). The unitary transformations  used to recast ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ as ND induce correlations between the onsite energies of the pairs of sites involved. In the models presented - Figs.\[fig:2bands\](b), \[fig:1D\_3bands\_DC\](b), \[fig:2D\_5bands\](b) - these correlations are between the sites within the same red-shaded area. These correlations depend of the parameters $z,w$ defining $U_{a,b}$ in Eq. . These parameters may also be chosen to vary upon changing $\kappa$ if the unitary transformations considered differ from unit cell to unit cell. Let us additionally observe that the breaking of translation invariance does not destroy the existence of the extensive set of local integrals of motion - they are given by the same operators as in the translationally invariant case. [*Conclusions* —]{} We have introduced a generator of Many-Body Localized disorder-free Hamiltonians by applying unitary transformations to suitably detangled Hamiltonians – a feature that assumes all-band-flat single particle Hamiltonians. This new phenomenon – coined Many-Body Flatband Localization – implies strict localization of any number of particles irrespective of dimensionality or interaction strength, and it does not require vastly different energy scales similar some models supposed to exhibit disorder-free MBL. Our work substantially extends previous studies of localization phenomena of interacting quantum many-body platforms with All-Band-Flat lattice single particle Hamiltonians [@vidal2000interaction; @junemann2017exploring; @mondaini2018pairing; @murad2018performed; @Barbarino2019topological; @roy2019compact; @kuno2020flat; @orito2020exact; @danieli2020cagingprep]. In particular, we propose a flexible and general set of many-body localized systems which may be experimentally feasible. A novel and unique feature of these systems is the existence of unitary mappings that recast them into a detangled form. This very property can be employed to study the impact of additional perturbations of the proposed networks which lift MBFBL and modify the proposed local integrals of motion in a systematic and analytical form. Hence, these systems offer innovative and powerful tools to potentially perform systematic analytical studies of conventional properties of MBL networks which typically relay on heavy numerical studies. [*Acknowledgments* —]{} This work was supported by the Institute for Basic Science (Project number IBS-R024-D1). SF acknowledges support by the New Zealand Institute for Advanced Study where part of this work was completed. [^1]: We point out that unlike 1D, the presence of several spatial directions may impose a constraint on the number of bands required to achieve MBFBL. For instance, the assumption of hopping between nearest neighbor unit cells in ${\ensuremath{\hat{{\mathcal{H}}}_\text{sp}}}$ implies that 2D MBFBL networks have to have three of more bands.

ArXiv

--- abstract: 'In this paper we investigate the LHC potential for discovering doubly-charged vector bileptons considering the measurable process $p,p$ $\rightarrow$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$. We perform the study using four different bilepton masses and three different exotics quark masses. Minimal LHC integrated luminosities needed for discovering and for setting limits on bilepton masses are obtained for both 7 TeV and 14 TeV center-of-mass energies. We find that these spectacular signatures can be observed at the LHC in the next years up to a bilepton mass of order of 1 TeV.' author: - 'B. Meirose' - 'A. A. Nepomuceno' title: | Searching for doubly-charged vector bileptons in the Golden Channel at the LHC --- \[sec:level1\]Introduction:\ ============================ Although bileptons [@CUYDAV] can be considered as “exotic” particles relative to their Standard Model (SM) cousins due to their unfamiliar quantum numbers, they are, in a sense, a conservative prediction. Even though they are suggested only by special extensions of the SM, their existence is, in our view, no more special than other proposals such as weak-scale extra dimensions or supersymmetry (SUSY). Indeed bileptons are a prediction employing only the model building rules for renormalizable gauge theories so successful in the Standard Model. The main motivation for expecting bileptons is that they explain three quark-lepton families. Generically speaking, a bilepton is a boson which couple to two leptons, but not to SM quarks, and which carries two units of lepton number. They are present in several beyond-SM scenarios, such as left-right symmetric models, technicolor and theories of grand unification. The bileptons in which we are interested are doubly charged vector bosons which couple to SM leptons, and are predicted when the Standard Model is embedded in a larger gauge group. The so-called 331 models [@PIPLEI; @FRA] fall into this category and in this article we restrict ourselves to this case. However, as will be explained in section II, we expect our results to hold in any model containing vector bileptons. In bilepton pair production, each of the bileptons will decay to two same-sign leptons. Therefore, they provide an exceptionally clean signature of four isolated high transverse momentum leptons, not necessarily of the same flavor. In this article we explore some of the consequences of this fact at the LHC. We study the actual collider signatures for the process $p,p$ $\rightarrow$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$ which has no SM background and is for this reason a “golden” channel for finding bileptons. In this process, a bilepton pair is produced via $s$ or $t$ channel where one of them decays into electrons while the other decays into muons. The most useful current lower bound on vector bileptons require these particles to be heavier than 740 GeV [@TUL]. This limit has been derived from experimental limits on fermion pair production at LEP and lepton-flavor charged lepton decays. Another useful lower bound is $M_Y >$ 850 GeV, a result which was established from muonium-antimuonium conversion [@WILL]. Although more stringent, this limit depends on the assumption that the bilepton coupling is flavor-diagonal. In this article we nevertheless consider bilepton masses as low as 400 GeV, following a similar line of reasoning as [@YARA], where the authors argued a lower bound of 350 GeV for doubly-charged vector bilepton masses, a limit that is compatible with other low energy bounds [@LOWE]. We also allow a larger upper bound for bilepton masses in 331 models than the usual 1 TeV considered by some authors. This article has been organized as follows. In section II we present our motivations to perform the present study. In section III we explain the numerical procedure for simulating the $p,p$ $\rightarrow$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$ reaction as well as its validation. In section IV we show relevant experimental observables for the bilepton golden channel. In section V we present the discovery potential for bileptons, calculating mass exclusion limits as a function of the LHC integrated luminosities, including a digression on the prospects for the accelerator’s 7 TeV run and for the super LHC (sLHC). We conclude in section VI. Motivations =========== Many interesting channels have been studied in the literature concerning bileptons, but curiously, there has been no systematic study on the phenomenology of the bilepton golden channel at the LHC. The authors in [@DION] did a fairly comprehensive study of bilepton phenomenology at hadron colliders. But in all cases they limited themselves to bilepton pair-production study, disregarding its decays. In the present study we go a step further in understanding the actual collider signatures, by considering measurable final states. 331 models ---------- The 331 models are based on the gauge symmetry $SU(3)_C \otimes SU(3)_L \otimes U(1)_X$, hence their name. Their first versions appeared in 1992 [@PIPLEI; @FRA]. There are many interesting aspects of the 331 models worth noticing but the most intriguing one is the explanation of three quark-lepton families, which is the main motivation for expecting bileptons in Nature. This is done via a nontrivial anomaly cancellation in the 331 Model that takes place between families, which is achieved by requiring the number of families to be equal to the number of quark colors. The explanation of the number of generations is also arguably one of the main reasons that keep model builders interested in 331 models, since they are one of the few which elegantly address this problem. The 331 models are also the simplest extension of the SM containing bileptons. Other interesting features of the 331 models include: a) they treat the third generation differently than the first two, this lead to an explanation of the heavy top quark mass; b) they have an automatic Peccei-Quinn symmetry [@PQ], hence they are also able to solve the strong CP problem. Moreover, gauge symmetry $SU(3)_C \otimes SU(3)_L \otimes U(1)_X$ is considered a subgroup of the popular $E_6$ [@E6] Grand Unified Theory (GUT), which can be itself derived from $E_8 \otimes E_8$ [@E8] heterotic string theory. Finally, it is worth mentioning that contrary to the SM, in 331 models lepton family number is not required to be conserved, only total lepton number. There is already experimental proof that lepton family number is not an exact symmetry via neutrino oscillations and one can regard this as circumstantial evidence for the non conservation of lepton family number in general. The combination of such intriguing aspects make bileptons desirable candidates to be found in Nature. Another important point to be discussed is the minimal version of the 331 models. There are different ways in which $SU(3)_C \otimes SU(3)_L \otimes U(1)_X$ can be broken down back to the $SU(3)_C \otimes SU(2)_L \otimes U(1)_Y$ SM gauge symmetry. The minimal version corresponds to use minimal Higgs structure to achieve this goal. In this version it is required that the new neutral vector boson $Z^\prime$ mass term to be coupled to the bilepton mass, like: $${M_Y \over M_{Z^\prime}} = {{\sqrt{3(1-4\sin^2\theta_W)} \over 2\cos \theta_W }}$$ Regarding theoretical upper bounds in 331 models there is no consensus in the literature. It was reasoned in [@FRA] that in 331 models, bileptons cannot be significantly heavier than 1 TeV, because of an upper limit in the symmetry breaking scale which is placed by requiring the sine squared of the Weinberg mixing angle ($\theta_W$) to be smaller than 1/4 , which is the same line of argumentation used in [@NG] to conclude that the $Z^\prime$ mass cannot be itself heavier than 3.1 TeV. Considering the mass relation between $Z^\prime$ and the bilepton, it could be argued that at least the minimal version of the 331 model could be excluded, should bileptons not be detected at the LHC, since any vector bilepton mass heavier than $\sim$ 840 GeV would violate the $Z^\prime$ mass upper limit via the mass relation between the two gauge bosons given by equation (1). This conclusion was challenged in [@PLEITEZ]. The argument is as follows. The 331 model predicts that there is an energy scale $\mu$ where the model loses its perturbative character. Should experimental data suggest a lower bound on the vector bilepton mass larger than $\mu$, the model would be ruled out. The value of $\mu$ is calculated through the condition $\sin^2\theta_W(\mu)=1/4$, but from this requirement alone it is not possible to know the real value of $\mu$. Then the upper limit on the vector bilepton mass could be, for instance, 3.5 TeV, as has been discussed by [@JAIN]. By the same token the 3.1 TeV upper limit in the $Z^\prime$ mass is automatically challenged. Therefore we do not believe it is possible to unambiguously discard *any* 331 model at the LHC (although they could be *discovered* at it), since we consider the bilepton mass upper bound to reasonably lie beyond the accelerator’s reach. For the lower bounds on vector bileptons we consider at least two mass points that violate the general 740 GeV limit imposed by LEP data. As explained by the authors in [@YARA], all the constraints on the 331 parameters coming from experiments involving leptonic interaction should be examined with caution. In the 331 model the leptons mix by a Cabibbo-Kobayashi-Maskawa-like mixing matrix whose elements have not yet been measured, so usually these experiments (and derived limits) apply only when the leptonic mixing matrix is diagonal. Also, in models with an extended Higgs sector some not unrealistic situations could exist in which the scalar bosons contribution to muonium to antimuonium conversion is not negligible. This puts also the possibility of strengthening bilepton experimental limits still at the LHC’s 7 TeV run in a new perspective, a possibility that we also discuss in this article. 331 Models and Supersymmetry ---------------------------- In recent years, a considerable fraction of both the experimental and theoretical communities has dedicated itself to supersymmetry. It is doubtlessly the mainstream subject in particle physics. The 331 models are not *necessarily* supersymmetric. But any renormalizable gauge theory can be extended to a globally supersymmetric model. The 331 models, being anomaly free, are renormalizable and fall of course in this category. Some authors have explored this possibility [@MSUSY331]. Furthermore, as it was argued also in [@PLEITEZ], in this model, the “hierarchy problem” is less severe than in the SM and its extensions since no arbitrary mass scale can be introduced. The masses of fundamental scalars are sensitive to the mass of the heaviest particles which couple directly or indirectly with them. Since in the 331 model the heaviest mass scale is of the order of a few TeVs there is not a “hierarchy problem” at all. This feature remains valid when supersymmetry is introduced. Thus, the breaking of the supersymmetry is also naturally at the TeV scale in the 331 model. Model independent vector bilepton searches at the LHC ----------------------------------------------------- Ideally one would like to study bileptons in hadron colliders as model independent as possible. For vector bileptons this is not possible as the non-inclusion of the $Z^\prime$ boson makes bilepton pair production to violate unitarity, in complete analogy to what happens with $e^+e^- \to W^+W^-$ using only photon exchange. This is another motivation to use the 331 model, although we do not restrain ourselves to the minimal version, allowing $M_Y$ and $M_{Z^\prime}$ to vary independently of one another. Even though model-dependent, our cross-sections should approximately be in the same order of magnitude with any other model containing bileptons, since at hadron colliders these particles have to be produced by the same Drell-Yan pair production process, regardless of the model. Exotic heavy quark exchange can influence this scenario, a possibility we do explore and which make our conclusions even more general. Numerical implementation and validation ======================================= To simulate the bilepton golden channel at the LHC we have implemented the 331 model in the Comphep generator [@COMPHEP]. We followed reference [@LONG] to implement the bilepton trilinear gauge interactions and [@DION] for the $Z^\prime$ couplings with fermions. For the bilepton interaction with leptons we have used the Lagrangian expression given in [@FRANG], generating the respective couplings using the Lanhep package [@LANHEP]. We also take into account bilepton interactions with exotic quarks in the 331 model. For this, we considered the following interactions with the exotic quark sector: $${\cal L}_Q = -\frac{g}{2\sqrt 2}[\overline{Q}^c\gamma^\mu\left(1 - \gamma_5\right)qY^{++}_\mu] + \mbox{H. c.}, \label{ll}$$ where Q ($D_1$,$D_2$,$T_1$) is an exotic heavy quark in the 331 model, q is a SM quark and $g = e\ / sin{\theta_W}$ as in the SM. We considered the respective interaction pairs for ($Q,q$) to be ($D_1$,$u$), ($D_2$,$c$) and ($T_1$,$b$), where $u$, $c$ and $b$ are SM quarks. Our purpose in including such interactions was, besides studying the influence of the heavy quark sector, to guarantee that all relevant quark subprocesses $q\overline{q} \to Y^{++}Y^{--}$ respect unitarity. Since the 331 model itself does not determine the elements of the mixing matrix (which determines how bileptons interact with exotic quarks), this is a reasonable criteria. This was needed for $u\overline{u} \to Y^{++}Y^{--}$, $b\overline{b} \to Y^{++}Y^{--}$ and $c\overline{c} \to Y^{++}Y^{--}$ which violate unitarity otherwise. In Figure 1, one can see the $u\overline{u} \to Y^{++}Y^{--}$ reaction cross-section as a function of the center of mass energy. Note that up to energies beyond the LHC designed center-of-mass-energy of 14 TeV, the cross-section dependence with energy behaves as expected. We tested this fact for all quarks involved in the proton parton distribution function (PDF), CTEQ6l1 [@CTEQ6l1] that was used for the complete $p,p$ $\to$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$ reaction simulation. Concerning particle parameters, we considered heavy quark masses to be equal to 400 GeV, 600 GeV and 800 GeV (lower bounds on exotic supersymmetric particles impose a lower bound on 331 exotic quark masses to be $\sim$ 250 GeV [@DAS]) and used 1 TeV for the $Z^\prime$ mass, since 331 $Z^{'}$ masses below 920 GeV were excluded using results from the CDF collaboration [@GUTI]. For the doubly charged bileptons $Y^{++}$ we considered four mass points: 400, 600, 800, GeV and 1 TeV. For each mass point, $10000$ $p,p$ $\to$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$ events were generated with Comphep. The $Y^{++}$, $Z^\prime$ and exotic quarks widths were calculated directly in Comphep for each mass point.The $Z^\prime$ width for $M_Q$ = 400 GeV is $\varGamma_{Z^\prime} \sim$ 360 GeV. For the other two exotics quark masses, $\varGamma_{Z^\prime}$ drops to $\sim$ 155 GeV, since Z’ decay into exotic quarks becomes kinematically forbidden. The variation of $\varGamma_{Z^\prime}$ with respect to $M_Y$ is of order of 1% between the highest and lowest bilepton mass considered. To further cross-check our implementation we reproduced the results from reference [@DION] Figure 4, for bilepton pair production at the LHC using $M_{Z^\prime}$ = 1 TeV. Minor numerical differences in the cross-sections can easily be explained by the use of different PDF’s. Observables =========== In what follows, all histograms were produced considering an integrated luminosity of $100\,\textnormal {fb}^{-1}$ and the nominal LHC energy of 14 TeV, unless otherwise stated. Cross section and Width ----------------------- Figure \[fig2\] shows the total cross-section of the $p,p$ $\rightarrow$ $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm} X$ process as a function of the doubly-charged bilepton mass, for three different exotic quark masses. Here one can see clear evidence on the problem of the influence of the heavy quark sector on bilepton production. Note that for a bilepton of $M_Y$ = 800 GeV the effect on the cross-section on having *heavier* exotic quark masses of $M_Q$ = 800 GeV as compared to $M_Q$ = 400 GeV is to *increase* the cross-section of the the process by a factor of $\sim$ 30. We can also see that the $M_{Q}$ = 600 and $M_{Q}$ = 800 curves split at $M_Y$ = 600 GeV. This happens because when $M_{Y} > M_ {Q}$, bilepton decays like $Y^{\pm \pm} \rightarrow qQ$ becomes kinematically allowed, which makes the value of $Br(Y^{\pm \pm} \rightarrow l^{\pm}l^{\pm}$) decrease for a given bilepton mass. Figure \[fig3\] shows the bilepton width as a function of bilepton mass for the same exotic quark masses as before. Here we can see how the width increases when new decays are allowed. It is also clear from the plot that the bilepton resonance is very narrow. We admittedly used a very simple approach to the problem of how bileptons and heavy quarks actually mix. However, different ways on determining the values of the mixing matrix between bileptons and heavy quarks could, in principle, intensify such effects even more drastically. An open problem that would deserve a separate study of its own. Pseudorapidity -------------- In order to investigate a more realistic scenario, we require the events that have been generated to pass some selection criteria according to the LHC detectors. First, the four leptons must be within the detectors geometrical acceptance, i.e., $\mid \eta \mid < 2.5$ [@ATLAS]. With this requirement, the fraction of selected events goes from 83% for $M_Y$ = 400 GeV to 93% for $M_Y >$ 900 GeV. Additionally, we also require each lepton $p_T$ be greater than 20 GeV. The loss of efficiency due this cut is negligible. Finally, we assume a reconstruction efficiency of 60%. Figure \[fig7\] illustrates the electron pair pseudorapidity distribution of the events before and after the selection for $M_Y$ = 600 GeV where we can see the fraction of surviving events. The acceptance (geometrical acceptance $\times$ efficiency) for this case is 53%, and this changes by $\pm$ 3% depending on the bilepton mass. The pseudorapidity distribution for muons is similar. Transverse Momentum ------------------- In Figures \[fig3b\] and \[fig4\] it is respectively shown the transverse momentum distribution for both the final state selected electrons and muons pairs of the golden channel, where both the doubly-charged bilepton and the heavy quarks have a mass of 600 GeV. Typically most of the events are produced in a region between $\sim$ 200 GeV and 400 GeV but the tail of the distribution has few events going up to 900 GeV for the electron pairs. Invariant Mass -------------- When doubly-charged vector bileptons are produced in pairs each bilepton can decay to a pair of same-sign leptons, not necessarily of the same flavor, where each lepton pair will have the same invariant mass distribution. This would be the most compelling evidence of a new resonance coming from a bilepton and very strong evidence of new physics. This is displayed in Figures 7 and 8 for the final electron and muon pairs respectively, considering both the bilepton and exotic quarks masses to be equal 600 GeV. The mean of the invariant mass for both the electrons and muon pairs in the final state unmistakably peaks at 600 GeV, the bilepton mass. Such a plot will of course demand several years of data taking such that enough statistics can be gathered, especially for higher bilepton masses, as we will address in the next section. Discovery Potential and Limits ============================== In order to determine the LHC potential to find bileptons, we calculate the minimal LHC integrated luminosity needed for a five-sigma bilepton discovery. For each bilepton mass, the detector acceptance as stated in section IV.B is considered and the $5 \sigma$ significance is obtained by requiring 5 events with two electrons and two muons in the final state to be produced. The minimal integrated luminosity $L_{int}$ is given by $$L_{int} = \frac{5}{\varepsilon(M_Y) \sigma(M_Y)}$$ where $\varepsilon(M_Y)$ is the detector acceptance and $\sigma(M_Y)$ is the cross seciton. Figure \[fig14\] shows the calculated values of $L_{int}$ as a function of bilepton mass. From the plot, we conclude that a integrated luminosity of order of 10 $\textnormal{pb}^{-1}$ is enough for discovering a bilepton of 400 GeV mass, which means that such signal can be observed at the very early days of LHC running with 14 TeV, even in a regime of low luminosity. Depending on the exotic quark mass, luminosities of order 10 $\textnormal{fb}^{-1}$ to 100 $\textnormal{fb}^{-1}$ are needed to discover bileptons if $M_{Y}$ = 800 GeV. These scenarios can be achieved after 1 year of LHC operation in low and high luminosity regimes, respectively. Finally, for $M_{Y}$ = 1 TeV, around 10 years of LHC operation at high luminosity would be needed for the resonance observation. This is in contrast with what would happen at the ILC where Møller and Bhabha scattering receive both huge corrections from virtual vector bileptons [@MEIROSE], so that the bilepton mass reach can be as high as $\sim$ 11 TeV, provided polarized beams are used. Figure \[fig15\] shows the integrated luminosity required for excluding bileptons at 95% CL as function of the bilepton mass. For this estimation, we have used the D0 limit calculator [@D0] to set upper limits on the cross sections that are consistent with the observation of zero events in data, assuming no background. The Bayesian technique is used to set this limits. In this approach, given a posterior probability density function for the signal cross-section, the upper limit on the signal cross section $\sigma_{\textnormal{up}}$, specified at some confidence level $100 \times \beta\%$, is given by $$\beta = \int_0^{\sigma_{\textnormal{up}}} d\sigma \rho(\sigma \arrowvert k,I)$$ where $\beta$ = 0.95, $\rho(\sigma \arrowvert k,I)$ is the posterior probability density function, $k$ is the number of events observed and $I$ represents prior information available. The acceptances values used in this calculation are the same as before. We have assumed an uncertainty of 5% on the acceptance and 10% on the integrated luminosity. It also is assumed that the errors on the acceptance and luminosity are uncorrelated. The obtained limits on the cross section are then translated to limits on the bilepton mass. Comparing Figure \[fig15\] with the discovery plot, we can see that around 60% of the discovery integrated luminosity is needed for excluding bileptons of a given mass. If a significant signal is observed, the next natural step is to determine the properties of the new particle. In order to check how well the bilepton mass can be reconstructed with the amount of data needed for discovering, we perform a unbinned maximum likelihood fit to the $M_{ee}$ and $M_{\mu \mu}$ distributions for $M_{Y}$ = 600 GeV and $M_{Y}$ = 800 GeV MC samples, with a fixed exotic quark mass of 600 GeV. The probability density function used in the fits is a Breit-Wigner and two parameters are fitted: the position of the invariant mass peak $m$ and the resonance width $\varGamma$. For each bilepton mass, fits are performed to 1000 MC experiments (i.e, 1000 $M_{ee}$ and $M_{\mu \mu}$ distributions) and the mean values of the fitted parameters are reported. The number of events in each MC experiment is fixed to 5. Table I shows the mean, $\bar{m}$, of the fitted mass values and the standard deviation of the $m$ distribution. For both bilepton masses, we see that there is a very good agreement between fitted and true masses in both channels. As expected, the spread of the distribution is larger for $M_{Y}$ = 800 GeV. The bilepton width can also be obtained from the fit at generator level, but it will be dominated by the detector resolution in a more realistic scenario, since bileptons are very narrow resonances. ---------- ---------- -------------- ---------- -------------- channel $\bar m$ $\sigma{_m}$ $\bar m$ $\sigma_{m}$ $ee$ $599.9$ $0.5$ $799.9$ $2.4$ $\mu\mu$ $600.1$ $0.6$ $799.9$ $2.6$ ---------- ---------- -------------- ---------- -------------- : Mean and standard deviation of fitted mass peak for dielectron and dimuon invariant mass distributions. \[tab:tab1\] LHC 7 TeV run potential {#lhc-7-tev-run-potential .unnumbered} ----------------------- Considering the LHC’s goals until the end of 2012 for a center-of-mass energy of 7 TeV, we estimate the potential for discovering or for setting limits on bilepton masses and couplings at this data taking stage. We consider three scenarios with 1, 5 and 10 $\textnormal{fb}^{-1}$ [@LHC] of integrated luminosity. Using the D0 limit calculator and using the same values for acceptance and uncertainties as in the previous sections, we obtain the respective bilepton masses consistent with the observation of zero events for three exotic quark masses: 427, 466 and 483 GeV for $M_{Q}$ = 400 GeV; 478, 534 and 566 GeV for both $M_{Q}$ = 600 GeV and $M_{Q}$ = 800 GeV, from the lowest to the highest integrated luminosity, respectively. These are the doubly-charged vector bilepton masses that can be excluded at 95% CL. The 7 TeV exclusion limits are summarized in Table II. --------- -------------------------- -------------------------- --------------------------- -- $M_{Q}$ 1 $\textnormal{fb}^{-1}$ 5 $\textnormal{fb}^{-1}$ 10 $\textnormal{fb}^{-1}$ $400$ $427$ $466$ $483$ $600$ $478$ $534$ $566$ $800$ $478$ $534$ $566$ --------- -------------------------- -------------------------- --------------------------- -- : Exclusion limits for doubly-charged bileptons with respect to LHC’s integrated luminosity and 331 exotic quark masses at the 7 TeV run. Masses are in GeV. \[tab:tab2\] The $5 \sigma$ discovery potentials at 5 and 10 $\textnormal{fb}^{-1}$ of integrated luminosity are respectively found to be 452, 535 GeV for $M_{Q}$ = 400 GeV, 511, 542 GeV for $M_{Q}$ = 600 GeV and 515, 544 GeV for $M_{Q}$ = 800 GeV. With 1 $\textnormal{fb}^{-1}$ the reach is 459 GeV using the highest exotic quark mass. This mass reach is way above the minimum bound of 350 GeV, so such a discovery at this phase, is not completely discarded. In any case, even if no discoveries are made at the 7 TeV run, the exclusion limits that will be established are still valuable for setting up the scenario for the 14 TeV run. The 7 TeV discovery reach results are summarized in Table III. --------- -------------------------- -------------------------- --------------------------- -- $M_{Q}$ 1 $\textnormal{fb}^{-1}$ 5 $\textnormal{fb}^{-1}$ 10 $\textnormal{fb}^{-1}$ $400$ 415 454 535 $600$ 459 511 542 $800$ 459 515 544 --------- -------------------------- -------------------------- --------------------------- -- : Discovery mass reach for doubly-charged bileptons with respect to LHC’s integrated luminosity and 331 exotic quark masses at the 7 TeV run. Masses are in GeV. \[tab:tabX\] sLHC {#slhc .unnumbered} ---- The upgrade of the LHC machine, also referred as the sLHC [@SLHC] project aims at increasing the peak luminosity by a factor of 10 and deliver approximately 3000 $\textnormal{fb}^{-1}$ to the experiments. Although it’s rather difficult to foresee what would be interesting to study at the sLHC without having the LHC run at its nominal luminosity first, here we assume that no vector bilepton signals were found at the LHC and explore the sLHC exclusion and discovery potentials for the exotic particles. Using the same techniques described in the previous sections we find the exclusion potential for three heavy quark masses: 1170 GeV for $M_{Q}$ = 400 GeV, 1220 GeV for $M_{Q}$ = 600 GeV and 1300 GeV for $M_{Q}$ = 800 GeV. The discovery potential is also increased: 1100 GeV for $M_{Q}$ = 400 GeV, 1150 GeV for $M_{Q}$ = 600 GeV and 1230 GeV for $M_{Q}$ = 800 GeV. This represents a gain of $\sim$ 200 GeV in terms of discovery mass reach compared to the default luminosity 14 TeV LHC run. This region is certainly worth exploring since it is still considerably below the upper limit of 3.5 TeV we have discussed in section II. The results for the sLHC are displayed in Table IV. $M_{Q}$ Discovery Exclusion --------- ----------- ----------- -- 400 1100 1170 600 1150 1220 800 1230 1300 : Discovery mass reach and exclusion limits for doubly-charged bileptons with respect to sLHC and 331 exotic quark masses. Masses are in GeV. \[tab:tab4\] Conclusions =========== We have investigated the LHC potential for discovering or setting limits on doubly-charged vector bileptons, in different scenarios considering important experimental aspects in the simulation like detector geometrical acceptance, luminosity uncertainty and lepton efficiency. By analysing the observable final state $e^{\mp}e^{\mp}\mu^{\pm}\mu^{\pm}$, we have found that bilepton signatures can already be observed at very early stages of LHC running with 14 TeV, if the bilepton mass is not much greater than 400 GeV. On the other hand, if the bilepton mass lies in the TeV scale, at least 10 years of the machine operation will be needed for discovering it, if $M_{Q} >$ 600 GeV. The observation of such signal, in combination with $Z^{\prime}_{331}$ searches in dilepton channel, would provide a very powerful way of discriminating between 331 models and other BSM scenarios which also predict heavy neutral gauge bosons. If no signal is observed, the LHC can extend considerably the currents limits on bilepton mass by direct search in the four lepton final state. At the current LHC energy, 7 TeV, masses up to 566 GeV can be excluded and if 10 $\textnormal{fb}^{-1}$ of data is recorded, vector bilepton masses up to 544 GeV could be discovered. We also found that the sLHC can expand the lower exclusion limits up to a mass of 1300 GeV. We also made a revision on current experimental bounds on bileptons in 331 models and the possibility of results from the LHC to exclude some versions of these models. We found that it is not possible to safely discard any 331 model, including its minimal version, at the LHC. Purely theoretical arguments taken from the literature are used to draw this conclusion. Furthermore, we investigated how the heavy quark sector of the 331 model influences our results. In some cases a substantial change in the process cross-section is observed by varying the value of the heavy quark masses. Since some of the best previous limits on bileptons were coming from experiments containing at least one leptonic beam, we conclude that new results from the LHC will be indispensable in determining to a more accurate extent which models like 331 can be disfavored or discovered. The final state studied in this article will be the best channel to experimentally determine this. [99]{} F. Cuypers and S. Davidson, Eur. Phy. J. C [**2**]{},503 (1998) \[hep-ph/9609487\] and references therein. F. Pisano and V. Pleitez, Phys. Rev. D [**46**]{}, 410 (1992); P. H. Frampton, Phys. Rev. Lett. [**69**]{}, 2889 (1992). M. B. Tully and G. C. Joshi, Phys. Lett. B466, 333 (1993); hep-ph/9905552. Willmann et al, Phys. Rev. Lett. 82, 49 (1999). Y. A. Coutinho, P. P. Queiróz Filho, M. D. Tonasse, Phys. Rev. D 60, 115001 (1999). P. H. Frampton, J. T. Liu, B. C. Rasco, and D. Ng, Mod. Phys. Lett. A 9, 1975 (1994); P. H. Frampton and D. Ng, Phys. Rev. D 45, 4240 (1992); E. D. Carlson and P. H. Frampton, Phys. Lett. B 283, 123 (1992). B. Dion, T. Grégoire, D. London, L. Marleau and H. Nadeau, Phys.Rev. D59 (1999). R. D. Peccei and H. Quinn, Phys. Rev. Lett. 38, 1440 (1977); Phys. Rev. D16, 1791 (1977). David L. Anderson and Marc Sher, Phys. Rev. D 72, 095014 (2005). David J. Gross, Jeffrey A. Harvey, Emil Martinec, and Ryan Rohm, Phys. Rev. Lett. 54, 502–505 (1985). Daniel Ng, Phys. Rev. D 49, 4805–4811 (1994). V. Pleitez, Phys. Rev. D 61, 057903 (2000). P. Jain and S. D. Joglekar, Phys. Lett. B 407, 151 (1997). J. C. Montero, V. Pleitez, M. C. Rodriguez, Phys.Rev. D65 (2002); J. C. Montero, V. Pleitez, M. C. Rodriguez, Phys.Rev. D70 (2004); M. C. Rodriguez, Int.J.Mod.Phys. A21 (2006); P. V. Dong, D. T. Huong, M. C. Rodriguez, H. N. Long, Nucl.Phys.B772:150-174 (2007). E. Boos et al, Nucl. Instrum. Meth. A534 (2004); A.Pukhov et al, INP MSU report 98-41/542 (arXiv:hep-ph/9908288). Hoang Ngoc Long and Dang Van Soa, Nucl. Phys. B601: 361-379, (2001). Paul H. Frampton and Daniel Ng, Phys. Rev. D 45, 4240–4245 (1992). A. Semenov. Nucl.Inst. and Meth. A393 (1997) p. 293. J. Pumplin et al, JHEP0207:012, (2002). P. Das, P. Jain, and D. W. McKay, Phys. Rev. D 59, 055011 (1999). J. G. Duenas, N. Gutierrez, R. Martinez and F. Ochoa, Eur. Phys. J. C 60, 653-659 (2009). ATLAS Collaboration, “Expected performance of the ATLAS experiment : detector, trigger and physics”, CERN-OPEN-2008-020 (2008) B. Meirose and A.J. Ramalho, Phys. Rev. D73, 075013 (2006); B. Meirose and A.J. Ramalho, J. Phys. G: Nucl. Part. Phys. 36, 095007 (2009). I. Bertram, G. Landsberg, J. Linnemann, R. Partridge M. Paterno and H.B. Prosper, Fermilab TM-2104 (2000); For the “Simple Limit Calculator” see http://www-d0.fnal.gov/Run2Physics/limit\_calc/limit\_calc.html. CERN Press Office (31 January 2011), “CERN announces LHC to run in 2012”. S. Fartoukh, sLHC Project Report 0049 (http://cdsweb.cern.ch/record/1301180) (2010). This work has been partially supported by the European Commission within the 7th framework programme (FP7-PEOPLE-2009-IEF, Grant Agreement no. 253339).

ArXiv

--- abstract: 'Comparing neuronal bursting models (NBM) with slow-fast autonomous dynamical systems (S-FADS), it appears that the specific features of a (NBM) do not allow a determination of the analytical slow manifold equation with the singular approximation method. So, a new approach based on Differential Geometry, generally used for (S-FADS), is proposed. Adapted to (NBM), this new method provides three equivalent manners of determination of the analytical slow manifold equation. Application is made for the three-variables model of neuronal bursting elaborated by Hindmarsh and Rose which is one of the most used mathematical representation of the widespread phenomenon of oscillatory burst discharges that occur in real neuronal cells.' author: - | Jean-Marc Ginoux and Bruno Rossetto,\ P.R.O.T.E.E. Laboratory, I.U.T. de Toulon,\ Université du Sud, B.P. 20132, 83957, La Garde cedex, France,\ E-mail: [email protected], [email protected] date: '[**Keywords**]{}: differential geometry; curvature; torsion; slow-fast dynamics; neuronal bursting models.' title: Slow Manifold of a Neuronal Bursting Model --- Slow-fast autonomous dynamical systems, neuronal bursting models {#sec1} ================================================================ Dynamical systems ----------------- In the following we consider a system of differential equations defined in a compact E included in $\mathbb{R}$: $$\label{eq1} \frac{d{\vec {X}}}{dt} = \vec \Im \left( \vec {X} \right)$$ with $$\vec {X} = \left[ {x_1 ,x_2 ,...,x_n } \right]^t \in E \subset \mathbb{R}^n$$ and $$\vec \Im \left( \vec {X} \right) = \left[ {f_1 \left( \vec {X} \right),f_2 \left( \vec {X} \right),...,f_n \left( \vec {X} \right)} \right]^t \in E \subset \mathbb{R}^n$$ The vector $\vec \Im \left( \vec {X} \right)$ defines a velocity vector field in E whose components $f_i $ which are supposed to be continuous and infinitely differentiable with respect to all $x_i $ and $t$, i.e., are $C^\infty $ functions in E and with values included in $\mathbb{R}$, satisfy the assumptions of the Cauchy-Lipschitz theorem. For more details, see for example [@2]. A solution of this system is an integral curve $\vec {X}\left( t \right)$ tangent to $\vec \Im $ whose values define the *states* of the *dynamical system* described by the Eq. (\[eq1\]). Since none of the components $f_i $ of the velocity vector field depends here explicitly on time, the system is said to be *autonomous*. Slow-fast autonomous dynamical system (S-FADS) ---------------------------------------------- A (S-FADS) is a *dynamical system* defined under the same conditions as previously but comprising a small multiplicative parameter $\varepsilon$ in one or several components of its velocity vector field: $$\label{eq2} \frac{d{\vec {X}}}{dt} = \vec \Im_\varepsilon \left( \vec {X} \right)$$ with $$\vec \Im_\varepsilon \left( \vec {X} \right) = \left[ {\frac{1}{\varepsilon}f_1 \left( \vec {X} \right),f_2 \left( \vec {X} \right),...,f_n \left( \vec {X} \right)} \right]^t \in E \subset \mathbb{R}^n$$ $$0 < \varepsilon \ll 1$$ The functional jacobian of a (S-FADS) defined by (\[eq2\]) has an eigenvalue called “fast”, i.e., great on a large domain of the phase space. Thus, a “fast” eigenvalue is expressed like a polynomial of valuation $ - 1$ in $\varepsilon $ and the eigenmode which is associated with this “fast” eigenvalue is said:\ - “evanescent” if it is negative, - “dominant” if it is positive.\ The other eigenvalues called “slow” are expressed like a polynomial of valuation $0$ in $\varepsilon $. Neuronal bursting models (NBM) ------------------------------ A (NBM) is a *dynamical system* defined under the same conditions as previously but comprising a large multiplicative parameter $\varepsilon^{ - 1}$ in one component of its velocity vector field: $$\label{eq3} \frac{d{\vec {X}}}{dt} = \vec \Im_\varepsilon \left( \vec {X} \right)$$ with $$\vec \Im_\varepsilon \left( \vec {X} \right) = \left[ { f_1 \left( \vec {X} \right),f_2 \left( \vec {X} \right),...,\varepsilon f_n \left( \vec {X} \right)} \right]^t \in E \subset \mathbb{R}^n$$ $$0 < \varepsilon \ll 1$$ The presence of the multiplicative parameter $\varepsilon^{ - 1}$ in one of the components of the velocity vector field makes it possible to consider the system (\[eq3\]) as a kind of *slow*-*fast* autonomous dynamical system (S-FADS). So, it possesses a *slow manifold*, the equation of which may be determined. But, paradoxically, this model is not *slow-fast* in the sense defined previously. A comparison between three-dimensional (S-FADS) and (NBM) presented in Table 1 emphasizes their differences. The dot $(\cdot)$ represents the derivative with respect to time and $\varepsilon \ll 1$. [|c|c|c|]{}\ & \ $ \hspace{0.1in} \dfrac{d{\vec {X}}}{dt} \left( {{\begin{array}{*{20}c} \dot {x} \hfill \\ \dot {y} \hfill \\ \dot {z} \hfill \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\dfrac{1}{\varepsilon } f\left( {x,y,z} \right)} \vspace{4pt} \\ {\mbox{ }g\left( {x,y,z} \right)} \vspace{4pt} \\ {h\left( {x,y,z} \right)} \\ \end{array} }} \right) \mbox{ }$ & $ \hspace{0.1in} \dfrac{d{\vec {X}}}{dt} \left( {{\begin{array}{*{20}c} \dot {x} \hfill \\ \dot {y} \hfill \\ \dot {z} \hfill \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\mbox{ }f\left( {x,y,z} \right)} \vspace{4pt} \\ {\mbox{ }g\left( {x,y,z} \right)} \vspace{4pt} \\ {\varepsilon h\left( {x,y,z} \right)} \\ \end{array} }} \right) \mbox{ } $ \ & \ & \ $ \hspace{0.1in} \dfrac{d{\vec {X}}}{dt} \left( {{\begin{array}{*{20}c} \dot {x} \hfill \\ \dot {y} \hfill \\ \dot {z} \hfill \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\mbox{ } fast} \vspace{4pt} \\ {\mbox{ }slow} \vspace{4pt} \\ {slow} \\ \end{array} }} \right) \mbox{ } $ & $\hspace{0.1in} \dfrac{d{\vec {X}}}{dt} \left( {{\begin{array}{*{20}c} \dot {x} \hfill \\ \dot {y} \hfill \\ \dot {z} \hfill \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\mbox{ }fast} \vspace{4pt} \\ {\mbox{ }fast} \vspace{4pt} \\ {slow} \\ \end{array} }} \right) \mbox{ }$ \ & \ \[tab1\] Analytical slow manifold equation {#sec2} ================================= There are many methods of determination of the analytical equation of the slow manifold. The classical one based on the singular perturbations theory [@1] is the so-called *singular approximation method*. But, in the specific case of a (NBM), one of the hypothesis of the Tihonov’s theorem is not checked since the *fast dynamics* of the *singular approximation* has a periodic solution. Thus, another approach developed in [@4] which consist in using *Differential Geometry* formalism may be used. Singular approximation method ----------------------------- The *singular approximation* of the *fast* dynamics constitutes a quite good approach since the third component of the velocity is very weak and so, $z$ is nearly constant along the periodic solution. In dimension three the system (\[eq3\]) can be written as a system of differential equations defined in a compact E included in $\mathbb{R}$: $$\hspace{0.1in} \frac{d{\vec {X}}}{dt} = \left( {{\begin{array}{*{20}c} {\frac{d{x}}{dt}} \vspace{4pt} \\ {\frac{d{y}}{dt}} \vspace{4pt} \\ {\frac{d{z}}{dt}} \\ \end{array} }} \right) = \vec \Im_\varepsilon \left( {{\begin{array}{*{20}c} {\mbox{ }f\left( {x,y,z} \right)} \vspace{4pt} \\ {\mbox{ }g\left( {x,y,z} \right)} \vspace{4pt} \\ {\varepsilon h\left( {x,y,z} \right)} \\ \end{array} }} \right)$$ On the one hand, since the system (\[eq3\]) can be considered as a (S-FADS), the *slow* dynamics of the *singular approximation* is given by: $$\label{eq4} \hspace{0.4in} \left\{ {{\begin{array}{*{20}c} {f\left( {x,y,z} \right) = 0} \hfill \\ {g\left( {x,y,z} \right) = 0} \hfill \\ \end{array} }} \right.$$ The resolution of this reduced system composed of the two first equations of the right hand side of (\[eq3\]) provides a one-dimensional *singular manifold*, called *singular curve*. This curve doesn’t play any role in the construction of the periodic solution. But we will see that there exists all the more a *slow dynamics*. On the other hands, it presents a *fast* dynamics which can be given while posing the following change: $$\tau = \varepsilon t\mbox{ } \Leftrightarrow \mbox{ }\frac{d}{dt} = \varepsilon \frac{d}{d{\tau} }$$ The system (\[eq3\]) may be re-written as: $$\label{eq5} \frac{d{\vec {X}}}{d{\tau} } = \left( {{\begin{array}{*{20}c} {\frac{d{x}}{d{\tau} }} \vspace{4pt} \\ {\frac{d{y}}{d{\tau} }} \vspace{4pt} \\ {\frac{d{z}}{d{\tau} }} \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\varepsilon ^{ - 1}f\left( {x,y,z} \right)} \vspace{4pt} \\ {\varepsilon ^{ - 1}g\left( {x,y,z} \right)} \vspace{4pt} \\ {\mbox{ }h\left( {x,y,z} \right)} \\ \end{array} }} \right)$$ So, the *fast* dynamics of the *singular approximation* is provided by the study of the reduced system composed of the two first equations of the right hand side of (\[eq5\]). $$\label{eq6} \left. {\frac{d{\vec {X}}}{d{\tau} }} \right|_{fast} = \left( {{\begin{array}{*{20}c} {\frac{d{x}}{d{\tau} }} \vspace{4pt} \\ {\frac{d{y}}{d{\tau} }} \\ \end{array} }} \right) = \vec{\Im_\varepsilon} \left( {{\begin{array}{*{20}c} {\varepsilon ^{ - 1}f\left( {x,y,z^\ast } \right)} \vspace{4pt} \\ {\varepsilon ^{ - 1}g\left( {x,y,z^\ast } \right)} \\ \end{array} }} \right)$$ Each point of the *singular curve* is a singular point of the *singular approximation* of the *fast* dynamics. For the $z$ value for which there is a periodic solution, the *singular approximation* exhibits an unstable focus, attractive with respect to the *slow* eigendirection. Differential Geometry formalism ------------------------------- Now let us consider a three-dimensional system defined by (\[eq3\]) and let’s define the instantaneous acceleration vector of the *trajectory curve* $\vec {X}\left( t \right)$. Since the functions $f_i $ are supposed to be $C^\infty$ functions in a compact E included in $\mathbb{R}$, it is possible to calculate the total derivative of the vector field $\vec{\Im_\varepsilon}$. As the instantaneous vector function $\vec V \left( t \right)$ of the scalar variable $t$ represents the velocity vector of the mobile M at the instant $t$, the total derivative of $\vec V \left( t \right)$ is the vector function $\vec \gamma \left( t \right)$ of the scalar variable $t$ which represents the instantaneous acceleration vector of the mobile M at the instant $t$. It is noted: $$\label{eq7} \vec \gamma \left( t \right)\mbox{ } = \mbox{ }\frac{d{\vec V} \left( t \right)}{dt}$$ Even if neuronal bursting models are not exactly slow-fast autonomous dynamical systems, the new approach of determining the *slow manifold* equation developed in [@4] may still be applied. This method is using *Differential Geometry* properties such as *curvature* and *torsion* of the *trajectory curve* $\vec {X}\left( t \right)$, integral of *dynamical systems* to provide their *slow manifold* equation. \[prop un\] The location of the points where the local torsion of the trajectory curves integral of a dynamical system defined by (\[eq3\]) vanishes, provides the analytical equation of the slow manifold associated with this system. $$\label{eq8} \frac{1}{\Im } = - \frac{\dot {\vec {\gamma }} \cdot \left( {\vec \gamma \times \vec V } \right)}{\left\| {\vec \gamma \times \vec V } \right\|^2} = 0\mbox{ } \Leftrightarrow \mbox{ }\dot {\vec {\gamma }} \cdot \left( {\vec \gamma \times \vec V } \right) = 0$$ Thus, this equation represents the *slow manifold* of a neuronal bursting model defined by (\[eq3\]). The particular features of neuronal bursting models (\[eq3\]) will lead to a simplification of this Proposition 1. Due to the presence of the small multiplicative parameter $\varepsilon $ in the third components of its velocity vector field, instantaneous velocity vector $\vec V \left( t \right)$ and instantaneous acceleration vector $\vec \gamma \left( t \right)$ of the model (\[eq3\]) may be written: $$\label{eq9} \hspace{0.4in} \vec V \left( {{\begin{array}{*{20}c} \dot {x} \hfill \\ \dot {y} \hfill \\ \dot {z} \hfill \\ \end{array} }} \right) = \vec \Im_\varepsilon \left( {{\begin{array}{*{20}c} {O\left( {\varepsilon ^0} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^0} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^1} \right)} \vspace{4pt} \\ \end{array} }} \right)$$ and $$\label{eq10} \hspace{0.3in} \vec {\gamma }\left( {{\begin{array}{*{20}c} \ddot {x} \hfill \\ \ddot {y} \hfill \\ \ddot {z} \hfill \\ \end{array} }} \right) = \frac{d{\vec \Im_\varepsilon }}{dt}\left( {{\begin{array}{*{20}c} {O\left( {\varepsilon ^1} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^1} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^2} \right)} \vspace{4pt} \\ \end{array} }} \right)$$ where ${O\left( {\varepsilon ^n} \right)}$ is a polynomial of $n$ degree in $\varepsilon$ Then, it is possible to express the vector product $\vec V \times \vec \gamma $ as: $$\label{eq11} \hspace{0.4in} \vec V \times \vec \gamma = \left( {{\begin{array}{*{20}c} {\dot {y}\ddot {z} - \ddot {y}\dot {z}} \hfill \\ {\ddot {x}\dot {z} - \dot {x}\ddot {z}} \hfill \\ {\dot {x}\ddot {y} - \ddot {x}\dot {y}} \hfill \\ \end{array} }} \right)$$ Taking into account what precedes (\[eq9\], \[eq10\]), it follows that: $$\label{eq12} \hspace{0.4in}\vec V \times \vec \gamma = \left( {{\begin{array}{*{20}c} {O\left( {\varepsilon ^2} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^2} \right)} \vspace{4pt} \\ {O\left( {\varepsilon ^1} \right)} \vspace{4pt} \\ \end{array} }} \right)$$ So, it is obvious that since $\varepsilon $ is a small parameter, this vector product may be written: $$\label{eq13} \hspace{0.4in}\vec V \times \vec \gamma \approx \left( {{\begin{array}{*{20}c} 0 \\ 0 \\ {O\left( {\varepsilon ^1} \right)} \\ \end{array} }} \right)$$ Then, it appears that if the third component of this vector product vanishes when both instantaneous velocity vector $\vec V \left( t \right)$ and instantaneous acceleration vector $\vec \gamma \left( t \right)$ are collinear. This result is particular to this kind of model which presents a small multiplicative parameter in one of the right-hand-side component of the velocity vector field and makes it possible to simplify the previous Proposition 1.\ \[prop deux\] The location of the points where the instantaneous velocity vector $\vec V \left( t \right)$ and instantaneous acceleration vector $\vec \gamma \left( t \right)$ of a neuronal bursting model defined by (\[eq3\]) are collinear provides the analytical equation of the *slow manifold* associated with this dynamical system. $$\label{eq14} \vec V \times \vec \gamma = \vec 0 \mbox{ } \Leftrightarrow \mbox{ }\dot {x}\ddot {y} - \ddot {x}\dot {y} = 0$$ Another method of determining the *slow manifold* equation proposed in [@14] consists in considering the so-called *tangent linear system approximation*. Then, a coplanarity condition between the instantaneous velocity vector $\vec V \left( t \right)$ and the *slow* eigenvectors of the *tangent linear system* gives the *slow manifold* equation. $$\label{eq15} \vec V .\left( {\vec {Y_{\lambda _2 } } \times \vec {Y_{\lambda _3 } } } \right) = 0$$ where $\vec {Y_{\lambda _i } } $ represent the *slow* eigenvectors of the *tangent linear system*. But, if these eigenvectors are complex the *slow manifold* plot may be interrupted. So, in order to avoid such inconvenience, this equation has been multiplied by two conjugate equations obtained by circular permutations. $$\left[ {\vec V \cdot \left( {\vec {Y_{\lambda _2 } } \times \vec {Y_{\lambda _3 } } } \right)} \right] \cdot \left[ {\vec V \cdot \left( {\vec {Y_{\lambda _1 } } \times \vec {Y_{\lambda _2 } } } \right)} \right] \cdot \left[ {\vec V \cdot \left( {\vec {Y_{\lambda _1 } } \times \vec {Y_{\lambda _3 } } } \right)} \right] = 0$$ It has been established in [@4] that this real analytical *slow manifold* equation can be written: $$\label{eq16} \left( {J^2\vec V } \right) \cdot \left( {\vec \gamma \times \vec V } \right) = 0$$ since the the *tangent linear system approximation* method implies to suppose that the functional jacobian matrix is stationary. That is to say $$\frac{d J}{dt} = 0$$ and so, $$\dot {\vec {\gamma }} = J\frac{d{\vec V }}{d t} + \frac{d J}{d t}\vec V = J\vec \gamma + \frac{d J}{d t}\vec V = J^2\vec V + \frac{d J}{d t}\vec V \approx J^2\vec V$$ \[prop trois\] The coplanarity condition (\[eq15\]) between the instantaneous velocity vector and the slow eigenvectors of the tangent linear system transformed into the real analytical equation (\[eq16\]) provides the *slow manifold* equation of a neuronal bursting model defined by (\[eq3\]). Application to a neuronal bursting model {#sec3} ======================================== The transmission of nervous impulse is secured in the brain by action potentials. Their generation and their rhythmic behaviour are linked to the opening and closing of selected classes of ionic channels. The membrane potential of neurons can be modified by acting on a combination of different ionic mechanisms. Starting from the seminal works of Hodgkin-Huxley [@7; @11] and FitzHugh-Nagumo [@3; @12], the Hindmarsh-Rose [@6; @13] model consists of three variables: $x$, the membrane potential, $y$, an intrinsic current and $z$, a *slow* adaptation current. Hindmarsh-Rose model of bursting neurons ---------------------------------------- $$\label{eq17} \hspace{0.4in}\left\{ {{\begin{array}{{ll}} {\frac{d{x}}{dt} = y - f\left( x \right) - z + I} \vspace{4pt} \hfill \\ {\frac{d{y}}{dt} = g\left( x \right) - y} \vspace{4pt} \hfill \\ {\frac{d{z}}{dt} = \varepsilon \left( {h\left( x \right) - z} \right)} \hfill \\ \end{array}}} \right.$$ $I$ represents the applied current, $f\left( x \right) = ax^3 - bx^2$ and $g\left( x \right) = c - dx^2$ are respectively cubic and quadratic functions which have been experimentally deduced [@5]. $\varepsilon$ is the time scale of the slow adaptation current and $h\left( x \right) = x - x^\ast$ is the scale of the influence of the *slow* dynamics, which determines whether the neuron fires in a tonic or in a burst mode when it is exposed to a sustained current input and where $\left( {x^\ast ,y^\ast } \right)$ are the co-ordinates of the leftmost equilibrium point of the model (1) without adaptation, i.e., $I = 0$.\ Parameters used for numerical simulations are:\ $a = 1$, $b = 3$, $c = 1$, $d = 5$, $\varepsilon = 0.005$, $s = 4$, $x^\ast = \frac{ - 1 - \sqrt 5 }{2}$ and $I = 3.25$.\ While using the method proposed in the section 2 it is possible to determine the analytical *slow manifold* equation of the Hindmarsh-Rose 84’model [@6]. Slow manifold of the Hindmarsh-Rose 84’model -------------------------------------------- In Fig. 1 is presented the *slow manifold* of the Hindmarsh-Rose 84’model determined with the Proposition 1.\ ![*Slow manifold* of the Hindmarsh-Rose 84’model with the Prop. 1.[]{data-label="fig:1"}](epnads1.eps){width="10cm" height="10cm"} The *slow manifold* provided with the use of the *collinearity condition* between both instantaneous velocity vector and instantaneous acceleration vector, i.e., while using the Proposition 2 is presented in Fig. 2.\ ![*Slow manifold* of the Hindmarsh-Rose 84’model with the Prop. 2.[]{data-label="fig:2"}](epnads2.eps){width="10cm" height="10cm"} Figure 3 presents the *slow manifold* of the Hindmarsh-Rose 84’model obtained with the *tangent linear system approximation*, i.e., with the use of Proposition 3. ![*Slow manifold* of the Hindmarsh-Rose 84’model with the Prop. 3.[]{data-label="fig:3"}](epnads3.eps){width="10cm" height="10cm"} Discussion {#sec4} ========== Since in the case of neuronal bursting model (NBM) bursting models one of the Tihonov’s hypothesis is not checked, the classical *singular approximation method* can not be used to determine the analytical *slow manifold* equation. In this work the application of the *Differential Geometry* formalism provides new alternative methods of determination of the *slow manifold* equation of a neuronal bursting model (NBM). - the *torsion method*, i.e., the location of the points where the local *torsion* of the *trajectory curve*, integral of *dynamical systems* vanishes, - the *collinearity condition* between the instantaneous velocity vector $\overrightarrow V $, the instantaneous acceleration vector $\overrightarrow \gamma $, - the *tangent linear system approximation*, i.e., the coplanarity condition between the instantaneous velocity vector eigenvectors transformed into a real analytical equation. The striking similarity of all figures due to the smallness of the parameter $\varepsilon$ highlights the equivalence between all the propositions. Moreover, even if the presence of this small parameter $\varepsilon$ in one of the right-hand-side component of the instantaneous velocity vector field of a (NBM) prevents from using the *singular approximation method*, it clarifies the Proposition 1 and transforms it into a *collinearity condition* in dimension three, i.e., Proposition 2. Comparing (S-FADS) and (NBM) in Table 1 it can be noted that in a (S-FADS) there is one fast component and two fast while in a (NBM) the situation is exactly reversed. Two fast components and one slow. So, considering (NBM) as a particular class of (S-FADS) we suggest to call (NBM) *fast-slow* instead of *slow-fast* in order to avoid any confusion. Further research should highlight other specific features of (NBM). Acknowledgements {#acknowledgements .unnumbered} ================ Authors would like to thank Professors M. Aziz-Alaoui and C. Bertelle for their useful collaboration. [08]{} Andronov AA, Khaikin SE, [&]{} Vitt AA (1966) Theory of oscillators, Pergamon Press, Oxford Coddington EA [&]{} Levinson N, (1955) Theory of Ordinary Differential Equations, Mac Graw Hill, New York Fitzhugh R (1961) Biophys. J 1:445–466 Ginoux JM [&]{} Rossetto B (2006) Int. J. Bifurcations and Chaos, (in print) Hindmarsh JL [&]{} Rose RM (1982) Nature 296:162–164 Hindmarsh JL [&]{} Rose RM (1984) Philos. Trans. Roy. Soc. London Ser. B 221:87–102 Hodgkin AL [&]{} Huxley AF (1952) J. Physiol. (Lond.) 116:473–96 Hodgkin AL [&]{} Huxley AF (1952) J. Physiol. (Lond.) 116: 449–72 Hodgkin AL [&]{} Huxley AF (1952) J. Physiol. (Lond.) 116: 497–506 Hodgkin AL [&]{} Huxley AF (1952) J. Physiol. (Lond.) 117: 500–44 Hodgkin AL, Huxley AF [&]{} Katz B (1952) B. Katz J. Physiol. (Lond.) 116: 424–48 Nagumo JS, Arimoto S [&]{} Yoshizawa S (1962) Proc. Inst. Radio Engineers 50:2061–2070 Rose RM [&]{} Hindmarsh JL (1985) Proc. R. Soc. Ser. B 225:161–193 Rossetto B, Lenzini T, Suchey G [&]{} Ramdani S (1998) Int. J. Bifurcation and Chaos, vol. 8 (11):2135-2145

ArXiv

--- abstract: 'We have been able to observe directly extended instantons on the lattice, with a new method that does not require dislocations to measure them, and where we do not perform cooling. We showed, based on the simple Abelian Higgs model in $1+1$ dim., that one can extract the instanton and anti-instanton density and their size, by measuring the topological charge, $Q_v$ , on sub-volumes $v$ larger than the instanton sizes, but smaller than the periodic lattice of size $V$. We are working on the generalization for non-abelian models.' address: | FB Physik, Carl von Ossietzky Universität Oldenburg\ 26111 Oldenburg, Germany author: - 'E. Mendel [^1] and G. Nolte [^2]' title: Extended instantons generated on the lattice --- Introduction ============ Instantons [@Bela; @tHoo] are expected to play an important role in such diverse phenomena as net Baryon number generation, the $U(1)$ problem or even quark confinement. In order to extract their Quantum contribution to the path integral it is desirable to use a lattice formulation. We will propose a method that enables the observation of these topological non trivial configurations, even for relatively smooth configurations that don’t present lattice artifacts called dislocations. (The topological charge is usually only measured on the whole volume $V$ where, as a consequence of periodic boundary conditions (PBC), we can only see a net charge if we have dislocations). Furthermore, we will show that one can extract information on the instanton density $\rho_I$ and size $L_I$, without the need of cooling [@Hoek] to suppress the quantum fluctuations. Note that one will still need improved actions [@Garc] or cooling, for $4$ dimensional gauge theories, in order to avoid small instantons comparable with the lattice spacing $a$. If eventually “perfect actions” are found that would avoid dislocations, our method could be very useful to enable the extraction of topological information. The method [@Our], which we present here, consists in measuring the topological charge on sub-volumes $v$ (of $V$), larger than or of the order of the relevant instantons (I, AI). Let us first assume that one can reach a coupling regime where the gauge configurations are smooth enough so that there are no individual sites where the topological density gets even close to $\pm 1/2$. This can be reached for the abelian Higgs model in $1+1$ dimensions at weak coupling. Then one has safely configurations without dislocations and if we take PBC the total geometric topological charge is exactly zero. This just means that the charge for a number of I and AI has to add to zero. We will use as our main observable, the probability distribution to get some topological charge $Q$ in sub-volumes $v$, $P_v (Q)$. For very large $v$, this distribution will have lumps peaked at integer values, representing a net number of I minus AI in $v$. As we reduce the volume $v$ to sizes smaller than the instanton, we can only see a distribution peaked at $Q=0$ as we have just pieces of instantons in $v$ or quantum fluctuations. Interestingly, these are $Q$ fluctuations in regions of stable topological vacua and therefore tend to cancel out over fairly short distances. Due to their topological character the total dispersion $\sigma$, that they produce in $Q$, only grows as $\sqrt {s}$ with $s$ being the free surface. For larger volumes, of the order or somewhat larger than the instantons, we start getting relevant contributions to $P_v (Q)$ from instantons, where the charge adds coherently to $\pm 1$ (if they are fully in $v$) in contrast to the fluctuations. The idea is then to look at scales $v$, where: $$a^D \ll v_I < \ \ v \ \ < V$$ so that we have mainly suppressed the fluctuations, and can study the ensemble of I and AI with effectively almost free spatial BC. For the abelian Higgs model on fine lattices, we have found clearly identifiable extended instantons over several lattice spacings, while the charge over the whole $V$ stays always zero. Assuming an almost free ensemble of I and AI, we have been able to extract an instanton average density $\rho_I$ and size $L_I$ (also as function of temperature, $T=1/N_\tau$). The density $\rho_I$ which can also be extracted from the second moment of $P_v (Q)$ for large $v$, falls steeply at higher $T$ as the instantons get squeezed in the time direction. We have also studied [@Our] the evolution of the position of the instantons in Monte Carlo time and found that they get created locally as I - AI pairs and annihilate later when they encounter another pairing possibility. As for some models it is hard to avoid dislocations, we have analyzed for our case what happens if one takes rougher lattices where dislocations start to appear. Only the net number $N$ of dislocations matters (pairs can be gauged away), effectively behaving as modified boundary conditions with $N$ instantons. We have found in this coupling regime, that the probability to be in a sector with $N$ net instantons in the full volume gives a compatible distribution to the case without dislocations in sub-volumes of the same size. The previous discussion has been done using the geometrical definition for the topological density [@Lusch; @Pana] but we have also investigated the goodness of the field theoretic definition which is easier to implement [@DiVec] and gives the right naive continuum limit, but does have only approximately right addition properties and so does not give exactly integers on the full volume $V$. Topological Q for Abelian Higgs model ====================================== The lattice formulation of the abelian Higgs model in 1+1 dim., which is one of the simplest allowing non-trivial topology, has the action $S_n$: $$\lambda\big(\Phi_n^\ast\Phi_n-1\big)^2-\beta Re(U_n) -2\kappa Re(\Phi_n^\ast U_{n,\mu}\Phi_{n+\mu})$$ Here $\Phi_n$ is the Higgs field, $U_{n,\mu}= \exp(i\theta_{n,\mu}) $ are the links and $U_n= \Pi_{\Box}U_{n,\mu} $ the $2$-d plaquettes. In the continuum the topological charge in $v$ is $$Q_v=\frac{1}{4\pi}\int_v d^2x\ \epsilon_{\mu\nu}F_{\mu\nu} =\frac{1}{2\pi}\int_{\partial v} ds\ n_\mu A_\mu \ . \label{bound}$$ so that on the whole $V$ it is always zero for PBC. On the lattice the field-theoretic definition, giving the right naive continuum limit, $$Q^F_v :=\frac{1}{2\pi}\sum_{n\in v} Im(U_n) , \label{def.f}$$ has the disadvantage of not being a proper topological density, not satisfying exactly the second equality in eq.(\[bound\]) and so it isn’t strictly $0$ in $V$. The geometrical charge definition, which is the simplest example of the fibre bundle construction used in the non-abelian case [@Lusch; @Pana], is $$Q^G_v :=\frac{1}{2\pi i}\sum_{n\in v}\log(U_n). \label{def.g1}$$ This $Q^G_v$ may still violate eq.(\[bound\]) by an integer amount in cases where the phases in the links of a plaquette add up to more than $\pm \pi$, this value being reduced by $\pm 2 \pi$, by the $\log$ to its principal branch. These lattice artifacts, called dislocations, are the reason for obtaining net charge on a periodic lattice. These artifacts can be avoided on fine enough lattices at least for this model. In order to keep track of these dislocations we will directly work with the phases $\theta_{n,\mu}$ of the links as the fundamental variables. For the plaquettes, $\theta_n=\theta_{n,\hat{x}}+\theta_{n+\hat{x},\hat{\tau}} -\theta_{n+\hat{\tau},\hat{x}} -\theta_{n,\hat{\tau}}$. We then define$$Q_v:=\frac{1}{2\pi }\sum_{n\in v}[\theta_n] \ , \label{def.g2}$$ where the brackets shift $\theta_n$ by an integer multiple of $2\pi$ into the interval $(-\pi,\pi]$. This definition is clearly equivalent to (\[def.g1\]), however, using $\theta_n$ as the fundamental variable we are now able to locate dislocations, namely places for which $|\theta_n| > \pi $. Note that in our Metropolis algorithm we update the $\theta_{n,\mu}$ by small changes from the old values, otherwise we could trivially generate dislocations by adding $2 \pi$ to a link. We find that if we choose parameters corresponding to a fine enough lattice, so that the physical instantons are much larger than $a$, we can avoid obtaining any dislocations at all over the whole Monte Carlo run. Most simulations have been done then around the “scaling region ” [@Our], $( \lambda=.2, \kappa=.37, \beta > 7. )$. =8.2cm Identifying Instantons ====================== In order to be able to identify instantons clearly, we have chosen lattices with spatially very elongated $V$ and with subvolumes $v$ which wrap around the $\tau$ axis in order to suppress the charge fluctuations along that long “boundary”. The finite temperature also serves to get a diluter system of instantons. We took for example $V = 320 \times 6$ and $v = (20-160) \times 6$ and could clearly see the creation of I - AI pairs. The charge probability distribution $P_v(Q)$ is shown in fig. 1, with clear peaks at integer charge. For a quantitative analysis to extract $\rho_I$, $L_I$ and the dispersion in the fluctuations $\sigma$, we developed a simple model of almost noninteracting instantons. Assuming a number density $\rho_I = \rho_{AI}$, the probability to have $n$ pairs in the total $V$ is $$p_n(\rho V)=\frac{(\rho V)^{2n}}{n!^2}\frac{1}{N} \ ,$$ where $N$ is the normalization. This $p_n$ has to be folded with the number of ways to get a charge $Q$ in $v$, assuming that the $n$ instantons of size $L_I$ can be located anywhere in $V$. With this distribution $f_n^v(Q)$ in a given sector, we get $$P_v(Q) =\sum_n p_n f_n^v(Q) \ .$$ This ideal distribution can be smeared by a gaussian with a width $\sigma$ due to quantum fluctuations. The best fit for the parameters in lattice units is $\rho \approx .0011$, $L_I \approx28$ and $\sigma \approx .27$. The topological susceptibility can also be modeled [@Our] for large $V$ being $<Q^2_v> \approx 2 v \rho_I$. We have checked compatibility for various $v$, and calculated at some T. Beyond optimal conditions ========================= As for some models we cannot reach couplings where we can suppress dislocations, it is interesting to compare for a $\beta=6.3$, where they just start to appear, the $P_{Vd} (Q)$ for sectors with net number of dislocations (and net instantons) with the $P_v (Q)$ without dislocations with $v=V_d$. The agreement is shown in Table 1, telling us that the net instantons come with the right $S$ and entropy. In this sense the traditional way to measure topology is fine, at least for large instantons. We also used the field theoretic $Q_v^F$ and found good agreement [@Our] with $Q_v^G$ for fine lattices. For $SU(2)$ we have observed much smaller fluctuations for $Q_v^F$ if, instead of symmetrizing loops as in Ref.[@DiVec], we reorder terms (same $Q$ on full $V$) taking for each hypercube all loops wrapping it. n 0 1 2 3 -------------- ---- -------- ---------- --------- with disloc. 1. .51(4) .11(1.5) .005(2) no disloc. 1. .53(2) .10(1) .007(3) : Relative weights of sectors with charge $|n|$. [00100]{} A. Belavin and A. Polyakov JETP Lett. 22 (1975) 245. G. ’t Hooft, Phys. Rev. Lett. 37 (1976) 8. J. Hoek, M. Teper and J. Waterhouse, Nucl. Phys. B288 (1987) 589. M. Garcia et al, Nucl. Phys.B34(Proc.Suppl.) (1994) 222; P. de Forcrand, M. Garcia and  I. Stamatescu, hep-lat/9509064. E. Mendel and G. Nolte, hep-lat/9511030. M. Lüscher, Com. Math. Phys. 85 (1982) 39. C. Panagiotakopoulos, Nucl. Phys. B251 (1985) 61. P. diVecchia et al, Nucl.Phys.B192(1981)392 [^1]: Dedicated to the memory of my beloved brother,                       Prof. Roberto Mendel. [^2]: New Address:   Abteilung für Neurobiologie, Universitätsklinikum Steglitz, 12200 Berlin.

ArXiv

In the low-energy minimal supersymmetric standard model (MSSM) there exist a large number of $D$-flat directions along which squark, slepton and Higgs fields may get expectation values. We collectively denote them by $\varphi$. In flat space at zero temperature exact supersymmetry guarantees that the effective potential $V(\varphi)$ along these $D$-flat directions vanishes to all orders in perturbation theory (besides the possible presence of nonrenormalizable terms in the superpotential). In the commonly studied supergravity scenario, supersymmetry breaking may take place in isolated hidden sectors [@nilles] and then gets transferred to the other sectors by gravity. The typical curvature of $D$-flat directions resulting from this mechanism is $M^2_\varphi \sim\frac{\left|F\right|^2}{M_{\rm P}^2}$, where $F$ is the vacuum expectation value (VEV) of the $F$-term breaking supersymmetry in the hidden sector. In order to generate soft masses of order of $M_W$ in the matter sector, $F$ is to be of order of $(M_WM_{\rm P})$ and sfermion and Higgs masses along flat directions turn out to be in the TeV range. $D$-flat directions are extremely important in the Affleck-Dine (AD) scenario for baryogenesis [@ad] if large expectation values along flat vacua are present during the early stages of the evolving Universe. When calculating high temperature corrections to the effective potential at large $\varphi$ one can often neglect the renormalizable interactions of the field $\varphi$ with other fields, because the fields directly interacting with $\varphi$ become too heavy to be excited by thermal effects. Therefore thermal effects do not push the field $\varphi$ towards $\varphi = 0$ in the early universe if initially it was large enough. For this reason, in the original version of the AD scenario [@ad; @adlinde] it was assumed that the initial value of the scalar field $\varphi$ was different in different parts of the universe, as in the chaotic inflation scenario. However, recently it was pointed out that initial position of the field $\varphi$ could be fixed by corrections to the effective potential which are proportional to the Hubble parameter $H$. Indeed, in generic supergravity theories moduli masses are of order of the Hubble parameter $H$ [@dine], [@moduli]. One of the reasons is that the effective potential provides a nonzero energy density $V\sim\left|F\right|^2$ which breaks supersymmetry. If the vacuum energy dominates, the Hubble parameter is given by $H^2=(V/3 M_{\rm P}^2)$ and therefore the curvature along the $D$-flat directions becomes $M_\varphi^2=c H_{{\rm I}}^2$, where $c$ may be either positive or negative. (Here we use the reduce Planck mass $M_{\rm P} \sim 2\times 10^{18}$ GeV.) A similar contribution appears after inflation as well. For a mechanism of stronger supersymmetry violation see [@Dvali]. The most relevant part of the entire AD potential in presence of nonrenormalizable superpotentials of the type $\delta W=(\lambda/n M^{n-3})\varphi^n$ somewhat schematically can be represented as follows [@dine]: $$\begin{aligned} \label{AD} V(\varphi) &=& {c_1 }\,m_{3/2}^2\, |\varphi|^2 - {c_2 } \:H^2|\varphi|^2 + \frac{A\lambda H \varphi^n e^{i \,n\theta} + h.c.}{n\:M^{n-3}}\nonumber \\ &+& \lambda^2{|\varphi|^{2n-2}\over M^{2n-6}} \ .\end{aligned}$$ Here $c_i$ are some positive constants which we suppose to be $O(1)$, $m_{3/2} = O(1)$ TeV is the gravitino mass, $M$ is some large mass scale such as the GUT or Planck mass, $n$ is some integer, which may take such values as $4$, $6$, etc., and the $A$-term violates the baryon or the lepton number and has a definite $CP$-violating phase relative to $\varphi$. This leads to symmetry breaking with $ \varphi \sim (H M ^{n-3}/\lambda)^{1/n-2}$. For definiteness we will take here $n = 4$ and $M = M_{\rm P}$. The main idea of ref. [@dine] is that in the early universe the field $\varphi$ will be fixed at $ \varphi \sim \sqrt {H M_{\rm P} /\lambda}$. When $H$ drops down to $m_{3/2}$, the terms proportional to $H$ become smaller than the first term in (\[AD\]), and the field $\varphi$ begins oscillating near $\varphi = 0$. If the field $\varphi$ can be associated with flat directions for squark-slepton fields, then the oscillations of the field $\varphi$ near $\varphi = 0$ in the presence of the A-term of eq. (\[AD\]) may produce baryon asymmetry of the universe. The aim of this Letter is to show that there is a much stronger mechanism of supersymmetry breaking in the early universe. This mechanism may lead to corrections to the quadratic terms in the effective potential of the AD field which are much greater than $ H^2\varphi^2$ and to $A$-terms much greater than the one of eq. (\[AD\]). As a result, under certain conditions the Affleck-Dime mechanism of baryon asymmetry generation may become quite different from the version proposed in [@dine]. The mechanism of supersymmetry breaking which we are going to discuss is related to particle production at the first stage of reheating after inflation. Indeed, Kofman, Linde and Starobinsky have recently pointed out that a broad parametric resonance, which under certain conditions occurs soon after the end of inflation, may lead to a very rapid decay of the inflaton field [@explosive]. The inflaton energy is released in the form of inflaton decay products, whose occupation number is extremely large. They have energies much smaller than the temperature that would have been obtained by an instantaneous conversion of the inflaton energy density into radiation. Since it requires several scattering times for the low-energy decay products to form a thermal distribution, it is rather reasonable to consider the period in which most of the energy density of the Universe was in the form of the nonthermal quanta produced by inflaton decay as a separate cosmological era, dubbed as preheating to distinguish it from the subsequent stages of particle decay and thermalization which can be described by the techniques developed in [@tec]. Several aspects of the theory of explosive reheating have been studied in the case of slow-roll inflation [@noneq] and first-order inflation [@kolb] and it has been recently proposed that non-thermal production and decay of Grand Unified Theory bosons at the intermediate stage of preheating may generate the observed baryon asymmetry [@bario]. One of the most important consequences of the stage of preheating is the possibility of nonthermal phase transitions with symmetry restoration [@KLSSR], [@tkachev]. These phase transitions appear due to extremely strong quantum corrections induced by particles produced at the stage of preheating. What is crucial for our considerations is that parametric resonance is a phenomenon peculiar of particles obeying Bose-Einstein statistics. Parametric resonant decay into fermions is very inefficient because of Pauli’s exclusion principle. This means that during the preheating period the Universe is populated exclusively by a huge number of soft bosons, and the occupation numbers of bosons and fermions belonging to the supermultiplet coupled to the inflaton superfield are completely unbalanced. Supersymmetry is then strongly broken during the preheating era and large loop corrections may arise since the usual cancellation between diagrams involving bosons and fermions within the same supermultiplet is no longer operative. Therefore all results obtained in [@KLSSR], [@tkachev] apply to the modification of the effective potential along the flat directions. As we will see, the curvature $V^{\prime\prime}(\varphi)$ during the preheating era often is much larger than the effective mass $H^2$ that $D$-flat directions acquire in the inflationary stage. Let us consider, [*e.g*]{}., chaotic inflation scenario [@linde] where the inflaton field $\phi$ couples to a complex $\chi$-field, $$\label{w} V=\frac{M^2_\phi}{2}\:\phi^2+ g^2\:\phi^2\:|\chi^2| \ .$$ We take the inflaton mass $M_\phi\sim 10^{13}$ GeV in order for for the density perturbations generated during the inflationary era to be consistent with COBE data. Inflation occurs during the slow rolling of the scalar field $\phi$ from its very large value. Then it oscillates with an initial amplitude $\phi_0\sim M_{\rm P}$. Within a dozen oscillations the initial energy $\rho_\phi\sim M_\phi^2\:\phi_0^2$ is transferred through the interaction $g^2\:\phi^2 \chi^2$ to [*bosonic*]{} $\chi$-quanta in the regime of parametric resonance [@explosive]. At the end of the broad parametric resonance the field $\phi$ drops down to $\phi_e \sim 10^{17}$ GeV. An exact number depends logarithmically on coupling constants; we will use $\phi_e \sim 10^{17}$ GeV $\sim 5\times 10^{-2} M_{\rm P}$ for our estimates. After this stage, the universe is expected to be filled up with $\chi$-bosons with very large occupation numbers $n_k \sim g^{-2}$, with relatively small energy per particle, $E_\chi \sim \sqrt{gM_\phi \phi_e} \sim 0.2 \sqrt{gM_\phi M_{\rm P}}$. It is especially important that the amplitude of field fluctuations produced at that stage is very large [@explosive], [@KLSSR], $$\label{r1} \langle \chi^2 \rangle \sim 5\times10^{-2} g^{-1} M_\phi M_{\rm P} \ .$$ If, for example, the energy of the inflaton field after preheating were instantly thermalized, we would obtain a much smaller value $\langle \chi^2 \rangle \sim 10^{-4} M_\phi M_{\rm P}$. In realistic models thermalization typically takes a lot of time, and the value of $\langle \chi^2 \rangle$ after complete thermalization is many orders of magnitude smaller than (\[r1\]). Note, that large fluctuations (\[r1\]) occur only in the bosonic sector of the theory, thus breaking supersymmetry at the quantum level. Anomalously large fluctuations of $\langle \chi^2 \rangle$ lead to specific nonthermal phase transitions in the early universe [@KLSSR], [@tkachev]. It will be interesting to study possible consequences of this effect for supersymmetric theories with flat directions. Let us make the natural assumption that the field $\varphi$ labelling the $D$-flat direction couples to the generic supermultiplet $\chi$ by a renormalizable interaction of the form ${\cal L}_{{\rm int}}=h^2|\varphi|^2|\chi|^2$ coming from some $F$-term in the potential. Reheating gives an additional contribution to the effective mass along the $\varphi$-direction : $$\label{broad} \Delta M^2_\varphi \sim 2 \:h^2 \langle \chi^2 \rangle \sim 10^{-1}\:\frac{h^2}{g}\: M_\phi\:M_{\rm P}.$$ As a result, the curvature of the effective potential becomes large and positive, and symmetry rapidly restores for $10^{-1}\:\frac{h^2}{g}\: M_\phi\:M_{\rm P} > c_2\:H^2$. At the end of preheating in our model $H \sim 2\times10^{-2} M_\phi$. Thus, for $c_2 = O(1)$ symmetry becomes restored for $h^2 {\ \lower-1.2pt\vbox{\hbox{\rlap{$>$}\lower5pt\vbox{\hbox{$\sim$}}}}\ } 2\times 10^{-8} g$. In other words, supersymmetry breaking due to preheating is much greater than the supersymmetry breaking proportional to $H$unless the coupling constant $h^2$ is anomalously small. This leads to symmetry restoration in the AD model soon after preheating. Let us check the applicability of our results. First of all, parametric resonance takes place for $g\phi > M_\phi$, where $\phi \sim 10^{-2} M_{\rm P}$ at the end of preheating. This implies that $g > 10^{-4}$. Secondly, even though self-interactions of the $\chi$-field do not terminate the resonance effect since particles remain inside the resonance shell, creation of quanta different from $\chi$ may remove the decay products of the inflaton away from the resonance shell. This leads us to consider two different possibilities. The resonance does not stop if scatterings are suppressed by kinematical reasons, [*i.e.*]{} if the non-thermal plasma mass of the final states is larger than the initial energy of the $\chi$’s. If the final states are identified with the $\varphi$-quanta, this happens for $h>g$. Otherwise, scatterings occur, but are slow enough to not terminate the resonance if the interaction rate $\Gamma\sim n_\chi\sigma$, where $\sigma$ denotes the scattering cross section, is smaller than the typical frequency of oscillations at the end of the preheating stage $\sim g\langle \chi^2\rangle^{1/2}$. Identifying again the final states with the $\varphi$’s, this condition translates into $h^2 {\ \lower-1.2pt\vbox{\hbox{\rlap{$<$}\lower5pt\vbox{\hbox{$\sim$}}}}\ } 10^2 g$, which is compatible with the condition for symmetry restoration discussed above. There is another necessary condition in order to have broad resonance and a rapid production of $\chi$-particles: the constant contribution $h|\varphi|$ to effective mass of the field $\chi$ induced by the condensate $|\varphi_0|$ should be smaller than the typical energy of the decay products $E_\chi$. This translates into a bound on the coupling $\lambda$ when taking into account the constraint on $h$ discussed previously. The exact bound depends upon the choice of the index $n$ and the mass $M$. For $n = 4$ and $M = M_{\rm P}$ one has the condition $h^2 < 3 g\lambda$. This condition simultaneously guarantees that the energy density of the AD field $\varphi$ is smaller than the inflaton energy density at the end of the broad resonance. All these constraints are not very restrictive. For example, one may take $h = 2g$ to satisfy the condition $h > g$. Then one has the constraints $g > 10^{-8}$ and $\lambda > g$. Let us now describe the dynamics of the $\varphi$-field during preheating. Under the conditions described above, the AD field $\varphi$ rapidly rolls towards the origin $\varphi = 0$ and makes fast oscillations about it with an initial amplitude given by $\varphi\sim |\varphi_0|$. The frequency of the oscillations is of order of $(\Delta M^2_\varphi)^{1/2}$ and is much higher than the Hubble parameter at preheating $\sim 10^{-1}\:M_\phi$. The field is underdamped and relaxes to the origin after a few oscillations. The crucial observation we would like to make here is that baryon asymmetry can be efficiently produced during this fast relaxation of the field to the origin. Our picture is similar to that of ref. [@dine], but there are considerable differences. According to [@dine], the baryon asymmetry is produced at very late times when the Hubble parameter $H$ becomes of order of the gravitino mass $m_{3/2}\sim$ TeV. At this epoch the field becomes underdamped and the $CP$-violating $A$-terms in the Lagrangian are comparable to the baryon-conserving ones. The field feels a torque from the $A$-terms and spirals from the initial point $\varphi=|\varphi_0|\:{\rm e}^{i\theta}$ inward in the harmonic potential, producing a baryon asymmetry $n_B/n_\varphi={\cal O}(1)$. In our case baryon number production occurs at very early stages, during the preheating era. This is possible because the same nonthermal effects generating a large curvature at the origin give rise also to sizable $CP$-violating terms along the AD flat direction. Indeed, under very general assumptions, one can expect the presence in the Lagrangian of nonrenormalizable $CP$-violating terms of the type $\alpha (\chi^2/M_{\rm P}^{m-2})\:\varphi^m+{\rm h.c.}$, where $m>2$. During the preheating period large amplitudes of the $\chi$-field give rise to the $CP$-violating term $$\label{A} \alpha \frac{\langle\chi^2\rangle}{m M_{\rm P}^{m-2}}\:\varphi^m e^{i m \theta_\chi }\:\:+\:{\rm h.c.},$$ For $m=3$ we get a $CP$-violating term ${A\varphi^3 e^{i m \theta_\chi }\over 3} +{\rm h.c.}$ with $A\sim 10^{-1}\:g^{-1}\alpha\,M_\phi$. Note that fluctuations $\langle\chi^2\rangle$ were absent during inflation; they increase very sharply (exponentially) during preheating, and therefore they begin strongly affecting the position of the scalar field only at the end of the broad parametric resonance, when they reach their maximal value. Thus, to the first approximation one has the following initial conditions for the field $\varphi$ at the end of the broad resonance: $|\varphi_0| \sim \sqrt{HM_{\rm P}/\lambda} \sim \sqrt{M_\phi M_{\rm P}/10\lambda}$, $\dot\varphi_0 = 0$. As for the complex phase $\theta$, its initial value is determined by the $H$-dependent A-term (\[AD\]). In general, this phase is quite different from the phase determined by eq. (5). As a result, the field $\varphi$ will spiral down to $\varphi = 0$ acquiring baryon charge density $n_B = 2|\varphi^2|\dot\theta$. After the end of the stage of broad parametric resonance the value of $\langle\chi^2\rangle$ decreases approximately as $a^{-2} \sim t^{-1}$. As a result, both $\Delta M_\varphi$ and $A$ after preheating decrease as $t^{-1}$, i.e. in the same way as the Hubble constant $H$ in the radiation dominated universe after preheating. Therefore the equation describing the relaxation of the $\varphi$-field toward the origin $$\ddot{\varphi}+ 3\:H\:\dot{\varphi}+\Delta M^2_\varphi\:\varphi+A\:(\varphi^*)^2=0$$ takes the following form: $$\label{motion} \ddot{\varphi}+ {3\over 2t} \dot{\varphi}+{1\over 2 gt} (h^2 M_{\rm P}\varphi + \alpha \varphi^{*2})=0 \ .$$ Before going further, let us compare the magnitudes of the last two terms for $|\varphi_0| \sim \sqrt{HM_{\rm P}/\lambda} \sim \sqrt{M_\phi M_{\rm P}/10\lambda}$. One can easily see that the $B$-violating term is greater than the mass term for $\alpha > 10^3 h^2\sqrt\lambda$. If, e.g., one takes $h \sim 10^{-7}$, $\lambda \sim 10^{-2}$, one finds that baryon number violation may be substantial even if the coefficient $\alpha$ in the A-term (\[A\]) is extremely small, $\alpha \sim 10^{-12}$. It is convenient to introduce new variables $y = \varphi\sqrt {\lambda t /M_{\rm P}}$ and $\tau = h\sqrt{2tM_{\rm P}/g}$. In these variables eq. (\[motion\]) simplifies: $$\label{motion2} y''+ y+{\alpha\sqrt 2\over \tau\, h\, \sqrt {g\lambda}}\, y^{*2}=0 \ .$$ In the new variables the motion of the field $y= |y|e^{i \theta}$ begins at $|y_0| = 1$, at the initial moment $\tau_0 \sim 10^4 h/\sqrt g$. The condition of vanishing initial velocity of the field $\varphi$ translates into the condition $y'_0 = y_0/\tau_0$. To calculate the ratio $n_B/s$, where $n_B= 2|\varphi|^2\dot{\theta}$, and $s$ is the entropy density, we will introduce a fictitious reheating temperature $T_R \sim 10^{-2} \sqrt {M_\phi M_{\rm P}} \sim 5\times 10^{13}$ GeV. This is the temperature which would be reached by our system if thermalization would occur instantaneously after the end of the broad parametric resonance at $\phi_e \sim 10^{-1} M_{\rm P}$. Even though thermalization may occur much later, this concept may be quite useful because at the radiation dominated stage the energy density of the universe decreases in such a way that at the moment when thermalization actually occurs the resulting temperature $T$ will be equal to the redshifted value of the fictitious reheating temperature, $T \sim T_R\sqrt{\tau_0/\tau}$. This leads to the following expression for the baryon number $B = n_B/s$ soon after preheating: $$\label{motion3} B = {n_B\over s} \sim \frac{n_B}{10^2\,T_R^3} \sim 2\times10^{-2}\, |y|^2\, \theta' {h \over \lambda \sqrt {g}} \ .$$ We solved eq. (\[motion2\]) numerically for various values of parameters and calculated the ratio of $n_B$ to the entropy density $s$. We have found that this ratio oscillates and approaches a constant at large times. For ${\alpha \over h \sqrt {g\lambda}} < 1$ and $h > 10^{-4} \sqrt g$ the typical value of the baryon asymmetry is given by $$\label{motion4} B = {n_B\over s} \sim 10^{-2} {\alpha\over g\lambda\sqrt \lambda}f(\tau_0)\ .$$ Here $f(\tau_0)$ is a certain function of $\tau_0 \sim 10^4 h/\sqrt g$: $f(1) \sim 1$, $f(10) \sim 0.1$, $f(100) \sim 0.05$. Validity of this result depends on details of the theory; typically it gives a reliable estimate of the baryon asymmetry only for $B \ll 1$ (which is what we need), but even in this case some care should be taken. For example, one can show that for $h^2 < g\lambda$ the temperature $T$ remains much greater than the time-dependent effective mass of the field $\varphi$ until this mass approaches the constant value $O(m_{3/2})$. In this regime quarks acquire large effective mass $O(T)$, and the field $\varphi$ cannot decay and transfer its baryon asymmetry to fermions until temperature drops to $O(m_{3/2})$. This may lead to some corrections to eq. (\[motion4\]) [@adlinde]. Also, the simple scaling rules used in the derivation of eq. (\[motion4\]) may break if, e.g., at some intermediate stage the universe becomes dominated by nonrelativistic particles. A detailed investigation of baryon asymmetry production in realistic theories including all of the effects mentioned above should become a subject of a separate investigation. The main purpose of our paper was to show that parametric resonance and nonthermal phase transitions found in [@explosive; @KLSSR] may lead to strong supersymmetry breaking in the early universe and to considerable modifications in the theory of baryogenesis in supersymmetric models. We have found that if the inflaton field $\phi$ couples in a renormalizable way to $\chi$-bosons, which are weakly coupled to the AD field $\varphi$, then the effect of parametric resonance may induce very large masses for particles corresponding to flat directions of the AD potential. The same effect may induce large terms violating baryon conservation (\[A\]). As a result, one can obtain large baryon asymmetry even in the models where all relevant coupling constants are extremely small. 0.3 cm We would like to thank D. Lyth, L.A. Kofman and E. Kolb for many useful discussions. GA is supported by the DOE under contract DE-AC02-76CH03000; AR is supported by the DOE and NASA under Grant NAG5–2788; AL is supported in part by the NSF grants PHY-9219345 and AST-9529225. H.P. Nilles, Phys. Rept. [**110**]{}, 1 (1984). I. Affleck and M. Dine, Nucl. Phys. [**B249**]{}, 361 (1985). A.D. Linde, Phys.Lett. [**160B**]{}, 243 (1985). M. Dine, L. Randall and S. Thomas, Phys. Rev. Lett. [**75**]{}, 398 (1995); [*ibid*]{} Nucl. Phys. [**458**]{}, 291 (1996). G. Dvali, IFUP-TH-09-95 (1995), hep-ph/9503259. The fact that during inflation fields generically have masses of order of $H$ has been noted by many authors, see e.g. M. Dine, W. Fischler and D. Nemeschnsky, Phys. Lett. [**B136**]{}, 169 (1984); G.D. Coughlan, R. Holman, P. Ramond and G.G. Ross, Phys. Lett. [**B140**]{}, 44 (1984); A.S. Goncharov, A.D. Linde, and M.I. Vysotsky, Phys. Lett. [**147B**]{}, 279 (1984). G. Dvali, CERN-TH-96-129, hep-ph/9605445. L. Kofman, A.D. Linde and A.A. Starobinsky, Phys. Rev. Lett. [**73**]{}, 3195 (1994). A.D. Dolgov and A.D. Linde, Phys.Lett. [**116B**]{}, 329 (1982); L.F. Abbott, E. Fahri and M. Wise, Phys. Lett. [**B117**]{}, 29 (1982). J. Traschen and R. Brandenberger, Phys. Rev. D [**42**]{}, 2491 (1990); Y. Shtanov, J. Traschen, and R. Brandenberger, Phys.Rev. D [**51**]{}, 5438 (1995); D. Boyanovsky, H.J. de Vega, R. Holman, D.S. Lee, and A. Singh, Phys. Rev. D [**51**]{}, 4419 (1995); H. Fujisaki, K. Kumekawa, M. Yamaguchi, and M. Yoshimura, TU/95/493, hep-ph/9511381; S.Yu. Khlebnikov and I. Tkachev, OSU-TA-08/96, hep-ph/9603378. E.W. Kolb and A. Riotto, FERMILAB-Pub-96-036-A, astro-ph 9602095. E.W. Kolb, A.D. Linde and A. Riotto, FERMILAB-Pub-133-A, hep-ph/9606260, submitted to Phys. Rev. Lett. L. Kofman, A.D. Linde and A.A. Starobinsky, Phys. Lett. Lett. [**76**]{}, 1011 (1996). I. Tkachev, Phys. Lett. [**B376**]{}, 35 (1996). A.D. Linde, Phys. Lett. [**B129**]{}, 177 (1983).

ArXiv

--- abstract: 'We provide a characterization of multiqubit entanglement monogamy and polygamy constraints in terms of negativity. Using the square of convex-roof extended negativity (SCREN) and the Hamming weight of the binary vector related to the distribution of subsystems proposed in Kim (Phys Rev A 97: 012334, 2018), we provide a new class of monogamy inequalities of multiqubit entanglement based on the $\alpha$th power of SCREN for $\alpha\geq1$ and polygamy inequalities for $0\leq\alpha\leq1$ in terms of squared convex-roof extended negativity of assistance (SCRENoA). For the case $\alpha<0$, we give the corresponding polygamy and monogamy relations for SCREN and SCRENoA, respectively. We also show that these new inequalities give rise to tighter constraints than the existing ones.' author: - | Long-Mei Yang$^1$, Bin Chen$^2$, Shao-Ming Fei $^{3,4}$, Zhi-Xi Wang$^3$[^1]\ [$^1$State Key Laboratory of Low-Dimensional Quantum Physics and Department of Physics, Tsinghua University, Beijing 100084, China]{}\ [$^2$School of Mathematical Sciences, Tianjin Normal University, Tianjin 300387, China]{}\ [$^3$School of Mathematical Sciences, Capital Normal University, Beijing 100048, China]{}\ [$^4$Max-Planck-Institute for Mathematics in the Sciences, 04103 Leipzig, Germany]{} title: Tighter constraints of multiqubit entanglement for negativity --- Introduction ============ Quantum entanglement [@F.M; @K.Chen; @HPB1; @HPB2; @Vicente; @CJZ] is one of the most intrinsic features of quantum mechanics, which distinguishes the quantum from the classical world. A distinct property of quantum entanglement is that a quantum system entangled with one of the other systems limits its shareability with the remaining ones, known as the monogamy of entanglement (MoE) [@BMT; @JSK]. Being a useful resource, MoE plays a significant role in many quantum information and communication processing tasks such as the security proof in quantum cryptographic scheme [@CHB]. For a given tripartite quantum state $\rho_{ABC}$, MoE can be characterized in a quantitative way known as monogamy inequality, $$E(\rho_{ABC})\geq E(\rho_{AB})+E(\rho_{AC}),$$ where $\rho_{AB}={\rm tr}_C(\rho_{ABC})$ and $\rho_{AC}={\rm tr}_B(\rho_{ABC})$ are the reduced density matrices. In Ref. [@VC], Coffman-Kundu-Wootters (CKW) established the first monogamy inequality based on the bipartite entanglement measure defined by tangle. Later, Osborne [*et al.*]{} generalize the three-qubit CKW inequality to arbitrary multiqubit systems [@TJO]. Monogamy inequalities in higher-dimensional quantum systems also have been deeply investigated by the use of various bipartite entanglement measures [@JSK1; @JSK2; @JSK3; @JSK4]. The assisted entanglement is a dual amount to bipartite entanglement measures, which accordingly has a dually monogamous property in multipartite quantum systems. This dually monogamous property gives rise to a dual monogamy inequality known as polygamy inequality [@G.G1; @G.G2]. For a tripartite state $\rho_{ABC}$, one has $$\label{tauABC} \tau^a(\rho_{A|BC})\leq\tau^a (\rho_{AB})+\tau^a(\rho_{AC}),$$ where $\tau^a(\rho_{A|BC})$ is the tangle of assistance. In Ref. [@JSK3; @F.B], the authors generalized the inequality to the cases of multiqubit quantum systems and some class of higher-dimensional quantum systems. By using the entanglement of assistance, a general polygamy inequality of multipartite entanglement in arbitrary-dimensional quantum systems has been also established [@JSK5; @JSK6]. Recently, based on the $\alpha$th power of entanglement measures, many generalized classes of monogamy inequalities were proposed [@Oliveira; @Luo; @SM.Fei1; @SM.Fei2; @SM.Fei3]. In Ref. [@JSK7], Kim investigated multiqubit entanglement constraints related to the negativity. By using the $\alpha$th power of squared convex-roof extended negativity (SCREN) and the squared convex-roof extended negativity of assistance (SCRENoA) for $\alpha\geq1$ and $0\leq\alpha\leq 1$, respectively, both monogamy and polygamy inequalities were established. These inequalities involve the notion of Hamming weight of the binary vector related to the distribution of subsystems and are shown to be tighter than the previous ones. In this paper, we show that both the monogamy inequalities with $\alpha\geq1$ and polygamy inequalities with $0\leq\alpha\leq 1$ given in Ref. [@JSK7] can be further improved to be tighter. Even for the case of $\alpha<0$, we can also provide tight constraints in terms of SCREN and SCRENoA. Thus, a complete characterization for the full range of the power $\alpha$ is given. These tighter constraints of multiqubit entanglement give rise to finer characterizations of the entanglement distributions among the multiqubit systems. Preliminaries ============= We first consider the monogamy inequalities and polygamy inequalities related to the negativity. The tangle of a bipartite pure states $|\psi\rangle_{AB}$ is defined as [@VC] $$\tau(|\psi\rangle_{A|B})=2(1-{\rm tr}\rho_A^2)$$ where $\rho_A={\rm tr}_B|\psi\rangle_{AB}\langle\psi|$. The tangle of a bipartite mixed state $\rho_{AB}$ is defined as $$\label{tauAB} \tau(\rho_{A|B})=\Bigg[\min\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\sqrt{\tau(|\psi_k\rangle_{A|B})}\Bigg]^2,$$ and the tangle of assistance (ToA) of $\rho_{AB}$ is defined as $$\label{taua} \tau^a(\rho_{A|B})=\Bigg[\max\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\sqrt{\tau(|\psi_k\rangle_{A|B})}\Bigg]^2,$$ where the minimization in and the maximum in are taken over all possible pure state decompositions of $\rho_{AB}=\sum\nolimits_{k}p_k|\psi_k\rangle_{AB}\langle\psi_k|$. For any bipartite quantum state $\rho_{AB}$, the negativity is defined as [@JSK7; @G.V], $\mathcal{N}(\rho_{A|B})=\|\rho_{AB}^{T_B}\|_1-1$, where $\rho_{AB}^{T_B}$ is the partial transposition of $\rho_{AB}$, and $\|\cdot\|_1$ is the trace norm. Then the notion of tangle and ToA for two-qubit state $\rho_{AB}$ in and can be rewritten as [@JSK7] $$\tau(\rho_{A|B})=\Bigg[\min\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\mathcal{N}(|\psi_k\rangle_{A|B})\Bigg]^2$$ and $$\tau^a(\rho_{A|B})=\Bigg[\max\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\mathcal{N}(|\psi_k\rangle_{A|B})\Bigg]^2,$$ respectively, due to the fact that $\mathcal{N}^2(|\psi\rangle_{A|B})=4\lambda_1\lambda_2=\tau(|\psi\rangle_{A|B})$ for any bipartite pure state $|\psi\rangle_{AB}$ with Schmidt rank 2, $|\psi\rangle_{AB}=\sqrt{\lambda_1}|e_0\rangle_A\otimes|f_0\rangle_B+\sqrt{\lambda_2}|e_1\rangle_A\otimes|f_1\rangle_B$. For higher-dimensional quantum systems, a rather natural generalization of two-qubit tangle is proposed, known as SCREN, $$\label{sc1} \mathcal{N}_{sc}(\rho_{A|B})=\Bigg[\min\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\mathcal{N}(|\psi_k\rangle_{A|B})\Bigg]^2.$$ The dual quantity to SCREN can also be defined as $$\label{dualsc1} \mathcal{N}_{sc}^a(\rho_{A|B})=\Bigg[\max\limits_{\{p_k,|\psi_k\rangle\}}\sum\limits_{k}p_k\mathcal{N}(|\psi_k\rangle_{A|B})\Bigg]^2,$$ which is called the SCREN of assistance (SCRENoA). Then, the tangle-based multiqubit monogamy and polygamy inequalities become as $$\label{ine1} \mathcal{N}_{sc}(|\psi\rangle_{A_1|A_2\cdots A_n})\geq\sum\limits_{j=2}^n\mathcal{N}_{sc}(\rho_{A_1|A_j}),$$ and $$\label{ine2} \mathcal{N}_{sc}^a(|\psi\rangle_{A_1|A_2\ldots A_n})\leq\sum\limits_{j=2}^n\mathcal{N}_{sc}^a(\rho_{A_1|A_j}),$$ where $\rho_{A_1|A_j}$ is two-qubit reduced density matrices $\rho_{A_1A_j}$ of subsystems $A_1A_j$ for $j=2,3,\ldots,n$ [@JSK7]. Recently, Kim provided a class of monogamy and polygamy inequalities of multiqubit entanglement by the use of powered SCREN and the Hamming weight of the binary vector related to the distribution of subsystems [@JSK7]. For any nonnegative integer $j$ and its binary expansion $j=\sum\nolimits_{i=0}^{n-1}j_i2^i$, where $\log_2^j<n$ and $j_i\in\{0,1\}$ for $i=0,1,\ldots,n-1$, one can define a binary vector $\vec{j}$ as $\vec{j}=\{j_0,j_1,\ldots,j_{n-1}\}$. The number of $1$’s in its coordinates is denoted as $\omega_H(\vec{j})$, called the Hamming weight of $\vec{j}$ [@MAN]. Based on these notions, Kim proposed tight constraints of multiqubit entanglement as follows [@JSK7]: $$\label{Sc1} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha\geq \sum\limits_{j=0}^{N-1}\alpha^{\omega_{H}(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha,$$ for $\alpha\geq1$, and $$\label{ine3} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha\leq \sum\limits_{j=0}^{N-1}\alpha^{\omega_{H}(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ for $0\leq\alpha\leq1$. Inequalities and are then further written as: $$\label{Sc2} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha\geq \sum\limits_{j=0}^{N-1}\alpha^{j}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha,$$ for $\alpha\geq1$, and $$\label{ine4} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha\leq \sum\limits_{j=0}^{N-1}\alpha^{j}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ for $0\leq\alpha\leq1$. However, these inequalities can be further improved to be much tighter under certain conditions, thus providing tighter constraints of multiqubit entanglement. Tighter constraints for SCREN ============================= In this section, we first provide a tighter monogamy inequality related to the $\alpha$th power of SCREN for $\alpha\geq1$. For $\alpha<0$, a polygamy inequality is also proposed. We need the following lemma. [@Yang] Suppose $k$ is a real number satisfying $0< k\leq1$, then for any $0\leq x\leq k$, we have $$\label{Negativity1} (1+x)^\alpha\geq1+\frac{(1+k)^\alpha-1}{k^\alpha}x^\alpha,$$ for $\alpha\geq1$. We have the following theorem. \[Nsc1\] For $\alpha\geq 1$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$, if the N-qubit subsystems $B_0, \ldots, B_{N-1}$ satisfy the following condition $$\label{order1} k\mathcal{N}_{sc}(\rho_{A|B_j})\geq\mathcal{N}_{sc}(\rho_{A|B_{j+1}})\geq 0,$$ where $j=0,1,\ldots,N-2$ and $0< k\leq1$, then we have $$\label{nsc1} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha \geq\sum\limits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha.$$ Similar to the proof in [@JSK7], from Eq. , we only need to prove $$\label{nc1} \left[\sum\limits_{j=0}^{N-1}\mathcal{N}_{sc}(\rho_{A|B_j})\right]^\alpha \geq\sum\limits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_{H}(\vec{j})} [\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha.$$ We first show that the inequality holds for the case of $N=2^n$. For $n=1$ and a three-qubit pure state $|\psi\rangle_{AB_0B_1}$, from and , one has $$\begin{array}{rl} &[\mathcal{N}_{sc}(\rho_{A|B_0})+\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha =[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha\Big(1+\frac{\mathcal{N}_{sc}(\rho_{A|B_1})}{\mathcal{N}_{sc}(\rho_{A|B_0})}\Big)^\alpha\\[4.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq [\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha \Bigg[1+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}\Bigg(\frac{\mathcal{N}_{sc}(\rho_{A|B_1})}{\mathcal{N}_{sc}(\rho_{A|B_0})}\Bigg)^\alpha\Bigg]\\[4.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ = [\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha, \end{array}$$ Thus, holds for $n=1$. Assume that inequality holds for $N=2^{n-1}$ with $n\geq 1$. We consider the case of $N=2^n$. For arbitrary $(N + 1)$-qubit pure state $|\psi\rangle_{AB_0B_1\ldots B_{N-1}}$ and its two-qubit reduced density matrices $\rho_{AB_j}$, $j=0,1,\ldots,N-1$, one has $\mathcal{N}_{sc}(\rho_{A|B_{j+2^{n-1}}})\leq k^{2^{n-1}}\mathcal{N}_{sc}(\rho_{A|B_j})$ from . Then, we find $$0\leq\frac{\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})}{\sum\nolimits_{j=0}^{2^{n-1}-1} \mathcal{N}_{sc}(\rho_{A|B_j})}\leq k^{2^{n-1}}\leq k,$$ which implies that $$\Bigg(1+\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})}{\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc} (\rho_{A|B_j})}\Bigg)^\alpha\geq 1+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha} \Bigg(\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})} {\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})}\Bigg)^\alpha.$$ Thus, $$\begin{array}{rl} &\Bigg(\sum\nolimits_{j=0}^{N-1}\mathcal{N}(\rho_{A|B_j})\Bigg)^\alpha =\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})+\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha \Bigg(1+\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})}{\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc} (\rho_{A|B_j})}\Bigg)^\alpha\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha \Bigg[1+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha} \Bigg(\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})} {\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})}\Bigg)^\alpha\Bigg]\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha +\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}\Bigg(\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha. \end{array}$$ Since we have assumed that $$\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha\geq \sum\nolimits_{j=0}^{2^{n-1}-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})-1}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha,$$ by relabeling the subsystems, we can always have $$\Bigg(\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha\geq \sum\nolimits_{j=2^{n-1}}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})-1}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha.$$ Then we have $$\Bigg(\sum\nolimits_{j=0}^{2^n-1}\mathcal{N}_{sc}(\rho_{A|B_j})\Bigg)^\alpha\geq\\[5mm] \sum\nolimits_{j=0}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha.$$ As there always exists an positive integer $n$ such that $0\leq N\leq 2^n$ for some positive integer $N$, we consider a $(2^n+1)$-qubit pure state, $$\label{gamma1} |\Gamma\rangle_{AB_0B_1\ldots B_{2^n-1}}=|\psi\rangle_{AB_0B_1\ldots B_{N-1}}\oplus |\phi\rangle_{B_N\ldots B_{2^n-1}},$$ which is a product of $|\psi\rangle_{AB_0B_1\ldots B_{N-1}}$ and an arbitrary $(2^n-N)$-qubit pure state $|\phi\rangle_{B_N\ldots B_{2^n-1}}$ [@JSK7]. Then we have $$[\mathcal{N}_{sc}(|\Gamma\rangle_{AB_0B_1\ldots B_{2^n-1}})]^\alpha \geq\sum\nolimits_{j=0}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\sigma_{A|B_j})]^\alpha$$ with $\sigma_{A|B_j}$ being the two-qubit reduced density matrix of $|\Gamma\rangle_{AB_0B_1\ldots B_{2^n-1}}$ for each $j=0,1,\ldots,2^n-1$. Thus, $$\begin{array}{rl} &[\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha =[\mathcal{N}_{sc}(|\Gamma\rangle_{A|B_0B_1\ldots B_{2^n-1}})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq\sum\nolimits_{j=0}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\sigma_{A|B_j})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha, \end{array}$$ since $|\Gamma\rangle_{A|B_0B_1\ldots B_{2^n-1}}$ is separable with respect to the bipartition between $AB_0\ldots B_{N-1}$ and $B_N\ldots B_{2^n-1}$. As $\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}\geq\alpha^{\omega_H(\vec{j})}$ when $\alpha\geq1$, we find that for any multiqubit pure state $|\psi\rangle_{A|B_0B_1\ldots B_{N-1}}$, $[\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha \geq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha \geq\sum\nolimits_{j=0}^{N-1}\alpha^{\omega_H(\vec{j})}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha$ with $\alpha\geq1$. Thus, inequality of Theorem \[Nsc1\] is tighter than inequality for any multiqubit pure state. Here, we give an example to show that our new monogamy inequality is indeed tighter than the previous one given in [@JSK7]. $\mathbf{Example} \ \ $ Let us consider a tripartite quantum state $$\label{psi} |\psi\rangle_{ABC}=\frac{1}{\sqrt{6}}(|012\rangle-|021\rangle+|120\rangle-|102\rangle+|201\rangle-|210\rangle).$$ Then we have $\mathcal{N}_{sc}(|\psi\rangle_{A|BC})=4$ and $\mathcal{N}_{sc}(|\psi\rangle_{A|B})=\mathcal{N}_{sc}(|\psi\rangle_{A|C})=1$ [@JSK7]. Note that in this case $k=1$, and $[\mathcal{N}_{sc}(|\psi\rangle_{A|B})]^\alpha+\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}(|\psi\rangle_{A|c})]^\alpha =1+\frac{(1+k)^\alpha-1}{k^\alpha} =2^\alpha\geq [\mathcal{N}_{sc}(|\psi\rangle_{A|B}]^\alpha+\alpha[\mathcal{N}_{sc}(|\psi\rangle_{A|c}]^\alpha=1+\alpha$ for $\alpha\geq1$. Furthermore, by using Lemma 1, we can also improve inequality to be a tighter one under certain condition. \[Nsc2\] Suppose $k$ is a real number satisfying $0< k\leq1$. For $\alpha\geq 1$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$, $$\label{nsc2} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha \geq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^j[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha,$$ if $k\mathcal{N}_{sc}(\rho_{A|B_j})\geq\sum\nolimits_{j=i+1}^{N-1}\mathcal{N}_{sc}(\rho_{A|B_j})$ for $i=0,1,\ldots, N-2$. The proof is similar to the one given in [@JSK7]. In the next, we discuss the polygamy of entanglement related to the $\alpha$th power of SCREN for $\alpha<0$. We have the following theorem. \[Nsc2\] For any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$ with $\mathcal{N}_{sc}(\rho_{AB_i})\neq0$, $i=0,1,\ldots,N-1$, we have $$\label{SCREN} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha \leq\frac{1}{N}\sum\nolimits_{j=0}^{N-1}[\mathcal{N}_{sc}(\rho_{A|B_j})]^\alpha,$$ for all $\alpha<0$. We follow the proof given in [@SM.Fei2]. For arbitrary tripartite state, we have $$\label{SCREN1} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1})]^\alpha \leq[\mathcal{N}_{sc}(\rho_{A|B_0})+\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha =\mathcal{N}_{sc}(\rho_{A|B_0})^\alpha\Big(1+\frac{\mathcal{N}_{sc}(\rho_{A|B_1})}{\mathcal{N}_{sc}(\rho_{A|B_0})}\Big)^\alpha <[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha,$$ where the first inequality is due to $\alpha<0$ and the second inequality is due to $\Big(1+\frac{\mathcal{N}_{sc}(\rho_{A|B_1})}{\mathcal{N}_{sc}(\rho_{A|B_0})}\Big)^\alpha<1$. Similarly, we get $$\label{SCREN2} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1})]^\alpha<[\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha.$$ From and , we obtain $$\label{SCREN3} [\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1})]^\alpha <\frac{1}{2}\{[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha+[\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha\}.$$ One can get $$\label{SCREN4} \begin{array}{rl} &[\mathcal{N}_{sc}(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})]^\alpha <\frac{1}{2}\{[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha+[\mathcal{N}_{sc}(\rho_{A|B_1\ldots B_{N-1}})]^\alpha\}\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ <\frac{1}{2}[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha+(\frac{1}{2})^2[\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha +(\frac{1}{2})^2[\mathcal{N}_{sc}(\rho_{A|B_2\ldots B_{N-1}})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ <\ldots \\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ <\frac{1}{2}[\mathcal{N}_{sc}(\rho_{A|B_0})]^\alpha+(\frac{1}{2})^2[\mathcal{N}_{sc}(\rho_{A|B_1})]^\alpha+\ldots\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +(\frac{1}{2})^{N-2}[\mathcal{N}_{sc}(\rho_{A|B_{N-2}})]^\alpha +(\frac{1}{2})^{N-2}[\mathcal{N}_{sc}(\rho_{A|B_{N-1}})]^\alpha. \end{array}$$ By cyclically permuting the subindices $B_0$, $B_1,$ $\ldots$, $B_{N-1}$ in , we can get a set of inequalities. Summing up these inequalities, we have . Tighter constraints for SCRENoA =============================== In this section, we provide a class of tighter polygamy inequalities of multiqubit entanglement in terms of the $\alpha$-powered SCRENoA and the Hamming weight of the binary vector related to the distribution of subsystems for $0\leq\alpha\leq1$. For the case of $\alpha<0$, we also propose a monogamy relation for SCRENoA. We need the following Lemma. [@Yang] Suppose $k$ is a real number satisfying $0< k\leq1$, then for any $0\leq x\leq k$, we have $$\label{Negativity2} (1+x)^\alpha\leq1+\frac{(1+k)^\alpha-1}{k^\alpha}x^\alpha,$$ for $0\leq \alpha\leq 1$. We have the following theorem. \[Nsc3\] Suppose $k$ is a real number satisfying $0< k\leq1$. For $0\leq\alpha\leq1$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$ satisfying $$\label{inequality3} k\mathcal{N}_{sc}^a(\rho_{A|B_j})\geq\mathcal{N}_{sc}^a(\rho_{A|B_{j+1}})\geq0$$ with $j=0,1,\ldots,N-2$, we have $$\label{nc6} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha \leq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha.$$ From inequality , we only need to show that $$\label{nc3} \Bigg(\sum\nolimits_{j=0}^{N-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha \leq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha.$$ First, we prove inequality for $N=2^n$. For $n=1$ and a three-qubit pure state $|\psi\rangle_{AB_0B_1}$ with two-qubit reduced density $\rho_{AB_0}$ and $\rho_{AB_1}$, one has $$\begin{array}{rl} &[\mathcal{N}_{sc}^a(\rho_{A|B_0})+\mathcal{N}_{sc}^a(\rho_{A|B_1})]^\alpha =[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha\Big(1+\frac{\mathcal{N}_{sc}^a(\rho_{A|B_1})}{\mathcal{N}_{sc}^a(\rho_{A|B_0})}\Big)^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \leq[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha \Bigg[1+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}\Bigg(\frac{\mathcal{N}_{sc}^a(\rho_{A|B_1})}{\mathcal{N}_{sc}^a(\rho_{A|B_0})}\Bigg)^\alpha\Bigg]\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}^a(\rho_{A|B_1})]^\alpha, \end{array}$$ where the inequality is due to . Assume is true for $N=2^{n-1}$ with $n\geq1$. We consider the case of $N=2^n$. From , we find $\mathcal{N}_{sc}^a(\rho_{A|B_{j+2^{n-1}}})\leq k^{2^{n-1}}\mathcal{N}_{sc}^a(\rho_{A|B_j})$ for $j=0,1,\ldots,2^{n-1}-1$. Then $$0\leq\frac{\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})}{\sum\nolimits_{j=0}^{2^{n-1}-1} \mathcal{N}_{sc}^a(\rho_{A|B_j})}\leq k^{2^{n-1}}\leq k.$$ Thus, $$\begin{array}{rl} &\Bigg(\sum\nolimits_{j=0}^{N-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha =\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha \Bigg(1+\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})}{\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})}\Bigg)^\alpha\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \leq\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha \Bigg[1+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}\Bigg(\frac{\sum_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})} {\sum_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})}\Bigg)^\alpha\Bigg]\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha +\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha. \end{array}$$ Since we have assumed that $$\Bigg(\sum\nolimits_{j=0}^{2^{n-1}-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha\leq \sum\nolimits_{j=0}^{2^{n-1}-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})-1}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ we obtain $$\Bigg(\sum\nolimits_{j=2^{n-1}}^{2^n-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha\leq \sum\nolimits_{j=2^{n-1}}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})-1}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ Thus, $$\begin{array}{rl} &\Bigg(\sum\nolimits_{j=0}^{N-1}\mathcal{N}_{sc}^a(\rho_{A|B_j})\Bigg)^\alpha \leq\sum\nolimits_{j=0}^{2^{n-1}-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha +\frac{(1+k)^\alpha-1}{k^\alpha} \sum\nolimits_{j=2^{n-1}}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})-1}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha\\[4mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\sum\nolimits_{j=0}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha. \end{array}$$ For an arbitrary nonnegative integer $N$ and an $(N+1)$-qubit pure state $|\psi\rangle_{AB_0B_1\ldots B_{N-1}}$, let us consider the $(2^n+1)$-qubit $|\Gamma\rangle_{AB_0B_1\ldots B_{N-1}}$ defined in . We have $$\begin{array}{rl} &\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0B_1\ldots B_{N-1}})=\mathcal{N}_{sc}^a(|\Gamma\rangle_{A|B_0B_1\ldots B_{2^n-1}})\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \leq\sum\nolimits_{j=0}^{2^n-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\sigma_{A|B_j})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =\sum\nolimits_{j=0}^{N}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha. \end{array}$$ It can be seen that is tighter than since $\frac{(1+k)^\alpha-1}{k^\alpha}\leq\alpha$ for $0\leq\alpha\leq1$. Moreover, the polygamy inequality of Theorem \[Nsc3\] can be further improved under some conditions. \[Nsc4\] Suppose $k$ is a real number satisfying $0< k\leq1$. For $0\leq\alpha\leq 1$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$, we have $$\label{nc5} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha \leq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^j[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ if $$k\mathcal{N}_{sc}^a(\rho_{A|B_i})\geq\sum\nolimits_{j=i+1}^{N-1}\mathcal{N}_{sc}^a(\rho_{A|B_j}),$$ for $i=0,1,\ldots,N-2$. The proof is similar to the one given in [@JSK7]. It should be noted that Theorems \[Nsc2\] and \[Nsc4\] provide the upper bound and the lower bound for $\mathcal{N}_{sc}(|\psi\rangle_{A|B_o\ldots B_{N-1}})$, since $\mathcal{N}_{sc}(|\psi\rangle_{A|B_o\ldots B_{N-1}})=\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_o\ldots B_{N-1}})$. The following lemma is useful for deriving monogamy relation in terms of $\alpha$-powered SCRENoA when $\alpha<0$. Suppose $k$ is a real number satisfying $0<k\leq 1$. For $0\leq x\leq k$ and $\alpha<0$, we have $$\label{Negativity3} (1+x)^\alpha\geq1+\frac{(1+k)^\alpha-1}{k^\alpha} x^\alpha.$$ Let us consider the function $f(t,\alpha)=(1+t)^\alpha-t^\alpha$ with $t\geq\frac{1}{k}$ and $\alpha<0$. Then $f_t(t,\alpha)=\alpha[(1+t)^{\alpha-1}-\alpha^{\alpha-1}]>0$, i.e., $f(t,\alpha)$ is an increasing function with respect to $t$. Thus, $$\label{Negativity4} f(t,\alpha)\geq f\Big(\frac{1}{k},\alpha\Big)=\Big(1+\frac{1}{k}\Big)^\alpha-\frac{1}{k}=\frac{(1+k)^\alpha-1}{k^\alpha}.$$ Set $x=\frac{1}{t}$ in , we get . \[SCRENoA3\] Suppose $k$ is a real number satisfying $0<k\leq 1$. For $\alpha<0$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$, we have $$\label{SCRENoA2} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha\geq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^j[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ if $$k\mathcal{N}_{sc}^a(\rho_{A|B_i})\geq\mathcal{N}_{sc}^a(\rho_{A|B_{i+1}\ldots B_{N-1}})$$ for $i=0,1,\ldots,N-2$. From , for arbitrary tripartite pure state $|\psi\rangle_{A|B_0B_1}$, we get $$\label{SCRENoA1} \begin{array}{rl} &[\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0B_1})]^\alpha\geq[\mathcal{N}_{sc}^a(\rho_{A|B_0})+\mathcal{N}_{sc}^a(\rho_{A|B_1})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha\Big(1+\frac{\mathcal{N}_{sc}^a(\rho_{A|B_1})}{\mathcal{N}_{sc}^a(\rho_{A|B_0})}\Big)^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha+\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}^a(\rho_{A|B_1})]^\alpha. \end{array}$$ For arbitrary pure state $|\psi\rangle_{A|B_0\ldots B_{N-1}}$, we obtain $$\begin{array}{rl} &[\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha \geq[\mathcal{N}_{sc}^a(\rho_{A|B_0})+\mathcal{N}_{sc}^a(\rho_{A|B_1\ldots B_{N-1}})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ =[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha\Big(1+\frac{\mathcal{N}_{sc}^a(\rho_{A|B_1\ldots B_{N-1}})}{\mathcal{N}_{sc}^a(\rho_{A|B_0})}\Big)^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}^a(\rho_{A|B_1\ldots B_{N-1}})]^\alpha\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq\ldots\\[1.5mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \geq[\mathcal{N}_{sc}^a(\rho_{A|B_0})]^\alpha+\displaystyle\frac{(1+k)^\alpha-1}{k^\alpha}[\mathcal{N}_{sc}^a(\rho_{A|B_1})]^\alpha+\ldots\\[2.0mm] &\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{N-1}[\mathcal{N}_{sc}^a(\rho_{A|B_{N-1}})]^\alpha, \end{array}$$ where the first inequality is due to $\alpha<0$, the second inequality is due to , and the rest inequalities are due to . Just like polygamy inequalities in Theorem 4 and Theorem 5, the following theorems give rise to the tighter monogamy relations in terms of $\alpha$-powered SCRENoA for $\alpha<0$, with the notion of weighted constraint also involved. \[poly1\] Suppose $k$ is a real number satisfying $0< k\leq1$. For $\alpha<0$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$ satisfying $$\label{inequality4} k\mathcal{N}_{sc}^a(\rho_{A|B_j})\geq\mathcal{N}_{sc}^a(\rho_{A|B_{j+1}})\geq0$$ with $j=0,1,\ldots,N-2$, we have $$\label{nc7} [\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha \geq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^{\omega_H(\vec{j})}[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha.$$ \[SCRENoA5\] Suppose $k$ is a real number satisfying $0< k\leq1$. For $\alpha<0$ and any multiqubit pure state $|\psi\rangle_{AB_0\ldots B_{N-1}}$, we have $$[\mathcal{N}_{sc}^a(|\psi\rangle_{A|B_0\ldots B_{N-1}})]^\alpha \geq\sum\nolimits_{j=0}^{N-1}\Big(\frac{(1+k)^\alpha-1}{k^\alpha}\Big)^j[\mathcal{N}_{sc}^a(\rho_{A|B_j})]^\alpha,$$ if $$k\mathcal{N}_{sc}^a(\rho_{A|B_i})\geq\sum\nolimits_{j=i+1}^{N-1}\mathcal{N}_{sc}^a(\rho_{A|B_j}),$$ for $i=0,1,\ldots,N-2$. Conclusion ========== Entanglement monogamy is a fundamental property of multipartite entangled systems. We have proposed tighter weighted monogamy inequalities related to the $\alpha$th power of SCREN for $\alpha\geq1$. We also have investigated the polygamy relations in terms of $\alpha$-powered SCRENoA for the case of $0\leq\alpha\leq 1$. Moreover, by using the $\alpha$th power of SCREN and SCRENoA for $\alpha<0$ respectively, the corresponding weighted polygamy and monogamy inequalities have also been established. These new tighter monogamy and polygamy relations give rise to finer characterizations of the entanglement distributions and capture better the intrinsic feature of multiparty quantum entanglement. Acknowledgements ================ This work is supported by the National Natural Science Foundation of China under Nos. 11805143 and 11675113, and NSF of Beijing under No. KZ201810028042. [00]{} F. Mintert, Marek Kuś and A. Buchleitner, Concurrence of mixed bipartite quantum states in arbitrary dimensions, Phys. Rev. Lett. $\mathbf{92}$, 167902 (2004). K.Chen, S. Albeverio, and S.M. Fei, Concurrence of arbitrary dimensional bipartite quantum states, Phys. Rev. Lett. **95**, 040504 (2005). H. P. Breuer, Separability criteria and bounds for entanglement measures, J. Phys. A Math. Gen. **39**, 11847 (2006). H. P. Breuer, Optimal entanglement criterion for mixed quantum states, Phys. Rev. Lett. **97**, 080501 (2006). J. I. de Vicente, Lower bounds on concurrence and separability conditions, Phys. Rev. A **75**, 052320 (2007). C. J. Zhang, Y. S. Zhang, S. Zhang, and G. C. Guo, Optimal entanglement witnesses based on local orthogonal observables, Phys. Rev. A **76**, 012334 (2007). B. M. Terhal, Is entanglement monogamous, IBM J. Res. Dev. **48**, 71 (2004). J. S. Kim, G. Gour, and B. C. Sanders, Limitations to sharing entanglement, Contemp. Phys. **53**, 417 (2012). C. H. Bennett, Quantum cryptography using any two nonorthogonal states, Phys. Rev. Lett. **68**, 3121 (1992). V. Coffman, J. Kundu, and W. K. Wootters, Distributed entanglement, Phys. Rev. A **61**, 052306 (2000). T. J. Osborne and F. Verstraete, General monogamy inequality for bipartite qubit entanglement, Phys. Rev. Lett. **96**, 220503 (2006). J. S. Kim, A. Das, and B. C. Sanders, Entanglement monogamy of multipartite higher-dimensional quantum systems using convex-roof extended negativity, Phys. Rev. A **79**, 012329 (2009). J. S. Kimand B. C. Sanders, Monogamy of multi-qubit entanglement using Rényi entropy, J. Phys.A Math. Theor. **43**, 445305 (2010). J. S. Kim, Tsallis entropy and entanglement constraints in multiqubit systems, Phys. Rev. A **81**, 062328 (2010). J. S. Kim and B. C. Sanders, Unified entropy, entanglement measures and monogamy of multi-party entanglement, J. Phys.A Math. Theor. **44**, 295303 (2011). G. Gour, D. A. Meyer, and B. C. Sanders, Deterministic entanglement of assistance and monogamy constraints, Phys. Rev. A **72**, 042329 (2005). G. Gour, S. Bandyopadhay, and B. C. Sanders, Dual monogamy inequality for entanglement, J. Math. Phys. **48**, 012108 (2007). F. Buscemi, G. Gour, and J. S. Kim, Polygamy of distributed entanglement, Phys. Rev. A **80**, 012324 (2009). J. S. Kim, General polygamy inequality of multiparty quantum entanglement, Phys. Rev. A **85**, 062302 (2012). J. S. Kim, Tsallis entropy and general polygamy of multiparty quantum entanglement in arbitrary dimensions, Phys. Rev. A **94**, 062338 (2016). T. R. de Oliveira, M. F. Cornelio, and F. F. Fanchini, Monogamy of entanglement of formation, Phys. Rev. A **89**, 034303 (2014). Y. Luo and Y. Li, Monogamy of $\alpha$th power entanglement measurement in qubit systems, Ann. Phys. (NY) **362**, 511 (2015). X. N. Zhu and S. M. Fei, Entanglement monogamy relations of qubit systems, Phys. Rev. A **90**, 024304 (2014). Z. X. Jin and S. M. Fei, Tighter entanglement monogamy relations of qubit systems, Quantum Inf. Process. **16**, 77 (2017). Z. X. Jin, J. Li, T. Li and S. M. Fei, Tighter monogamy relations in multiqubit systems, Phys. Rev. A **97**, 032336 (2018). J. S. Kim, Negativity and tight constraints of multiqubit entanglement, Phys. Rev. A **97**, 012334 (2018). G. Vidal and R. F. Werner, Computable measure of entanglement, Phys. Rev. A **65**, 032314 (2002). M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information (Cambridge University Press, Cambridge, UK, 2000). L.-M. Yang, B. Chen, S.-M. Fei, and Z.-X. Wang, Tighter constraints of multiqubit entanglement, Commun. Theor. Phys. **71**, 545-554 (2019). [^1]: Corresponding author: [email protected]

ArXiv

--- abstract: 'We present a new inner bound for the rate region of the $t$-stage successive-refinement problem with side-information. We also present a new upper bound for the rate-distortion function for lossy-source coding with multiple decoders and side-information. Characterising this rate-distortion function is a long-standing open problem, and it is widely believed that the tightest upper bound is provided by Theorem 2 of Heegard and Berger’s paper “Rate Distortion when Side Information may be Absent,” *IEEE Trans. Inform. Theory*, 1985. We give a counterexample to Heegard and Berger’s result.' author: - 'Roy Timo, Terence Chan and Alex Grant[^1][^2]' title: 'Rate Distortion with Side-Information at Many Decoders' --- Rate distortion, side-information, successive refinement. Introduction {#Sec:1} ============ of the most important and celebrated results in multi-terminal information theory is Wyner and Ziv’s solution to the problem of *lossy source coding with side-information at the decoder* [@Wyner-Jan-1976-A] – the Wyner-Ziv problem (fig. \[fig:Wyner-Ziv\]). The main objective of this problem is to find a computable characterisation [@Csiszar-1981-B Pg. 259] of the rate-distortion function $R(d)$. This function describes the smallest rate at which the encoder can compress an iid random sequence ${\mathbf{X}}$ so that the decoder, which has side-information ${\mathbf{Y}}$, can produce a replica $\Hat{{\mathbf{X}}}$ of ${\mathbf{X}}$ that satisfies the average distortion constraint $$\label{Sec:1:Eqn:Distortion-1} \mathbb{E}\left[\frac{1}{n}\sum_{i=1}^n \delta\big(X_i,\Hat{X}_i\big)\right] \leq d\ ,$$ where $\delta$ is a real-valued distortion measure [@Cover-1991-B] and $\mathbb{E}[\cdot]$ is the expectation operation. In [@Wyner-Jan-1976-A Thm. 1], Wyner and Ziv famously showed that $$\begin{aligned} \label{Sec:1:Eqn:Wyner-Ziv-1} R(d) &= \min_U \big\{ I(X;U) - I(U;Y)\big\}\ ,\end{aligned}$$ where the minimization is taken over all choices of an auxiliary random variable $U$ that is jointly distributed with $(X,Y)$ and which satisfies the following two properties: (1) $U$ is conditionally independent of $Y$ given $X$; and (2) there exists a function $\Hat{X}(U,Y)$ with $\mathbb{E}\delta(X,\Hat{X}(U,Y)) \leq d$. In this paper, we study the following two extensions of the Wyner-Ziv problem: (1) the Wyner-Ziv problem with multiple decoders (fig. \[Sec:1:Fig:mKHB\]); and (2) the successive-refinement problem with side-information (fig. \[Sec:1:Fig:tSR\]). A brief history of the literature on these problems is as follows. ![The Wyner-Ziv Problem: $({\mathbf{X}},$ ${\mathbf{Y}})$ $=$ $(X_1,$ $Y_1)$, $(X_2,$ $Y_2)$, $\ldots$, $(X_n,$ $Y_n)$ is an iid random sequence emitted by a source $q(x,y) = \text{Pr}[X=x,Y=y]$. The encoder maps ${\mathbf{X}}$ to an index $M$, which belongs to a finite set ${\mathscr{M}}$, at a rate $r$. Using $M$ and ${\mathbf{Y}}$, the decoder is required to generate a replica $\Hat{{\mathbf{X}}} = \Hat{X}_1,\Hat{X}_2,\ldots,\Hat{X}_n$ of ${\mathbf{X}}$ to within an average distortion $d$, according to . The rate-distortion function $R(d)$ is defined as the smallest rate for which such a reconstruction is possible. A single-letter expression for this function was first given in [@Wyner-Jan-1976-A Thm. 1].[]{data-label="fig:Wyner-Ziv"}](WynerZivF){width="0.6\columnwidth"} The Wyner-Ziv Problem with $t$-Decoders --------------------------------------- Suppose that the side-information ${\mathbf{Y}}$ in Figure \[fig:Wyner-Ziv\] is unreliable in the sense that it may or may not be available to the decoder. If the encoder does not know *a priori* when ${\mathbf{Y}}$ is available, then Wyner and Ziv’s coding argument for  fails, and a more sophisticated argument is required to exploit ${\mathbf{Y}}$. This observation inspired Kaspi [@Kaspi-Nov-1994-A] in 1980 (published by Wyner on behalf of Kaspi in 1994) as well as Heegard and Berger [@Heegard-Nov-1985-A] in 1985 to independently study the problem shown in fig. \[Sec:1:Fig:KHB\] – the Kaspi/Heegard-Berger problem. As with the Wyner-Ziv problem, the objective of this problem is to characterise the corresponding rate-distortion function $R(d_1,d_2)$. That is, to find the smallest rate such that decoders 1 and 2 can produce replicas $\Hat{{\mathbf{X}}}_1$ and $\Hat{{\mathbf{X}}}_2$ of ${\mathbf{X}}$ to within average distortions $d_1$ and $d_2$, respectively. To this end, Heegard and Berger [@Heegard-Nov-1985-A Thm. 1] showed that[^3] $$R(d_1,d_2) = \min_{U,W} \big\{ I\left(X;{W}\right) + I\left(X;{U}\mid Y, {W}\right) \big\}\ ,$$ where the minimization is taken over all choices of two auxiliary random variables, ${U}$ and ${W}$, that are jointly distributed with $(X,Y)$ and which satisfy the following two properties: (1) $({U},{W})$ is conditionally independent of $Y$ given $X$; and (2) there exist functions $\Hat{X}_{1}({W})$ and $\Hat{X}_{2}(Y,{U},{W})$ with $\mathbb{E}\delta(X,\Hat{X}_1({W})) \leq d_1$ and $\mathbb{E}\delta(X,\Hat{X}_2(Y,{U},{W})) \leq d_2$, respectively. ![The Kaspi/Heegard-Berger Problem: The encoder compresses ${\mathbf{X}}$ in a manner suitable for two decoders – one of which has side-information ${\mathbf{Y}}$. The rate-distortion function $R(d)$ defines the smallest rate at which decoders 1 and 2 can generate replicas $\Hat{{\mathbf{X}}}_1 = \Hat{X}_{1,1},\Hat{X}_{1,2},\ldots,\Hat{X}_{1,n}$ and $\Hat{{\mathbf{X}}}_2 = \Hat{X}_{2,1},\Hat{X}_{2,2},\ldots,\Hat{X}_{2,n}$ of ${\mathbf{X}}$ to within average distortions $d_1$ and $d_2$, respectively. A single-letter expression for this function was independently given in [@Kaspi-Nov-1994-A] and [@Heegard-Nov-1985-A].[]{data-label="Sec:1:Fig:KHB"}](KaspiHeegardBergerF){width=".75\columnwidth"} The Kaspi/Heegard-Berger problem in Figure \[Sec:1:Fig:KHB\] was further generalised by Heegard and Berger in [@Heegard-Nov-1985-A Sec. VII] to the problem shown in Figure \[Sec:1:Fig:mKHB\]. There are $t$-decoders, each with different side-information, and the objective is to characterise the corresponding rate-distortion function $R({\mathbf{d}})$. Unfortunately, this function has eluded characterisation for all but a few special cases. For example, Heegard and Berger [@Heegard-Nov-1985-A Thm. 3] have characterised $R({\mathbf{d}})$ for stochastically degraded side-information[^4]; Tian and Diggavi [@Tian-Aug-2007-A; @Tian-Dec-2008-A] have characterised $R({\mathbf{d}})$ for a quadratic Gaussian source with jointly Gaussian side-information; and Sgarro’s result [@Sgarro-Mar-1977-A Thm. 1] subsumes the corresponding lossless problem. Notwithstanding this difficulty, however, this problem has helped stimulate a number of important results [@Kaspi-Nov-1994-A; @Steinberg-Aug-2004-A; @Tian-Oct-2006-C; @Tian-Dec-2008-A; @Tian-Aug-2007-A]. ![The Wyner-Ziv problem with $t$-decoders. The encoder compresses ${\mathbf{X}}$ in a manner suitable for $t$-decoders – each of which has different side-information. The rate-distortion function $R({\mathbf{d}})$, where ${\mathbf{d}} = (d_1,d_2,\dots,d_t)$, defines the smallest rate at which decoder $l$, for all $l =1,2,\ldots,t$, can generate a replica $\Hat{{\mathbf{X}}}_l$ of ${\mathbf{X}}$ to within an average distortion $d_l$. This problem is open for $t \geq 2$. We present an upper bound for $R({\mathbf{d}})$ in Theorem \[Sec:4:Thm:HB\].[]{data-label="Sec:1:Fig:mKHB"}](HeegardBergerF){width=".75\columnwidth"} In [@Heegard-Nov-1985-A Thm. 2], Heegard and Berger claimed that a certain functional, $R_0({\mathbf{d}})$, is an upper bound for $R({\mathbf{d}})$. (The expression for $R_0({\mathbf{d}})$ is given in equation  of Section \[Sec:2\]; however, this expression requires notation from Section \[Sec:2\].) For twenty-five years, $R_0({\mathbf{d}})$ has been universally considered to be the tightest upper bound for $R({\mathbf{d}})$ in the literature. In Example \[Sec:2:Counterexample\] of Section \[Sec:2\], we present a counterexample to [@Heegard-Nov-1985-A Thm. 2] that shows $R_0({\mathbf{d}})$ is *not* an upper bound for $R({\mathbf{d}})$. The invalidity of [@Heegard-Nov-1985-A Thm. 2] is by no means obvious as it involves a difficult minimization over $(2^t - 1)$-auxiliary random variables. Indeed, we note that this theorem has been cited with modest frequency in the literature, and all the while this error appears to have gone unnoticed. We present a new upper bound for $R({\mathbf{d}})$ in Theorem \[Sec:4:Thm:HB\] of Section \[Sec:4\]. The Successive-Refinement Problem with Side-Information ------------------------------------------------------- The aforementioned counterexample led us to study the $t$-stage (or, $t$-decoder) successive-refinement problem shown in Figure \[Sec:1:Fig:tSR\]. The encoder maps ${\mathbf{X}}$ to $t$ indices: $M_1, M_2, \dots, M_t$. It is required that decoder $l$ uses indices $M_1$ through $M_l$ together with its side-information ${\mathbf{Y}}_l$ to produce a replica $\Hat{{\mathbf{X}}}_l = X_{l,1},X_{l,2},\ldots,X_{l,n}$ of ${\mathbf{X}}$ to within an average distortion $d_l$. The objective of this problem is to characterise the resulting admissible-rate region ${\mathscr{R}}({\mathbf{d}})$. That is, to determine the set of all rate tuples ${\mathbf{r}} = (r_1,r_2,\ldots,r_t)$ for which each decoder can reconstruct ${\mathbf{X}}$ to within its desired distortion level. ![The successive-refinement problem with $t$ stages. The encoder compresses ${\mathbf{X}}$ in $t$-stages. At stage $l$, decoder $l$ generates a replica $\Hat{{\mathbf{X}}}_l$ of ${\mathbf{X}}$. This problem is open for $t \geq 2$. We present an inner bound for the admissible-rate region in Section \[Sec:3\] (Theorem \[Sec:3:Thm:Successive-Refinement\]). []{data-label="Sec:1:Fig:tSR"}](SuccessiveRefinementF "fig:"){width=".75\columnwidth"}\ Assuming the side-information is stochastically degraded, Steinberg and Merhav [@Steinberg-Aug-2004-A] characterised ${\mathscr{R}}(d_1,d_2)$ for $t=2$ decoders. Shortly thereafter, Tian and Diggavi [@Tian-Aug-2007-A] extended this problem to $t$-decoders and proved the following result. \[Sec:1:Pro:Degraded-Side-Information\] If the side-information is stochastically degraded, then ${\mathscr{R}}({\mathbf{d}})$ is equal to the set of all rate tuples ${\mathbf{r}}$ for which there exists $t$ auxiliary random variables $U_1$, $U_2$, $\ldots$, $U_t$ such that $$\sum_{k=1}^l r_k \geq \sum_{k=1}^l I\big(X;U_k\big|U_1,U_2,\ldots,U_{k-1},Y_k\big)\ ,$$ for all $l = 1,2,\ldots,t$, where 1. $(U_1,U_2,\ldots,U_t)$ is conditionally independent of $(Y_1,$ $Y_2,$ $\ldots,$ $Y_t)$ given $X$; and 2. there exist $t$ functions $\Hat{X}_l(U_l,Y_l)$, $l = 1,2,\ldots,t$, with $\mathbb{E}\delta_l(X,\Hat{X}(U_l,Y_l)) \leq d_l$. More recently, Tian and Diggavi [@Tian-Dec-2008-A] gave the following non-trivial inner bound for ${\mathscr{R}}(d_1,d_2)$ under the assumption that $X$ and $Y_2$ are conditionally independent given $Y_1$ – the scalable side-information source coding problem. Note, this conditional independence is the reverse of the stochastic degradedness used in Proposition \[Sec:1:Pro:Degraded-Side-Information\]. \[Sec:1:Pro:SI-Scalable\] If $X$ and $Y_2$ are conditionally independent given $Y_1$, then a rate pair $(r_1,r_2)$ is $(d_1,d_2)$-admissible if there exists three auxiliary random variables, $U_{12}$, $U_1$ and $U_2$, such that $$\begin{aligned} r_1 &\geq I\big(X;U_1,U_{12}\big|Y_1\big)\\ r_1 + r_2 &\geq I\big(X;U_2,U_{12}\big|Y_2) + I\big(X;U_1\big|Y_1,U_{12}\big)\ ,\end{aligned}$$ where 1. $(U_{12},U_1,U_2)$ is conditionally independent of $(Y_1,Y_2)$ given $X$; 2. there exist functions $\Hat{X}_1(U_1,Y_1)$ and $\Hat{X}_2(U_2,Y_2)$ such that $\mathbb{E}\delta_1(X,\Hat{X}_1(U_1,Y_1))\leq d_1$ and $\mathbb{E}\delta_2(X,$ $\Hat{X}_2(U_2,Y_2))$ $\leq$ $d_2$. We present a new inner bound for ${\mathscr{R}}({\mathbf{d}})$ for the general $t$-decoder problem with arbitrarily correlated side-information in Theorem \[Sec:3:Thm:Successive-Refinement\] of Section \[Sec:3\]. Paper Outline $\&$ Notation --------------------------- In Section \[Sec:2\], we formally define ${\mathscr{R}}({\mathbf{d}})$ and $R({\mathbf{d}})$ and give the counterexample to [@Heegard-Nov-1985-A Thm. 2]. In Sections \[Sec:3\] and \[Sec:4\], we respectively present new achievability results for ${\mathscr{R}}({\mathbf{d}})$ and $R({\mathbf{d}})$. We describe a new lossless source coding problem in Section \[Sec:5\], and the paper is concluded in Section \[Sec:6\]. The non-negative real numbers and the natural numbers are written as $\mathbb{R}_+$ and $\mathbb{N}$, respectively. For $s,t \in \mathbb{N}$ with $s\leq t$, we let $[s,t] \triangleq \{s,s+1,s+2,\dots,t\}$. When $s = 1$, we drop $s$, i.e. $[t] \triangleq \{1,2,\ldots,t\}$. Proper subsets and subsets are identified by $\subset$ and $\subseteq$, respectively. Random variables and random sequences are identified by upper case and bolded uppercase letters, respectively. For example, ${\mathbf{X}} = X_1,X_2,\ldots,X_n$ denotes the random sequence to be replicated at the decoders, and ${\mathbf{Y}}_l = Y_{l,1},Y_{l,2},\ldots,Y_{l,n}$ denotes the side-information at decoder $l$. The letter $U$ is always used to represent auxiliary random variables. The alphabets of random variables are identified by matching calligraphic typeface, e.g. ${\mathscr{X}}$ and ${\mathscr{U}}$ are the respective alphabets of $X$ and $U$. A generic element of an alphabet is identified by a matching lowercase letter, e.g. $x \in {\mathscr{X}}$ and $u \in {\mathscr{U}}$. The Cartesian product operation is denoted by $\times$, e.g. ${\mathscr{X}} \times {\mathscr{Y}}$. The $t$-fold Cartesian product of a single alphabet/set is identified with a superscript, e.g. ${\mathscr{X}}^t$ and $\mathbb{R}_+^t$. Tuples from product spaces are identified by boldfaced lowercase letters, e.g. ${\mathbf{x}} = (x_1,x_2,\ldots,x_n) \in {\mathscr{X}}^n$. For notational convenience, the same letter is used to represent a joint pmf and its marginals, e.g. if $(X,Y)$ on ${\mathscr{X}} \times {\mathscr{Y}}$ is defined by $p(x,y) \triangleq \text{Pr}[X=x,Y=y]$, then $p(x) \triangleq \sum_{x\in{\mathscr{X}}} p(x,y)$. The symbol $\minuso$ is used to denote Markov Chains, e.g. if $(X,Y,Z)$ on ${\mathscr{X}} \times {\mathscr{Y}} \times {\mathscr{Z}}$ is defined by $p(x,y,z) \triangleq \text{Pr}[X=x,Y=y,Z=z]$ where $$p(x,y,z) = \left\{ \begin{array}{ll} p(x,y)p(y,z)/p(y), & \hbox{ if } p(y) > 0 \\ 0, & \hbox{ otherwise,} \end{array} \right.$$ then we write $X \minuso Y \minuso Z\ [p]$. Mutual information and entropy are written in the standard fashion [@Cover-1991-B] using $I$ and $H$, respectively. We sometimes use subscripts for $I$ and $H$ to emphasize that random variables under consideration are defined by a particular pmf, e.g. if $(X,Y)$ is defined by $p(x,y) = \text{Pr}[X=x,Y=y]$, then we write $I_p(X;Y)$. Definitions $\&$ Counterexample {#Sec:2} =============================== Successive Refinement with Side-Information {#Sec:2:A} ------------------------------------------- Consider Figure \[Sec:1:Fig:tSR\]. Let ${\mathscr{X}}$, ${\mathscr{Y}}_1$, ${\mathscr{Y}}_2$, $\ldots$, ${\mathscr{Y}}_t$ be finite alphabets and set ${\mathscr{Y}}^* \triangleq {\mathscr{Y}}_1 \times {\mathscr{Y}}_2 \times \cdots \times {\mathscr{Y}}_t$. Let $$\big({\mathbf{X}},{\mathbf{Y}}_{1},{\mathbf{Y}}_{2},\dots,{\mathbf{Y}}_{t}\big) \triangleq \Big\{\big(X_i, Y_{1,i},Y_{2,i},\dots,Y_{t,i})\Big\}_{i=1}^n$$ denote $n$ $(t+1)$-tuples of random variables that are drawn in an iid manner from ${\mathscr{X}}\times {\mathscr{Y}}^*$ according to a generic pmf $q$, where $$q\left(x,y_{1},\dots,y_{t}\right) \triangleq \Pr\big[X_1 = x_1, Y_{1} = y_{1},\ldots, Y_{t} = y_{t}\big].$$ We assume that ${\mathbf{X}} = X_1,X_2,\ldots,X_n$ is known to encoder and ${\mathbf{Y}}_l = Y_{l,1},Y_{l,2},\ldots,Y_{l,n}$ is known to decoder $l$. The encoder compresses ${\mathbf{X}}$ with $$f^{(n)} : {\mathscr{X}}^n \rightarrow {\mathscr{M}}_1 \times {\mathscr{M}}_2 \times \cdots \times {\mathscr{M}}_t\ ,$$ where ${\mathscr{M}}_1$, ${\mathscr{M}}_2$, $\ldots$, ${\mathscr{M}}_t$ are finite sets. The resulting $t$ indices $$(M_{1},M_{2},\ldots,M_{t}) = f^{(n)}\left({\mathbf{X}}\right)$$ are sent over channels $1$ through $t$, respectively. The rate of the encoder on channel $l$ (in bits per source symbol) is given by $$\kappa_l^{(n)} \triangleq \frac{1}{n} \log_2 |{\mathscr{M}}_l|\ ,$$ where $|{\mathscr{M}}_l|$ is the cardinality of ${\mathscr{M}}_l$. Consider decoder $l$. Let $\Hat{{\mathscr{X}}}_{l}$ be a finite reconstruction alphabet, and let $$\delta_{l} : {\mathscr{X}} \times \Hat{{\mathscr{X}}}_{l} \rightarrow {\mathbb{R}}_+$$ be a per-letter distortion measure. Observe that $\Hat{{\mathscr{X}}}_l$ and $\delta_l$ can be different to those used at the other decoders. We assume that $\delta_l$ is normal[^5] in sense that $\delta_l(x,\Hat{x}^*(x)) = 0$ for all $x \in {\mathscr{X}}$, where $$\Hat{x}^*(x) \triangleq \underset{\Hat{x} \in \Hat{{\mathscr{X}}}}{\operatorname{argmin}}\ \delta_l(x,\Hat{x})\ .$$ This decoder is required to generate a replica $\Hat{{\mathbf{X}}}_{l}\triangleq \Hat{X}_{l,1},\Hat{X}_{l,2},\ldots,\Hat{X}_{l,n}$ of ${\mathbf{X}}$ using $$g^{(n)}_l : {\mathscr{M}}_1 \times {\mathscr{M}}_2 \cdots \times {\mathscr{M}}_l \times {\mathscr{Y}}^n_{l} \rightarrow \Hat{{\mathscr{X}}}_{l}^n\ ;$$ that is, $$\Hat{{\mathbf{X}}}_{l} = g_l^{(n)}(M_1, M_2, \ldots,M_l,{\mathbf{Y}}_{l})\ .$$ Finally, the quality of this replica is measured by the average distortion $$\Delta_l^{(n)} \triangleq \operatorname{\mathbb{E}}\left[ \frac{1}{n}\sum_{i=1}^n \delta_l(X_i,\Hat{X}_{l,i})\right]\ .$$ \[Sec:2:Def:d-admissible\] Suppose ${\mathbf{d}}$ $=$ $(d_1,$ $d_2,$ $\ldots,$ $d_t)$ $\in$ ${\mathbb{R}}_+^t$. A rate tuple ${\mathbf{r}}$ $=$ $(r_1$, $r_2$, $\ldots$, $r_t)$ $\in$ ${\mathbb{R}}_+^t$ is said to be [*${\mathbf{d}}$-admissible*]{} if, for arbitrary $\epsilon > 0$, there exists an $n_\epsilon \in \mathbb{N}$, an encoder $f^{(n_\epsilon)}$ and $t$-decoders $g^{(n_\epsilon)}_1$, $g^{(n_\epsilon)}_2$, $\ldots$, $g^{(n_\epsilon)}_t$ such that $$\begin{aligned} d_l + \epsilon &\geq \Delta_l^{(n_\epsilon)},\quad \forall l \in [t], \text{ and}\\ r_l + \epsilon &\geq \kappa_l^{(n_\epsilon)}, \quad \forall l \in [t] .\end{aligned}$$ We let ${{\mathscr{R}}}({\mathbf{d}})$ denote the set of all ${\mathbf{d}}$-admissible rate tuples. We note that Definition \[Sec:2:Def:d-admissible\] matches Tian and Diggavi [@Tian-Aug-2007-A] in that the $l^{\text{th}}$ channel (or, refinement) rate $\kappa_l^{(n)}$ is characterised in an individual (or, incremental) manner. In contrast, Steinberg and Merhav [@Steinberg-Aug-2004-A] define the $l^{\text{th}}$ refinement rate in a cumulative manner, e.g. $(1/n)\log(|{\mathscr{M}}_1||{\mathscr{M}}_2|\cdots|{\mathscr{M}}_l|)$. We also note that ${{\mathscr{R}}}({\mathbf{d}})$ is dependent on the successive-refinement decoding order [@Tian-Dec-2008-A]. That is, if we interchange decoders (keeping the same side-information and distortion constraints at each decoder), then ${\mathscr{R}}({\mathbf{d}})$ will change. We conclude this section with a summary of some fundamental properties of ${\mathscr{R}}({\mathbf{d}})$. These properties can all be deduced directly from Definition \[Sec:2:Def:d-admissible\]. See  [@Gray-Nov-1974-A; @Effros-Sep-1999-A; @Steinberg-Aug-2004-A; @Tian-Aug-2007-A; @Tian-Jul-2009-C] for similar discussions. \[Sec:2:Pro:Marginal-Property\] The region ${\mathscr{R}}({\mathbf{d}})$ is completely defined by the pair-wise marginal distributions of $X$ with each side-information. Let $q'$ and $q''$ be pmfs on ${\mathscr{X}}\times{\mathscr{Y}}^*$, and let ${\mathscr{R}}(\mathbf{d})[q']$ and ${\mathscr{R}}(\mathbf{d})[q'']$ denote their respective ${\mathbf{d}}$-admissible rate regions (assuming the same distortion measures). If $q'(x,y_l) = q''(x,y_l)$ for all $(x,y_l) \in {\mathscr{X}} \times {\mathscr{Y}}_l$ and $l \in [t]$, then ${\mathscr{R}}(\mathbf{d})[q'] = {\mathscr{R}}(\mathbf{d})[q'']$. \[Sec:2:Pro:Convex\] The region ${\mathscr{R}}({\mathbf{d}})$, for every ${\mathbf{d}} \in {\mathbb{R}}_+^t$, is a closed convex subset of ${\mathbb{R}}_+^t$ that is uniquely determined by its lower boundary $$\Big\{{\mathbf{r}} \in {\mathscr{R}}({\mathbf{d}})\ :\ \forall \tilde{{\mathbf{r}}} \in {\mathscr{R}}({\mathbf{d}}),\ \tilde{r}_l \leq r_l\ , \forall l \in [t]\ \Rightarrow \tilde{r}_l = r_l\ \forall l \in [t]\Big\}\ .$$ \[Sec:2:Pro:Sum-Incremental\] The region ${\mathscr{R}}({\mathbf{d}})$ is sum incremental in the sense that rate can always be transferred from higher-index channels to lower-index channels. If ${\mathbf{r}} \in {\mathscr{R}}({\mathbf{d}})$, then $$\label{Sec:2:Eqn:Latent} {\mathscr{R}}({\mathbf{d}}) \supseteq {\mathscr{L}}({\mathbf{r}}) \triangleq \left\{\tilde{{\mathbf{r}}} \in {\mathbb{R}}_+^t\ :\ \sum_{k=1}^l \tilde{r}_k \geq \sum_{k=1}^l r_k\ \quad \forall l \in [t]\right\} \ .$$ We note in passing that Proposition \[Sec:2:Pro:Sum-Incremental\] also holds in a more universal setting. Suppose ${\mathbf{r}} \in {\mathbb{R}}_+^t$. Consider all combinations of the source distribution, distortion measures and distortion tuple (e.g., $\tilde{{\mathscr{X}}}$, $\tilde{{\mathscr{Y}}}^*$, $\tilde{q}$, $\{\tilde{\delta}_l\}_{l=1}^t$ and $\tilde{{\mathbf{d}}} \in {\mathbb{R}}_+^t$) such that the resulting $\tilde{{\mathbf{d}}}$-admissible rate region ${\mathscr{R}}(\tilde{{\mathbf{d}}})[\tilde{q}]$ contains ${\mathbf{r}}$. The proposition shows that ${\mathscr{L}}({\mathbf{r}})$ is an inner bound for every such region. In addition, it can be shown that ${\mathscr{L}}({\mathbf{r}})$ is maximal in the sense that ${\mathscr{L}}({\mathbf{r}}) = {\mathscr{R}}(\tilde{{\mathbf{d}}})[\tilde{q}]$ for some choice of $\tilde{{\mathscr{X}}}$, $\tilde{{\mathscr{Y}}}^*$, $\tilde{q}$, $\{\tilde{\delta_l}\}$ and $\tilde{{\mathbf{d}}}$. Therefore, the ${\mathbf{d}}$-admissibility of $\tilde{{\mathbf{r}}} \notin {\mathscr{L}}({\mathbf{r}})$ cannot be inferred from the ${\mathbf{d}}$-admissibility of ${\mathbf{r}}$ without specific consideration of the source distribution, distortion measures and distortion tuple. For this reason, ${\mathscr{L}}({\mathbf{r}})$ can be called the latent admissible rate region implied by ${\mathbf{r}}$. See, for example, [@Tian-Jul-2009-C]. We give an inner bound for ${\mathscr{R}}({\mathbf{d}})$ in Theorem \[Sec:3:Thm:Successive-Refinement\] of Section \[Sec:3\]. However, before giving this bound, it is useful to formally define the rate-distortion function $R({\mathbf{d}})$ (fig. \[Sec:1:Fig:mKHB\]) and then review Heegard and Berger’s functional $R_0({\mathbf{d}})$. Rate Distortion with Side-Information at $t$-decoders {#Sec:2:B} ----------------------------------------------------- The rate-distortion function $R({\mathbf{d}})$ for the problem shown in Figure \[Sec:1:Fig:mKHB\] can be efficiently recovered from ${\mathscr{R}}({\mathbf{d}})$ by restricting the code rate on channels $2$ through $t$ to be zero. \[Sec:2:Def:Heegard-Berger\] The rate-distortion function for lossy source coding with side-information at $t$-decoders (fig. \[Sec:1:Fig:mKHB\]) is defined by $$R({\mathbf{d}}) \triangleq \min \big\{ r \in \mathbb{R}_+ : (r,0,0,\cdots,0) \in {\mathscr{R}}({\mathbf{d}}) \big\}\ ,$$ where the indicated minimum exists because ${\mathscr{R}}({\mathbf{d}})$ is closed and bounded from below. It should be noted that Definition \[Sec:2:Def:Heegard-Berger\] technically permits the use of codes with *asymptotically-vanishing rates* on channels $2$ through $t$. That is, the ${\mathbf{d}}$-admissibility of rates approaching $R({\mathbf{d}})$ from above can be proved using a sequence of codes where $\epsilon_i \rightarrow 0$ and $\kappa^{(n_{\epsilon_i})}_l \rightarrow 0$ for all $l \in [2,t]$. Such codes, however, are not permitted in the single-channel rate-distortion problem (fig. \[Sec:1:Fig:mKHB\]); we can only use codes with $\kappa^{(n)}_l = 0$ for all $l \in [2,t]$. Despite this subtle difference, Definition \[Sec:2:Def:Heegard-Berger\] is equivalent to the definition used in [@Heegard-Nov-1985-A] because any message transmitted on channels $2$ through $t$ can be transferred[^6] to channel $1$ (see Proposition \[Sec:2:Pro:Sum-Incremental\]). As mentioned in Section \[Sec:2:A\], ${\mathscr{R}}({\mathbf{d}})$ depends on the successive-refinement decoding order. This dependence, of course, is not shared by $R({\mathbf{d}})$. Indeed, the aforementioned rate-transfer argument can be used show that the decoding order (used to define ${\mathscr{R}}({\mathbf{d}})$ in Definition \[Sec:2:Def:Heegard-Berger\]) can be interchanged with any other decoding order without altering $R({\mathbf{d}})$. Using the time-sharing principle, it can be shown that $R({\mathbf{d}})$ is convex on $\mathbb{R}_+^t$. This convexity ensures that $R({\mathbf{d}})$ is continuous on the interior of $\mathbb{R}_+^t$ [@Rockafellar-1997-B Thm. 10.1]. Moreover, it can also be verified that $R({\mathbf{d}})$ is continuous whenever $d_l = 0$ for some $l \in [t]$; see, for example,  [@Wyner-Jan-1976-A Pg. 2]. The rate-distortion function $R({\mathbf{d}})$ is continuous, non-increasing (i.e., $R({\mathbf{d}}) \leq R(\tilde{\bf{d}})$ when $d_l \geq \tilde{d}_l$ for all $l\in[t]$) and convex on $\mathbb{R}_+^t$. The following proposition for lossless reconstructions can be obtained as an extension to the Slepian-Wolf Theorem [@Slepian-Jul-1973-A Thm. 2], a variant of a more general result by Sgarro [@Sgarro-Mar-1977-A Thm. 2], or a special case of Bakshi and Effros [@Bakshi-Jul-2008-C Thm. 1]. \[Sec:2:Pro:Slepian-Wolf\] If, for every $l \in [t]$, $\Hat{{\mathscr{X}}}_l = {\mathscr{X}}$ and $\delta_l$ satisfies $$\delta_l(x,x) = 0 \text{ and}$$ $$\delta_l(x,\Hat{x}) > 0,\quad x \neq \Hat{x}\ ,$$ then $$R(0,0,\ldots,0) = \max_{l \in [t]} H(X|Y_l)\ .$$ To review Heegard and Berger’s work on $R({\mathbf{d}})$ for generic distortion tuples, we first need to define $(2^t-1)$-auxiliary random variables – one for every non-empty subset of decoders. For this purpose, arrange the non-empty subsets of $[t]$ into a list ${\mathscr{S}}_1,{\mathscr{S}}_2,\ldots,{\mathscr{S}}_{2^t-1}$ (the ordering is not important). For each $j\in[2^t-1]$, let ${\mathscr{U}}_{{\mathscr{S}}_j}$ be a finite alphabet. Define ${\mathscr{U}}^* \triangleq {\mathscr{U}}_{{\mathscr{S}}_1} \times {\mathscr{U}}_{{\mathscr{S}}_2} \times \cdots \times {\mathscr{U}}_{{\mathscr{S}}_{2^t-1}}$. Let ${\mathscr{P}}$ denote all those pmfs $p$ on ${\mathscr{U}}^* \times {\mathscr{X}} \times {\mathscr{Y}}^*$ whose $({\mathscr{X}} \times {\mathscr{Y}}^*)$-marginal is equal to the source distribution $q$: $$\begin{aligned} p(x,y_1,\ldots,y_t) &\triangleq \sum_{(u_1,u_2\ldots,u_{2^t-1}) \in {\mathscr{U}}^{*}} p(u_1,u_2\ldots,u_{2^t-1},x,y_1,\ldots,y_t)\\ &=q(x,y_1,y_2,\ldots,y_t)\ .\end{aligned}$$ Each $p \in {\mathscr{P}}$ specifies a joint pmf for $(2^t - 1)$-auxiliary random variables. We denote these variables by $U_{{\mathscr{S}}_1}$, $U_{{\mathscr{S}}_2}$, $\ldots$, $U_{{\mathscr{S}}_{2^t-1}}$, where $U_{{\mathscr{S}}_j}$ takes values from ${\mathscr{U}}_{{\mathscr{S}}_j}$. Let ${\mathscr{A}}$ $\triangleq$ $\{U_{{\mathscr{S}}_1}$, $U_{{\mathscr{S}}_2}$, $\ldots$, $U_{{\mathscr{S}}_{2^t-1}}\}$, and let $$\begin{aligned} {\mathscr{A}}^\supset_{{{\mathscr{S}}}_j} &\triangleq \Big\{U_{{\mathscr{S}}_k} \in {\mathscr{A}} :\ {\mathscr{S}}_k \supset {{\mathscr{S}}}_j \Big\}\end{aligned}$$ denote those auxiliary random variables associated with supersets of ${\mathscr{S}}_j$. Let ${\mathscr{P}}({\mathbf{d}})$ denote the set of all $p \in {\mathscr{P}}$ for which the following two properties are satisfied: 1. $p$ factors to form the Markov chain: $$\big(U_{{{\mathscr{S}}}_1},U_{{{\mathscr{S}}}_2},\ldots,U_{{{\mathscr{S}}}_{2^t-1}}\big) \minuso X \minuso \big(Y_1,Y_2,\ldots,Y_t\big)\ [p]\ ; \text{ and}$$ 2. for every decoder $l \in [t]$ there exists a function $\Hat{X}_l(Y_l,U_{\{l\}},{\mathscr{A}}^\supset_{\{l\}})$ with $$\mathbb{E}_p \delta_l \Big(X,\Hat{X}_l\big(Y_l,U_{\{l\}},{\mathscr{A}}^\supset_{\{l\}}\big)\Big) \leq d_l\ .$$ Heegard and Berger claimed [@Heegard-Nov-1985-A Thm. 2] that the functional $$\label{Sec:2:Eqn:HB-Upper-Bound} R_{0}({\mathbf{d}}) = \min_{p \in {\mathscr{P}}({\mathbf{d}})} \sum_{j=1}^{2^t-1} \max_{l\in{\mathscr{S}}_j} I_p\big(X;U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j},Y_l\big)$$ is an upper bound for $R({\mathbf{d}})$ for all finite alphabets ${\mathscr{U}}_{{\mathscr{S}}_1}$, ${\mathscr{U}}_{{\mathscr{S}}_2}$, $\ldots$, ${\mathscr{U}}_{{\mathscr{S}}_{2^t-1}}$ such that ${\mathscr{P}}({\mathbf{d}})$ is non-empty. In the next two examples, we confirm that $R_{0}({\mathbf{d}})$ is an upper bound for $R({\mathbf{d}})$ when there is one or two decoders $(t = 1 \text{ or } 2)$; however, in the third example we show that $R_{0}({\mathbf{d}})$ is *not* an upper bound for $R({\mathbf{d}})$ when there is three or more decoders ($t \geq 3$). For brevity, we drop set notation for each auxiliary random in the following three examples. For example, we write $U_1$, $U_{12}$ and $U_{123}$ in place of $U_{\{1\}}$, $U_{\{1,2\}}$ and $U_{\{1,2,3\}}$, respectively. If $t=1$, then reduces to $$\begin{aligned} \notag R_0(d_1) &= \min_{p \in {\mathscr{P}}(d_1)} I_p\big(X;U_1\big|Y_1\big) \\ \label{Sec:2:Eqn:HB-WZ} &= \min_{p \in {\mathscr{P}}(d_1)} \Big\{I_p\big(X;U_1\big) - I_p\big(U_1;Y_1\big)\Big\}\ ,\end{aligned}$$ where the equality in  follows from the chain rule for mutual information and the Markov chain $U_1 \minuso X \minuso Y_1\ [q]$. If the cardinality of ${\mathscr{U}}_{{\mathscr{S}}_1}$ is limited to $|{\mathscr{U}}_{{\mathscr{S}}_1}| \leq |{\mathscr{X}}| + 1$, then the right hand side of  reduces to the Wyner-Ziv formula . $\square$ If $t=2$, then reduces to $$\begin{aligned} \label{Sec:2:Eqn:2-decoder} R_0(d_1,d_2) &= \min_{p \in {\mathscr{P}}(d_1,d_2)} \Big\{ \max_{l \in \{1,2\}} I_p\big(X;U_{12}\big|Y_l\big) + I_p\big(X;U_{1}\big|Y_1,U_{12}\big) + I_p\big(X;U_{2}\big|Y_2,U_{12}\big) \Big\}\ .\end{aligned}$$ One may invoke the Support Lemma [@Csiszar-1981-B Pg. 310] to show that imposing the cardinality constraints $|{\mathscr{U}}_{\{1,2\}}| \leq |{\mathscr{X}}| + 5$, $|{\mathscr{U}}_{\{1\}}| \leq |{\mathscr{X}}|\ |{\mathscr{U}}_{{\{1,2\}}}| + 1$, and $|{\mathscr{U}}_{\{2\}}| \leq |{\mathscr{X}}|\ |{\mathscr{U}}_{{\{1,2\}}}| + 1$, does not alter the minimization in . It can be shown, see Theorem \[Sec:4:Thm:HB\], that $R_0(d_1,d_2) \geq R(d_1,d_2)$. $\square$ \[Sec:2:Counterexample\] If $t = 3$ and $|{\mathscr{Y}}_{1}|$ $=$ $|{\mathscr{Y}}_{2}|$ $=$ $|{\mathscr{Y}}_{3}|$ $=$ $1$, then reduces to $$\begin{aligned} \notag R_0(d_1,d_2,d_3) = \min_{p \in {\mathscr{P}}(d_1,d_2,d_3)} \Big\{ & I_p\big(X;U_{123}\big) + I_p\big(X;U_{12}\big|U_{123}\big) + I_p\big(X;U_{13}\big|U_{123}\big)\\ \notag &\quad + I_p\big(X;U_{23}\big|U_{123}\big) + I_p\big(X;U_{1}\big|U_{12},U_{13},U_{123}\big)\\ \label{Sec:2:Eqn:HB-CE-1} &\quad + I_p\big(X;U_{2}\big|U_{12},U_{23},U_{123}\big) + I_p\big(X;U_{3}\big|U_{13},U_{23},U_{123}\big)\Big\}\ .\end{aligned}$$ Suppose that ${\mathscr{X}}$ $=$ $\Hat{{\mathscr{X}}}_1$ $=$ $\Hat{{\mathscr{X}}}_2$ $=$ $\Hat{{\mathscr{X}}}_3$ $=$ $\{0,1,2\}$, and let $X$ be uniform on ${\mathscr{X}}$. Finally, set $$\label{Sec:3:Eqn:Hamming} \delta_l(x,\Hat{x}) = \left\{ \begin{array}{ll} 0, & \hbox{if } x = \Hat{x} \\ 1, & \hbox{otherwise,} \end{array} \right.$$ for $l = 1,2,3$ and require that $d_1 = d_2 = d_3 = 0$. We now choose the following auxiliary random variables. Set $${\mathscr{U}}_{\{1,2\}} = {\mathscr{U}}_{\{1,3\}} = {\mathscr{U}}_{\{2,3\}} = \{0,1,2\}\ ,\ \text{and}$$ $$\label{Sec:2:BE:Assumption-1} \big|{\mathscr{U}}_{\{1\}}\big| = \big|{\mathscr{U}}_{\{2\}}\big| = \big|{\mathscr{U}}_{\{3\}}\big| = \big|{\mathscr{U}}_{\{1,2,3\}}\big| = 1\ .$$ Let $C$ be independent of $X$ and uniform on $\{0,1,2\}$. Using modulo-3 arithmetic, choose $$\label{Sec:2:BE:Assumption-2} {U}_{12} = C, \quad {U}_{13} = X + C,\ \text{ and}\quad {U}_{23} = X + 2C\ .$$ Note, $X$ can be written as a function of any pair of $U_{12}$, $U_{13}$ and $U_{23}$, and the Markov chain $(U_1$, $U_2$, $U_3$, $U_{12}$, $U_{13}$, $U_{23}$, $U_{123})$ $\minuso$ $X$ $\minuso$ $(Y_1$, $Y_2$, $Y_3)$ is trivially satisfied. It follows that these auxiliary random variables are defined by some $p' \in {\mathscr{P}}(0,0,0)$. From , it follows that  is bound from above by $$\label{Sec:2:BE:rate-2} R_{0}(0,0,0) \leq I_{p'}\left(X; {U}_{12}\right) + I_{p'}\left(X; {U}_{13}\right) + I_{p'}\left(X;{U}_{23}\right)\ .$$ Furthermore, every mutual information term on the right hand side of  is zero from . Since $R_0(0,0,0)$ is non-negative, it follows that $R_0(0,0,0) = 0$; however, from Proposition \[Sec:2:Pro:Slepian-Wolf\] we have that $R(0,0,0) = H(X) > 0$. This counterexample demonstrates that $R_{0}({\mathbf{d}})$ is not an upper bound for $R({\mathbf{d}})$. $\square$ It appears that this counterexample does not invalidate any results in the rate-distortion literature. In particular, those papers that cite [@Heegard-Nov-1985-A Thm. 3] are either concerned with the special case of $2$ decoders or stochastically degraded side-information. See, for example,  [@Kaspi-Nov-1994-A; @Steinberg-Aug-2004-A; @Tian-Oct-2006-C; @Tian-Dec-2008-A; @Tian-Aug-2007-A]. The case of stochastically degraded side-information is discussed in the next section. When $t = 3$, we can force  to become an upper bound for $R(d_1,d_2,d_3)$ by modifying the set ${\mathscr{P}}(d_1,d_2,d_3)$ on which the minimization takes place. Namely, if we define $$\label{Sec:2:Eqn:Markov-Constraints-2} {\mathscr{P}}^*(d_1,d_2,d_3) \triangleq \Bigg\{ p \in {\mathscr{P}}(d_1,d_2,d_3) : \begin{array}{l} U_{13} \minuso (X,U_{123}) \minuso U_{12}\ [p] \\ U_{23} \minuso (X,U_{123}) \minuso (U_{12},U_{13})\ [p] \end{array}\Bigg\}\ ,$$ then it can be shown that $$R^*_0(d_1,d_2,d_3) \triangleq \min_{p \in {\mathscr{P}}^*(d_1,d_2,d_3)} \sum_{j=1}^{7} \max_{l\in{\mathscr{S}}_j} I_p\big(X;U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j},Y_l\big)$$ is an upper bound for $R(d_1,d_2,d_3)$. The additional Markov chains in  are sufficient to verify, via classical random coding techniques, the admissibility of rates approaching $R_{0}^*(d_1,d_2,d_3)$ from above. In general, this approach can be extended to $t \geq 3$ decoders by carefully choosing appropriate Markov chains for each of the $(2^t-1)$-auxiliary random variables[^7]. For example, if $U_{{\mathscr{S}}_j}$ is chosen to be degenerate (constant) whenever ${\mathscr{S}}_j$ is not of the form $[l,t]$ for some $l \in [t]$, then one obtains appropriate Markov chains and a valid upper bound for $R({\mathbf{d}})$. In fact, this particular choice of auxiliary random variables is optimal when the side-information is stochastically degraded. Rate-Distortion with Degraded Side-Information {#Sec:2:C} ---------------------------------------------- The side-information, as defined by $q$, is said to be *degraded* if $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1\ [q]$ forms a Markov chain. The side-information $q$ is said to be *stochastically degraded* if there exists a pmf $q'$ on ${\mathscr{X}} \times {\mathscr{Y}}^*$ where $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1\ [q']$ forms a Markov chain and $q(x,y_l) = q'(x,y_l)$ for every $(x,y_l) \in {\mathscr{X}} \times {\mathscr{Y}}_l$ and $l \in [t]$. If ${\mathscr{R}}({\mathbf{d}})[q]$ and ${\mathscr{R}}({\mathbf{d}})[q']$ are the respective ${\mathbf{d}}$-admissible rate regions for $q$ and $q'$, then this condition and Proposition \[Sec:2:Pro:Marginal-Property\] ensures that ${\mathscr{R}}({\mathbf{d}})[q] = {\mathscr{R}}({\mathbf{d}})[q']$. Thus, it is sufficient to consider degraded side-information. When the side-information is degraded, $R({\mathbf{d}})$ can be characterised using $t$ auxiliary random variables. These variables are $U_{[1,t]}$, $U_{[2,t]}$, $\ldots$, $U_{\{t\}}$, and the corresponding subsets of decoders are $[1,t]$, $[2,t]$, $\ldots$, $\{t\}$. To formally define these variables using the notation of Section \[Sec:2:B\], choose $|{\mathscr{U}}_{{\mathscr{S}}_j}|=1$ whenever ${\mathscr{S}}_j \neq [l,t]$ for some $l \in [t]$, and let ${\mathscr{P}}_{deg}({\mathbf{d}})$ denote the resultant set of $p \in {\mathscr{P}}$ that satisfy properties **(P1)** and **(P2)**. \[Sec:2:Pro:HB-Lower-Bound-Degraded\] If $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1\ [q]$ forms a Markov chain, then $$\label{Sec:2:Eqn:HB-Degraded-1} R({\mathbf{d}}) = \min_{p \in {\mathscr{P}}_{deg}({\mathbf{d}})} \sum_{l=1}^t I_p\big(X;U_{[l,t]}\big|Y_l,U_{[1,t]},U_{[2,t]},\ldots,U_{[l-1,t]}\big)\ ,$$ where the cardinality of each set ${\mathscr{U}}_{[l,t]}$ is bound by $$\big|{\mathscr{U}}_{[l,t]}\big| \leq \big|{\mathscr{X}}\big| \prod_{l'=1}^{l-1} \big|{\mathscr{U}}_{[l',t]}\big| - 1 + t - l + \frac{(t-l+1)(t-l+2)}{2} \ .$$ The converse theorem for this result can be found on [@Heegard-Nov-1985-A Pgs. 733-734]. Note, however, that the use of $R_0({\mathbf{d}})$ in [@Heegard-Nov-1985-A Thm. 3] is incorrect. For example, the side-information used in Example \[Sec:2:Exa:Counterexample\] is trivially degraded. Finally, we note that the Markov chain $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1\ [q]$ appears to be essential for the converse theorem [@Heegard-Nov-1985-A Pgs. 733-734]. In contrast, the coding theorem that proves the admissibility of rates approaching  is less dependent on this assumption. Indeed, this Markov chain can be disregarded provided there is an appropriate increase in rate. For example, the functional $$\min_{p\in{\mathscr{P}}_{deg}({\mathbf{d}})} \sum_{l=1}^t \max_{l' \in [l,t]} I_p\big(X;U_{[l,t]}\big|Y_{l'},U_{[1,t]},\ldots,U_{[l-1,t]}\big)$$ is an upper bound for $R({\mathbf{d}})$. We will extend this idea in the next section to give an inner bound for ${\mathscr{R}}({\mathbf{d}})$. Main Results for ${\mathscr{R}}({\mathbf{d}})$ {#Sec:3} ============================================== An Inner Bound for ${\mathscr{R}}({\mathbf{d}})$ {#Sec:3:A} ------------------------------------------------ We now present a new inner bound for ${\mathscr{R}}({\mathbf{d}})$. This bound will require an auxiliary random variable for each non-empty subset of decoders. For this purpose, arrange the non-empty subsets of $[t]$ into an ordered list $v = {\mathscr{S}}_1,{\mathscr{S}}_2,\ldots,{\mathscr{S}}_{2^t-1}$ with decreasing cardinality. That is, $|{\mathscr{S}}_j| \geq |{\mathscr{S}}_k|$ whenever $j \leq k$. Let ${\mathscr{V}}$ denote the set of all such lists. Fix $v \in {\mathscr{V}}$. Let ${\mathscr{U}}_{{\mathscr{S}}_1}$, ${\mathscr{U}}_{{\mathscr{S}}_2}$, $\ldots$, ${\mathscr{U}}_{{\mathscr{S}}_{2^t-1}}$ be finite alphabets and define ${\mathscr{U}}_v^* \triangleq {\mathscr{U}}_{{\mathscr{S}}_1}$ $\times$ ${\mathscr{U}}_{{\mathscr{S}}_2}$ $\times$ $\cdots$ $\times$ ${\mathscr{U}}_{{\mathscr{S}}_{2^t-1}}$. Let ${\mathscr{P}}_v$ denote the set of all distributions on ${\mathscr{U}}^*_v \times {\mathscr{X}} \times {\mathscr{Y}}^*$ whose $({\mathscr{X}} \times {\mathscr{Y}}^*)$-marginal is equal to $q$; that is, $p\big(x,y_1,\ldots,y_y\big) = q\big(x,y_1,\ldots,y_y\big)$. As before, each $p \in {\mathscr{P}}_v$ specifies a joint distribution for $(2^t-1)$-auxiliary random variables. We denote these variables by $U_{{\mathscr{S}}_j}$, $j=1,2,\ldots,2^t-1$, where $U_{{\mathscr{S}}_j}$ takes values from ${\mathscr{U}}_{{\mathscr{S}}_j}$. Let ${\mathscr{A}} \triangleq \{U_{{\mathscr{S}}_1},\ U_{{\mathscr{S}}_2},\ \ldots, U_{{\mathscr{S}}_{2^t-1}}\}$, and define $$\begin{aligned} {\mathscr{A}}^{-}_{{\mathscr{S}}_j} &\triangleq \Big\{U_{{\mathscr{S}}_i} \in {\mathscr{A}}\ :\ i < j,\ {\mathscr{S}}_i \nsupseteq {\mathscr{S}}_j \Big\} \text{ and}\\ {\mathscr{A}}^{\supset}_{{\mathscr{S}}_j} &\triangleq \Big\{ U_{{\mathscr{S}}_i} \in {\mathscr{A}}\ :\ {\mathscr{S}}_i \supset {\mathscr{S}}_j \Big\}\ .\end{aligned}$$ We note that the union of ${\mathscr{A}}^{-}_{{\mathscr{S}}_j}$ and ${\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}$ is the set of all those auxiliary random variables associated with subsets that appear before ${\mathscr{S}}_j$ in $v$. Let us further define $$\begin{aligned} {\mathscr{A}}^{+}_{{\mathscr{S}}_j} &\triangleq \Big\{U_{{\mathscr{S}}_k} \in {\mathscr{A}}\ :\ k > j,\ {\mathscr{S}}_k \cap {\mathscr{S}}_j \neq \emptyset \Big\} \ , \\ {\mathscr{A}}^{\dag}_{{\mathscr{S}}_j} &\triangleq \left\{ U_{{\mathscr{S}}_i} \in {\mathscr{A}}^-_{{\mathscr{S}}_j} : \begin{array}{ll} \exists U_{{\mathscr{S}}_k} \in {\mathscr{A}}^+_{{\mathscr{S}}_j}, \\ {\mathscr{S}}_i \cap {\mathscr{S}}_k \neq \emptyset \end{array}\right\}\ \text{ and}\\ {\mathscr{A}}^\ddag_{{\mathscr{S}}_j,l} &\triangleq \Big\{U_{{\mathscr{S}}_i} \in {\mathscr{A}}^\dag_{{\mathscr{S}}_j}\ :\ {\mathscr{S}}_i\ni l \Big\}\ \text{ when } l \in {\mathscr{S}}_j\ .\end{aligned}$$ Finally, let ${\mathscr{P}}_{v}({\mathbf{d}})$ denote the set of all $p \in {\mathscr{P}}_v$ satisfying properties $(\textbf{P1})$ and $(\textbf{P2})$ from Section . Our inner bound for ${\mathscr{R}}({\mathbf{d}})$ will be built using the following functional. For each subset ${\mathscr{S}}_j \subseteq [t]$ and $l \in [t]$ such that ${\mathscr{S}}_j\cap[l]\neq\emptyset$, let $$\label{Sec:3:Eqn:Phi} \Phi_p\big({\mathscr{S}}_j,l\big) \triangleq I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big) -\min_{l'\in{\mathscr{S}}_j \cap [l]}I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,l'},Y_{l'}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big)\ .$$ Finally, for each $p \in {\mathscr{P}}_{v}({\mathbf{d}})$, define[^8] $${\mathscr{R}}_{p,v}({\mathbf{d}}) \triangleq \left\{ {\mathbf{r}} \in \mathbb{R}_+^t\ :\sum_{i=1}^l r_i \geq \sum_{\tiny \begin{array}{c} {\mathscr{S}}_j \subseteq [t], \\ {\mathscr{S}}_j \cap [l] \neq \emptyset \end{array} } \Phi_p\big({\mathscr{S}}_j,l\big),\ \forall l \in [t]\right\}\ ,$$ and let $${\mathscr{R}}_{in}({\mathbf{d}}) \triangleq \text{co} \left( \bigcup_{v\in{\mathscr{V}}} \bigcup_{p\in{\mathscr{P}}_v({\mathbf{d}})} {\mathscr{R}}_{p,v}({\mathbf{d}})\right)\ ,$$ where $\text{co}(\cdot)$ denotes the closure of the convex hull. \[Sec:3:Thm:Successive-Refinement\] If ${\mathbf{d}} \in \mathbb{R}_+^t$, then every rate tuple within ${\mathscr{R}}_{in}({\mathbf{d}})$ is ${\mathbf{d}}$-admissible; that is, $${\mathscr{R}}_{in}({\mathbf{d}}) \subseteq {\mathscr{R}}({\mathbf{d}})\ .$$ Our proof of this result is given in Appendix \[App:1\]. Stochastically Degraded Side-Information {#Sec:3:B} ---------------------------------------- Assuming that the side-information is stochastically degraded, Tian and Diggavi gave a single-letter characterisation of ${\mathscr{R}}({\mathbf{d}})$ in [@Tian-Aug-2007-A Thm. 1] (see Proposition \[Sec:1:Pro:Degraded-Side-Information\]). We now show that the forward (coding) part of this result can be obtained as a special case of Theorem \[Sec:3:Thm:Successive-Refinement\]. We can assume that $X$ $\minuso$ $Y_t$ $\minuso$ $Y_{t-1}$ $\minuso$ $\cdots$ $\minuso$ $Y_1$ $[q]$ forms a Markov chain. Recall ${\mathscr{P}}_{deg}({\mathbf{d}})$ from Section \[Sec:2:C\]. Each $p \in {\mathscr{P}}_{deg}$ specifies a joint distribution for $t$ non-degenerate auxiliary random variables. These variables are $U_{[1,t]}$, $U_{[2,t]}$, $\ldots$, $U_{\{t\}}$ and the associated subsets are $[1,t]$, $[2,t]$, $\ldots$, $\{t\}$, respectively. We can ignore the degenerate random variables in ${\mathscr{A}}$, so that for all $l \in [1,t]$ we have $$\label{Sec:3:Eqn:Degraded-SI-ARV-Supset} {\mathscr{A}}^\supset_{[l,t]} = \Big\{U_{[1,t]},U_{[2,t]},\ldots,U_{[l-1,t]}\Big\}\ ,$$ $$\label{Sec:3:Eqn:Degraded-SI-ARV-Dag} {\mathscr{A}}^\dag_{[l,t]} = \emptyset\quad \text{ and}$$ $$\label{Sec:3:Eqn:Degraded-SI-ARV-DDag} {\mathscr{A}}^\ddag_{[l,t],l'} = \emptyset\quad \forall l'\in[l,t].$$ On combining the Markov chain $(U_{[1,t]}$, $U_{[2,t]},$ $\ldots,$ $U_{\{t\}})$ $\minuso$ $X$ $\minuso$ $(Y_1,$ $Y_2,$ $\ldots,$ $Y_t)$ $[p]$ with the Markov chain $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1$ $[p]$, we obtain the following Markov chains: $$\label{Sec:3:Eqn:MC-1} U_{[l,t]} \minuso \Big({\mathscr{A}}^{\supset}_{[l,t]},Y_{l'}\Big) \minuso Y_l\ [p],\quad \forall\ l' \in [l,t]\ .$$ On substituting , and  into , we obtain $$\label{Sec:3:Eqn:Phi-Degraded} \Phi_p([l,t],j) = I_p\Big(X;U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big) - \min_{l' \in [l,j]} I_p\Big(U_{[l,t]};Y_{l'}\Big|{\mathscr{A}}^\supset_{[l,t]} \Big)\ .$$ The second term on the right hand side of  can be rewritten as $$\begin{aligned} \notag \min_{l' \in [l,j]} I_p\Big(U_{[l,t]};Y_{l'}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big) &=H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big)- \max_{l' \in [l,j]} H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]},Y_{l'}\Big)\\ \label{Sec:3:Eqn:Phi-Degraded-2} &=H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big)- \max_{l' \in [l,j]} H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]},Y_{l'},Y_l\Big)\\ \label{Sec:3:Eqn:Phi-Degraded-3} &=H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big)- H\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]},Y_l\Big)\\ \label{Sec:3:Eqn:Phi-Degraded-4} &=I_p\Big(U_{[l,t]};Y_l\Big|{\mathscr{A}}^\supset_{[l,t]}\Big)\ ,\end{aligned}$$ where  follows from the Markov chain , and  follows since $$H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]},Y_l\Big) \geq H_p\Big(U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]},Y_{l'},Y_l\Big)\ , \quad \forall l' \in [l,j]\ .$$ On combining  and , we get $$\begin{aligned} \notag \Phi_p([l,t],j) &=I_p\Big(X;U_{[l,t]}\Big|{\mathscr{A}}^\supset_{[l,t]}\Big) - I_p\Big(U_{[l,t]};Y_{l}\Big|{\mathscr{A}}^\supset_{[l,t]} \Big)\\ \label{Sec:3:Eqn:Phi-Degraded-5} &= I_p\Big(X,{\mathscr{A}}^\supset_{[l,t]};U_{[l,t]}\Big) - I_p\Big(U_{[l,t]};Y_{l},{\mathscr{A}}^\supset_{[l,t]}\Big).\end{aligned}$$ From  and since $U_{[l,t]} \minuso (X,{\mathscr{A}}^{\supset}_{[l,t]}) \minuso Y_l\ [p]$ forms a Markov chain, further simplifies to $$\label{Sec:3:Eqn:Phi-Degraded-6} \Phi_p([l,t],j) = I_p\Big(X;U_{[l,t]}\Big|U_{[1,t]},U_{[2,t]},\ldots,U_{[l-1,t]},Y_l\Big)\ .$$ Finally, substituting  into the definition of ${\mathscr{R}}_{p,v}({\mathbf{d}})$ proves the ${\mathbf{d}}$-admissibility of every rate tuple ${\mathbf{r}} \in \mathbb{R}_+^t$ for which there exists some $p \in {\mathscr{P}}_{deg}({\mathbf{d}})$ with $$\sum_{i=1}^j r_i \geq \sum_{l=1}^j I_p\Big(X;U_{[l,t]}\Big|U_{[1,t]},U_{[2,t]},\ldots,U_{[l-1,t]},Y_l\Big)\ ,$$ for $j = 1,2,\ldots,t$. Side-Information Scalable Source Coding {#Sec:3:C} --------------------------------------- If $t=2$ and the side-information is degraded ($X \minuso Y_2 \minuso Y_1 [q]$), then an optimal compression strategy should satisfy the distortion constrains of decoder $2$ after the distortion constraints of decoder $1$ have been satisfied. See, for example, Section \[Sec:2:C\]. However, this ordering may not be optimal when the side-information is not degraded. This observation led Tian and Diggavi in [@Tian-Dec-2008-A Thm. 1] (see Proposition \[Sec:1:Pro:SI-Scalable\]) to propose and study the side-information scalable source coding problem. In the context of this paper, this problem is a special case of the successive-refinement problem where $X \minuso Y_1 \minuso Y_2$ $[q]$ is assumed to form a Markov chain. We now show that this result can be obtained as a special case of Theorem \[Sec:3:Thm:Successive-Refinement\]. Choose the list $v$ as follows: ${{\mathscr{S}}}_1 = \{1,2\}$, ${{\mathscr{S}}}_2 = \{1\}$ and ${{\mathscr{S}}}_3 = \{2\}$. For each $p \in {\mathscr{P}}_v(d_1,d_2)$, we have the chains $X \minuso Y_1 \minuso Y_2$ $[p]$ and $(U_{12},U_1,U_2) \minuso X \minuso (Y_1,Y_2)$, therefore simplifies to $$\begin{aligned} \Phi_p(\{1,2\},1) &= I_p(X;U_{12}) - I_p(U_{12};Y_1)\\ &= I_p(X;U_{12}|Y_1)\\ \Phi_p(\{1,2\},2) &= I_p(X;U_{12}) - \min_{l'\in\{1,2\}} I_p(U_{12};Y_{l'})\\ &= I_p(X;U_{12}) - I_p(U_{12};Y_{2})\\ &= I_p(X;U_{12}|Y_2)\\ \Phi_p(\{1\},1) &= I_p(X;U_1|U_{12}) - I_p(U_1;Y_1|U_{12})\\ &= I_p(X;U_1|U_{12},Y_1)\\ \Phi_p(\{1\},2) &= I_p(X;U_1|U_{12},Y_1)\\ \Phi_p(\{2\},2) &= I_p(X;U_2|U_{12}) - I_p(U_2;Y_2|U_{12})\\ &= I_p(X;U_2|U_{12},Y_2)\ .\end{aligned}$$ On substituting these equalities into the definition of ${\mathscr{R}}_{v,p}(d_1,d_2)$, it can been seen from Theorem \[Sec:3:Thm:Successive-Refinement\] that any rate pair $(r_1,r_2)$ satisfying $$\begin{aligned} r_1 &\geq \Phi_p(\{1,2\},1) + \Phi_p(\{1\},1)\\ &= I_p(X;U_{12}|Y_1) + I_p(X;U_1|U_{12},Y_1)\\ &= I_p(X;U_1,U_{12}|Y_1)\ ,\end{aligned}$$ and $$\begin{aligned} r_1 + r_2 &\geq \Phi_p(\{1,2\},2) + \Phi_p(\{1\},2) + \Phi_p(\{2\},2)\\ &= I_p(X;U_{12}|Y_2) + I_p(X;U_1|U_{12},Y_1) + I_p(X;U_2|U_{12},Y_2) \\ &= I_p(X;U_2,U_{12}|Y_2) + I_p(X;U_1|U_{12},Y_1)\end{aligned}$$ for some $p \in {\mathscr{P}}_v(d_1,d_2)$ is ${\mathbf{d}}$-admissible. This condition matches the desired inner bound [@Tian-Dec-2008-A Thm. 1] (Proposition \[Sec:1:Pro:SI-Scalable\]). Main Results for the Wyner-Ziv Problem with $t$-Decoders {#Sec:4} ======================================================== An Upper Bound for $R({\mathbf{d}})$ ------------------------------------ Recall Figure \[Sec:1:Fig:mKHB\] and the rate-distortion function $R({\mathbf{d}})$. \[Sec:4:Thm:HB\] $$\label{Sec:4:Eqn:HB-Upper-Bound} R({\mathbf{d}}) \leq \min_{ \begin{tiny} \begin{array}{c} v\in{\mathscr{V}},\\ p \in {\mathscr{P}}_v({\mathbf{d}}) \end{array} \end{tiny} } \sum_{j =1}^{2^t-1} \left[I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big) - \min_{l' \in {\mathscr{S}}_j} I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,l'},Y_{l'}|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big)\right]\ .$$ We note the following special cases where this upper bound known to be tight. For one decoder, the right hand side of  gives the Wyner-Ziv formula . For $t$-decoders and degraded side-information, the right hand side of  is equal to the right hand side of . (Set $|{\mathscr{U}}_{{\mathscr{S}}_j}|=1$ whenever ${\mathscr{S}}_j \neq [l,t]$ for some $l \in [t]$, and following the reasoning given in Section \[Sec:3:B\].) In fact, this upper bound is tight whenever $X$ $\minuso$ $Y_{\alpha_1}$ $\minuso$ $Y_{\alpha_2}$ $\minuso$ $\cdots$ $\minuso$ $Y_{\alpha_t}$, where $\alpha_l$, $l = 1,2,\ldots,t$ each take unique values from $[t]$ (see Remark \[Sec:5:Rem:DeterministicDistortionMeasures\]). Most importantly, however, this bound avoids those problems suffered by $R_0({\mathbf{d}})$ in Example \[Sec:2:Counterexample\]. Proof of Theorem \[Sec:4:Thm:HB\] --------------------------------- The following lemma will be useful for the proof of Theorem \[Sec:4:Thm:HB\]. \[Sec:4:Lem:Phi\] Suppose $p \in {\mathscr{P}}_v({\mathbf{d}})$, and recall the functional $\Phi_p({\mathscr{S}}_j,l)$ defined in . For every ${\mathscr{S}}_j \subseteq [t]$ and $l,l' \in [t]$ such that ${{\mathscr{S}}}_j \cap [l] \neq \emptyset$ and ${{\mathscr{S}}}_j \cap [l'] \neq \emptyset$, we have: 1. $\Phi_p({\mathscr{S}}_j,l) \leq \Phi_p({\mathscr{S}}_j,l')$ when $l' > l$, and 2. $\Phi_p({\mathscr{S}}_j,l) \geq 0$. The fact that $\Phi_p({\mathscr{S}}_j,l) \leq \Phi_p({\mathscr{S}}_j,l')$ follows because ${\mathscr{S}}_j \cap [l] \subseteq {\mathscr{S}}_j \cap [l']$. To see that $\Phi_p({\mathscr{S}}_j,l) \geq 0$, consider the following. Let $$\tilde{l} \triangleq \underset{i \in {\mathscr{S}}_j \cap [l]}{\operatorname{argmin}}\ I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,i},Y_i\big|{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}\big)\ ,$$ then $$\begin{aligned} \notag \Phi_p\big({\mathscr{S}}_j,l\big) &\equiv I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}\big) -I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,\tilde{l}},Y_{\tilde{l}}\big|{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}\big)\\ \label{Sec:3:Eqn:Phi-1} &= I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j},Y_{\tilde{l}};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}\big) -I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,\tilde{l}},Y_{\tilde{l}}\big|{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j}\big)\\ \label{Sec:3:Eqn:Phi-2} &=I_p\big(X,{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j},{\mathscr{A}}^\dag_{{\mathscr{S}}_j},Y_{\tilde{l}};U_{{\mathscr{S}}_j}\big) -I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j},{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,\tilde{l}},Y_{\tilde{l}}\big)\\ \label{Sec:3:Eqn:Phi-3} &\geq 0\ ,\end{aligned}$$ where  follows because $Y_{\tilde{l}} \minuso (X,{\mathscr{A}}^{\supset}_{{\mathscr{S}}_j},{\mathscr{A}}^{\dag}_{{\mathscr{S}}_j}) \minuso U_{{\mathscr{S}}_j}$ $[p]$ forms a Markov chain, follows from the chain rule for mutual information, and  follows from ${\mathscr{A}}^\dag_{{\mathscr{S}}_j} \supseteq {\mathscr{A}}^\ddag_{{\mathscr{S}}_j,\tilde{l}}$. We now prove Theorem \[Sec:4:Thm:HB\]. First, note that the minimum on the right hand side of  exists. Suppose that $v$ and $p$ achieve this minimum, and choose any $r \in {\mathbb{R}}_+$ such that $$\label{Sec:4:Eqn:HB-Bound-Proof-3} r \geq \sum_{j =1}^{2^t-1} \left[I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big) - \min_{l' \in {\mathscr{S}}_j} I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,l'},Y_{l'}|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big)\right]\ .$$ In the following, we prove the ${\mathbf{d}}$-admissibility of $r$ using Theorem \[Sec:3:Thm:Successive-Refinement\]. Consider the successive refinement problem shown in Figure \[Sec:1:Fig:tSR\], the corresponding ${\mathbf{d}}$-admissible rate region ${\mathscr{R}}({\mathbf{d}})$ (defined in Section \[Sec:2:A\]), and the inner bound ${\mathscr{R}}_{in}({\mathbf{d}})$ given in Theorem \[Sec:3:Thm:Successive-Refinement\]. In particular, consider the region ${\mathscr{R}}_{p,v}({\mathbf{d}})$, where $v$ and $p$ achieve the aforementioned minimum. Define the $t$-tuple $\tilde{{\mathbf{r}}} \triangleq (r,0,0,\ldots,0)$. It is clear that $r \geq R({\mathbf{d}})$ iff $\tilde{{\mathbf{r}}}\in {\mathscr{R}}({\mathbf{d}})$, therefore the result will follow if it can be shown that $\tilde{{\mathbf{r}}}\in {\mathscr{R}}_{p,v}({\mathbf{d}})$. For every $l \in [t]$, we have $$\begin{aligned} \notag \sum_{i=1}^l \tilde{r}_i &\geq \sum_{j =1}^{2^t-1} \left[I_p\big(X,{\mathscr{A}}^\dag_{{\mathscr{S}}_j};U_{{\mathscr{S}}_j}\big|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big) - \min_{l' \in {\mathscr{S}}_j} I_p\big(U_{{\mathscr{S}}_j};{\mathscr{A}}^\ddag_{{\mathscr{S}}_j,l'},Y_{l'}|{\mathscr{A}}^\supset_{{\mathscr{S}}_j}\big)\right]\\ \label{Sec:4:Eqn:HB-Bound-Proof-1a} &= \sum_{{{\mathscr{S}}}_j \subseteq [t]}\Phi_p\big({\mathscr{S}}_j,t\big)\\ \label{Sec:4:Eqn:HB-Bound-Proof-1} &\geq \sum_{\tiny \begin{array}{c} {\mathscr{S}}_j \subseteq [t], \\ {\mathscr{S}}_j \cap [l] \neq \emptyset \end{array} } \Phi_p\big({\mathscr{S}}_j,t\big)\end{aligned}$$ $$\begin{aligned} \label{Sec:4:Eqn:HB-Bound-Proof-2} &\geq \sum_{\tiny \begin{array}{c} {\mathscr{S}}_j \subseteq [t], \\ {\mathscr{S}}_j \cap [l] \neq \emptyset \end{array} } \Phi_p\big({\mathscr{S}}_j,l\big)\ ,\end{aligned}$$ where  follows from , and Lemma \[Sec:4:Lem:Phi\] gives  and . From Theorem \[Sec:3:Thm:Successive-Refinement\] we have that $\tilde{{\mathbf{r}}} \in {\mathscr{R}}_{v,p}({\mathbf{d}})$ and $\tilde{{\mathbf{r}}} \in {\mathscr{R}}({\mathbf{d}})$, therefore $r \geq R({\mathbf{d}})$. $\square$ Theorem \[Sec:4:Thm:HB\] is a consequence of the inner bound ${\mathscr{R}}_{in}({\mathbf{d}})$ given in Theorem \[Sec:3:Thm:Successive-Refinement\]. Like ${\mathscr{R}}({\mathbf{d}})$, ${\mathscr{R}}_{in}({\mathbf{d}})$ depends on the successive-refinement decoding order: if we interchange the decoders (keeping the same side-information and distortion constraints at each decoder), then the resulting inner bound ${\mathscr{R}}_{in}({\mathbf{d}})$ will change. One might, therefore, be inspired to pursue a stronger version of Theorem \[Sec:4:Thm:HB\] wherein the choice of successive-refinement order is optimized. Note, however, that the proof of Theorem \[Sec:4:Thm:HB\] requires only the bound for $r_1 + r_2 + \cdots + r_t$ in ${\mathscr{R}}_{p,v}({\mathbf{d}})$, and this bound is independent of the successive-refinement decoding order. Lossless Source Coding with Private Messages {#Sec:5} ============================================ In Proposition \[Sec:2:Pro:Slepian-Wolf\], we reviewed a broadcast problem wherein ${\mathbf{X}}$ is reconstructed losslessly at every decoder. This lossless problem can be easily solved as a variant of existing work by Slepian and Wolf [@Slepian-Jul-1973-A Thm. 2]; Sgarro [@Sgarro-Mar-1977-A Thm. 2]; or Bakshi and Effros [@Bakshi-Jul-2008-C Thm. 1]. In this section, we consider a more complex scenario wherein each decoder is required to decode one part of ${\mathbf{X}}$ losslessly. ![Lossless Source Coding with Private Messages. The encoder compresses $({\mathbf{W}}_1,{\mathbf{W}}_2,\ldots,{\mathbf{W}}_t)$ to $(M_1,M_2,\ldots,M_t)$. It is required that decoder $l$ uses $M_1$ through $M_l$ together with ${\mathbf{Y}}_l$ to produce a lossless replica $\Hat{{\mathbf{W}}_l}$ of ${\mathbf{W}}_l$. In Theorem \[Sec:4:Thm:Lossless\], we give an explicit characterisation of the ${\mathscr{R}}(0,0,\ldots,0)$ for degraded side-information $(W_1,$ $W_2,$ $\ldots,$ $W_t)$ $\minuso$ $Y_t$ $\minuso$ $Y_{t-1}$ $\minuso$ $\cdots$ $\minuso$ $Y_1$.[]{data-label="Sec:4:Fig:Lossless"}](Lossless){width="0.75\columnwidth"} Let ${\mathscr{W}}_1$, ${\mathscr{W}}_2$, $\ldots$, ${\mathscr{W}}_t$ be finite alphabets, and consider the problem shown Figure \[Sec:4:Fig:Lossless\]. In the nomenclature of previous sections, set ${\mathscr{X}}$ $\triangleq$ ${\mathscr{W}}_1$ $\times$ ${\mathscr{W}}_2$ $\times$ $\cdots$ $\times$ ${\mathscr{W}}_t$, $X \triangleq (W_1,$ $W_2,$ $\ldots,$ $W_t)$, and let $({\mathbf{W}}_1$, ${\mathbf{W}}_2$, $\ldots$, ${\mathbf{W}}_t$, ${\mathbf{Y}}_1$, ${\mathbf{W}}_2$, $\ldots$, ${\mathbf{W}}_t)$ be drawn iid according to $q(w_1,w_2,\ldots,w_t,y_1,y_2,\ldots,y_t)$. It is required that decoder $l$ reconstructs ${\mathbf{W}}_l$ with vanishing probability of symbol error. To this end, set $\Hat{{\mathscr{X}}}_l \triangleq {\mathscr{W}}_l$ and define the average symbol error probability at decoder $l$ to be $$\begin{aligned} P^l_{e} &\triangleq \frac{1}{n} \sum_{i=1}^n P^l_{e,i}\end{aligned}$$ where $P^l_{e,i} \triangleq \mathbb{E} [\delta_l(W_{1,i},W_{2,i},\ldots,W_{t,i},\Hat{W}_{l,i})]$, $$\label{Sec:5:Eqn:Distortion-Measure} \delta_l\big(w_1,w_2,\ldots,w_t,\Hat{w}_l\big) \triangleq \left\{ \begin{array}{ll} 0, & \hbox{ if } w_l = \Hat{w}_l \\ 1, & \hbox{ otherwise,} \end{array} \right.$$ defines the probability of error for the $i^{\text{th}}$-symbol. A computable characterisation of ${\mathscr{R}}(0,0,\ldots,0)$ has yet to be found. A direct application of Theorem \[Sec:3:Thm:Successive-Refinement\] yields an inner bound for ${\mathscr{R}}(0,0,\ldots,0)$; however, it is not clear if this bound is tight. The next theorem shows that this bound is tight when the side-information is degraded. Although this result is a special case of Proposition \[Sec:1:Pro:Degraded-Side-Information\], we state it here in an explicit form – without auxiliary random variables – to highlight the generality of this problem. \[Sec:4:Thm:Lossless\] If $(W_1,$ $W_2,$ $\ldots,$ $W_t)$ $\minuso$ $Y_t$ $\minuso$ $Y_{t-1}$ $\minuso$ $\cdots$ $\minuso$ $Y_1$ $[q]$ and $\delta_l$ is given by , then $${\mathscr{R}}(0,0,\ldots,0) = \Bigg\{ {\mathbf{r}} \in {\mathbb{R}}_+^t : \sum_{k=1}^l r_k \geq \sum_{k=1}^l H\big(W_k\big|W_1,W_2,\ldots,W_{k-1},Y_k\big)\Bigg\}$$ The lossless one-channel version of Theorem \[Sec:4:Thm:Lossless\] follows immediately. \[Sec:4:Cor:Lossless\] If $(W_1,$ $W_2,$ $\ldots,$ $W_t)$ $\minuso$ $Y_t$ $\minuso$ $Y_{t-1}$ $\minuso$ $\cdots$ $\minuso$ $Y_1$ $[q]$ and $\delta_l$ is given by , then $$R(0,0,\ldots,0) = \sum_{l=1}^t H\big(W_l\big|W_1,W_2,\ldots,W_{l-1},Y_l\big)\ .$$ \[Sec:5:Rem:DeterministicDistortionMeasures\] The lossless problems considered in this section are equivalent to the concept of deterministic distortion measures [@Fu-July-2002-A; @Tian-Dec-2008-A], wherein certain functions $\{\phi_i(X)\}$ of the source $X$ are to be reconstructed with vanishing symbol error probability at the receivers. If $t = 2$, $Z_1 = \phi_1(X)$ is to be reconstructed at receiver $1$, $Z_2 = \phi_2(X)$ is to be reconstructed at receiver $2$, and the side-information is reversibly degraded (i.e. $X \minuso Y_1 \minuso Y_2$ $[q]$ forms a Markov chain), then Tian and Diggavi have shown that [@Tian-Dec-2008-A Cor. 4] $$R(0,0) = H(Z_2|Y_2) + H(Z_1|Y_1,Z_2)\ .$$ This result is consistent with Corollary \[Sec:4:Cor:Lossless\] in the following sense. The achievability of Corollary \[Sec:4:Cor:Lossless\] follows from Theorem \[Sec:4:Thm:HB\] by setting $U_{{{\mathscr{S}}}_j} = W_l$ whenever ${{\mathscr{S}}}_j = [l,t]$ for some $l \in [t]$ and $U_{{{\mathscr{S}}}_j} = \text{constant}$ otherwise. The bound in Theorem \[Sec:4:Thm:HB\] is equal to the rate-distortion function $R({\mathbf{d}})$ for every order of degraded side-information. For example, suppose that $X$ $\minuso$ $Y_{\alpha_t}$ $\minuso$ $Y_{\alpha_{t-1}}$ $\minuso$ $\cdots$ $\minuso$ $Y_{\alpha_1}$ $[q]$ forms a Markov chain, where $\alpha_l$, $l = 1,2,\ldots,t$ each take unique values from $[t]$. This markov condition is simply a relabelling of the degradedness considered in Section \[Sec:2:C\], so it is appropriate to choose the $t$ non-trivial auxiliary random variables to be $U_{[\alpha_1,\alpha_t]}$, $U_{[\alpha_2,\alpha_t]}$, $\ldots$, $U_{\{\alpha_t\}}$, where $[\alpha_i,\alpha_t] = \{\alpha_i,\alpha_{i+1},\ldots,\alpha_{t}\}$. Thus, we can set $U_{[\alpha_i,\alpha_t]} = W_{\alpha_i}$ to restate Corollary \[Sec:4:Cor:Lossless\] for an arbitrary order of degraded side-information. Tian and Diggavi also characterise the successive-refinement region ${\mathscr{R}}(0,0)$ in [@Tian-Dec-2008-A Thm. 4] for $t = 2$ and reversibly degraded side-information. This result is not captured by Theorem \[Sec:4:Thm:Lossless\], and it would be interesting to see if a similar result can be obtained for $t$-receivers and arbitrary ordering of degraded side-information. The forward (coding) part follows from by setting $U_{l} = W_l$ in Proposition \[Sec:1:Pro:Degraded-Side-Information\]. The converse theorem requires some work and is given below. For brevity, we use the following notation: $M_{\leq l} \triangleq \{M_1,M_2,\ldots,M_l\}$, ${\mathbf{W}}_{\leq l} \triangleq \{{\mathbf{W}}_1,{\mathbf{W}}_2,\ldots,{\mathbf{W}}_l\}$ and ${\mathbf{Y}}_{\leq l} \triangleq \{{\mathbf{Y}}_1,{\mathbf{Y}}_2,\ldots,{\mathbf{Y}}_l\}$. By definition, we have $$\begin{aligned} \label{Sec:4:Eqn:Con-Proof-Lossless-1} \sum_{k=1}^l \big(r_k + \epsilon\big) &\geq \frac{1}{n}\sum_{k=1}^l \log_2|{\mathscr{M}}_l| \\ &\geq \frac{1}{n} H\big(M_{\leq l}\big) \\ \label{Sec:4:Eqn:Con-Proof-Lossless-2} &\geq \frac{1}{n} I\big({\mathbf{W}}_{\leq l},{\mathbf{Y}}_{\leq l};M_{\leq l}\big)\\ \label{Sec:4:Eqn:Con-Proof-Lossless-3} &= \frac{1}{n} \sum_{k=1}^l I\big({\mathbf{W}}_k,{\mathbf{Y}}_k;M_{\leq l}\big|{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k-1}\big)\\ \label{Sec:4:Eqn:Con-Proof-Lossless-4} &\geq \frac{1}{n} \sum_{k=1}^l I\big({\mathbf{W}}_k;M_{\leq l}\big|{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k}\big)\\ \label{Sec:4:Eqn:Con-Proof-Lossless-6} &= \frac{1}{n} \sum_{k=1}^l \Big[H\big({\mathbf{W}}_k\big|{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k}\big) - H\big({\mathbf{W}}_k\big|M_{\leq l} ,{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k}\big)\Big]\\ \notag &= \frac{1}{n} \sum_{k=1}^l \sum_{i=1}^n \Big[H\big(W_{k,i}\big|{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k},W_{k,1},W_{k,2},\ldots,W_{k,i-1}\big)\\ \label{Sec:4:Eqn:Con-Proof-Lossless-7} &\qquad \qquad \qquad \qquad -H\big(W_{k,i}\big|M_{\leq l},{\mathbf{W}}_{\leq k-1},{\mathbf{Y}}_{\leq k},W_{k,1},W_{k,2},\ldots,W_{k,i-1}\big)\Big]\\ \label{Sec:4:Eqn:Con-Proof-Lossless-9} &\geq \sum_{k=1}^l H\big(W_{k}\big|W_{1},W_{2},\ldots,W_{k-1},Y_{k}\big) -\frac{1}{n} \sum_{k=1}^l \sum_{i=1}^n H\big(W_{k,i}\big|\Hat{W}_{k,i}\big)\\ \label{Sec:4:Eqn:Con-Proof-Lossless-10} &\geq \sum_{k=1}^l H\big(W_{k}\big|W_{1},W_{2},\ldots,W_{k-1},Y_{k}\big) -\frac{1}{n} \sum_{k=1}^l \sum_{i=1}^n \Big[h\big(P^k_{e,i}\big)+P^k_{e,i}\log_2|{\mathscr{W}}_k|\Big]\\ \label{Sec:4:Eqn:Con-Proof-Lossless-11} &\geq \sum_{k=1}^l H\big(W_{k}\big|W_{1},W_{2},\ldots,W_{k-1},Y_{k}\big) -\sum_{k=1}^l \Big[h\big(P^k_{e}\big)+P^k_{e}\log_2|{\mathscr{W}}_k|\Big]\\ \label{Sec:4:Eqn:Con-Proof-Lossless-12} & \geq \sum_{l=1}^t H\big( W_l \big| W_1, W_2,\ldots,W_{l-1},Y_l\big) - \Big[l\ h(\epsilon) + \epsilon \log_2|{\mathscr{W}}_1|\ |{\mathscr{W}}_2|\ \cdots |{\mathscr{W}}_l|\Big]\ ,\end{aligned}$$ where  through  follow from standard Shannon inequalities; follows because $({\mathbf{W}}_1$, $\ldots$, ${\mathbf{W}}_t$, ${\mathbf{Y}}_1$, $\ldots$, ${\mathbf{Y}}_t)$ is iid, $W_{k}$ $\minuso$ $(W_{1}$, $W_{2}$, $\ldots$, $W_{k-1}$, $Y_{k})$ $\minuso$ $(Y_{1}$, $Y_{2}$, $\ldots$, $Y_{k-1})$ forms a Markov chain; conditioning reduces entropy and $\Hat{W}_{k,i}$ is a function of $M_1,M_2,\ldots,M_k$ and ${\mathbf{Y}}_k$; follows from Fano’s Inequality where $h(\cdot)$ is the binary entropy function [@Cover-1991-B]; follows from the concavity of $h(\cdot)$ and Jensen’s inequality; follows by assuming $\epsilon$ is small (i.e. $0 < P^k_e < \epsilon < 1/2$). Finally, $l\ h(\epsilon)$ $+$ $\epsilon \log_2|{\mathscr{W}}_1|\ |{\mathscr{W}}_2|\ \cdots |{\mathscr{W}}_l|\rightarrow 0$ as $\epsilon \rightarrow 0$. Conclusion {#Sec:6} ========== We studied the rate-distortion function $R({\mathbf{d}})$ and the rate region ${\mathscr{R}}({\mathbf{d}})$ for the problems shown in Figures \[Sec:1:Fig:mKHB\] and \[Sec:1:Fig:tSR\], respectively. In [@Heegard-Nov-1985-A Thm. 2], Heegard and Berger claimed that a certain functional, $R_0({\mathbf{d}})$, is an upper bound for $R({\mathbf{d}})$. By way of a counterexample, we demonstrated that $R_0({\mathbf{d}})$ is not an upper bound for $R({\mathbf{d}})$. In Theorem \[Sec:4:Thm:HB\], we gave a new upper bound for $R({\mathbf{d}})$. This bound followed from a new inner bound for ${\mathscr{R}}({\mathbf{d}})$ that we presented in Theorem \[Sec:3:Thm:Successive-Refinement\]. Finally, we gave an explicit characterisation of the rates needed to losslessly reconstruct private messages at each decoder (assuming degraded side-information) in Theorem \[Sec:4:Thm:Lossless\]. Acknowledgements {#acknowledgements .unnumbered} ================ The authors would like to thank Dr. Chao Tian and an anonymous reviewer whose comments helped generalise Theorem \[Sec:3:Thm:Successive-Refinement\] from a more restricted statement into its current form. Proof of Theorem \[Sec:3:Thm:Successive-Refinement\] {#App:1} ==================================================== Fix $v\in{\mathscr{V}}$ and $p\in{\mathscr{P}}_v({\mathbf{d}})$ arbitrarily. It is sufficient to prove the ${\mathbf{d}}$-admissibility of rate tuples within ${\mathscr{R}}_{p,v}({\mathbf{d}})$. (The ${\mathbf{d}}$-admissibility of tuples within ${\mathscr{R}}_{in}({\mathbf{d}})$ follows by standard time-sharing arguments.) Our proof uses a random-coding argument that is based on the concept of $\epsilon$-letter typical sequences[^9]. This argument employs $(2^t-1)$-randomly generate codebooks; one codebook for every non-empty subset of receivers. The encoder selects a codeword from each codebook and sends some information (the bin indices of each codeword) to the decoders. Each decoder tries to recover those codewords where it is a member of the corresponding subset. To help elucidate the main ideas of the proof, we present the special case of four decoders as a series of examples in parallel to the main proof. For notational convenience, we impose the natural ordering on the elements of each subset ${{\mathscr{S}}}_j$, and we let ${\mathscr{S}}_j[i]$ denote the $i^{\text{th}}$-smallest element of ${\mathscr{S}}_j$. For example, if ${{\mathscr{S}}}_j = \{1,3,5\}$, then ${{\mathscr{S}}}_j[1] = 1$, ${{\mathscr{S}}}_j[2] = 3$ and ${{\mathscr{S}}}_j[3] = 5$. Code Construction ----------------- For each subset ${{\mathscr{S}}}_j$, construct an $|{{\mathscr{S}}}_j|$-layer nested codebook in the following manner. For each vector-valued index ${\mathbf{k}}_{{{\mathscr{S}}}_j} \triangleq ( k_{{{\mathscr{S}}}_j,1},k_{{{\mathscr{S}}}_j,2},\ldots, k_{{{\mathscr{S}}}_j,|{{\mathscr{S}}}_j|},k'_{{{\mathscr{S}}}_j} )$, where $$k_{{{\mathscr{S}}}_j,i} = 1,2,\ldots,2^{nR_{{{\mathscr{S}}}_j,i}}\ , \quad i = 1,2,\ldots,|{{\mathscr{S}}}_j|\ ,$$ $$k'_{{{\mathscr{S}}}_j} = 1,2,\ldots,2^{nR'_{{{\mathscr{S}}}_j}}\ ,$$ generate a length $n$ codeword ${\mathbf{u}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j}) \in {\mathscr{U}}^n_{{{\mathscr{S}}}_j}$ by selecting $n$ symbols from ${\mathscr{U}}_{{{\mathscr{S}}}_j}$ in an iid manner using $p(u_{{\mathscr{S}}_j})$ – the $U_{{{\mathscr{S}}}_j}$-marginal of $p$. The values of $R_{{{\mathscr{S}}}_j,i}$ and $R_{{{\mathscr{S}}}_j}'$ will be defined shortly. Choose the list $v$ as follows: ${{\mathscr{S}}}_1 = \{1,2,3,4\}$, ${{\mathscr{S}}}_2 = \{1,2,3\}$, ${{\mathscr{S}}}_3 = \{1,2,4\}$, ${{\mathscr{S}}}_4 = \{1,3,4\}$, ${{\mathscr{S}}}_5 = \{2,3,4\}$, ${{\mathscr{S}}}_6 = \{1,2\}$, ${{\mathscr{S}}}_7 = \{1,3\}$, ${{\mathscr{S}}}_8 = \{1,4\}$, ${{\mathscr{S}}}_9 = \{2,3\}$, ${{\mathscr{S}}}_{10} = \{2,4\}$, ${{\mathscr{S}}}_{11} = \{3,4\}$, ${{\mathscr{S}}}_{12} = \{1\}$, ${{\mathscr{S}}}_{13} = \{2\}$, ${{\mathscr{S}}}_{14} = \{3\}$ and ${{\mathscr{S}}}_{15} = \{4\}$. Figure \[Sec:7:Fig:Code-Layers\] shows the $3$-layer nested codebook associated with the subset $\{1,2,3\}$. In the first layer, there are $2^{nR_{123,1}}$ bins (labelled with the index $k_{123,1}$) each of which contain $2^{n(R'_{123} + R_{123,2} + R_{123,3})}$ codewords. The set of codewords inside a particular layer one bin define the second layer of the codebook. Specifically, each layer one index $k_{123,1} \in [2^{nR_{123,1}}]$ identifies $2^{nR_{123,2}}$ layer two bins. These bins are labelled with the index $k_{123,2}$, and each bin contains $2^{n(R'_{123} + R_{123,3})}$ codewords. Similarly, each pair $k_{123,1} \in [2^{nR_{123,1}}]$ and $k_{123,2} \in [2^{nR_{123,2}}]$ identifies $2^{nR_{123,3}}$ layer three bins. There are $2^{n(R'_{123})}$ codewords in each one of the layer three bins. \ Encoding -------- Encoding proceeds sequentially over $(2^t - 1)$-stages using $\epsilon$-letter typical-set encoding rules. For this purpose, choose $0< \epsilon_0 < \epsilon_1 < \cdots < \epsilon_{2^t}$ to be arbitrarily small real numbers. The encoder is given ${\mathbf{x}} \in {\mathscr{X}}^n$. At encoding stage $j$ it selects the codebook with label ${{\mathscr{S}}}_j$ and looks for an index vector ${\mathbf{k}}_{{{\mathscr{S}}}_j}$ where the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j})$ is $\epsilon_{j}$-letter typical with ${\mathbf{x}}$ and \[Sec:Proof:Eqn:Typ-Enc-1\] $$\begin{aligned} \label{Sec:Proof:Eqn:Typ-Enc-1a} {\mathbf{u}}_{{{\mathscr{S}}}_j}^{\supset} &\triangleq \big\{ {\mathbf{u}}_{{{\mathscr{S}}}_i}({\mathbf{k}}_{{{\mathscr{S}}}_i}) : i < j,\ {{\mathscr{S}}}_i \supset {{\mathscr{S}}}_j \big\}, \text{ and}\ \\ \label{Sec:Proof:Eqn:Typ-Enc-1b} {\mathbf{u}}^\dag_{{{\mathscr{S}}}_j} &\triangleq \big\{ {\mathbf{u}}_{{{\mathscr{S}}}_i}({\mathbf{k}}_{{{\mathscr{S}}}_i}) : i<j,\ {{\mathscr{S}}}_i \nsupseteq {{\mathscr{S}}}_j,\ \exists {{\mathscr{S}}}_{i'},\ i' > j,\ {{\mathscr{S}}}_{i'} \cap {{\mathscr{S}}}_j \neq \emptyset,\ {{\mathscr{S}}}_{i} \cap {{\mathscr{S}}}_{i'}\neq \emptyset\big\}\ .\end{aligned}$$ If successful[^10], the encoder sends the bin index $k_{{{\mathscr{S}}}_j,i}$ over channel ${{\mathscr{S}}}_j[i]$ for every $i = 1,2,\dots,|{{\mathscr{S}}}_j|$. If unsuccessful, the encoder sends $k_{{{\mathscr{S}}}_j,i}=1$ over each of these channels. Note the correspondence between the sets ${\mathbf{u}}_{{{\mathscr{S}}}_j}^{\supset}$ and ${\mathbf{u}}^\dag_{{{\mathscr{S}}}_j}$ and the sets of auxiliary random variables ${\mathscr{U}}_{{{\mathscr{S}}}_j}^{\supset}$ and ${\mathscr{U}}_{{{\mathscr{S}}}_j}^{\dag}$, respectively. Finally, note that when $|{{\mathscr{S}}}_j|\geq 3$, then ${\mathbf{u}}_{{{\mathscr{S}}}_j}^{\supset} \cup {\mathbf{u}}^\dag_{{{\mathscr{S}}}_j} = \{ {\mathbf{u}}_{{{\mathscr{S}}}_i}({\mathbf{k}}_{{{\mathscr{S}}}_i}) : i < j \}$; that is, the encoder chooses ${\mathbf{u}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j})$ to be jointly typical with every codeword it has previously selected. The situation is more complex when $|{{\mathscr{S}}}_j|\leq 2$. $$\begin{array}{|c|c|c|c|} \hline \text{Subset } {{\mathscr{S}}}_j & {\mathbf{u}}^{\supset}_{{{\mathscr{S}}}_j}, \Hat{{\mathbf{u}}}^{\supset}_{{{\mathscr{S}}}_j} & {\mathbf{u}}^{\dag}_{{{\mathscr{S}}}_j} & \Hat{{\mathbf{u}}}^{\ddag}_{{{\mathscr{S}}}_j,l} \\ \hline {{\mathscr{S}}}_1 = \{1,2,3,4\} & \emptyset & \emptyset & \emptyset \\ \hline {{\mathscr{S}}}_2 = \{1,2,3\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1}\} & \emptyset & \emptyset \\ \hline {{\mathscr{S}}}_3 = \{1,2,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1}\} &\{{\mathbf{u}}_{{{\mathscr{S}}}_2}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_3,1} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_3,2} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_3,4} = \emptyset \end{array} \\ \hline {{\mathscr{S}}}_4 = \{1,3,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1}\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_4,1} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3}\} \\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_4,3} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2}\} \\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_4,4} = \{{\mathbf{u}}_{{{\mathscr{S}}}_3}\} \end{array}\\ \hline {{\mathscr{S}}}_5 = \{2,3,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1}\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_5,2} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_5,3} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_4}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_5,4} = \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4}\} \end{array} \\ \hline {{\mathscr{S}}}_6 = \{1,2\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3}\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_5}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_6,1} = \{{\mathbf{u}}_{{{\mathscr{S}}}_4}\} \\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_6,2} = \{{\mathbf{u}}_{{{\mathscr{S}}}_5}\} \end{array} \\ \hline {{\mathscr{S}}}_7 = \{1,3\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_4}\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_5},{\mathbf{u}}_{{{\mathscr{S}}}_6}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_7,1} = \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_6}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_7,3} = \{{\mathbf{u}}_{{{\mathscr{S}}}_5}\} \end{array} \\ \hline {{\mathscr{S}}}_8 = \{1,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4}\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_5},{\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_7}\} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_8,1} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_7}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_8,4} = \{{\mathbf{u}}_{{{\mathscr{S}}}_5}\} \end{array} \\ \hline {{\mathscr{S}}}_9 = \{2,3\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_5}\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_7}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_8}\} \end{array} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_9,2} = \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_6}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_9,3} = \{{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_7}\} \end{array} \\ \hline {{\mathscr{S}}}_{10} = \{2,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_5}\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_7}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_8},{\mathbf{u}}_{{{\mathscr{S}}}_9}\} \end{array} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_{10},2} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_9}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_{10},4} = \{{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_8}\} \end{array} \\ \hline {{\mathscr{S}}}_{11} = \{3,4\} & \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_5}\} & \begin{array}{c} \{{\mathbf{u}}^\ddag_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_7},{\mathbf{u}}_{{{\mathscr{S}}}_8}, \\ {\mathbf{u}}^\ddag_{{{\mathscr{S}}}_9},{\mathbf{u}}_{{{\mathscr{S}}}_{10}}\} \end{array} & \begin{array}{c} \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_{11},3} = \{{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_7},{\mathbf{u}}_{{{\mathscr{S}}}_9}\}\\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_{11},4} = \{{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_8},{\mathbf{u}}_{{{\mathscr{S}}}_{10}}\} \end{array} \\ \hline {{\mathscr{S}}}_{12} = \{1\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_7},{\mathbf{u}}_{{{\mathscr{S}}}_8}\} \end{array} & \emptyset & \emptyset\\ \hline {{\mathscr{S}}}_{13} = \{2\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_5}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_6},{\mathbf{u}}_{{{\mathscr{S}}}_9},{\mathbf{u}}_{{{\mathscr{S}}}_{10}}\} \end{array} & \emptyset & \emptyset\\ \hline {{\mathscr{S}}}_{14} = \{3\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_2},{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_5}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_7},{\mathbf{u}}_{{{\mathscr{S}}}_9},{\mathbf{u}}_{{{\mathscr{S}}}_{11}}\} \end{array} & \emptyset & \emptyset\\ \hline {{\mathscr{S}}}_{15} = \{4\} & \begin{array}{c} \{{\mathbf{u}}_{{{\mathscr{S}}}_1},{\mathbf{u}}_{{{\mathscr{S}}}_3},{\mathbf{u}}_{{{\mathscr{S}}}_4},{\mathbf{u}}_{{{\mathscr{S}}}_5}, \\ {\mathbf{u}}_{{{\mathscr{S}}}_8},{\mathbf{u}}_{{{\mathscr{S}}}_{10}},{\mathbf{u}}_{{{\mathscr{S}}}_{11}}\} \end{array} & \emptyset & \emptyset\\ \hline \end{array}$$ Table \[Sec:Proof:Tab:1\] lists the fifteen encoding sets ${\mathbf{u}}^{\supset}_{{{\mathscr{S}}}_j}$ and ${\mathbf{u}}^{\dag}_{{{\mathscr{S}}}_j}$ and Figure \[Sec:App:Fig:4DecodersIndexAssignement\] depicts the index to channel assignments for the four decoder example. In stage $1$, the encoder considers subset ${{\mathscr{S}}}_1$ and looks for an index vector ${\mathbf{k}}_{{{\mathscr{S}}}_1}$ such that the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_1}({\mathbf{k}}_{{{\mathscr{S}}}_1})$ is jointly typical with ${\mathbf{x}}$. (The sets ${\mathbf{u}}^{\supset}_{{{\mathscr{S}}}_1}$ and ${\mathbf{u}}^{\dag}_{{{\mathscr{S}}}_1}$ are empty – see Table \[Sec:Proof:Tab:1\].) The resulting indices $k_{{{\mathscr{S}}}_1,1}$, $k_{{{\mathscr{S}}}_1,2}$, $k_{{{\mathscr{S}}}_1,3}$ and $k_{{{\mathscr{S}}}_1,4}$ are sent over channels $1$, $2$, $3$ and $4$, respectively. In the eleventh encoding stage, takes the codebook for ${{\mathscr{S}}}_{11} = \{3,4\}$ and looks for a index vector ${\mathbf{k}}_{{{\mathscr{S}}}_{11}} = (k_{{{\mathscr{S}}}_{11},1},k_{{{\mathscr{S}}}_{11},2},k'_{{{\mathscr{S}}}_{11}})$ such that the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_{11}}({\mathbf{k}}_{{{\mathscr{S}}}_{11}})$ is jointly typical with ${\mathbf{x}}$, ${\mathbf{u}}_{{{\mathscr{S}}}_{1}}({\mathbf{k}}_{{{\mathscr{S}}}_{1}})$ through to ${\mathbf{u}}_{{{\mathscr{S}}}_{5}}({\mathbf{k}}_{{{\mathscr{S}}}_{5}})$ and ${\mathbf{u}}_{{{\mathscr{S}}}_{7}}({\mathbf{k}}_{{{\mathscr{S}}}_{7}})$ through to ${\mathbf{u}}_{{{\mathscr{S}}}_{10}}({\mathbf{k}}_{{{\mathscr{S}}}_{10}})$. (Note, that this codeword need not be jointly typical with ${\mathbf{u}}_{{{\mathscr{S}}}_{6}}({\mathbf{k}}_{{{\mathscr{S}}}_{6}})$.) The resulting indices $k_{{{\mathscr{S}}}_{11},1}$, $k_{{{\mathscr{S}}}_{11},2}$ are sent over channels $3$ and $4$, respectively. ![$4$-Decoders Example: Assignment of bin indices to channels. Subsets are ${{\mathscr{S}}}_1 = \{1,2,3,4\}$, ${{\mathscr{S}}}_2 = \{1,2,3\}$, ${{\mathscr{S}}}_3 = \{1,2,4\}$, ${{\mathscr{S}}}_4 = \{1,3,4\}$, ${{\mathscr{S}}}_5 = \{2,3,4\}$, ${{\mathscr{S}}}_6 = \{1,2\}$, ${{\mathscr{S}}}_7 = \{1,3\}$, ${{\mathscr{S}}}_8 = \{1,4\}$, ${{\mathscr{S}}}_9 = \{2,3\}$, ${{\mathscr{S}}}_{10} = \{2,4\}$, ${{\mathscr{S}}}_{11} = \{3,4\}$, ${{\mathscr{S}}}_{12} = \{1\}$, ${{\mathscr{S}}}_{13} = \{2\}$, ${{\mathscr{S}}}_{14} = \{3\}$ and ${{\mathscr{S}}}_{15} = \{4\}$. Bin index $k_{{{\mathscr{S}}}_j,i}$ is sent over channel ${{\mathscr{S}}}_j[i]$.[]{data-label="Sec:App:Fig:4DecodersIndexAssignement"}](4DecoderExampleF "fig:"){width=".75\columnwidth"}\ Decoding -------- Consider decoder $l$. Like the encoding procedure, decoder $l$ forms its reconstruction $\Hat{{\mathbf{X}}}_l$ of ${\mathbf{X}}$ using $(2^t-1)$-decoding stages. Recall, this decoder recovers every bin index transmitted on channels $1$ through $l$; it does not have access to any index transmitted on channels $l + 1$ through $t$. In stage $j$ decoder $l$ considers subset ${{\mathscr{S}}}_j$. If $l \notin {{\mathscr{S}}}_j$, then it does nothing and moves to decoding stage $j + 1$. If $l \in {\mathscr{S}}_j$, then the decoder forms a reconstruction ${\mathbf{u}}_{{\mathscr{S}}_j}(\Hat{{\mathbf{k}}}_{{\mathscr{S}}_j})$ of the codeword ${\mathbf{u}}_{{\mathscr{S}}_j}({\mathbf{k}}_{{\mathscr{S}}_j})$, which was selected by the encoder, using the following procedure. Note, decoder $l$ will have reconstructed the following codewords in decoding stages $1$ through $j-1$: \[eq:decoder-1\] $$\Hat{{\mathbf{u}}}_{{{\mathscr{S}}}_j}^{\supset} \triangleq \big\{ {\mathbf{u}}_{{{\mathscr{S}}}_i}(\Hat{{\mathbf{k}}}_{{{\mathscr{S}}}_i}) : i < j,\ {{\mathscr{S}}}_i \supset {{\mathscr{S}}}_j \big\}, \text{ and}$$ $$\Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_j,l} \triangleq \big\{ {\mathbf{u}}_{{{\mathscr{S}}}_i}(\Hat{{\mathbf{k}}}_{{{\mathscr{S}}}_i}) : i<j,\ {{\mathscr{S}}}_i \ni l,\ {{\mathscr{S}}}_i \nsupseteq {{\mathscr{S}}}_j\big\}\ .$$ Note the correspondence between the decoding sets in  and the sets of auxiliary random variables ${\mathscr{A}}^{\supset}_{{{\mathscr{S}}}_j}$ and ${\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,l}$. To form its reconstruction ${\mathbf{u}}_{{\mathscr{S}}_j}(\Hat{{\mathbf{k}}}_{{\mathscr{S}}_j})$, decoder $l$ takes the bin indices $$\left\{ k_{{{\mathscr{S}}}_j,i} ;\ i = 1,2,\ldots,|[l] \cap {{\mathscr{S}}}_j| \right\}$$ from channels $1$ through $l$. It then looks for an index vector $\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j}$, with $\tilde{k}_{{{\mathscr{S}}}_j,i} = k_{{{\mathscr{S}}}_j,i}$ for all $i = 1,2,\ldots,|[l]\cap{{\mathscr{S}}}_j|$, such that the corresponding codeword ${\mathbf{u}}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j})$ is $\epsilon_{j+1}$-letter typical with ${\mathbf{y}}_l$ as well as the codewords in  that were decoded in the first $(j-1)$-stages: $$\label{eq:decoder} \left(\Hat{{\mathbf{u}}}_{{{\mathscr{S}}}_j}^{\supset},\ \Hat{{\mathbf{u}}}^\ddag_{{{\mathscr{S}}}_j,l},\ {\mathbf{u}}_{{{\mathscr{S}}}_j}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j}), {\mathbf{y}}_{l} \right) \in T_{\epsilon_{j+1}}^{(n)}(p) .$$ Note that there are $$\exp_2\left[ {n\left(R'_{{{\mathscr{S}}}_j} + \sum_{i = |[l]\cap {{\mathscr{S}}}_j|+1}^{|{{\mathscr{S}}}_j|} R_{{{\mathscr{S}}}_j,i}\right)} \right]$$ codewords in the bin specified by the indices $\{k_{{{\mathscr{S}}}_j,i} : i = 1,2,\ldots,|[l] \cap {{\mathscr{S}}}_j|\}$. If one or more of these codewords satisfy this typicality condition, then decoder $l$ selects one arbitrarily and sets $\Hat{{\mathbf{k}}}_{{{\mathscr{S}}}_j} = \tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j}$. If there is no such codeword, it sets each of the unknown indices equal to $1$. Consider the second decoder $(l = 2)$. In stage one, take $k_{{{\mathscr{S}}}_1,1}$ (from channel $1$) and $k_{{{\mathscr{S}}}_1,2}$ (from channel $2$) and look for a vector $\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_1}$ $=$ $(k_{{{\mathscr{S}}}_1,1}$, $k_{{{\mathscr{S}}}_1,2}$, $\tilde{k}_{{{\mathscr{S}}}_1,3},\tilde{k}'_{{{\mathscr{S}}}_1})$ such that the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_1}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_1})$ is typical with ${\mathbf{y}}_2$. Similarly, in stage nine take $k_{{{\mathscr{S}}}_9,1}$ (from channel $2$) and look for $\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_9} = (k_{{{\mathscr{S}}}_9,1},\tilde{k}_{{{\mathscr{S}}}_9,2},\tilde{k}'_{{{\mathscr{S}}}_9})$ such that the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_9}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_9})$ is jointly typical with ${\mathbf{y}}_2$ and ${\mathbf{u}}_{{{\mathscr{S}}}_1}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_1})$, ${\mathbf{u}}_{{{\mathscr{S}}}_2}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_2})$, ${\mathbf{u}}_{{{\mathscr{S}}}_3}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_3})$, ${\mathbf{u}}_{{{\mathscr{S}}}_5}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_5})$ and ${\mathbf{u}}_{{{\mathscr{S}}}_6}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_6})$, which were decoded during stages one through six. Finally, in stage thirteen take $k_{{{\mathscr{S}}}_{13},1}$ (from channel $2$) and look for $\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_{13}} = (k_{{{\mathscr{S}}}_{12},1},\tilde{k}'_{{{\mathscr{S}}}_{13}})$ such that the corresponding codeword ${\mathbf{u}}_{{{\mathscr{S}}}_{13}}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_6})$ is jointly typical with ${\mathbf{y}}_2$ and ${\mathbf{u}}_{{{\mathscr{S}}}_1}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_1})$, ${\mathbf{u}}_{{{\mathscr{S}}}_2}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_2})$, ${\mathbf{u}}_{{{\mathscr{S}}}_3}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_3})$, ${\mathbf{u}}_{{{\mathscr{S}}}_5}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_5})$, ${\mathbf{u}}_{{{\mathscr{S}}}_6}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_6})$, ${\mathbf{u}}_{{{\mathscr{S}}}_9}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_9})$ and ${\mathbf{u}}_{{{\mathscr{S}}}_{10}}(\Hat{\mathbf{k}}_{{{\mathscr{S}}}_{10}})$, which were decoded during stages one through ten. Error Analysis: Encoding ------------------------ The coding scheme is based on $\epsilon$-letter typical set encoding and decoding techniques. As such, the distortion criteria at each decoder will not be satisfied when $\left({\mathbf{x}}, {\mathbf{y}}_{1}, {\mathbf{y}}_{2}, \dots, {\mathbf{y}}_{t}\right) \notin T^{(n)}_{\epsilon_0}(p)$. We denote this event by $E_1$. From Lemma \[App:Lem:1\], the probability of this event may be bound by $$\Pr\left[E_{1} \right] \leq \delta_1 \left(n,\epsilon_0,\mu(p)\right)\ ,$$ where $\delta_1 \left(n,\epsilon_0,\mu(p)\right) {\rightarrow} 0$ as $n\rightarrow\infty$. Assume $E_1$ does not occur. Let $E_{2,{{\mathscr{S}}}_j}$ denote the event that the encoder fails to find an $\epsilon_j$-letter typical codeword during stage $j$ of encoding procedure given that it found an $\epsilon_i$-letter typical codeword for every stage $i\in [j-1]$. From Lemma \[App:Lem:2\] and the inequality $(1 - x)^t \leq e^{-tx}$ we have $$\begin{aligned} \notag \Pr \left[E_{2,{{\mathscr{S}}}_j} \right] &= \Bigg[ 1 -\text{Pr}\Big[\big({\mathbf{u}}^{\supset}_{{{\mathscr{S}}}_j},{\mathbf{u}}^{\dag}_{{{\mathscr{S}}}_j},{\mathbf{U}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j}),{\mathbf{x}}\big) \in T^{(n)}_{\epsilon_{j+1}}(p)\Big]\Bigg]^{2^{n \left( R'_{{{\mathscr{S}}}_j} + \sum_{i=1}^{|S_j|} R_{{{\mathscr{S}}}_j,i}\right)}}\\ \label{Sec:7:Eqn:Error:E2-4} & \leq \exp \Bigg(-\big(1 - \delta_2\big) 2^{n \left( R'_{{{\mathscr{S}}}_j} + \sum_{i=1}^{|S_j|} R_{{{\mathscr{S}}}_j,i}\right)} \cdot 2^{-n \Big(I\left({\mathscr{A}}^\supset_{{{\mathscr{S}}}_j},{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j},X;{U}_{{{\mathscr{S}}}_j}\right) +2\epsilon_{j} H\left({U}_{{{\mathscr{S}}}_j}\right)\Big)}\Bigg)\end{aligned}$$ where we have written the function $\delta_2\left(n,\epsilon_{j-1},\epsilon_{j},\mu(p)\right)$ as $\delta_2$ for compact representation. Let $E_2$ denote the event where a typical codeword cannot be found at any one of the encoding stages. By the union bound we get the following upper bound for $\text{Pr}[E_2]$: $$\begin{aligned} \Pr \left[E_{2}\right] \leq \sum_{j=1}^{2^t-1} \exp \Bigg[&-\big(1 - \delta_2\big) 2^{n \left( R'_{{{\mathscr{S}}}_j} + \sum_{i=1}^{|S_j|} R_{{{\mathscr{S}}}_j,i}\right)} \cdot 2^{-n \Big[I\left({\mathscr{A}}^\supset_{{{\mathscr{S}}}_j},{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j},X;U_{{{\mathscr{S}}}_j}\right) +2\epsilon_{j} H\left(U_{{{\mathscr{S}}}_j}\right)\Big]}\Bigg]\ .\end{aligned}$$ Finally, note that if $$\label{Sec:7:Eqn:Error:E2-5} R'_{{{\mathscr{S}}}_j} + \sum_{i=1}^{|{{\mathscr{S}}}_j|} R_{{{\mathscr{S}}}_j,i} > I\left({\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}, {\mathscr{A}}^\dag_{{{\mathscr{S}}}_j},X;{U}_{{{\mathscr{S}}}_j}\right) + 2\epsilon_{j} H\left({U}_{{{\mathscr{S}}}_j}\right)$$ for every $j = 1,2,\ldots,2^t - 1$, then $\Pr[E_{2}] \rightarrow 0$ as $n \rightarrow \infty$. Error Analysis: Decoding ------------------------ Assume $E_1$ and $E_2$ do not occur. Consider decoder $l$ and a non-trivial decoding stage $j$ where ${{\mathscr{S}}}_j \ni l$. Let $D_{l,{{\mathscr{S}}}_j}$ be the event that it cannot find a unique codeword that satisfies the typicality condition given that at every stage $i < j$ (where ${{\mathscr{S}}}_i \ni l$) it found a unique codeword ${\mathbf{u}}(\Hat{{\mathbf{k}}}_{{{\mathscr{S}}}_i})$ satisfying this typicality condition. By the Markov lemma (Lemma \[App:Lem:3\]), the probability that the codewords $\Hat{{\mathbf{u}}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j})$, $\Hat{{\mathbf{u}}}^{\supset}_{{{\mathscr{S}}}_j}$, ${\mathbf{u}}^{\ddag}_{{{\mathscr{S}}}_j,l}$ are not jointly typical with ${\mathbf{y}}_{l}$ is small for large $n$: $$\Pr\left[{\mathbf{Y}}_{l} \notin T^{(n)}_{\epsilon_{j+1}}\left(p \mid\Hat{{\mathbf{u}}}^{\supset}_{{{\mathscr{S}}}_j}, \Hat{{\mathbf{u}}}^{\ddag}_{{{\mathscr{S}}}_j,l} {\mathbf{u}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j}),{\mathbf{x}}\right)\right] \leq \delta_2 \left(n,\epsilon_j,\epsilon_{j+1},\mu(p)\right)\ .$$ An upper bound for the probability that there exists one or more codewords ${\mathbf{u}}_{{{\mathscr{S}}}_j}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j}) \neq {\mathbf{u}}_{{{\mathscr{S}}}_j}({\mathbf{k}}_{{{\mathscr{S}}}_j})$, which satisfy , is $$\begin{gathered} \label{Sec:Proof:Eqn:Decoder-Error-Upperbound} \Pr \left[\bigcup_{{\mathscr{K}}_j} \left\{ \Hat{{\mathbf{u}}}_{{{\mathscr{S}}}_j}^{\supset},\Hat{{\mathbf{u}}}_{{{\mathscr{S}}}_j,l}^{\ddag}, {\mathbf{y}}_{l},{\mathbf{u}}_{{{\mathscr{S}}}_j}(\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j}) \right\} \in T^{(n)}_{\epsilon_{j+1}}(p)\right]\\ < \exp_2 \Bigg[ n\Big(R'_{{{\mathscr{S}}}_j}+\sum_{i= |[l]\cap {{\mathscr{S}}}_j|+1}^{|{{\mathscr{S}}}_j|} R_{{{\mathscr{S}}}_j,i} - I\big({U}_{{{\mathscr{S}}}_j} ;{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}, {\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,l}, Y_{l}\big) + 2\epsilon_{j+1}H\big({U}_{{{\mathscr{S}}}_j}\big)\Big)\Bigg]\ ,\end{gathered}$$ where we have taken the union over $${\mathscr{K}}_j = \left\{\tilde{{\mathbf{k}}}_{{{\mathscr{S}}}_j} \neq {\mathbf{k}}_{{{\mathscr{S}}}_j},\ \{\tilde{k}_{{{\mathscr{S}}}_j,i} = k_{{{\mathscr{S}}}_j,i}\}_{i=1}^{|\{1,2,\ldots,l\}\cap {{\mathscr{S}}}_j|}\right\}\ .$$ Applying the union bound we get $$\begin{aligned} \Pr \left[ D_{l,{{\mathscr{S}}}_j} \right] &<\delta_2 +\exp_2\Bigg[n\Bigg(R_{{{\mathscr{S}}}_j}'+\sum_{i=|[l]\cap{{\mathscr{S}}}_j|}^{{{\mathscr{S}}}_j} R_{{{\mathscr{S}}}_j,i}\Bigg) - n \Bigg( I\Big(U_{{{\mathscr{S}}}_j};{\mathscr{A}}^\supset_{{{{\mathscr{S}}}_j}},{\mathscr{A}}^\ddag_{{{{\mathscr{S}}}_j,l}},Y_l\Big) - 2 \epsilon_{j+1} H(U_{{{\mathscr{S}}}_j}) \Bigg) \Bigg]\ .\end{aligned}$$ Thus, if $$\label{Sec:7:Eqn:RL-SJ-Unique} R'_{{{\mathscr{S}}}_j} + \sum_{i = |[l]\cap {{\mathscr{S}}}_j|+1}^{|{{\mathscr{S}}}_j|} R_{{{\mathscr{S}}}_j,i} <I\left({U}_{{{\mathscr{S}}}_j}; {\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}, {\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,l},Y_{l}) - 2\epsilon_{j+1} H\left({U}_{{{\mathscr{S}}}_j}\right)\right)$$ then $\Pr[D_{l,{{\mathscr{S}}}_j}] \rightarrow 0$ as $n \rightarrow \infty$. Rate Constraints ---------------- Consider decoder $l$ and any subset ${{\mathscr{S}}}_j$ where $l \in {{\mathscr{S}}}_j$. On combining the rate constraints  and  we get $$\begin{aligned} \notag \sum_{i = 1}^{|{{\mathscr{S}}}\cap [l]|} R_{{{\mathscr{S}}},i} &> I\big(X,{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j},{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j};U_{{{\mathscr{S}}}}\big) - I\big(U_{{{\mathscr{S}}}_j};{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j},{\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,l},Y_l\big)\\ \label{Sec:Proof:Eqn:SJ-Unique-2} &= I\big(X,{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j};U_{{{\mathscr{S}}}}\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big) - I\big(U_{{{\mathscr{S}}}_j};{\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,l},Y_l\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big)\ .\end{aligned}$$ Since $\epsilon_j$ and $\epsilon_{j+1}$ may be selected arbitrarily small, we can ignore the $2(\epsilon_j + \epsilon_{j+1})H({{\mathscr{S}}}_j)$ term. Consider the other decoders in $[l] \cap {{\mathscr{S}}}_j$. Since $R_{{{\mathscr{S}}}_j,i} \geq 0$ for all $i$, it must be true that $$\begin{aligned} \notag \sum_{i=1}^{|[l]\cap {{\mathscr{S}}}|} R_{{{\mathscr{S}}}_j,i} &> \max_{\tilde{l} \in [l] \cap {{\mathscr{S}}}_j} \Big[I\big(X,{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j};U_{{{\mathscr{S}}}_j}\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big) - I\big(U_{{{\mathscr{S}}}_j};{\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,\tilde{l}},Y_{\tilde{l}}\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big)\Big]\\ \label{Sec:7:Eqn:RL-SJ-Cor-Dec} &= I\big(X,{\mathscr{A}}^\dag_{{{\mathscr{S}}}_j};U_{{{\mathscr{S}}}_j}\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big) - \min_{\tilde{l} \in [l] \cap {{\mathscr{S}}}_j} I\big(U_{{{\mathscr{S}}}_j};{\mathscr{A}}^\ddag_{{{\mathscr{S}}}_j,\tilde{l}},Y_{\tilde{l}}\big|{\mathscr{A}}^\supset_{{{\mathscr{S}}}_j}\big)\ ;\end{aligned}$$ that is, the rate constraint for decoder $l$ must be at least as large as the rate constraint for decoder $\tilde{l}$ (for every $\tilde{l} \in [l]\cap{\mathscr{S}}_j$). The rate constraint  is valid for any set ${{\mathscr{S}}}_j$ where $l \in {{\mathscr{S}}}_j$. For such subsets, define $l^* \triangleq \max_{i \in [l] \cap {{\mathscr{S}}}_j} i$. Since $l^* \in {{\mathscr{S}}}_j$ and $[l^*] \cap {{\mathscr{S}}}_j = [l] \cap {{\mathscr{S}}}_j$, it follows that  is also valid for any set ${{\mathscr{S}}}_j$ where $[l] \cap {{\mathscr{S}}}_j \neq \emptyset$. Finally, consider the sum rate $\sum_{i=1}^l R_i$ for the first $l$ channels. By construction, we have that $$\label{Sec:Proof:Eqn:Sum-Rate-1} \sum_{i=1}^l R_i = \sum_{{{\mathscr{S}}}_j\subseteq [t] \atop{{{\mathscr{S}}}_j \cap [l] \neq\emptyset}} \sum_{i=1}^{|[l]\cap{{\mathscr{S}}}_j|} R_{{{\mathscr{S}}}_j,i}\ .$$ Substituting the rate constraint  into yields the desired result. $\epsilon$-Letter Typicality {#App:B} ============================ For $\epsilon \geq 0$, a sequence $x^n \in {\mathscr{X}}^n$ is said to be $\epsilon$-letter typical with respect to a discrete memoryless source $({\mathscr{X}},p_X)$ if $$\left|\frac{1}{n}N(a|x^n) - p_X(a) \right| \leq \epsilon \cdot p_X(a)\quad \forall a \in {\mathscr{X}}\ ,$$ where $N(a|x^n)$ is the number of times the letter $a$ occurs in the sequence $x^n$. The collection of all $\epsilon$-letter typical sequences is denoted by $T^{(n)}_{\epsilon}(p_X)$. In a similar fashion, a pair of sequences $x^n$ and $y^n$ are said to jointly $\epsilon$-letter typical with respect to a discrete memoryless two source $({\mathscr{X}} \times {\mathscr{Y}},p_{XY})$ if $$\left|\frac{1}{n}N(a,b|x^n,y^n) - p_{XY}(a,b) \right| \leq \epsilon \cdot p_{XY}(a,b)\quad \forall (a,b) \in {\mathscr{X}} \times {\mathscr{Y}}\ ,$$ where $N(a,b|x^n,y^n)$ is the number of times the pair of letters $(a,b)$ occurs in the pair $(x^n,y^n)$. The collection of all joint $\epsilon$-typical sequence pairs is denoted by $T_\epsilon^{(n)}(p_{XY})$. Given $({\mathscr{X}} \times {\mathscr{Y}},p_{XY})$ and $x^n \in {\mathscr{X}}^n$, the set $$T_\epsilon^{(n)}\left(p_{XY}\mid x^n\right) = \big\{ y^n\ :\ (x^n,y^n) \in T_\epsilon^{(n)}(p_{XY}) \big\}$$ is called the set of conditionally $\epsilon$-letter typical sequences. Let $\mu({\mathscr{X}},p_{X}) = \min \{ p_{X}(x) :\ x \in \text{support}(p_{X})\}$ and define $$\delta_1\left(n,\epsilon,\mu(p_X)\right) = 2|{\mathscr{X}}| \cdot e^{-n\epsilon^2\mu(p_{X})}.$$ Note, $\delta\big(n,\epsilon,\mu(p_X)\big) \rightarrow 0$ as $n \rightarrow \infty$. \[App:Lem:1\] Suppose $X^n$ is emitted by a discrete memoryless source $({\mathscr{X}},p_X)$. If $0 < \epsilon \leq \mu(p_{X})$, then $$1 - \delta_1\left(n,\epsilon,\mu(p_X)\right) \leq \Pr\left[X^n \in T^{(n)}_{\epsilon}(p_{X}) \right] \leq 1\ .$$ Now consider a discrete memoryless two-source $({\mathscr{X}} \times {\mathscr{Y}},p_{XY})$, let $$\delta_2\big(n,\epsilon_1,\epsilon_2,\mu(p_{XY})\big) = 2|{\mathscr{X}}||{\mathscr{Y}}| \cdot e^{-n\frac{(\epsilon_2 - \epsilon_1)^2}{1+\epsilon_1}\mu(p_{XY})},$$ and note that $\delta_2\big(n,\epsilon_1,\epsilon_2,\mu(p_X)\big) \rightarrow 0$ as $n \rightarrow \infty$. \[App:Lem:2\] Suppose $Y^n$ is emitted by $({\mathscr{Y}},p_Y)$ where $p_Y$ is equal to the $Y$-marginal of $p_{XY}$. If $0 < \epsilon_1 < \epsilon_2 \leq \mu(p_{XY})$ and $x^n \in T^{(n)}_{\epsilon_1}(p_{X})$, then $$\begin{gathered} \left(1 - \delta_2\left(n,\epsilon_1,\epsilon_2,\mu(p_{XY})\right)\right) 2^{-n\left(I(X;Y) + 2\epsilon_2H(Y)\right)} \\ \leq \Pr\left[ Y^n \in T^{(n)}_{\epsilon_2}\left(p_{XY}\mid x^n\right) \right] \leq 2^{-n\left(I(X;Y)-2 \epsilon_2 H(Y) \right)} .\end{gathered}$$ Finally, a direct consequence of Lemma \[App:Lem:2\] for Markov sources is the following result. \[App:Lem:3\] Suppose $(X^n,Y^n,Z^n)$ is emitted by a discrete memoryless three-source $({\mathscr{X}} \times {\mathscr{Y}} \times {\mathscr{Z}},p_{XYZ})$ where $X {\minuso}Y {\minuso}Z$. If $0 < \epsilon_1 < \epsilon_2 \leq \mu(p_{XYZ})$ and $(x^n,y^n) \in T^{(n)}_{\epsilon_1}(p_{XY})$, then $$\begin{aligned} \Pr&\left[Z^n \in T_{\epsilon_2}^{(n)} \left( p_{XYZ} \mid x^n,y^n \right) \mid Y^n = y^n\right]\\ &= \Pr\left[Z^n \in T_{\epsilon_2}^{(n)}\left(p_{XYZ} \mid x^n,y^n\right) \mid X^n = x^n,Y^n = y^n\right]\\ &\geq 1-\delta_2\left(n,\epsilon_1,\epsilon_2,\mu(p_{XYZ})\right).\end{aligned}$$ [10]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{} A. Wyner and J. Ziv, “The [R]{}ate-[D]{}istortion [F]{}unction for [S]{}ource [C]{}oding with [S]{}ide [I]{}nformation at the [D]{}ecoder,” *IEEE Transactions on Information Theory*, vol. 22, no. 1, pp. 1–10, 1976. I. Csisz$\acute{\text{a}}$r and J. K$\ddot{\text{o}}$rner, *Information [T]{}heory: [C]{}oding [T]{}heorems for [D]{}iscrete [M]{}emoryless [S]{}ystems*. 1em plus 0.5em minus 0.4emAcademic Press, 1981. T. Cover and J. Thomas, *Elements of [I]{}nformation [T]{}heory*.1em plus 0.5em minus 0.4emNew York: Wiley, 1991. A. H. Kaspi, “Rate-[D]{}istortion [F]{}unction when [S]{}ide-[I]{}nformation [M]{}ay [B]{}e [P]{}resent at the [D]{}ecoder,” *IEEE Transactions on Information Theory*, vol. 40, no. 6, pp. 2031–2034, 1994. C. Heegard and T. Berger, “Rate [D]{}istortion when [S]{}ide [I]{}nformation [M]{}ay [B]{}e [A]{}bsent,” *IEEE Transactions on Information Theory*, vol. 31, no. 6, pp. 727–734, 1985. C. Tian and S. Diggavi, “On [M]{}ultistage [S]{}uccessive [R]{}efinement for [W]{}yner-[Z]{}iv [S]{}ource [C]{}oding with [D]{}egraded [S]{}ide [I]{}nformations,” *IEEE Transactions on Information Theory*, vol. 53, no. 8, pp. 2946–2960, 2007. C. Tian and S. N. Diggavi, “Side-[I]{}nformation [S]{}calable [S]{}ource [C]{}oding,” *IEEE Transactions on Information Theory*, vol. 54, no. 12, pp. 5591–5608, 2008. A. Sgarro, “Source [C]{}oding with [S]{}ide [I]{}nformation at [S]{}everal [D]{}ecoders,” *IEEE Transactions on Information Theory*, vol. 23, no. 2, pp. 179–182, 1977. Y. Steinberg and N. Merhav, “On [S]{}uccessive [R]{}efinement for the [W]{}yner-[Z]{}iv [P]{}roblem,” *IEEE Transactions on Information Theory*, vol. 50, no. 8, pp. 1636–1654, 2004. C. Tian and S. N. Diggavi, “A [C]{}alculation of the [H]{}eegard-[B]{}erger [R]{}ate-[D]{}istortion [F]{}unction for a [B]{}inary [S]{}ource,” in *proceedings of the IEEE Information Theory Workshop*, Chengdu, China, 2006, pp. 342–346. R. Yeung, *A [F]{}irst [C]{}ourse in [I]{}nformation [T]{}heory*.1em plus 0.5em minus 0.4emKluwer Academic/Plenum Publishers, 2002. R. Gray and A. Wyner, “Source [C]{}oding for a [S]{}imple network,” *Bell System Technical Journal*, vol. 53, no. 9, pp. 1681–1721, 1974. M. Effros, “Distortion-[R]{}ate [B]{}ounds for [F]{}ixed-and [V]{}ariable-[R]{}ate [M]{}ultiresolution [S]{}ource [C]{}odes,” *IEEE Transactions on Information Theory*, vol. 45, no. 6, pp. 1887–1910, 1999. C. Tian, “Latent [C]{}apacity [R]{}egion: [A]{} [C]{}ase [S]{}tudy on [S]{}ymmetric [B]{}roadcast with [C]{}ommon [M]{}essages,” in *proceedings of the IEEE International Symposium on Information Theory*, Seoul, Korea, 2009, pp. 1834–1838. B. N. Vellambi and R. Timo, “Multi-[T]{}erminal [S]{}ource [C]{}oding: [C]{}an [Z]{}ero-[R]{}ate [E]{}ncoders [E]{}nlarge the [R]{}ate [R]{}egion?” in *proceedings of the International Zurich Seminar on Communications*, Zurich, Switzerland, 2010. R. T. Rockafellar, *Convex [A]{}nalysis*.1em plus 0.5em minus 0.4emPrinceton [U]{}niversity [P]{}ress, 1997. D. Slepian and J. Wolf, “Noiseless [C]{}oding of [C]{}orrelated [I]{}nformation [S]{}ources,” *IEEE Transactions on information Theory*, vol. 19, no. 4, pp. 471–480, 1973. M. Bakshi and M. Effros, “On [A]{}chievable [R]{}ates for [M]{}ulticast in the [P]{}resence of [S]{}ide [I]{}nformation,” in *proceeding of the IEEE International Symposium on Information Theory*, Toronto, Canada, 2008, p. 1661–1665. F. Fang-Wei and R. W. Yeung, “On the [R]{}ate-[D]{}istortion [R]{}egion for [M]{}ultiple [D]{}escriptions,” *IEEE Transactions on Information Theory*, vol. 48, no. 7, pp. 2012–2021, 2002. G. Kramer, “Topics in [M]{}ulti-[U]{}ser [I]{}nformation [T]{}heory,” *Foundations and Trends in Communications and Information Theory*, vol. 4, no. 4–5, pp. 265–444, 2008. [^1]: This work was funded by the Australian Research Council Grant DP0880223. [^2]: R. Timo, T. Chan and A. Grant are with the Institute for Telecommunications Research, University of South Australia. Email: {roy.timo, terence.chan, alex.grant}@unisa.edu.au. [^3]: Kaspi’s result, [@Kaspi-Nov-1994-A Thm. 2], gives an alternative characterisation of $R(d_1,d_2)$ that uses one auxiliary random variable. [^4]: The joint probability distribution of $(X,$ $Y_1,$ $Y_2,$ $\ldots,$ $Y_t)$ can be manipulated to form the Markov chain $X \minuso Y_t \minuso Y_{t-1} \minuso \cdots \minuso Y_1$ without altering $R({\mathbf{d}})$. We discuss this problem in detail in Section \[Sec:2:C\]. [^5]: It is possible to remove this assumption and extend the results of this paper to general reconstruction alphabets and per-letter distortion measures using the procedure given in [@Yeung-2002-B Sec. 9.1]. [^6]: In general, it is difficult to prove the equivalence of *asymptotically-vanishing rates* and *zero-capacity channels* (i.e. “deleting the channel”) without such a rate-transfer argument. See, for example,  [@Vellambi-Mar-2010-C]. [^7]: In Section \[Sec:4\], we will take a slightly more general approach wherein the mutual information terms in  – rather than the minimization set ${\mathscr{P}}({\mathbf{d}})$ – are modified to produce an upper bound for $R({\mathbf{d}})$. We would like to thank Dr. Chao Tian as well as an anonymous reviewer for suggesting this more general approach. [^8]: One can invoke the Support Lemma [@Csiszar-1981-B Pg.310] to upper bound the cardinality of each set ${\mathscr{U}}_{{\mathscr{S}}_j}$. Note, these bounds will depend on the particular choice of list $v$. [^9]: We have reviewed the relevant $\epsilon$-letter typical results in Appendix \[App:B\] for convenience; a more detailed treatment can be found in [@Kramer-2008-A]. [^10]: If there are two-or-more such codewords, we assume that the encoder selects one codeword arbitrarily and sends the corresponding indices.

ArXiv

--- abstract: | Numerical simulations of spin glass models with continuous variables set the problem of a reliable but efficient discretization of such variables. In particular, the main question is how fast physical observables computed in the discretized model converge toward the ones of the continuous model when the number of states of the discretized model increases. We answer this question for the XY model and its discretization, the $q$-state clock model, in the mean-field setting provided by random graphs. It is found that the convergence of physical observables is exponentially fast in the number $q$ of states of the clock model, so allowing a very reliable approximation of the XY model by using a rather small number of states. Furthermore, such an exponential convergence is found to be independent from the disorder distribution used. Only at $T=0$ the convergence is slightly slower (stretched exponential). Thanks to the analytical solution to the $q$-state clock model, we compute accurate phase diagrams in the temperature versus disorder strength plane. We find that, at zero temperature, spontaneous replica symmetry breaking takes place for any amount of disorder, even an infinitesimal one. We also study the one step of replica symmetry breaking (1RSB) solution in the low-temperature spin glass phase. author: - Cosimo Lupo - 'Federico Ricci-Tersenghi' bibliography: - 'myBiblio.bib' title: 'Approximating the XY model on a random graph with a $q$-state clock model' --- Introduction ============ The theory of spin glasses, and disordered systems in general, has been mostly developed using the Ising variables [@Book_MezardEtAl1987]. And the same is true also in different but related fields of research, such as discrete combinatorial optimization, where the Boolean variables play a prominent role [@Book_MezardMontanari2009]. The reason for this is not surprising: even with simple dichotomous variables, these models are very complex and show a very rich behavior. Thus the interest in moving to $m$-component variables living in a higher-dimensional space (e.g. the XY or Heisenberg spins), which would make computations harder, has been limited and this is witnessed by the scarcer literature available with respect to the Ising case. Nonetheless, it is known from the experiments performed on spin glass materials having different degrees of anisotropy that the behavior of Ising-like models is different from Heisenberg-like models [@BertEtAl2004]. This difference is particularly evident in some properties, like rejuvenation and memory effects, that are considered as a trademark of the complex and hierarchical organization of spin glass long-range order. Numerical simulations of Ising spin glasses at present have not found any clear evidence for rejuvenation [@PiccoEtAl2001; @MaioranoEtAl2005] and new extensive numerical simulations of XY or Heisenberg spin glasses will be required in the near future. Up to now, most of the numerical simulations of XY or Heisenberg spin glasses have been performed on three-dimensional lattices with the main aim of understanding the role of the chiral long-range order [@Villain1977; @OliveEtAl1986; @KawamuraTanemura1987; @Kawamura1992; @HukushimaKawamura2000; @KawamuraLi2001; @LeeYoung2003; @CamposEtAl2006; @FernandezEtAl2009], but $m$-component variables naturally appear in other interesting problems, like the synchronization problems [@Book_Kuramoto1975; @AcebronEtAl2005; @BandeiraEtAl2014; @JavanmardEtAl2016] and the very recent field of random lasers [@GordonFischer2002; @AngelaniEtAl2006; @AntenucciEtAl2015], just for citing a few. From the analytical point of view, models with $m$-component variables have been mostly studied on fully connected topologies [@KirkpatrickSherrington1978; @deAlmeidaEtAl1978; @ElderfieldSherrington1982a; @ElderfieldSherrington1982b], often giving results quite different with respect to the ones of the Ising case. For example, the Gabay-Toulose critical line [@GabayToulouse1981] exists only for $m>1$. A limiting case that has been studied in some detail is the one where the number of components diverges ($m\to\infty$). In this limit, some analytical computations can be performed [@AspelmeierMoore2004; @beyer2012one; @JavanmardEtAl2016], although it is worth stressing that in the $m\to\infty$ limit the free energy landscape becomes much less complex (since the energy function is convex), and thus the phase diagrams simplify. Also, the way the low-temperature physics changes by increasing the number of components is a very interesting problem that deserves specific studies [@baity2015inherent; @ricci2016performance]. Here, we are interested in models with continuous variables, but with a small number of components: this is the reason why we choose to study the XY model ($m=2$). A key difference between Ising ($m=1$) models and vector ($m>1$) models is that variables in the latter are continuous. This may be bothersome both in analytical and numerical computations: in the former case dealing with probability distributions on an $m$-dimensional unit sphere can not be done exactly and requires strong approximations [@JavanmardEtAl2016], while in numerical simulations working with discrete variables often allows one to better optimize the simulation code (e.g. by using look-up tables). It is thus natural to ask how good can be a discrete approximation to a vector model. It is well known that the discretization of a ferromagnetic system of vector spins in the low-temperature region works very badly, e.g. as it happens for SU(3) symmetry in lattice gauge theories [@GrosseKuhnelt1982]. The main reason for this failure is the fact small thermal fluctuations around the fully ordered ground state are not well described by the discretized model. However, when quenched disorder is introduced in the model, the situation may dramatically change. Indeed, in this case the presence of frustration makes low-energy configurations not fully ordered and much more abundant: the inability of the discretized model to correctly describe small fluctuations may be not so relevant as long as it can cope with the many low-energy configurations. To understand how good a discretized model can reproduce the physics of a vector model, in the present work we study the $q$-state clock model, which is a discretized version of the XY model. We will mainly consider how physical observables change when increasing the number $q$ of states and check how fast the XY model is approached in the $q\to\infty$ limit. A similar question has been answered in a very recent work [@MarruzzoLeuzzi2015] for the XY model with four-spin interactions, while here we only consider models with two-spin interactions, that belong to a different universality class with respect to those studied in Ref. [@MarruzzoLeuzzi2015]. More precisely, we are going to consider the $q$-state clock model in both the ferromagnetic and the spin glass versions with different kinds of disorder. We will focus on models defined on random regular graphs. We will study phase diagrams in the temperature versus disorder strength plane at many values of $q$. Most of the computations are analytic, within the replica symmetric (RS) ansatz. Finally, by considering the ansatz with one step of replica symmetry breaking (1RSB), we will try to understand whether the universality class changes by varying $q$. A short comment on chiral ordering, which is a very debated issue in models with continuous variables defined on regular lattices [@kawamura1991chiral; @LeeYoung2003; @obuchi2013monte; @baity2014phase]. When a model with continuous spin variables is defined on a random graph the topological defects play no longer any role. Indeed, typical loops in random graphs are $O(\ln{N})$ long, while short loops are rare, with a $O(1/N)$ density. Chiral ordering can not take place on random graphs, so we do not enter at all into the debate on the coupling/decoupling between spin and chiral degrees of freedom. The structure of this paper is the following. In Sec. \[sec:the\_models\] we introduce the models we are going to study, the XY model and its discretized version, the $q$-state clock model. Then in Sec. \[sec:bp\_cavity\_method\_rs\], we recall the basic features of cavity method in the replica symmetric framework and discuss about its validity. In Sec. \[sec:rs\_solution\_xy\_clock\], we solve the XY model and the $q$-state clock model on random regular graphs within the replica symmetric ansatz, by exploiting the cavity method both analytically and numerically. Then, in Sec. \[sec:convergence\], we actually study the convergence of physical observables for the $q$-state clock model when $q$ is increased, reaching the limiting values given by the XY model. In the end, in Sec. \[sec:bp\_cavity\_method\_1rsb\], we extend the cavity method to one step of replica symmetry breaking and we apply it to the $q$-state clock model, in order to give an insight of the exact solution (which should be full replica symmetry breaking) and to see if and when the universality class changes when $q$ increases. The models {#sec:the_models} ========== XY model -------- The simplest case of continuous variables models — or vector spin models — is the *XY model* whose Hamiltonian can be written either as $$\mathcal{H}[\{{\vec{\sigma}}\}] = -\sum_{\braket{ij}}J_{ij}\,{\vec{\sigma}}_i\cdot{\vec{\sigma}}_j\,,$$ where spins have $m=2$ components (${\vec{\sigma}}_i\in\mathbb{R}^2$) and unit norm ($|{\vec{\sigma}}_i|=1$), either as $$\mathcal{H}[\{\theta\}] = -\sum_{\braket{ij}}J_{ij} \cos(\theta_i-\theta_j)\,, \label{eq:H_xy_pm_J}$$ with $\theta_i\in[0,2\pi)$. The sums run over the pairs of nearest-neighbor vertices on a generic graph. Couplings $J_{ij}$ can be all positive in ferromagnetic models or can be extracted from a distribution $\mathbb{P}_J(J_{ij})$ having support on both positive and negative regions in the spin glass case. Despite of its simplicity, the XY model shows a lot of interesting features and allows new phenomena to rise up with respect to the Ising case: for example, thanks to the continuous nature of its variables, at very low temperatures small fluctuations are allowed in the XY model. In turn, this may produce null or very small eigenvalues in the spectral density of the Hessian matrix of the model. Other interesting features of the XY model — and of vector spin models in general — regard its behavior in an external field [^1], where e.g. one can observe different kinds of transitions: the de Almeida-Thouless transition [@deAlmeidaThouless1978] — which is present also in the Ising case — and the Gabay-Toulouse transition [@GabayToulouse1981; @CraggEtAl1982], which does not show up in Ising models. $q$-state clock model --------------------- The problem of discretizing the continuous variables of the XY model may naturally arise when one wishes to simulate very efficiently the model (e.g. by the use of look-up tables) or even when a reduction in the variables domain is required in the search for an analytically treatable solution [@JavanmardEtAl2016]. The simplest way to discretize the XY model is to allow each spin ${\vec{\sigma}}_i$ to take only a finite number $q$ of directions along the unit circle, equally spaced by the elementary angle $2\pi/q$. Taking $q$ large enough the error committed by the discretization should be negligible. The Hamiltonian of the $q$-state clock model has formally the same expression in (\[eq:H\_xy\_pm\_J\]), but with angles taking value in a finite set $\theta_i \in \{0,2\pi/q,4\pi/q,\ldots,2\pi(q-1)/q\}$. While the XY model is recovered in the $q\to\infty$ limit, very small values of $q$ are expected to produce a rather different behavior: in particular, $q=2$ corresponds to the Ising model, the $q=3$ clock model can be mapped to a $3$-state Potts model and the $4$-state clock model is nothing but a *double* Ising model, apart from a rescaling of couplings $J_{ij}$ by a factor $1/2$ [@NobreSherrington1986]. For larger $q$ values, the XY model is approached and we will try to understand how fast is this process in the different parts of the phase diagram. In particular we are interested in the role played by the disorder. Indeed, the slowest convergence is expected at very low temperatures, and in this region the different kinds of long range order (ferromagnetic or spin glass) may vary sensibly the convergence to the XY model. Our naive expectation is that the strong frustration present in a spin glass phase may produce low-energy configurations in the clock model which are less rigid (with respect to the ferromagnetic case) and thus may have small energy fluctuations, so making the disordered $q$-state clock model more similar to the disordered XY model, with respect to the corresponding ferromagnetic versions. Random graphs ------------- Mean-field approximations are correct for models defined on fully connected graphs and on sparse random graphs; the latter case, apart from being more general (the fully connected topology can be recovered in the limit of large mean degree), is much more interesting, because the variables have a number of neighbors $O(1)$ and this produces fluctuations in the local environment of each variable, that closely resembles what happens in systems defined in a finite dimensional space. In this work for simplicity we focus on models defined on random $c$-regular graphs (RRG), where each vertex has exactly $c$ neighbors. Given that we are going to solve the cavity equations via the population dynamics algorithm, we do not need to specify the algorithm to generate any specific realization of the RRG (e.g. the configurational model). Replica symmetric cavity method {#sec:bp_cavity_method_rs} =============================== Exact solutions of spin glass models on random graphs is highly non-trivial, due to the spatial fluctuations naturally induced by the topology of such graphs. Furthermore, as in the Sherrington-Kirkpatrick model, we expect the XY model to require a full replica symmetry breaking (fRSB) scheme to be exactly solved [@Book_MezardEtAl1987]. Anyway, a large class of models defined on random graphs can be solved just by exploiting a single step of replica symmetry breaking (1RSB), as, e.g. the diluted $p$-spin [@MezardEtAl2003], the coloring problem, and the Potts model [@ZdeborovaKrzakala2007; @KrzakalaZdeborova2008]. In some cases, it can be demonstrated that such solutions are stable towards further steps of replica symmetry breaking [@MontanariEtAl2004]. Here we start from the simplest replica symmetric (RS) ansatz, which actually corresponds to the well known Bethe approximation [@Bethe1935]. The RS ansatz is always correct for models defined on a tree and for models defined on a random graph if model correlations decay fast enough [@Book_MezardMontanari2009]. The locally tree-like structure of random graphs allows one to use the cavity method, either at the RS level [@MezardParisi1987] or at the 1RSB level [@MezardParisi2001; @MezardParisi2003]. The belief propagation (BP) algorithm [@Book_Pearl1988; @YedidiaEtAl2003] is a convenient recursive algorithm to solve the RS self-consistency equations on a given graph. However, here we are interested in understanding the physical properties of typical random graphs, and this can be better achieved by the use of the population dynamics algorithm [@Book_MezardMontanari2009]. Cavity equations at finite temperature -------------------------------------- The basic idea of cavity method is that on a tree the local marginals $$\eta_i(\theta_i) = \sum_{\underline\theta \setminus \theta_i} \mu(\underline\theta) \qquad , \quad \mu(\underline\theta) = \frac1Z e^{-\beta {\mathcal{H}}(\underline\theta)}$$ can be written in term of cavity marginals $\eta_{i\to j}(\theta_i)$, namely the marginal probability of variable $\theta_i$ in the graph where the coupling $J_{ij}$ has been removed (hence the name cavity), $$\eta_{i}(\theta_i)=\frac{1}{{\mathcal{Z}}_i}\,\prod_{k\in\partial i}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\eta_{k\to i}(\theta_k)\,, \label{eq:fullMarginals}$$ where ${\mathcal{Z}}_i$ is a normalization constant $${\mathcal{Z}}_i=\int d\theta_i\prod_{k\in\partial i}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\eta_{k\to i}(\theta_k)$$ and $\partial i$ is the set of neighbors of spin $i$. ![Due to the *Bethe approximation*, when all edges around site $i$ are removed from the graph, then its neighbors become uncorrelated. So marginal probability distribution $\eta_{i\to j}(\theta_i)$ of site $i$ when edge $\braket{ij}$ has been removed from the graph can be computed iteratively from marginals $\eta_{k\to i}(\theta_k)$’s, $k\in\partial i \setminus j$. Alternatively, the *belief* coming out from site $i$ to site $j$ is given by all other beliefs entering site $i$ from sites $k$’s belonging to $\partial i\setminus j$.[]{data-label="fig:random_graphs"}](treelike_graph_STANDALONE){width="0.85\columnwidth"} Referring to the notation in Fig. \[fig:random\_graphs\], it is not difficult to write down self-consistency equations among the cavity marginals [@YedidiaEtAl2003; @Book_MezardMontanari2009]: $$\begin{split} \eta_{i\to j}(\theta_i) &= {\mathcal{F}}(\{\eta_{k\to i},J_{ik}\})\\ &= \frac{1}{{\mathcal{Z}}_{i\to j}}\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,\eta_{k\to i}(\theta_k)\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,, \label{eq:self_cons_eta} \end{split}$$ where ${\mathcal{Z}}_{i\to j}$ is just the normalization constant $${\mathcal{Z}}_{i\to j}=\int d\theta_i\prod_{k\in\partial i\setminus j}\int d\theta_k\,\eta_{k\to i}(\theta_k)\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,.$$ We call the above BP equations, where BP stands both for Bethe-Peierls and Belief Propagation. The addition of site $i$ and couplings $J_{ik}$ with $k\in\partial i \setminus j$, causes a free energy shift $\Delta F_{i\to j}$ in the system, which is directly related to the normalization constant ${\mathcal{Z}}_{i\to j}$ via $${\mathcal{Z}}_{i\to j}=e^{\,-\beta\,\Delta F_{i\to j}}\,.$$ Once the cavity marginals satisfying Eq. (\[eq:self\_cons\_eta\]) have been computed, the free energy density in a graph of $N$ vertices is given by $$f_N(\beta)=-\frac{1}{\beta N}\,\Bigl(\sum_i\ln{{\mathcal{Z}}_i}-\sum_{\braket{ij}}\ln{{\mathcal{Z}}_{ij}}\Bigr)\,, \label{eq:f_rs_bp}$$ where ${\mathcal{Z}}_{ij}$ is given by $${\mathcal{Z}}_{ij}=\int d\theta_i d\theta_j\,e^{\,\beta J_{ij}\cos{(\theta_i-\theta_j)}}\,\eta_{i\to j}(\theta_i)\,\eta_{j\to i}(\theta_j)\,.$$ As usual, the true free energy density $f(\beta)$ is obtained in the thermodynamic limit, $f(\beta)=\lim_{N\to\infty}f_{N}(\beta)$. From previous expressions of free energy density $f(\beta)$ a crucial consequence of the Bethe approximation on sparse graphs rises up: extensive quantities can be computed as a sum of local terms involving sites and edges of the graphs [@Book_MezardMontanari2009]. Indeed, the same holds for internal energy density $e(\beta)=\partial_{\beta}(\beta f(\beta))$, while magnetizations can be easily computed from the marginal probability distributions $\eta_i(\theta_i)$. So in the end it is enough to solve self-consistency equations (\[eq:self\_cons\_eta\]) in order to be able to compute all the physical observables of the system. So far, we have considered a given instance for the underlying RRG, and hence for the set of couplings $\{J_{ij}\}$. But when dealing with random topologies, in general, physical observables have to be computed by averaging over all the possible realizations of the graph and the disorder. This task can be accomplished by noting that cavity messages $\eta_{i\to j}$’s arriving from a branch of the tree are distributed according to a probability distribution $P[\eta_{i\to j}]$, and so BP equations (\[eq:self\_cons\_eta\]) can be reinterpreted as a single distributional equation: $$\begin{gathered} P[\eta_{i\to j}]=\mathbb{E}_{G,J}\int\prod_{k=1}^{d_i-1}{\mathcal{D}}\eta_{k\to i}\,P[\eta_{k\to i}]\\ \times\delta\Bigl[\eta_{i\to j}-{\mathcal{F}}[\{\eta_{k\to i},J_{ik}\}]\Bigr], \label{eq:self_cons_eta_distrib}\end{gathered}$$ where $\mathbb{E}_{G,J}$ stands for the expectation value over the realization of the graph and of the disorder, and ${\mathcal{F}}$ is defined in Eq. (\[eq:self\_cons\_eta\]). In particular, for the RRG ensemble, all the degrees $\{d_i\}$ are equal to $c$ and so the corresponding average can be ignored. Accordingly, the free energy density averaged over the RRG ensemble is given by $$f(\beta)=-\frac{1}{\beta}\,\mathbb{E}_{\eta,J}\Bigl[\ln{{\mathcal{Z}}_i}\Bigr]+\frac{c}{2\beta}\,\mathbb{E}_{\eta,J}\Bigl[\ln{{\mathcal{Z}}_{ij}}\Bigr]\,, \label{eq:f_rs_distrib}$$ where the average over $\eta$ is made according to $P[\eta]$ satisfying Eq. (\[eq:self\_cons\_eta\_distrib\]). Cavity equations at zero temperature ------------------------------------ When temperature $T$ goes to zero, the inverse temperature $\beta$ diverges and the integrals in Eq. (\[eq:self\_cons\_eta\]) can be solved by the saddle point method. We rewrite the cavity marginals $\eta_{i\to j}$ as large deviation functions in $\beta$: $$\eta_{i\to j}(\theta_i)\equiv e^{\,\beta h_{i\to j}(\theta_i)}\,,$$ with cavity fields $h_{i\to j}(\theta_i)$ being nonpositive functions. The normalization on $\eta_{i\to j}$ requires to appropriately shift $h_{i\to j}$ such that its maximum has zero height: $$\max\nolimits_{\theta_i} h_{i\to j}(\theta_i) = 0\,. \label{eq:norm_h_T_0}$$ In the $T\to 0$ limit, the BP equations become $$\begin{split} h_{i\to j}(\theta_i) &= {\mathcal{F}}_0[\{h_{k\to i}, J_{ik}\}]\\ &= \sum_{k\in\partial i\setminus j}\max_{\theta_k}{\left[h_{k\to i}(\theta_k)+J_{ik}\cos{(\theta_i-\theta_k)}\right]}\,, \label{eq:self_cons_h_T_0} \end{split}$$ up to an additive constant due to the normalization condition in Eq. (\[eq:norm\_h\_T\_0\]). Taking the average over the disorder (couplings and RRGs of fixed degree $c$), we get again a self-consistency equation for functional probability distribution $P[h]$ of cavity fields $h$: $$P[h]=\mathbb{E}_{J}\int \prod_{i=1}^{c-1} {\mathcal{D}}h_i\,P[h_i]\,\delta\Bigl[h-{\mathcal{F}}_0[\{h_i, J_i\}]\Bigr]\,. \label{eq:self_cons_h_T_0_distrib}$$ The zero-temperature expression for the free energy density $f$ can be written in terms of the $P[h]$ satisfying Eq. (\[eq:self\_cons\_h\_T\_0\_distrib\]), $$f=-\mathbb{E}_{h,J}\,\bigl[f_i\bigr]+\frac{c}{2}\,\mathbb{E}_{h,J}\,\bigl[f_{ij}\bigr]\,, \label{eq:f_rs_T_0_distrib}$$ where $$\begin{aligned} f_i &=& \max_{\theta_i}{\biggl[\,\sum_{k\in\partial i}\max_{\theta_k}{\bigl[h_{k\to i}(\theta_k)+J_{ik}\cos{(\theta_i-\theta_k)}\bigr]}\biggr]}\,,\\ f_{ij} &=& \max_{\theta_i ,\theta_j}{\Bigl[h_{i\to j}(\theta_i)+h_{j\to i}(\theta_j)+J_{ij}\cos{(\theta_i-\theta_j)}\Bigr]}\,.\end{aligned}$$ Validity of the RS cavity method {#subsec:validity_rs_method} -------------------------------- So far, we have implicitly assumed that the set of BP equations always admit a (unique) solution. However, in general, it is not always true, and so it is not obvious to get at a solution to eq. (\[eq:self\_cons\_eta\]). This fact is very intimately related to the presence of loops of finite size or strong correlations in the model. Indeed, if the graph is a tree or locally tree-like, in the cavity graph where vertex $i$ has been removed the following factorization holds: $$\mu_{\partial i}\bigl(\{\theta_k\}_{k\in\partial i}\bigr)=\prod_{k\in\partial i}\eta_{k\to i}(\theta_k)\,, \label{eq:factor_mu}$$ which in turn allows one to derive Eq. (\[eq:self\_cons\_eta\]). But when the graph is not locally tree-like, i.e. it has short loops, or when correlations in the model are so strong that even in the cavity graphs without vertex $i$ the marginals $\eta_{k\to i}(\theta_k)$ are correlated, then factorization in Eq. (\[eq:factor\_mu\]) does not hold anymore, and the error committed in assuming it can be non negligible, even in the case of very large graphs. From the formal point of view, a rigorous proof of the conditions under which the replica symmetric cavity method is correct does not yet exist. A possible condition that has to be fulfilled regards the uniqueness of the Gibbs measure, $$\mu(\{{\vec{\sigma}}\})=\frac{1}{{\mathcal{Z}}}e^{\,-\beta{\mathcal{H}}[\{{\vec{\sigma}}\}]}, \label{eq:gibbs_measure}$$ meaning that the clustering property holds and that each spin ${\vec{\sigma}}_i$ in the bulk of the system is independent from any choice of boundary conditions. However, this is a very strict condition, and often it is observed that RS cavity method still provides a correct result even when Gibbs measure (\[eq:gibbs\_measure\]) ceases to be unique. So a weaker condition to be fulfilled regards the extremality of the Gibbs measure (\[eq:gibbs\_measure\]), so that even when it is no longer unique but extremal, then in the thermodynamic limit the unique relevant solution is still the RS one [@KrzakalaEtAl2007]. Roughly speaking, the extremality of the Gibbs measure means that the behavior of a spin ${\vec{\sigma}}_i$ in the bulk of the system depends only on a set of boundary conditions with null measure. From the analytical and numerical point of view, instead, there are several and equivalent approaches for the study of the stability of the RS solution of a given model. For example, one can apply the 1RSB cavity method (Sec. \[sec:bp\_cavity\_method\_1rsb\]) and check if it reduces to the RS solution, which is then exact in this case. From a more physical point of view, one can compute the spin glass susceptibility $\chi_{SG}$, $$\chi_{SG}=\frac{1}{N}\,\sum_{i,j}\braket{{\vec{\sigma}}_i\cdot{\vec{\sigma}}_j}^2_c,$$ and see where it diverges, signaling a phase transition. The computation of $\chi_{SG}$ can be done in an iterative way via the so-called *susceptibility propagation* algorithm, which in practice corresponds to check the growth rate of small perturbations around the fixed-point cavity messages: if these perturbations tend to grow, then the RS fixed point is no longer stable and the RS ansatz is only approximate. In Appendix \[app:susc\_propag\] we deal with this topic in a deeper way, explicitly linearizing the BP equations — both at finite and zero temperature — and so obtaining the analytical expression of the growth rate of such perturbations. The study of the growth rate of perturbations can be done either on *given instances* either in *population dynamics* taking the average over the ensemble of graphs and coupling realizations. We exploit the second approach, redirecting the reader to Ref. [@Thesis_Zdeborova2009] for the demonstration of the equivalence between them. RS solution of the XY model and the $q$-state clock model {#sec:rs_solution_xy_clock} ========================================================= Let us now apply the cavity method to the XY model and to the $q$-state clock model, in order to find their RS solution. We use first the bimodal distribution $J_{ij}=\pm 1$, with different weights $p$ and $1-p$, and we derive analytical phase diagrams in the $p$ versus $T$ plane. We move then to the low-temperature region and numerically identify, when varying $p$, three phases: the ferromagnetic (RS) one, the spin glass (RSB) one and also a *mixed* phase between them, characterized by both a non vanishing magnetization and a breaking of the replica symmetry [@CastellaniEtAl2005]. Finally, we use a different disorder distribution, usually called *gauge glass*, such that each interaction $J_{ij}$ acts through a rotation of a random angle $\omega_{ij}$, still belonging to one of the allowed directions of the $q$-state clock model. This further choice of disorder allows us to study the behavior of the clock model also for odd values of $q$, since now there is no reflection symmetry to take care of. Furthermore, it should be more “physical” for the XY model, where in principle disorder can not only cause a complete inversion of a spin with respect to the direction it had in the ferromagnetic case, but also a small rotation. Analytical solution with the bimodal distribution of couplings {#sec:analyticalBimodal} -------------------------------------------------------------- Let us consider the following disorder distribution: $$\mathbb{P}_{J}(J_{ij})=p\,\delta(J_{ij}-1)+(1-p)\,\delta(J_{ij}+1)\;, \label{eq:disord_distr_pm_J}$$ with $p\in[1/2,1]$, such that $p=1$ corresponds to a pure ferromagnet and $p=1/2$ to an unbiased spin glass. Let us start our analytical computation from the XY model. In order to find a solution to BP equations (\[eq:self\_cons\_eta\]) for the XY model, it is useful to expand cavity marginals $\eta_{k\to i}$ in Fourier series: $$\eta_{k\to i}(\theta_k)= \frac{1}{2\pi}\biggl\{1+\sum_{l=1}^{\infty}\Bigl[a_{l}^{(k\to i)}\cos{(l\theta_k)}+b_{l}^{(k\to i)}\sin{(l\theta_k)}\Bigr]\biggr\},$$ where Fourier coefficients are defined as usual: $$\left\{ \begin{aligned} &a_l^{(i\to j)}=2\int d\theta_i\,\eta_{i\to j}(\theta_i)\,\cos{(l\theta_i)},\\ &b_l^{(i\to j)}=2\int d\theta_i\,\eta_{i\to j}(\theta_i)\,\sin{(l\theta_i)}. \end{aligned} \right. \label{eq:def_fourier_coeff_XY}$$ Note that, in general, Fourier coefficients are different for each cavity marginal. Substituting this expansion in the right-hand side of BP equations (\[eq:self\_cons\_eta\]), we get $$\begin{split} &\eta_{i\to j}(\theta_i)=\frac{1}{{\mathcal{Z}}_{i\to j}}\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\\ &\,\,\,\,\frac{1}{2\pi}\biggl\{1+\sum_{l=1}^{\infty}\Bigl[a_{l}^{(k\to i)}\cos{(l\theta_k)}+b_{l}^{(k\to i)}\sin{(l\theta_k)}\Bigr]\biggr\}, \label{eq:BPclock} \end{split}$$\ where ${\mathcal{Z}}_{i\to j}$ now reads $$\begin{split} &{{\mathcal{Z}}_{i\to j}} = \int d\theta_i \prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\\ &\,\,\frac{1}{2\pi}\biggl\{1+\sum_{l=1}^{\infty}\Bigl[a_{l}^{(k\to i)}\cos{(l\theta_k)}+b_{l}^{(k\to i)}\sin{(l\theta_k)}\Bigr]\biggr\}\,. \label{eq:Zclock} \end{split}$$ Then, integrals in $d\theta_k$ can be performed by introducing modified Bessel functions of the first kind [@Book_AbramowitzStegun1964]: $$\begin{aligned} &&\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\cos{(l\theta_k)}=2\pi\,I_l(\beta J_{ik})\cos{(l\theta_i)}\,,\\ &&\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\sin{(l\theta_k)}=2\pi\,I_l(\beta J_{ik})\sin{(l\theta_i)}\,,\\ &&\text{where} \quad I_n(x)\equiv\frac{1}{2\pi}\int_0^{2\pi}d\theta\,e^{x\cos{\theta}}\cos{(n\theta)}\,. \label{eq:modif_Bes_func_xy}\end{aligned}$$ BP equations thus become $$\begin{split} &\eta_{i\to j}(\theta_i)=\frac{1}{{\mathcal{Z}}_{i\to j}}\,\prod_{k\in\partial i\setminus j}\biggl\{I_0(\beta J_{ik})\\ &\quad +\sum_{l=1}^{\infty}I_l(\beta J_{ik})\Big[a_{l}^{(k\to i)}\cos{(l\theta_i)}+b_{l}^{(k\to i)}\sin{(l\theta_i)}\Bigr]\biggr\}, \end{split}$$ where also the normalization constant ${\mathcal{Z}}_{i\to j}$ has to be rewritten in terms of Bessel functions by following the same steps. At this point, if we substitute this expression for $\eta_{i\to j}$ into (\[eq:def\_fourier\_coeff\_XY\]), we get a set of self-consistency equations for the Fourier coefficients: $$\left\{ \begin{aligned} &a_{l}^{(i\to j)}=\frac{2}{{\mathcal{Z}}_{i\to j}}\int d\theta\,\cos{(l\theta)}\prod_{k\in\partial i \setminus j} \biggl\{I_0(\beta J_{ik})+\sum_{p=1}^{\infty}I_p(\beta J_{ik})\Bigl[a_p^{(k\to i)}\cos{(p\,\theta)}+b_p^{(k\to i)}\sin{(p\,\theta)}\Bigr]\biggr\},\\ &b_{l}^{(i\to j)}=\frac{2}{{\mathcal{Z}}_{i\to j}}\int d\theta\,\sin{(l\theta)}\prod_{k\in\partial i \setminus j} \biggl\{I_0(\beta J_{ik})+\sum_{p=1}^{\infty}I_p(\beta J_{ik})\Bigl[a_p^{(k\to i)}\cos{(p\,\theta)}+b_p^{(k\to i)}\sin{(p\,\theta)}\Bigr]\biggr\},\\ &{\mathcal{Z}}_{i\to j} = \int d\theta\,\prod_{k\in\partial i \setminus j} \biggl\{I_0(\beta J_{ik})+\sum_{p=1}^{\infty}I_p(\beta J_{ik})\Bigl[a_p^{(k\to i)}\cos{(p\,\theta)}+b_p^{(k\to i)}\sin{(p\,\theta)}\Bigr]\biggr\}. \end{aligned} \right. \label{eq:self_cons_a_b}$$ It is straightforward to verify that, in absence of any external field, the uniform distribution over the $[0,2\pi]$ interval is a solution of Eq. (\[eq:self\_cons\_eta\]) at any temperature: $$\eta_{i\to j}(\theta_i)=\frac{1}{2\pi}\quad,\qquad\forall i\to j, \label{eq:para_sol_xy_1}$$ and obviously it corresponds to a vanishing solution for the self-consistency equations (\[eq:self\_cons\_a\_b\]): $$a_{l}^{(i\to j)}=b_{l}^{(i\to j)}=0\quad,\qquad\forall i\to j\,,\,\forall l. \label{eq:para_sol_xy_2}$$ This solution is nothing but the paramagnetic one, characterized by a set of null local magnetizations. Furthermore, it is very easy to compute the expression for the free energy density, $$f_{para}(\beta)=-\frac{1}{\beta}\ln{2\pi}-\frac{c}{2\beta}\ln{I_0(\beta)},$$ as well as the expression for the internal energy density: $$e_{para}(\beta)=-\frac{c}{2}\,\frac{I_1(\beta)}{I_0(\beta)}.$$ Assuming that, lowering the temperatures, a second-order phase transition takes place, we expect some of the Fourier coefficients to become nonzero in a continuous way. To identify the critical line $T_c(p)$ we expand Eq. (\[eq:self\_cons\_a\_b\]) to linear order in the Fourier coefficients. At linear order, the normalization constant is just $${\mathcal{Z}}_{i\to j} = 2\pi\,\prod_{k\in\partial i\setminus j}I_0(\beta J_{ik}).$$ Expanding also numerators in Eq. (\[eq:self\_cons\_a\_b\]), and restricting to $a$’s coefficients (since expressions for $b$’s coefficients are similar), we get $$a_l^{(i\to j)}=\sum_{k\in\partial i\setminus j}\frac{I_l(\beta J_{ik})}{I_0(\beta J_{ik})}a_l^{(k\to i)} \label{eq:self_cons_a_linear}$$ and, taking the average over the disorder distribution and the graph realization, one gets a self-consistency equation for the mean values of Fourier coefficients, $$\overline{a_l} = \mathbb{E}_{G,J}\,\left[a_l^{(i\to j)}\right],$$ that depends on the parity of the coefficient, namely $$\begin{aligned} \overline{a_l} = (c-1)(2p-1)\frac{I_l(\beta)}{I_0(\beta)} \overline{a_l} &&\qquad \text{for $l$ odd},\\ \overline{a_l} = (c-1)\frac{I_l(\beta)}{I_0(\beta)} \overline{a_l} &&\qquad \text{for $l$ even}. \label{mom1}\end{aligned}$$ For $\beta>0$, the ratios $I_l(\beta)/I_0(\beta)$ are increasing functions of $\beta$. Moreover, the inequality $I_{l+1}(\beta) < I_l(\beta)$ implies that the first Fourier coefficients to become nonzero lowering the temperature are $a_1$ (for $p$ close to $1$) and $a_2$ (for $p$ close to $1/2$). However, we have to consider that Fourier coefficients are random variables fluctuating from edge to edge. In strongly disordered models ($p$ close to $1/2$), the mean value of $a_1$ may stay zero, while fluctuations may become relevant and eventually diverge. To check for this, we compute the self-consistency equation for the second moments: $$\overline{a_l^2} =\mathbb{E}_{G,J}\,\left[\Bigl(a_l^{(i\to j)}\Bigr)^2\right] =(c-1)\frac{I_l^2(\beta)}{I_0^2(\beta)}\,\overline{a_l^2}. \label{mom2}$$ The comparison of Eqs. (\[mom1\]) and (\[mom2\]), together with the inequality $$\frac{I_2(x)}{I_0(x)} \le \frac{I_1^2(x)}{I_0^2(x)}\,,$$ lead to the conclusion that the instability of paramagnetic phase, lowering the temperature, is always driven by the instability in the first order Fourier coefficients $\{a_1^{(i\to j)}\}$. The low-temperature phase will be ferromagnetic in case the stability produces a nonzero mean value $\overline{a_1}$, which in turn leads to a nonzero magnetization. While a spin glass order prevails if the transition is such that $\overline{a_1^2}$ becomes nonzero, while $\overline{a_1}$ stays null. Which phase transition takes actually place depends on the highest critical temperature between $T_F=1/\beta_F$ and $T_{SG}=1/\beta_{SG}$, with $$\begin{aligned} (c-1)(2p-1)\frac{I_1(\beta_F)}{I_0(\beta_F)} &=& 1,\\ (c-1)\frac{I_1^2(\beta_{SG})}{I_0^2(\beta_{SG})} &=& 1.\end{aligned}$$ These results completely agree with those obtained for the XY model on Erdős-Rényi graphs in Refs. [@SkantzosEtAl2005; @CoolenEtAl2005] through a slightly different approach, namely a functional moment expansion around paramagnetic solution. Notice that $T_F$ does depend on density of ferromagnetic couplings $p$, while $T_{SG}$ does not. The *multicritical* point $(p_{mc},T_{mc})$ is located where $T_F(p)$ and $T_{SG}$ meet, i.e. $$p_{mc}=\frac{1+(c-1)^{-1/2}}{2} \qquad , \quad T_{mc} = T_{SG}\,. \label{eq:abscissa_p_mc}$$ Just below the critical temperature $$T_c=\max\big(T_F(p),T_{SG}\big)$$ the nonlinear terms in Eq. (\[eq:self\_cons\_a\_b\]) that couple different Fourier coefficients lead to the following scaling $$a_l^{(i\to j)} \propto \tau^{\,l/2} \label{eq:scalingCoeff}$$ with $\tau \equiv |T-T_c|/T_c$. Equation (\[eq:scalingCoeff\]) can be obtained noticing the following two aspects of Eq. (\[eq:self\_cons\_a\_b\]): *i)* for $l>1$ the first nonlinear term is $$a_l^{(i\to j)} \propto \prod_{k\in\partial i\setminus j} a_{p_k}^{(k \to i)},$$ with $p_k$ algebraically summing to $l$, implying $a_l \propto (a_1)^l$; *ii)* the first nonlinear term in the equation for $a_1$ is cubic and thus, close to the critical point, we have $$a_1 = (1+\tau) a_1 + A\, a_1^3$$ implying $a_1 \propto \sqrt{\tau}$. The above Fourier expansion can be used only for identifying the instability of the paramagnetic phase in absence of an external field. Unfortunately, in presence of a field or in the low-temperature phases all the Fourier coefficients become $O(1)$ and the above expansion becomes useless, since keeping few coefficients is a too drastic approximation. So, in order to complete the phase diagram and locate the critical line between the spin glass and the ferromagnetic phases that runs from $(p_{mc},T_{mc})$ to $(p_{SG},T=0)$, we will move to a numerical approach based on the RS cavity method (see Sec. \[subsec:numerical\_cavity\_method\]). Before moving to the low-temperature phase, let us conclude the analysis of the critical lines of the paramagnetic phase in the $q$-state clock model, following exactly the same Fourier expansion above. Since our aim is to understand how fast the clock model converges to the XY model increasing $q$, we would like to study a model where the $q$ dependence is smooth. Using the bimodal distribution for couplings, $J_{ij}=\pm1$, we are forced to work only with even values of $q$. Indeed, for odd values of $q$, couplings with $J_{ij}=1$ can be fully satisfied, while coupling with $J_{ij}=-1$ can not (there are no two states differing by a $\pi$ angle for $q$ odd). Although for large $q$ the differences between even and odd $q$ models vanish, for small $q$ values they lead to strong oscillations in most physical observables. For this reason we focus only on even values for $q$ as long as we use bimodal couplings. First of all, let us rewrite BP equations for the $q$-state clock model. They are still given by Eq. (\[eq:self\_cons\_eta\]), with a slight modification due to the discrete nature of the model: $$\begin{aligned} \eta_{i\to j}(\theta_{i,a})&=\frac{1}{{\mathcal{Z}}_{i\to j}}\,\prod_{k\in\partial i\setminus j}\,\sum_{b_k=0}^{q-1}\eta_{k\to i}(\theta_{k,b_k})\\ &\qquad\qquad\qquad\times e^{\,\beta J_{ik}\cos{\left(\theta_{i,a}-\theta_{k,b_k}\right)}}, \label{eq:self_cons_eta_qclock} \end{aligned}$$ where indices $a$ and $b_k$’s label the $q$ possible values for angles $\theta$’s, and ${\mathcal{Z}}_{i\to j}$ is given by $${\mathcal{Z}}_{i\to j}=\frac{2\pi}{q}\sum_{a=0}^{q-1}\prod_{k\in\partial i\setminus j}\sum_{b_k=0}^{q-1}\eta_{k\to i}(\theta_{k,b_k})\, e^{\,\beta J_{ik}\cos{\left(\theta_{i,a}-\theta_{k,b_k}\right)}}.$$ Since we are now dealing with a discrete model, in order to find a solution to BP equations (\[eq:self\_cons\_eta\_qclock\]) it is useful to expand cavity marginals $\eta_{i\to j}$ in discrete Fourier series, $$\eta_{i\to j}(\theta_{i,a})=\frac{1}{q}\sum_{b=0}^{q-1}c_b^{(i\to j)}e^{\,2\pi i\,ab/q}, \label{eq:discr_four_series}$$ where the complex coefficients $c_b^{(i\to j)}$ are given by $$c_b^{(i\to j)}=\sum_{a=0}^{q-1}\eta_{i\to j}(\theta_{i,a})\,e^{-2\pi i\,ab/q}.$$ The zero-order coefficient $c_0^{(i\to j)}$ is nothing but the sum of the values taken by cavity marginal $\eta_{i\to j}$ over the $q$ values of the angle $\theta_i$: $$c_0^{(i\to j)}=\sum_{a=0}^{q-1}\eta_{i\to j}(\theta_{i,a}).$$ If we choose to put the norm of probability distributions for the $q$-state clock model equal to $q/2\pi$, so that in the limit $q\to\infty$ we can exactly recover the marginal probability distributions for the XY model, then we have $c_0^{(i\to j)}=q/2\pi$ and we can write $$\eta_{i\to j}(\theta_{i,a})=\frac{1}{2\pi}\biggl[1+\frac{2\pi}{q}\,\sum_{b=1}^{q-1}c_b^{(i\to j)}e^{\,2\pi i\,ab/q}\biggr].$$ A different choice is to put the norm equal to $1$, so obtaining $c_0=1$. This choice gives the correct value for physical observables for a discrete model (e.g. an entropy which is always positive defined), but it will be less useful when studying the convergence of the $q$-state clock model toward the XY model. So from now on, we will always use the $q/2\pi$ normalization. It is worth noticing that $$c_m^{(i\to j)}=\left(c_{q-m}^{(i\to j)}\right)^*\,,$$ given that cavity messages $\eta_{i\to j}$ are real quantities. In particular $c_{q/2}$ is real, since we are using $q$ even. Expanding in discrete Fourier series both sides of BP equations (\[eq:self\_cons\_eta\_qclock\]), we get the following self-consistency equations for the Fourier coefficients: $$\begin{gathered} c_{m}^{(i\to j)}=\frac{1}{{\mathcal{Z}}_{i\to j}}\,\sum_{a=0}^{q-1} e^{-2\pi i\,am/q}\\ \prod_{k\in\partial i\setminus j}\,\sum_{b_k=0}^{q-1} c_{b_k}^{(k\to i)}\,I_{b_k}^{(q)}(\beta J_{ik})\,e^{\,2\pi i\,ab_k/q}, \label{eq:self_cons_c_qclock}\end{gathered}$$ where also ${\mathcal{Z}}_{i\to j}$ has to be expressed in terms of discrete Fourier coefficients. In order to keep a compact notation we have introduced the discrete analogous of modified Bessel functions of the first kind, $$I_n^{(q)}(x) \equiv \frac{1}{q}\sum_{a=0}^{q-1}\,e^{\,x\cos(2\pi a/q)}\cos\bigg(\frac{2\pi n a}{q}\bigg)\,,\quad n\in\mathbb{Z}, \label{eq:modif_Bes_func_qclock}$$ which converge to usual Bessel functions in the large $q$ limit: $$\lim_{q\to\infty} I_n^{(q)}(x) = I_n(x)\,.$$ In analogy with the XY model, also BP equations (\[eq:self\_cons\_eta\_qclock\]) admit the paramagnetic solution, given by the uniform distribution over the $q$ values (note the $q/2\pi$ normalization) $$\eta_{i\to j}(\theta_{i,a})=\frac{1}{2\pi}\quad,\quad\forall i\to j\,,\,\forall a\in\{1,\dots,q-1\} \label{eq:para_sol_qclock_1}$$ that corresponds to a vanishing solution for the self-consistency equations: (\[eq:self\_cons\_c\_qclock\]) $$c_{m}^{(i\to j)}=0\quad,\qquad\forall i\to j\,,\,\forall m\in\{1,\dots,q-1\}. \label{eq:para_sol_qclock_2}$$ The corresponding expressions for free energy density $f^{(q)}(\beta)$ and internal energy density $e^{(q)}(\beta)$ are $$f_{para}^{(q)}(\beta)=-\frac{1}{\beta}\ln{2\pi}-\frac{c}{2\beta}\ln{I_0^{(q)}(\beta)},$$ $$e_{para}^{(q)}(\beta)=-\frac{c}{2}\,\frac{I_1^{(q)}(\beta)}{I_0^{(q)}(\beta)}.$$ As expected, thank to the choice of the $q/2\pi$ normalization, these expressions converge to those for the XY model in the $q\to\infty$ limit. The next step is to study the stability of paramagnetic solution. An analysis analogous to the one made for the XY model tells us that most unstable coefficient are first order ones, satisfying at linear order the following equations $$c_{1}^{(i\to j)}=\sum_{k\in\partial i\setminus j}\frac{I_1^{(q)}(\beta J_{ik})}{I_0^{(q)}(\beta J_{ik})}\,c_{1}^{(k\to i)} \label{eq:self_cons_c_1_qclock}$$ By averaging over the disorder distribution $\mathbb{P}_J(J_{ij})$, we get self-consistency equations for the first two momenta, $$\left\{ \begin{aligned} &\overline{c_1}=(c-1)(2p-1)\frac{I_1^{(q)}(\beta)}{I_0^{(q)}(\beta)}\,\overline{c_1},\\ &\overline{c_1^2}=(c-1)\,\left[\frac{I_1^{(q)}(\beta)}{I_0^{(q)}(\beta)}\right]^2\,\overline{c_1^2}, \end{aligned} \right.$$ that identify critical temperatures for the phase transitions towards a ferromagnet, $T_F^{(q)}(p)=1/\beta_F^{(q)}(p)$, and towards a spin glass, $T_{SG}^{(q)}=1/\beta_{SG}^{(q)}$: $$\left\{ \begin{aligned} &(c-1)(2p-1)\frac{I_1^{(q)}\big(\beta_F^{(q)}(p)\big)}{I_0^{(q)}\big(\beta_F^{(q)}(p)\big)}=1,\\ &(c-1)\,\left[\frac{I_1^{(q)}\big(\beta_{SG}^{(q)}\big)}{I_0^{(q)}\big(\beta_{SG}^{(q)}\big)}\right]^2=1. \end{aligned} \right.$$ The paramagnetic phase is stable for temperatures larger than the critical one: $$T_c^{(q)}=\max\big(T_F^{(q)}(p),T_{SG}^{(q)}\big)\,.$$ In analogy with the XY model, for $p$ close to $1$, the clock model has a transition towards a ferromagnetic phase, while for $p$ close to $1/2$ the transition is towards a spin glass phase. Surprisingly, the abscissa $p_{mc}$ of the multicritical point in the $q$-state clock model has exactly the same expression in Eq. (\[eq:abscissa\_p\_mc\]) found for the XY model. ![Convergence of the discretized modified Bessel functions $I_n^{(q)}(x)$ toward their limiting values when $q\to\infty$, computed at $x=5$ (upper dataset) and $x=2$ (lower dataset). For each value of $n$, we plot the logarithm of $I_n^{(q)}(x)-I_n(x)$ together with a linear fit, to highlight the exponential convergence in $q$.[]{data-label="fig:conv_Bessel_log"}](convBes_beta2_000_and_beta5_000_STANDALONE){width="\columnwidth"} The only dependence of these critical lines on the number $q$ of states is through the discrete Bessel function $I_n^{(q)}$. So in order to understand how fast the clock model phase diagram converges to the one of the XY model, we need to study the rate of convergence of the functions $I_n^{(q)}(x)$ to the Bessel functions $I_n(x)$ in the large $q$ limit. We show in Fig. \[fig:conv\_Bessel\_log\] a numerical evidence that this convergence is exponentially fast in $q$, i.e. like $\exp(-q/q^*)$, with a characteristic scale $q^*$ increasing with the argument $x$: $$\begin{aligned} q^*(x=2) &\simeq 2.0\,,\\ q^*(x=5) &\simeq 2.5\,.\end{aligned}$$ Numerical evidence shows that $q^*$ is finite for any finite value $x$, but it seems to diverge in the $x\to\infty$ limit. In that limit, the convergence may follow a stretched exponential. {width="\linewidth"} Unfortunately, we have not been able to find a fully analytical proof of this statement. The following argument should, however, convince the reader that a power law decay in $q$ is not expected to take place every time one approximates the integral of a periodic function with a finite sum of $q$ terms. Let us suppose $f(x)$ is an infinitely differentiable function, $2\pi$-periodic, i.e. $f(x+2\pi)=f(x)$, and we are interested in approximating the integral $$I(f) = \frac{1}{2\pi} \int_0^{2\pi} dx\,f(x)$$ with the finite sum $$I^{(q)}(f) = \frac{1}{q} \sum_{a=0}^{q-1} f(2\pi a/q).$$ Rewriting $I^{(q)}$ as the integral of a step-wise function, the error $\Delta^{(q)}=I^{(q)}-I$ can be written as the sum of $q$ local terms, each one computed in a small interval $\Gamma_a\equiv [2\pi a/q-\pi/q,2\pi a/q+\pi/q]$ of size $2\pi/q$ around $2\pi a/q$: $$\Delta^{(q)}(f) = \frac{1}{2\pi} \sum_{a=0}^{q-1}\,\,\int_{\Gamma_a}dx\,\bigl[f(x)-f(2\pi a/q)\bigr].$$ For large $q$, we can Taylor expand the integrand around the central point of each interval $\Gamma_a$, $$f(x) - f(2\pi a/q) = \sum_{k=1}^\infty f^{(k)}(2\pi a/q) \frac{(x-2\pi a/q)^k}{k!}\,,$$ where $f^{(k)}$ is the $k$th derivative of $f$. Thus the error is given by the following series: $$\Delta^{(q)}(f) = \sum_{\substack{k \text{ even}\\k>0}} \frac{\pi^k}{q^{k+1} (k+1)!} \sum_{a=0}^{q-1} f^{(k)}(2\pi a/q).$$ For $q$ large, the internal sum can be approximated by the $q\to\infty$ limit, plus the error term, $$\begin{gathered} \frac1q \sum_{a=0}^{q-1} f^{(k)}(2\pi a/q) = \frac{1}{2\pi}\int_0^{2\pi}dx\,f^{(k)}(x) + \Delta^{(q)}(f^{(k)})=\\ \,\,\, = \frac{f^{(k-1)}(2\pi) - f^{(k-1)}(0)}{2\pi}+ \Delta^{(q)}(f^{(k)}) = \Delta^{(q)}(f^{(k)})\,,\end{gathered}$$ where the last inequality follows from the $2\pi$ periodicity. So, the equation for the error term is given by $$\Delta^{(q)}(f) = \sum_{\substack{k \text{ even}\\k>0}} \frac{\pi^k}{q^k (k+1)!} \,\Delta^{(q)}(f^{(k)})\,. \label{eq:error}$$ For a function $f$ smooth enough — like the one in the definition of the modified Bessel functions of first kind, $f(\theta)=\exp[x \cos(\theta)]\cos(n\theta)$ — we expect the error on the derivatives, $\Delta^{(q)}(f^{(k)})$, to decay with $q$ in the same way as the error on the function itself, $\Delta^{(q)}(f)$. This expectation is further confirmed by the data in Fig. \[fig:conv\_Bessel\_log\], where we see that the error on the function $I_0$ decays as the error on its derivative $I_1$. Noticing that the power-law ansatz $\Delta^{(q)}(f^{(k)}) \propto q^{-\alpha}$ is incompatible with Eq. (\[eq:error\]) for any value of the power $\alpha$, we conclude that the error $\Delta^{(q)}(f)$ decays faster that any power law. Apart from excluding a power law decay, the above argument is not able to provide the final answer: e.g. whether the decay is a simple exponential decay or a stretched one. The evidence presented in Sec. \[sec:convergence\] will suggest the decay is exponential for any positive temperature and changes to a stretched exponential at $T=0$. Numerical solution with the bimodal distribution of couplings {#subsec:numerical_cavity_method} ------------------------------------------------------------- As already explained above, the analytical expansion in Fourier series can be used only in the high temperature phase. The low-temperature region can be fully explored and understood only by using numerical tools. In particular, we will implement the cavity method at the RS stage, both at finite and zero temperature, by exploiting the *population dynamics algorithm*. This method, firstly introduced in Ref. [@AbouChacraEtAl1973] and then revisited and refined in Refs. [@MezardParisi2001; @MezardParisi2003], allows one to compute physical observables averaged over the disorder distribution and the graph realizations. To this purpose, we consider a population of ${\mathcal{N}}$ cavity marginals $\eta(\theta)$, randomly initialized, that evolve according to the iterative BP equations: at each step of the algorithm, each marginal in the population is updated according to the following equation: $$\eta_\ell(\theta) \leftarrow {\mathcal{F}}\big(\{\eta_{i_k}, J_k\}_{k=1,\ldots,c-1}\big)\,,$$ where ${\mathcal{F}}$ is defined in Eq. (\[eq:self\_cons\_eta\]), $J_k$’s are random variables generated according to the coupling distribution $\mathbb{P}_J$, and $i_k$’s are random indices uniformly drawn in $[1,{\mathcal{N}}]$, so as to choose $c-1$ random marginals in the population. {width="\linewidth"} Physical observables, which are functionals of the marginals in the population, usually change during the first part of the evolution, and then converge to an asymptotic value, corresponding to the thermodynamical expectation of that observable (within the replica symmetric ansatz). Being the algorithm of a stochastic nature and the population of finite size, we expect fluctuations of $O(1/\sqrt{{\mathcal{N}}})$. In what follows, if not stated otherwise, we use a population of ${\mathcal{N}}=10^6$ cavity marginals and a fixed degree $c=3$ for the underlying RRG. We have also checked that estimates of physical observables are compatible with what can be measured on a given samples of large size; however, the population dynamics algorithm is more efficient in computing physical observables averaged over the RRG ensemble and coupling distribution. At variance with the BP equations (\[eq:self\_cons\_eta\]) that may not have a solution, the population dynamics algorithm always converges to a fixed-point probability distribution of marginals ${\mathcal{P}}^*[\eta]$, independently from the initial conditions. Furthermore, this is true even when the RS assumption is no longer correct: when this happens, the distribution ${\mathcal{P}}^*[\eta]$ we get is no longer the exact one, and so we have to use (at least) the 1RSB ansatz. This will be done in Sec. \[sec:bp\_cavity\_method\_1rsb\]. Before searching for the transition lines between the different low-temperature phases in the $q$-state clock model and the XY model, we would like to verify the scaling of Fourier coefficients just below the critical temperature. In Fig. \[fig:Fourier\_coefficients\], we report the results of this check both for the para-ferro phase transition ($p=0.95$, left panel) and for the para-spin glass phase transition ($p=0.5$, right panel). In the former case we have $\overline{a_l} \propto \tau^{l/2}$, while in the latter $\overline{a_l} = 0$ and $\overline{a_l^2} \propto \tau^{l}$, as expected from the computation in Sec. \[sec:analyticalBimodal\]. In the low-temperature phases, each spin variable has a nonzero average: $$\vec{m}_i = (m_{i,x}, m_{i,y}) = \braket{\,(\cos{(\theta_i)},\sin{(\theta_i)})\,}\,,$$ where the angular brackets represent the average over a full marginal $\eta_i(\theta_i)$ defined in Eq. (\[eq:fullMarginals\]). From the local magnetizations $\vec{m}_i$’s we can build two order parameters: the norm of global magnetization vector, $$M \equiv \left| \frac1N \sum_i \vec{m}_i \right|\,,$$ and the overlap, $$Q \equiv \frac{1}{N} \sum_i |\vec{m}_i|^2\,,$$ which satisfy the inequality $M^2\le Q$. From the analysis of Fourier coefficients shown before, we expect both $Q$ and $M^2$ to grow linearly below the critical temperature $T_c$ (see Fig. \[fig:critical\_lines\_J\_Q008\], left panel). In the paramagnetic phase, all local magnetizations are null ($M=Q=0$), while in a pure ferromagnetic phase, all spins are perfectly aligned and so $M^2=Q>0$. In the more general case ($p<1$ and $T<T_c$), local magnetizations exist, but do not align perfectly and so we have $Q>0$ and $M^2<Q$: the unbiased spin glass phase ($M=0$, $Q>0$) belongs to this class, but also two other phases — the disordered ferromagnet and the magnetized spin glass, the so-called mixed phase — have $0<M^2<Q$ and can not be distinguished by just looking at these two order parameters. In order to distinguish this two phases we will need to check for the stability of the RS solution with respect to a breaking of the replica symmetry. This scenario is very similar to the one taking place in the Ising model [@CastellaniEtAl2005]. As long as we keep track only of the order parameters $Q$ and $M$, in the low-temperature phase ($T<T_c$), we have that $Q>0$ anywhere, while $M$ is non zero only for $p$ large enough. In central panel of Fig. \[fig:critical\_lines\_J\_Q008\], we show the typical behavior of $M^2$ as a function of $p$ at $T=0.2<T_c$ for the $q=8$ clock model. The critical $p_c$ estimated this way is an approximation to the true critical line separating the unbiased spin glass phase and the mixed phase; the right computation should be done within a full replica symmetry breaking ansatz, which is unfortunately unavailable for the diluted models we are studying here. We expect, however, the RS ansatz to provide a very good approximation. The critical line separating the RSB mixed phase from the RS disordered ferromagnet can be computed by studying the stability of the RS fixed point via the *susceptibility propagation* (SuscP) algorithm, that we run in population dynamics. As explained in Sec. \[subsec:validity\_rs\_method\] and in Appendix \[app:susc\_propag\], the SuscP algorithm amounts at studying the stability of the linearized BP equations around the RS solution. At any finite temperature, the perturbations around the fixed-point cavity marginals $\eta^*_{i\to j}$ evolve via the following linearized equations: $$\delta\eta_{i\to j} = \sum_{k\in\partial i \setminus j}\,\Biggl|\frac{\delta\,{\mathcal{F}}[\{\eta\}]}{\delta\,\eta_{k\to i}}\Biggr|_{\eta^*_{k\to i}}\delta\eta_{k\to i}\,,$$ where ${\mathcal{F}}$ is defined in Eq. (\[eq:self\_cons\_eta\]). We check for the growth of these perturbations by measuring the following norm: $$|\delta\eta|\equiv\sum_{(i\to j)}\,\sum_{a=0}^{q-1}\,|\delta\eta_{i\to j}(\theta_{i,a})|\,.$$ We define the growing rate as $$\lambda = \lim_{t\to\infty} \frac{\ln|\delta\eta|}{t}\,.$$ In the right panel of Fig. \[fig:critical\_lines\_J\_Q008\], we report the values of $\lambda$ as a function of $p$, for $T=0.2<T_c$ in the $q=8$ clock model. The $p_*$ value where $\lambda=0$ corresponds to the phase transition between a mixed RSB phase and an RS disordered ferromagnet: it is the point where the spin glass susceptibility diverges [@parisi2014diluted]. In the RSB phase the perturbations growing rate $\lambda$ is strictly positive, because the RS solution is unstable there. Having explained the way we compute the different critical lines, we can now draw in Fig. \[fig:phase\_diag\_J\] the full phase diagram for the $q$-state clock model, with several values of $q$. We notice that the convergence to the XY model in the $q\to\infty$ limit is very fast: in practice, critical lines with $q\ge 8$ are superimposed and coincide with those in the XY model. The only region where a $q$ dependence is still visible is that with $p$ close to $1$ and $T$ close to $0$. In this region we have a strong dependence on the discretization, such that for small $q$ values the RS disordered ferromagnet ($p$ is close to 1, but strictly smaller) is stable down to $T=0$, while in the XY model ($q\to\infty$ limit) there is always a phase transition to an RSB phase lowering the temperature with $p<1$. ![Phase diagram $p$ versus $T$ of the $q$-state clock model with bimodal coupling and different values of $q$. Full critical lines are exact, while dashed ones are approximated. The convergence to the XY model in the $q\to\infty$ limit is very fast and critical lines with $q \ge 8$ are practically superimposed, but for the region in the lower right corner.[]{data-label="fig:phase_diag_J"}](phase_diag_J_STANDALONE){width="\columnwidth"} This is an important new finding (to the best of our knowledge it was not known before). It suggests the XY model may show RSB effects much more easily than the Ising model, when the disorder is weak. The reason for this behavior is maybe due to the fact that, in presence of a weak disorder, which is not strong enough to force discrete variables in different directions, the continuous variables in the XY model can adapt more easily to several different orientations (states). An analogous behavior when increasing $q$ is also found in the study of the $q$-state clock model in the $d=3$ cubic lattice by means of Migdal-Kadanoff approximate renormalization group [@IlkerBerker2013]. Again, paramagnetic-ferromagnetic critical line converges very fast in $q$, while a stronger dependence in $q$ is found for the ferromagnetic-spin glass critical line (moving toward larger fractions of ferromagnetic couplings, as in our case) and the paramagnetic-spin glass critical line (moving toward the zero-temperature axis, unlike our case). Numerical solution of the gauge glass model ------------------------------------------- In order to discuss the gauge glass model, it is convenient to rewrite the Hamiltonian in a different form: $$\mathcal{H}[\{{\vec{\sigma}}\}]=-\sum_{\braket{ij}}\cos{(\theta_i-\theta_j-\omega_{ij})}\,, \label{eq:H_xy_gg}$$ where $\omega_{ij}$ are the preferred relative orientation between neighboring spins. The model with bimodal couplings studied above corresponds to $\omega_{ij} \in \{0,\pi\}$. In the gauge glass model, instead, the random rotations $\omega_{ij}$ take values uniformly in $[0,2\pi)$. The latter choice seems more in line with the continuous nature of the variables. The straightforward extension to the $q$-state clock model suggests to take $\omega_{ij} \in \{0,2\pi/q,\ldots,2\pi(q-1)/q\}$. Willing to interpolate with a single parameter between the pure ferromagnetic model ($\omega_{ij}=0$) and the unbiased spin glass model ($\omega_{ij}$ uniformly distributed), we choose the following coupling distribution: $$\mathbb{P}_{\omega}^{(q)}(\omega_{ij})=(1-\Delta)\,\delta(\omega_{ij})+\frac{\Delta}{q}\sum_{a=0}^{q-1} \delta\left(\omega_{ij}-\frac{2\pi\,a}{q}\right) \label{eq:disord_distr_gg}$$ with $\Delta\in[0,1]$. In this way, when $q\to\infty$ we exactly recover the uniform continuous distribution for the XY gauge glass model. Furthermore, in this model we can use any $q$ value, since we do not expect any difference between even and odd values: any pair of spins can in principle assume a configuration satisfying a coupling $\omega_{ij}$ for any $q$ value. ![Phase diagram $\Delta$ versus $T$ of the $q$-state clock model with the gauge glass couplings and different values of $q$. Full critical lines are exact, while dashed ones are approximated. The convergence to the XY model in the $q\to\infty$ limit is very fast and critical lines with $q \ge 8$ are practically superimposed, but for the region in the lower left corner.[]{data-label="fig:phase_diag_W"}](phase_diag_W_STANDALONE){width="\columnwidth"} {width="\linewidth"} By using the same techniques exposed above, we derive the phase diagram of this model as a function of $T$ and $\Delta$. The results are shown in Fig. \[fig:phase\_diag\_W\], where we draw critical lines for several different values of $q$. As in the bimodal case, the convergence to the XY model in the $q\to\infty$ limit is very fast and already for $q=8$ most of the phase diagram does not depend on $q$ anymore and provides the result for the XY model. Also in this case the only region where the $q$ dependence is stronger is the one where the temperature is close to zero and the disorder is very weak. In the $q\to\infty$ limit, the XY model seems again to show RSB effects for any infinitesimal amount of disorder in the $T=0$ limit. We notice *en passant* that for $q=2$ the two versions of the clock model are identical up to the transformation $p \leftrightarrow 1-\Delta/2$. This is clearly visible in Figs. \[fig:phase\_diag\_J\] and \[fig:phase\_diag\_W\], where the red lines corresponding to the $q=2$ clock model are identical up to a horizontal reflection. Convergence of physical observables {#sec:convergence} =================================== We face now the main task of this work, studying the convergence of physical observables of the $q$-state clock model toward those of the XY model. We measure physical observables via the population dynamics algorithm, but we need to use a population size ${\mathcal{N}}=10^7$ in order to achieve the required accuracy. The results presented in the previous sections about the fast convergence of the phase diagrams and the exponential convergence of discretized Bessel functions, strongly suggest an exponential convergence of physical observables as long as $T>0$. Indeed, as shown by the three upper panels in Fig. \[fig:multiplot\_freeEnergy\_J\_N1\_00e+07\], the free-energy of the clock model converges to the one of the XY model exponentially fast in $q$ as long as $T>0$. In these plots, we restrict the analysis to the low-temperature phases (spin glass, mixed, and ferromagnetic), because in the paramagnetic phase the convergence is so fast that it is hardly measurable. {width="\linewidth"} We fit data at $T>0$ via the exponential function $$\ln{\Delta f^{(q)}(T)} \equiv \ln{\left[f^{(q)}(T)-f^{(\infty)}(T)\right]}= A-q/q^*$$ estimating the following values for $q^*$ (all fits have an acceptable $\chi^2$ per degree of freedom, as shown by the values reported below on the right): $$\begin{gathered} q^*(p=0.50,\,T=0.02)=2.57(1)\,, \qquad \chi^2/dof=0.20/5\,,\\ q^*(p=0.95,\,T=0.02)=2.60(5)\,, \qquad \chi^2/dof=0.19/3\,,\\ q^*(p=0.99,\,T=0.30)=0.70(1)\,, \qquad \chi^2/dof=0.12/2\,. \end{gathered}$$ As soon as $q \gg q^*$ the clock model provides an extremely good approximation to the XY model physical observables, with a systematic error which is by far much smaller than the typical statistical uncertainty achieved in numerical simulations. In this sense, a Monte Carlo study of the clock model with $q$ large enough can be a much more efficient way of measuring physical observables in the XY model. At $T=0$, we observe a slower convergence in $q$, which is well fitted by a stretched exponential $$\ln{\Delta f^{(q)}(T=0)} = A-(q/q^*)^b\,.$$ For the two cases reported in Fig. \[fig:multiplot\_freeEnergy\_J\_N1\_00e+07\], we find that the $b$ exponent is very close to $1/2$ (and thus we fix it to that value in the fits), while values for $q^*$ are the following: $$\begin{gathered} q^*(p=0.50)=0.67(1)\,, \qquad \chi^2/dof=6.4/9\,,\\ q^*(p=0.95)=0.79(1)\,, \qquad \chi^2/dof=2.4/9\,. \end{gathered}$$ Although these values are slightly smaller than those in the $T>0$ case, the $b\simeq 1/2$ exponent makes the convergence at $T=0$ slower. So, it seems that even in the slowest case ($T=0$) the convergence of clock model observables to those of the XY model is fast enough to safely allow to use the clock model with moderately large values of $q$. In Fig. \[fig:multiplot\_freeEnergy\_W\_N1\_00e+07\], we report the analogous results for the clock model with the gauge glass coupling distribution in Eq. (\[eq:disord\_distr\_gg\]). Also in this case, convergence in $q$ is exponentially fast for $T>0$ with the following parameters: $$\begin{gathered} q^*(\Delta=1.0,\,T=0.02)=2.55(1)\,, \qquad \chi^2/dof=0.48/12\,,\\ q^*(\Delta=0.2,\,T=0.02)=2.66(1)\,, \qquad \chi^2/dof=1.35/13\,,\\ q^*(\Delta=0.1,\,T=0.20)=0.83(1)\,, \qquad \chi^2/dof=0.43/4\,. \end{gathered}$$ It is interesting to note that these values are similar to the ones found by using the bimodal coupling distribution. Again, at $T=0$, the convergence becomes a stretched exponential, with a $b$ exponent close to $1/2$ for any $\Delta$ value: $$\begin{gathered} q^*(\Delta=1.0)=0.67(1)\,, \qquad \chi^2/dof=12.5/19\,,\\ q^*(\Delta=0.2)=0.68(1)\,, \qquad \chi^2/dof=12.8/19\,. \end{gathered}$$ 1RSB Cavity Method {#sec:bp_cavity_method_1rsb} ================== We have shown that in both the XY model and the $q$-state clock model at low temperatures the replica symmetric ansatz breaks down if the disorder is strong enough (low $p$ or large $\Delta$ values). We observe this replica symmetry breaking (RSB) via the RS fixed point becoming unstable. In the RSB phase, we know that the RS result is just an approximation, although we expect it to be a rather good approximation for some observables (e.g. self-averaging observables, like the energy). Nonetheless, in order to keep track of the RSB effects and the many states present in a RSB phase, we can resort to a more complicated ansatz with one step of replica symmetry breaking (1RSB). In models with pairwise interactions, like those we are studying here, we expect a full RSB ansatz to be required in the strongly disordered and low-temperature phase. Nonetheless, even the 1RSB results can be illuminating on the true physical behavior. We thus solve the $q$-state clock model by means of the 1RSB cavity method derived by Mézard and Parisi [@MezardParisi2001]. The presence of many states breaks the validity of factorization in Eq. (\[eq:factor\_mu\]). Indeed, by adding nodes and links to the graph by following the RS cavity method prescriptions within each state, one realizes that each state gets a different free-energy shift and thus this leads to a reweighing of the different states [@MezardParisi2001]. We redirect the reader to book [@Book_MezardMontanari2009] and lecture notes [@Zamponi2008] for an exhaustive description of the 1RSB cavity method for solving sparse disordered models. Here we just sketch the key concepts involved in the 1RSB solution. At a given temperature $T=1/\beta$, the number of states with free energy density $f$ in a system of size $N$ can be written as $${\mathcal{N}}(f) = e^{\,N\Sigma_\beta(f)}\,,$$ where $\Sigma_\beta(f)$ is called *complexity* in the literature on spin glasses and *configurational entropy* in that on structural glasses. Introducing a *replicated partition function* $$\begin{split} {\mathcal{Z}}_\beta(x) \equiv \sum_\alpha e^{-\beta N x f_\alpha} &\simeq \int df \,e^{\,N[\Sigma_\beta(f) - \beta x f]}\\ &\simeq e^{\,N\big[\Sigma_\beta\big(f^*(\beta,x)\big)-\beta x f^*(\beta,x)\big]}\,, \end{split}$$ where $f_\alpha$ is the free energy of state $\alpha$ and $f^*(\beta,x)$ is the maximizer of $\Sigma(f) - \beta x f$, that depends on both $\beta$ and $x$, one can compute $\Sigma(f)$ as the Legendre transform of the *replicated free energy* $\phi_\beta(x)$: $$\begin{aligned} \phi_\beta(x) &\equiv& -\,\frac{1}{\beta x N} \ln{\mathcal{Z}}_\beta(x)\,,\\ f^*(\beta,x) &=& \phi_\beta(x) + x\,\partial_x\phi_\beta(x)\,,\\ \Sigma(\beta,x) &=& \beta x^2\,\partial_x\phi_\beta(x)\,.\end{aligned}$$ The complexity $\Sigma(f)$ can be obtained by plotting parametrically $\Sigma(\beta,x)$ versus $f^*(\beta,x)$ varying $x$ at fixed $\beta$. Thermodynamical quantities are obtained by setting $x=1$ if the corresponding complexity is positive, i.e. $\Sigma(\beta,x=1)\ge 0$. Otherwise, if $\Sigma(\beta,x=1)<0$, these states do not exists, and the partition function is dominated by the states with $x=x^*<1$ such that $\Sigma(\beta,x=x^*)=0$, where $x^*$ is called Parisi parameter. 1RSB equations and their solution by means of population dynamics algorithm --------------------------------------------------------------------------- The computation of the replicated partition function ${\mathcal{Z}}_\beta(x)$ must take into account the presence of many states, each one weighed by $\exp(-\beta x f_\alpha)$. In each state $\alpha$ BP equations (\[eq:self\_cons\_eta\]) are still valid; let us refer to them briefly as $\eta_{i\to j}={\mathcal{F}}[\{\eta_{k\to i}\}]$. Since now we have to reweigh the cavity messages $\eta$ according to the free energy shift they produce, it is necessary to introduce a probability distribution ${\mathcal{P}}[\cdot]$ over the RS probability distribution $P[\eta]$ of cavity messages. These two levels of populations come from the two different averages that we have to perform in the 1RSB approach: *i)* a first average in a given state, that gives the RS population $P[\eta]$, and *ii)* a second average over the states, that gives the 1RSB population ${\mathcal{P}}[P]$. If ${\mathcal{Z}}_{i\to j}$ is the normalization constant that comes from the computation of cavity message $\eta_{i\to j}$ via the RS BP equation $\eta_{i\to j}={\mathcal{F}}[\{\eta_{k\to i}\}]$, i.e. the free energy shift due to the addition of site $i$ and directed edges $(k\to i)$’s with $k\in\partial i \setminus j$, then the reweigh acts as follows $$\begin{split} P_{i\to j}[\eta_{i\to j}] &\equiv {\mathcal{G}}[\{P_{k\to i}\}]\\ &=\mathbb{E}_{G,J}\int\prod_{k\in\partial i \setminus j}{\mathcal{D}}\eta_{k\to i}\,P_{k\to i}[\eta_{k\to i}]\quad\quad\\ &\qquad\times\delta\Bigl[\eta_{i\to j}-{\mathcal{F}}[\{\eta_{k\to i}\}]\Bigr]\Bigl({\mathcal{Z}}_{i\to j}[\{\eta_{k\to i}\}]\Bigr)^x \label{eq:self_cons_eta_distrib_reweigh} \end{split}$$ In this way the RS solutions are reweighed by $\exp{(-\beta N m f_{\alpha})}$, as required in the computation of ${\mathcal{Z}}(x)$. The average over all the states yields the distributional equation for probability distribution ${\mathcal{P}}[P]$: $${\mathcal{P}}[P_{i\to j}]=\int \prod_{k\in\partial i \setminus j}{\mathcal{D}}P_{k\to i}\,{\mathcal{P}}[P_{k\to i}] \delta\Bigl[P_{i\to j}-{\mathcal{G}}[\{P_{k\to i}\}]\Bigr] \label{eq:self_cons_eta_distrib_1rsb}$$ Eqs. (\[eq:self\_cons\_eta\_distrib\_reweigh\]) and (\[eq:self\_cons\_eta\_distrib\_1rsb\]) are solved by a population dynamics algorithm, that considers the two levels of average. We store ${\mathcal{N}}$ populations each one made of ${\mathcal{M}}$ cavity messages $\eta$. Cavity messages evolve via Eq. (\[eq:self\_cons\_eta\_distrib\_reweigh\]), where populations entering the r.h.s. are randomly chosen according to Eq. (\[eq:self\_cons\_eta\_distrib\_1rsb\]). For each population, the reweighing of messages is performed as explained in Ref. [@Book_MezardMontanari2009], by first computing $r{\mathcal{M}}$ (with $r>1$) new cavity messages, and then selecting ${\mathcal{M}}$ among these with a probability proportional to ${\mathcal{Z}}^x$. Typical values for the $r$ parameter are contained in the range $[2,5]$. Physical observables in the thermodynamical limit can be written as averages over the two levels of probability distributions [@MezardParisi2001; @Book_MezardMontanari2009], namely over populations $P$’s and over cavity marginals $\eta$’s: $$\begin{aligned} \phi_\beta(x) &=& \frac{c}{2\beta x}\,\mathbb{E}_P\Bigl[ \ln{\mathbb{E}_{\eta}\left[{\mathcal{Z}}_{ij}^x\right]} \Bigr] - \frac{1}{\beta x}\,\mathbb{E}_P\Bigl[ \ln{\mathbb{E}_{\eta}\left[{\mathcal{Z}}_i^x\right]} \Bigr]\,,\\ f(\beta,x) &=& -\frac{1}{\beta}\,\mathbb{E}_P\left[ \frac{\mathbb{E}_{\eta}\left[{\mathcal{Z}}_i^x \ln{{\mathcal{Z}}_i}\right]} {\mathbb{E}_{\eta}\left[{\mathcal{Z}}_i^x\right]} \right] +\\ && + \frac{c}{2\beta}\,\mathbb{E}_P\left[ \frac{\mathbb{E}_{\eta}\left[{\mathcal{Z}}_{ij}^x \ln{{\mathcal{Z}}_{ij}}\right]} {\mathbb{E}_{\eta}\left[{\mathcal{Z}}_{ij}^x\right]}\right]\,,\\ \Sigma(\beta,x) &=& \beta x \Big(f(\beta,x) - \phi_\beta(x)\Big)\,.\end{aligned}$$ {width="\linewidth"} In the 1RSB ansatz, thanks to the two levels of average, we can define two different overlaps: an inner overlap $Q_1$, describing the similarity of local magnetizations inside a given state: $$Q_1 = \mathbb{E}_P\left[ \frac{\mathbb{E}_{\eta}\left[{\mathcal{Z}}_i^x\bigl(m_{i,x}^2+m_{i,y}^2\bigr)\right]} {\mathbb{E}_{\eta}\left[{\mathcal{Z}}_i^x\right]} \right]\,,$$ and an outer overlap $Q_0$, describing the similarity of magnetizations between different states: $$Q_0 = \mathbb{E}_P\left[ \frac{\mathbb{E}_{\eta}\bigl[{\mathcal{Z}}_i^x m_{i,x}\bigr]} {\mathbb{E}_{\eta}\bigl[{\mathcal{Z}}_i^x\bigr]} \right]^2 +\mathbb{E}_P\left[ \frac{\mathbb{E}_{\eta}\bigl[{\mathcal{Z}}_i^x m_{i,y}\bigr]} {\mathbb{E}_{\eta}\bigl[{\mathcal{Z}}_i^x\bigr]} \right]^2\,.$$ As expected, it holds $Q_1 \geqslant Q_0$. These two overlaps are nothing but the analogous of the ones introduced in the 1RSB solution to the SK model by Parisi [@Parisi1980a; @Parisi1980b]. We are going to use these overlaps to approximate the Parisi function, $Q(x)$, in the spin glass phase, given that the full RSB solution is not known for disordered models on sparse graphs. 1RSB solution of the $q$-state clock model ------------------------------------------ The 1RSB ansatz is known to provide the correct solution to many disordered models defined on random sparse graphs (e.g. $p$-spin models [@MezardEtAl2003], random $K$-SAT problems [@MontanariEtAl2004] and random coloring problems [@ZdeborovaKrzakala2007]), at least in a range of parameters. These models have either interactions involving $p>2$ variables ($p$-spin and $K$-SAT models) or variables taking $q>2$ values (coloring problems). For the $q$-state clock model on a sparse random graph (hereafter we restrict to symmetric bimodal couplings, $p=1/2$), it is less clear how close to the right solution the 1RSB ansatz is. We expect a continuous phase transition for most $q$ values, and so the 1RSB solution should be seen as an approximation to the correct full RSB solution. Our expectation comes from the following observations. For $q=2$, the clock model coincides with an Ising model, and for $q=4$, it is equivalent to a double Ising model. For $q=3$, the clock model is equivalent to a Potts model or $q=3$ coloring problem that has no thermodynamical phase transition on a random 3-regular graph [@KrzakalaEtAl2004], but only a dynamical phase transition is expected to happen (well described by a 1RSB ansatz). This case can be studied by fixing $x^*=1$ as in Ref. [@MarruzzoLeuzzi2016], where the $q$-state clock model with multi-body interactions is studied in the 1RSB frame. For $q\ge5$, there are no known results on sparse graphs and our results in Sec. \[sec:analyticalBimodal\] suggest a continuous transition for any $q$ value. This may look at variance with results for the $q$-state Potts model, but in the clock model the $q$ states have a specific ordering that eventually converge to the orientations of a continuous variable in the $q\to\infty$ limit. Given that in that limit the transition is again continuous, it is reasonable to expect that the transition in the disordered $q$-state clock model on a sparse random graph is continuous for any $q$ values. The literature provides one more piece of evidence in this direction. Nobre and Sherrington [@NobreSherrington1986] have studied the $q$-state clock model on the complete graph, finding a continuous phase transition for any $q \neq 3$ value. Moreover, they have expanded the replicated free energy close to the critical point, i.e. for $\tau=(T_c-T)/T_c\ll 1$, where the Parisi function $Q(x)$ can be well approximated by a linear function $Q(x)=a\,x$ for $x<b\,\tau$ and a constant $Q(x)=a\,b\,\tau$ for $x\ge b\,\tau$. The parameters $a$ and $b$ determine the universality class. Nobre and Sherrington [@NobreSherrington1986] found that $q=2$ and $q=4$ belong to the Ising universality class, while $q=3$ is in the $3$-state Potts one and for $q\ge 5$ the universality class is the same as the one of the XY model. {width="\linewidth"} Unfortunately, for diluted models with a continuous transition, the study of the 1RSB solution very close to the critical point is infeasible, because the replicated free energy $\phi_\beta(x)$ differs from the constant RS free energy $\phi_\beta(0)$ by a quantity going to zero linearly in $\tau$, and — given the complexity is very small (see Fig. \[fig:sigma\_1rsb\]) — computing numerically $\phi_\beta(x)$ at very small $\tau$ is too noisy. For this reason we have computed $\phi_\beta(x)$ at $\tau=1/2$, namely in the middle of the spin glass phase. We focus on values of $q$ ranging up to $8$, excluding the $q=3$ case that we know to be very different (it has just a 1RSB dynamical transition). The numerical evaluation has been performed with ${\mathcal{N}}=262\,144$ populations, each made of ${\mathcal{M}}=512$ marginals. This unbalanced choice (${\mathcal{M}}\ll {\mathcal{N}}$) is due to the observation that the complexity $\Sigma(\beta,x)$ shows much stronger finite size effects in ${\mathcal{N}}$ than in ${\mathcal{M}}$. The reweighing factor $r$ is *dynamically* adapted during the run in order to avoid the presence of “twins” in a population, that would reduce the effective size of the population; the range actually spanned is $r\in[2,10]$. In Fig. \[fig:sigma\_1rsb\], we plot the complexity $\Sigma(\beta=2\beta_c,x)$ for all the $q$ values studied. We have used the same plot ranges in all panels, in order to allow a direct comparison between different $q$ values. In each plot, we also draw the fitting quadratic curve that we use to estimate the Parisi parameter $x^*$, such that $\Sigma(x^*)=0$. The values of the estimated $x^*$ parameters are reported in Table \[table\]. Errors on $\Sigma$ are large due to the fact we are measuring a very small complexity, $\Sigma \sim O(10^{-4})$. $q$ $x^*$ $Q_0$ $Q_1$ ----- --------- ---------- ---------- 2 0.45(1) 0.497(4) 0.748(1) 4 0.47(2) 0.506(9) 0.750(2) 5 0.48(1) 0.427(5) 0.700(1) 6 0.51(3) 0.499(9) 0.685(3) 7 0.47(2) 0.447(8) 0.666(2) 8 0.46(2) 0.449(7) 0.664(2) : 1RSB parameters in the $q$-state clock model with $J_{ij}=\pm1$ couplings ($p=1/2$) on a random $3$-regular graph. \[table\] The overlaps $Q_0$ and $Q_1$ as a function of $x$ are shown in Fig. \[fig:overlap\_1rsb\] for all the $q$ values studied. Statistical errors on the overlaps at a given $x$ are smaller than the symbol size; thus the uncertainty reported in Tab. \[table\] is completely due to the error on the estimation of $x^*$. We notice that data for $q=2$ and $q=4$ in Fig. \[fig:overlap\_1rsb\] perfectly coincide (and this is expected for the reasons explained above). More interestingly is that also data for $q=7$ and $q=8$ coincide: this seems to suggests that, after the “transient” values $q=5$ and $q=6$, the $q$-state clock model converges immediately to its large $q$ limit, the XY model. A comparison with the 1RSB solution for the corresponding fully connected model can be done only in the Ising case ($q=2$); for the SK model the 1RSB solution returns at $T=T_c/2$ the following parameters [@Parisi1980a]: $$x^*=0.28 \, , \quad Q_0=0.213 \, , \quad Q_1=0.619\,,$$ which are rather different from the ones describing the 1RSB solution in the $q=2$ sparse case (see Table \[table\]). So the solutions we are studying are quite far from those on the fully connected topology. Nonetheless, similarly to what has been observed in the fully connected case [@NobreSherrington1986], we notice that the 1RSB parameters in Table \[table\] seem to vary little for $q \ge 5$; actually they are compatible with $q$-independent values within the error bars. Only the value $q=6$ shows a peculiar behavior, maybe due to some reminiscence of the $q=3$ case. Given the very fast convergence in $q$, and willing to have a precautionary attitude, we can safely take the $q=8$ clock model as a very good approximation to the XY model, even in the RSB low-temperature phase. Also, the $d=3$ cubic lattice case in the Migdal-Kadanoff RG approximation [@IlkerBerker2014] shows very similar results, with $q=2$ and $q=4$ having the same critical exponents, $q=3$ showing a peculiar behavior, and finally $q \gtrsim 5$ converging to the asymptotic values of the XY model. Conclusions =========== In this work we have studied analytically the disordered XY model on random regular graphs (Bethe lattice) with different disorder distributions. We have used the cavity method, that provides the correct answer for models with a locally tree-like topology. Given that the (replica symmetric) solution to the XY model requires to deal with $O(N)$ probability distributions on $[0,2\pi)$, with $N$ being the system size, we have chosen to approximate the XY model with the $q$-state clock model. The computational effort to solve the latter scales as $O(q^2 N)$. First of all we have shown that both the physical observables and the critical lines of the $q$-state clock model converge to those of the XY model very quickly in $q$: exponentially in $q$ for positive temperatures ($T>0$) and following a stretched exponential in $q$ for $T=0$. In practice we believe that the clock model with a not too large number of states can be safely used both in analytical computations and numerical simulations, especially in presence of quenched disorder and positive temperatures. Indeed, the situation where the discretization becomes more evident is the very low-temperature limit in presence of a very weak disorder. Avoiding this limit, the clock model perfectly mimics the XY model. Secondly, by using the clock model with large $q$ as a proxy for the XY model, we have computed accurate phase diagrams in the temperature versus disorder strength plane. We have used different disorder distributions and found some common features, that are apparently universal. Among these features, the one which is markedly different from disordered Ising models is the presence of spontaneous replica symmetry breaking at zero temperature for any disorder level: indeed, even an infinitesimal amount of disorder seems to bring the system into a spin glass phase. This new finding suggests that in the XY model it is much easier to create many different kinds of long range order (namely thermodynamic states), probably due to the continuous nature of XY variables. Finally, we tried to study the low-temperature spin glass phase within the ansatz with one step of replica symmetry breaking. Our original aim was to make connection to what is known in the fully connected clock model. The 1RSB analytical solution is computationally very demanding. Within the accuracy we managed to achieve, the clock model with finite connectivity studied here looks quantitatively far from its fully connected version. However, as in the fully connected case, the dependence on $q$ seems to be relevant only for very small values of $q$: as soon as $q \gtrsim 5$, the model properties resemble those of the XY model. The authors thank Giorgio Parisi for useful discussions. This research has been supported by the European Research Council (ERC) under the European Unions Horizon 2020 research and innovation programme (grant agreement No \[694925\]). Susceptibility propagation {#app:susc_propag} ========================== As stated in Sec. \[subsec:validity\_rs\_method\], one of the most used methods to detect the RS instability is to linearize the BP equations and see when RS fixed point becomes unstable under a small perturbation. This method is better known as susceptibility propagation, since the propagation of the BP perturbations allows one to compute the susceptibility in any given sample. Susceptibility propagation at a positive temperature ---------------------------------------------------- Let us start from the $J_{ij}=\pm 1$ case at $T>0$. We use the notation for the XY model, since the corresponding equations for the clock model are easily obtained by changing all integrals over $\theta_i$’s with discrete sums. To help the reader, we rewrite here the BP equations (\[eq:self\_cons\_eta\]): $$\eta_{i\to j}(\theta_i)=\frac{1}{{\mathcal{Z}}_{i\to j}}\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\eta_{k\to i}(\theta_k)\,, \label{eq:app_self_cons_eta}$$ with ${\mathcal{Z}}_{i\to j}$ given by: $${\mathcal{Z}}_{i\to j}=\int d\theta_i\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\eta_{k\to i}(\theta_k)\,.$$ The most generic perturbation to a cavity marginal $\eta_{i\to j}(\theta_i)$ must be such that the perturbed marginal $$\eta'_{i\to j}(\theta_i)=\eta_{i\to j}(\theta_i)+\delta\eta_{i\to j}(\theta_i)$$ remains well normalized, thus implying $$\int d\theta_i\,\delta\eta_{i\to j}(\theta_i)=0\,.$$ So, in the study of the instability of the RS fixed point, one has to search for the most unstable perturbation among those satisfying the above condition. Linearization of the BP equations (\[eq:app\_self\_cons\_eta\]) leads to $$\begin{aligned} \delta\eta_{i\to j}(\theta_i)&=\frac{1}{{\mathcal{Z}}_{i\to j}}\sum_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\delta\eta_{k\to i}(\theta_k)\prod_{k'\neq k}\int d\theta_{k'}\,e^{\,\beta J_{ik'}\cos{(\theta_i-\theta_{k'})}}\,\eta_{k'\to i}(\theta_{k'})\\ &\quad -\frac{1}{{\mathcal{Z}}_{i\to j}^2}\,\Biggl[\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\eta_{k\to i}(\theta_k)\Biggr]\\ &\qquad \times\int d\theta_i\,\Biggl[\,\sum_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J_{ik}\cos{(\theta_i-\theta_k)}}\,\delta\eta_{k\to i}(\theta_k)\prod_{k'\neq k}\int d\theta_{k'}\,e^{\,\beta J_{ik'}\cos{(\theta_i-\theta_{k'})}}\,\eta_{k'\to i}(\theta_{k'})\Biggr]\,. \label{eq:app_self_cons_eta_linear_pm_J} \end{aligned}$$ All cavity messages that appear in this expression — as well as normalization constant ${\mathcal{Z}}_{i\to j}$ — have to be evaluated on the RS BP fixed point. Eq. (\[eq:app\_self\_cons\_eta\_linear\_pm\_J\]) can be solved on a given graph or in population dynamics if one is interested only in computing the typical behavior. In the latter case, we need to evolve a population of pairs of functions $(\eta_i(\theta_i),\delta\eta_i(\theta_i))$: marginals $\eta_i(\theta_i)$ evolve according to BP equations (\[eq:app\_self\_cons\_eta\]), while perturbations $\delta\eta_i(\theta_i)$ evolve according to Eq. (\[eq:app\_self\_cons\_eta\_linear\_pm\_J\]). We then measure the growth rate of the perturbations via the norm $$|\delta\eta|\equiv\sum_{\{i\to j\}}\,\int d \theta_i\,|\eta_{i\to j}(\theta_i)|\,,$$ that evolves as $|\delta\eta| \propto \exp(\lambda\, t)$ for large times. The critical point leading to RS instability is identified by the condition $\lambda=0$. In the gauge glass the linearized BP equations read $$\begin{aligned} \delta\eta_{i\to j}(\theta_i)&=\frac{1}{{\mathcal{Z}}_{i\to j}}\sum_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J\cos{(\theta_i-\theta_k-\omega_{ik})}}\,\delta\eta_{k\to i}(\theta_k)\prod_{k'\neq k}\int d\theta_{k'}\,e^{\,\beta J\cos{(\theta_i-\theta_{k'}-\omega_{ik'})}}\,\eta_{k'\to i}(\theta_{k'})\\ &\quad -\frac{1}{{\mathcal{Z}}_{i\to j}^2}\,\Biggl[\,\prod_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J\cos{(\theta_i-\theta_k-\omega_{ik})}}\,\eta_{k\to i}(\theta_k)\Biggr]\\ &\qquad \times\int d\theta_i\,\Biggl[\,\sum_{k\in\partial i\setminus j}\int d\theta_k\,e^{\,\beta J\cos{(\theta_i-\theta_k-\omega_{ik})}}\,\delta\eta_{k\to i}(\theta_k)\prod_{k'\neq k}\int d\theta_{k'}\,e^{\,\beta J\cos{(\theta_i-\theta_{k'}-\omega_{ik'})}}\,\eta_{k'\to i}(\theta_{k'})\Biggr]\,. \label{eq:app_self_cons_eta_linear_gg} \end{aligned}$$ Susceptibility propagation at zero temperature ---------------------------------------------- At zero temperature the situation is more complicated and important differences arise between the $q$-state clock model with discrete variables and the XY model with continuous variables. Let us recall the zero temperature version of the BP equations for the $J_{ij}=\pm 1$ XY model: $$h_{i\to j}(\theta_i)\cong\sum_{k\in\partial i\setminus j}\max_{\theta_k}{\left[h_{k\to i}(\theta_k)+J_{ik}\cos{(\theta_i-\theta_k)}\right]} \label{eq:app_self_cons_h_T_0_pm_J}$$ where $h_{i\to j}(\theta_i)$ is defined up to an additive constant such that $\max_{\theta_i}[h_{i\to j}(\theta_i)]=0$, as we discuss in Sec. \[sec:bp\_cavity\_method\_rs\]. Linearizing Eq. (\[eq:app\_self\_cons\_h\_T\_0\_pm\_J\]) we get the following equation for the evolution of perturbations $$\delta h_{i\to j}(\theta_i)\cong\sum_{k\in\partial i\setminus j}\delta h_{k\to i}(\theta_k^*(\theta_i))\,, \label{eq:app_self_cons_h_T_0_linear_pm_J}$$ where $\theta_k^*(\theta_i)$ is given by: $$\theta_k^*(\theta_i)=\operatorname{arg}\,\max_{\theta_k}{\left[h_{k\to i}(\theta_k)+J_{ik}\cos{(\theta_i-\theta_k)}\right]}\,. \label{eq:theta_star}$$ In practice we never solve the equations for the XY model; we always solve those for the $q$-state clock model, where the only difference is that the maximum in Eqs. (\[eq:app\_self\_cons\_h\_T\_0\_pm\_J\]) and (\[eq:theta\_star\]) must be taken over the discrete set of $q$ possible values for $\theta_k$. A natural question is whether the solution to these equations changes smoothly with $q$ in the limit of very large $q$. We find that the marginals in the $q$-state clock model with $q$ large are very close to those in the corresponding XY model: thus Eq. (\[eq:app\_self\_cons\_h\_T\_0\_pm\_J\]) for the marginals can be used safely, and the XY model well approximated by the clock model with moderately large values for $q$. On the contrary, perturbations in the clock model at $T=0$ evolve in a completely different way with respect to what happens in the XY model. Indeed, due to the fact the maximum in Eq. (\[eq:theta\_star\]) is taken over a discrete set, it can happen that perturbations obtained from Eq. (\[eq:app\_self\_cons\_h\_T\_0\_linear\_pm\_J\]) have the same value for any $\theta_i$. And this in turn corresponds to a null perturbation. To understand the last statement, one has to remember that the correct normalization for each cavity marginal is enforced by the condition that the maximum of $h_{i\to j}(\theta_i)$ is zero, and this must be true also for the perturbed marginal. This implies that the perturbation must be zero in $\theta_i^\text{max}=\text{argmax}_{\theta_i} [h_{i\to j}(\theta_i)]$. We enforce this condition by shifting the perturbations obtained from Eq. (\[eq:app\_self\_cons\_h\_T\_0\_linear\_pm\_J\]) as follows $$\delta h_{i\to j}(\theta_i) \;\leftarrow\; \delta h_{i\to j}(\theta_i) - \delta h_{i\to j}(\theta_i^\text{max})\,.$$ Consequently a constant perturbation generated by Eq. (\[eq:app\_self\_cons\_h\_T\_0\_linear\_pm\_J\]) becomes a null perturbation after the shift. So, in the $T=0$ clock model, the perturbations evolve exactly as in the Ising model [@MoroneEtAl2014a], where perturbations divide in two groups: null perturbations and nonnull perturbations. Since the norm of the nonnull perturbations changes a little and remains finite during the evolution, the instability of the RS BP fixed point is mainly determined by the evolution of the *fraction* of nonnull perturbations. Nonetheless, we have observed that such a fraction shrinks with a rate that depends on $q$, and in the $q\to\infty$ limit this fraction seems to remain finite. So, we conclude that the shrinking of the fraction of nonnull perturbations is a direct consequence of the discretization, and it is not present in the XY model. The fact that perturbations evolve in a drastically different way in the XY model and in the clock model for any value of $q$, may suggest the latter model is unfit to describe the fluctuations of the XY model, and thus its physical behavior in the $T=0$ limit. Luckily enough there is a way out to this problem. The solution is to use the BP equations (\[eq:app\_self\_cons\_h\_T\_0\_pm\_J\]) in a fully discretized way, but computing the maximum in Eq. (\[eq:theta\_star\]) on the reals. Given that the argument of the argmax in Eq. (\[eq:theta\_star\]) is defined only on $q$ discrete points, we interpolate with a parabola the three points around the maximum, i.e. the point where the argument achieves its maximum and the two nearby. A similar interpolation is performed on the function $\delta h_{k\to i}(\cdot)$ that needs to be evaluated at the value $\theta_k^*(\theta_i)$ that no longer belongs to the discrete set. By proceeding this way, we obtain perturbations that shrink, but never become exactly zero. Thus the critical point can be computed just by checking the evolution of the norm of the perturbations as in the $T>0$ case. [^1]: C. Lupo, G. Parisi, and F. Ricci-Tersenghi, *work in progress*.

ArXiv

=15.5pt DESY 13-169, RUP-13-10, RESCEU-4/13 [**Graceful exit from Higgs G-inflation**]{} [ Kohei Kamada$^{a,}$[^1], Tsutomu Kobayashi$^{b,}$[^2], Taro Kunimitsu$^{c,d,}$[^3], Masahide Yamaguchi$^{e,}$[^4], and Jun’ichi Yokoyama$^{d,f,}$[^5] ]{} *$^{a}$ Deutsches Elektronen-Synchrotron DESY, Notkestrasse 85, D-22607 Hamburg, Germany*\ *$^{b}$ Department of Physics, Rikkyo University, Toshima, Tokyo 175-8501, Japan*\ *$^{c}$ Department of Physics, Graduate School of Science,\ The University of Tokyo, Tokyo 113-0033, Japan*\ *$^{d}$ Research Center for the Early Universe (RESCEU), Graduate School of Science,\ The University of Tokyo, Tokyo 113-0033, Japan*\ *$^{e}$ Department of Physics, Tokyo Institute of Technology, Tokyo 152-8551, Japan*\ *$^{f}$ Kavli Institute for the Physics and Mathematics of the Universe (Kavli IPMU),\ WPI, TODIAS, The University of Tokyo, Kashiwa, Chiba, 277-8568, Japan* Higgs G-inflation is a Higgs inflation model with a generalized Galileon term added to the standard model Higgs field, which realizes inflation compatible with observations. Recently, it was claimed that the generalized Galileon term induces instabilities during the oscillation phase, and that the simplest Higgs G-inflation model inevitably suffers from this problem. In this paper, we extend the original Higgs G-inflation Lagrangian to a more general form, namely introducing a higher-order kinetic term and generalizing the form of the Galileon term, so that the Higgs field can oscillate after inflation without encountering instabilities. Moreover, it accommodates a large region of the $n_s$ - $r$ plane, most of which is consistent with current observations, leading us to expect the detection of B-mode polarization in the cosmic microwave background in the near future. Introduction {#sec:intro} ============ Cosmic inflation [@Starobinsky:1980te; @Sato:1980yn; @Guth:1980zm], which is now considered to be an indispensable part of standard cosmology, requires an effective scalar field in order to keep the universe in a quasi-de Sitter phase. Such scalar fields are thought to be ubiquitous in theories beyond the standard model (SM) of particle physics, but the only scalar field identified so far is the 125 GeV scalar boson recently discovered at the LHC [@atlas:2012; @cms:2012], whose characteristics have up to date no deviation from the predictions of the SM Higgs. Although there are many indications of physics beyond the SM, no direct evidence has been reported until now except for neutrino oscillations [@Fukuda:1998mi; @GonzalezGarcia:2007ib]. This motivates us to consider the possibility that the Higgs field, the only scalar field in the SM, is responsible for inflation, assuming that the discovered particle is the SM Higgs. With no hints of physics beyond the SM at the LHC, this type of inflation models has to be taken seriously. It is known that inflation driven by the SM Higgs field, with the canonical kinetic term and renormalizable potential, does not reproduce the observable universe, since both the curvature and tensor perturbations produced are much larger than what we observe today. Consequently, there have been various proposals for Higgs inflation models by extending the structure of the SM Lagrangian, starting with the inclusion of a large non-minimal coupling term with gravity [@CervantesCota:1995tz; @Bezrukov:2007ep; @Barvinsky:2008ia]. Other models include a derivative coupling with the Einstein tensor (new Higgs inflation) [@Germani:2010gm], a galileon-like term (Higgs G-inflation) [@Kamada:2010qe], and a non-canonical kinetic term (running kinetic inflation) [@Nakayama:2010kt], all of which can be treated in a unified manner in the context of generalized G-inflation [@Kobayashi:2011nu] as generalized Higgs inflation [@Kamada:2012se][^6]. Among these, Higgs G-inflation [@Kamada:2010qe] is distinct from the other possibilities in the sense that it does not involve non-minimal couplings between the Higgs field and gravity, and although it includes higher derivative terms in the Lagrangian, it keeps the derivatives in the equation of motion at second order. The original Higgs G-inflation considered in Ref. [@Kamada:2010qe] has an additional term in the Higgs field Lagrangian of the form $$\frac{\varphi X}{M^4} \Box \varphi,$$ which is a generalization of the Galileon term [@Deffayet:2010qz; @Kobayashi:2010cm], and is the simplest possibility of this type. Here, $\varphi$ is the Higgs field in the unitary gauge, $X$ is its canonical kinetic function, $X:= -\partial_\mu\varphi\partial^\mu\varphi/2$ (we use the mostly plus sign convention for the metric), and $M$ is some mass scale. Recently, there was a claim that the Higgs field oscillation after inflation does not occur for the above Galileon-inspired term [@Ohashi:2012wf]. This is due to the large effect of the Galileon term on the oscillation dynamics, which makes the coefficient of $\ddot{\varphi}$ vanish in the equation of motion of the Higgs field, thereby ending up with a singularity. By taking a small value for the Higgs self-coupling constant $\lambda$, oscillation can be realized in this model. Even in this case, however, the sound speed squared of the curvature perturbation possibly becomes negative, which makes the curvature perturbations obey a Laplacian equation instead of a wave equation, leading to instabilities at small scales. As a result, a robust oscillation of the Higgs field occurs only if $$\lambda<2.9\times10^{-9},$$ which is far smaller than the value for the SM Higgs field. In this paper, we extend the above model to a broader framework, and propose a new Higgs G-inflation model in which the Higgs field oscillates after inflation keeping $c_s^2>0$ even with phenomenologically natural values of the Higgs quartic coupling. We will show that by extending the structure of the kinetic sector in Higgs G-inflation models, oscillation does occur in Higgs G-inflation models while keeping $c_s^2$ positive. We will also present a convenient method of calculating the predictions of the generalized model. We concentrate on inflation realized by the Higgs field, but the analysis here can easily be extended to general potential driven G-inflation models. Before we go on, we would like to note the value of the Higgs quartic coupling constant $\lambda$. For the 125 GeV SM Higgs field, the value of $\lambda$ at the electroweak scale is determined by the mass of the Higgs boson, which gives $$\lambda = \frac{m_h^2}{2v^2}\simeq 0.13.$$ When considering inflation, we need to take into account the running of the coupling constant. Using the renormalization group equations for the SM [@Casas:1994qy; @EliasMiro:2011aa], and assuming that the Galileon term does not change the equations substantially, $\lambda$ becomes logarithmically smaller at higher energy scales.[^7] The value could become negative, depending on the values of the top quark mass and the strong coupling constant, but once the Higgs field acquires an expectation value where $\lambda$ is negative, inflation never occurs and it cannot fall down to the electroweak vacuum. Thus, we assume here that the Higgs field was in an initial condition at which it can drive inflation with positive $\lambda$. At this scale, the value of $\lambda$ would be $\mathcal{O}(0.01)$ assuming no fine-tuning. For this reason, we will take $\lambda$ to be 0.01 in the numerical analyses of this paper, although the results of this paper would not strongly depend on the precise value. The paper is organized as follows. In the next section, we review the original Higgs G-inflation model, and show where the problem lies. In Sec. \[sec3\] we extend the form of the Lagrangian in order to realize inflation and subsequent field oscillation consistently, and present the results from numerical calculations. The final section is devoted to conclusions and discussions. In the Appendixes, we summarize the basic formulas for the equations of motion and the primordial perturbations. Higgs G-inflation and its reheating phase {#sec2} ========================================= In this section, we review the Higgs G-inflation model and its possible instabilities during the reheating phase. The SM Higgs Lagrangian is written as $$S= \int d^4 x \sqrt{-g}\left[ \frac{M_{P}^2}{2}R -|D_\mu \mathcal{H}|^2 - \lambda\left(|\mathcal{H}|^2 - v^2\right)^2 \right],$$ where $M_{P}$ is the reduced Planck mass, $R$ is the Ricci scalar, and $\mathcal{H}$ is the Higgs doublet. Higgs G-inflation is driven by the neutral component of the SM Higgs field, or the scalar Higgs boson in the unitary gauge, so we will focus on this component. The action for the SM Higgs field becomes of the form $$S= \int d^4 x \sqrt{-g}\left[ \frac{M_{P}^2}{2}R + X -V(\varphi) \right],$$ where $\varphi$ is the neutral component of the Higgs field. Here and hereafter we omit interactions with massive gauge fields that are irrelevant to the dynamics of inflation. We consider the case where the neutral component of the Higgs field has a large value compared to the electroweak vacuum expectation value, $v=246\ \mathrm{GeV}$. In this case, the Higgs potential can be approximated as a quartic potential: $$V(\varphi) \simeq \frac{1}{4}\lambda \varphi^4.$$ In order to realize inflation consistent with observations, we need to extend the structure of the Higgs field action. Higgs G-inflation [@Kamada:2010qe] is a model in which an additional self-interaction term of the form $$- G(\varphi, X) \Box \varphi$$ is introduced to the Higgs field Lagrangian. This term is a generalized Galileon term [@Deffayet:2010qz; @Kobayashi:2010cm], which keeps the derivatives in the equation of motion at second order. Adding this generalized Galileon term, the full action becomes $$S= \int d^4 x \sqrt{-g}\left[ \frac{M_{P}^2}{2}R + X - \frac{1}{4}\lambda \varphi^4 - G(\varphi, X) \Box \varphi \right].$$ $G$ is an arbitrary function of $\varphi$ and $X$, except that it should not contain any even powers of $\varphi$ in order to keep the gauge invariance of the action. In the original Higgs G-inflation model, $$G(\varphi, X) = -\frac{\varphi X}{M^4} \label{original}$$ was adopted. $M$ here is a constant with a dimension of mass, which would be related to the scale where new physics comes in[^8]. If $M$ is large enough compared to the electroweak scale, this term would not affect the predictions in collider experiments, and hence becomes compatible with the known characteristics of the Higgs field. The equation of motion for the Higgs field corresponds to Eq. (\[back13\]) in Appendix B with $m=1$, $n=0$, and $\widetilde{M} \rightarrow \infty$, $$\begin{aligned} 0&= \left(1-6H\frac{\varphi\dot{\varphi}}{M^4}+2\frac{\dot{\varphi}^2}{M^4} + \frac{3\varphi^2\dot{\varphi}^4}{2M_P^2M^8}\right)\ddot{\varphi} +3H\dot{\varphi} +\lambda \varphi^3 \notag\\ &\ \ -\left(9H^2-\frac{3\dot{\varphi}^2}{2M_P^2}-\frac{3\dot{\varphi}^4}{2M_P^2M^4}+\frac{3\lambda\varphi^4}{4M_P^2}\right) \frac{\varphi \dot{\varphi}^2}{2M^4} \label{HiggsGnum} \\ &=H\dot{\varphi}\left[3 - \eta - H\frac{ \varphi \dot{\varphi}}{M^{4}} \left(9- 3 \epsilon - 6\eta +2\eta\alpha \right) \right]+ \lambda \varphi^3, \label{HiggsGslow} \end{aligned}$$ where $\epsilon$, $\eta$, and $\alpha$ are the slow-roll parameters defined by $$\begin{aligned} \epsilon:=-\frac{\dot H}{H^2}, \quad \eta:=-\frac{\ddot\varphi}{H\dot\varphi}, \quad \alpha:=\frac{\dot \varphi}{H \varphi}.\end{aligned}$$ The original Higgs G-Inflation model is realized when the Higgs field has a large expectation value, where the slow-roll conditions $$|\epsilon|,\ |\eta|,\ |\alpha| \ll 1,$$ and the Galileon domination condition $$|H\dot{\varphi}G_X| =\left|H\frac{ \varphi \dot{\varphi}}{M^4}\right|\gg 1,$$ are satisfied. Here the dynamics of the Higgs field is described by the slow-roll equation of motion, $$- 9H^2\frac{\varphi \dot{\varphi}^2}{M^4} + \lambda \varphi^3 \simeq 0.$$ $M$ is determined by the observed amplitude of thecurvature perturbations. The second order action for the comoving curvature perturbation $\zeta$ is given by [@Kobayashi:2011nu] $$S_2 = M_{P}^2 \int d^4 x a^3\left[\mathcal{G}_S \dot{\zeta}^2 -\frac{\mathcal{F}_S}{a^2}(\vec{\nabla}\zeta)^2\right],$$ where $\mathcal{F}_S$ and $\mathcal{G}_S$ are described by the background quantities and are given in Eqs. (\[pertFs\]) and (\[pertGs\]) for more general cases in Appendix B. In the present case, from Eqs. (\[pertFs\]) and (\[pertGs\]) with $m=1$, $n=0$, and $\widetilde{M} \rightarrow \infty$, we see that $$\begin{aligned} \mathcal{F}_S &\simeq -\frac{2\varphi\dot{\varphi}^3}{HM_{Pl}^2M^4}, \\ \mathcal{G}_S &\simeq -\frac{3\varphi\dot{\varphi}^3}{HM_{Pl}^2M^4},\end{aligned}$$ which yield the sound speed squared, $$c_s^2 := \frac{\mathcal{F}_S}{\mathcal{G}_S} \simeq \frac23.$$ The power spectrum of the comoving curvature perturbation, $\zeta$, is estimated in Eq. (\[Gpower\]) as $$\mathcal{P}_\zeta = \frac{1}{8\pi^2 c_s \mathcal{F}_S} \left(\frac{H}{M_P}\right)^2 = -\frac{H^3M^4}{16\pi^2c_s \varphi \dot{\varphi}^3}. \label{eqn:zetanorm}$$ Using the Planck normalization, $\mathcal{P}_\zeta \simeq 2.2 \times 10^{-9}$ for $k=0.05\ \mathrm{Mpc}^{-1}$ yields $$M \simeq 4.5 \times 10^{-6} \lambda^{-\frac{1}{4}}M_P \simeq 10^{13} \; \mathrm{GeV}.$$ The spectral index and tensor-to-scalar ratio are given as $$n_s = 1-4\epsilon \simeq 0.967,$$ $$r = \frac{64}{3}\left(\frac{2}{3}\right)^{\frac{1}{2}} \epsilon \simeq 0.14,$$ respectively, and there is a consistency relation between $r$ and the tensor spectral index $n_T$: $$r = -\frac{32\sqrt{6}}{9}n_T. \label{consistency}$$ For the universe to transform into the hot big bang state after inflation, the Higgs field has to oscillate around the minimum of the potential and decay into the SM particles. But recently, it was found that the Higgs field could not oscillate after inflation for this model [@Ohashi:2012wf]. The coefficient of $\ddot{\varphi}$ in Eq. (\[HiggsGnum\]) vanishes before the Higgs field reaches the maximum point after passing the potential minimum, where $\ddot{\varphi}$ goes to infinity, leading to catastrophe. Moreover, before reaching the point where $\ddot{\varphi}$ diverges, the value of the sound speed squared of the fluctuations $c_s^2$ becomes negative. This occurs independent of the above catastrophe, and leads to instabilities of the perturbations at small scales. Even a very short period of $c_s^2<0$ would result in a disaster, since the growth rate of a mode is proportional to the wave number $k$, and all modes with wavelengths smaller than $\sqrt{-c_s^2}\Delta t$ (where $\Delta t$ is the duration of $c_s^2<0$), at least up to the Planck scale, would grow exponentially as $\exp(k\sqrt{-c_s^2}\Delta t)$ during that period. We have to resolve this issue in order to obtain a consistent inflationary scenario. These instabilities can be avoided by tuning the self-coupling coefficient $\lambda$ to a small value. The singularity of the equation of motion can be avoided in case $$\lambda<7.2\times10^{-9},$$ while $c_s^2>0$ is maintained if[^9] $$\lambda < 2.9 \times 10^{-9}.$$ This corresponds to making the value of $M$ larger, which in turn diminishes the effects of the Galileon term after inflation. But we are now considering the SM Higgs field, which has $\lambda$ determined by collider experiments, much larger than the above value. Thus, we consider ameliorating these issues by extending the structure of the Lagrangian in the next section. Extending the Higgs G-inflation model {#sec3} ===================================== In this section, we extend the Lagrangian of the Higgs G-inflation model, in order to avoid the instabilities stated in the previous section. Furthermore, we aim to construct a model with predictions preferred by the Planck satellite results. For the detailed calculations, we refer to the formulation of Generalized G-inflation [@Kobayashi:2011nu]. Adding a higher-order kinetic term ---------------------------------- The situation can be alleviated by adding a higher-order kinetic term to the Higgs field Lagrangian. We illustrate it first by adding a term of the form $$\frac{1}{2\widetilde{M}^4}X^2,$$ to the Higgs field Lagrangian. Here $\widetilde{M}$ represents the scale of new physics associated with this term. The resultant action becomes, $$S= \int d^4 x \sqrt{-g}\left[ \frac{M_{P}^2}{2}R + X + \frac{1}{2\widetilde{M}^4}X^2 - \frac{1}{4}\lambda \varphi^4 + \frac{\varphi X}{M^4} \Box \varphi \right].$$ By adding this term, from Eq. (\[back13\]) with $\ell=2$, $m=1$ and $n=0$, the full equation of motion for the homogeneous part becomes $$\begin{aligned} \left(1+\frac{3\dot{\varphi}^2}{2\widetilde{M}^4}-6H\frac{\varphi\dot{\varphi}}{M^4}+2\frac{\dot{\varphi}^2}{M^4} + \frac{3\varphi^2\dot{\varphi}^4}{2M_P^2M^8}\right)\ddot{\varphi} +3\left(1+\frac{{\dot \varphi}^2}{2 {\widetilde M}^4}\right)H\dot{\varphi} +\lambda \varphi^3& \notag\\ -\left(9H^2 - \frac{3\dot{\varphi}^4}{8M_P^2\widetilde{M}^4}-\frac{3\dot{\varphi}^2}{2M_P^2}-\frac{3\dot{\varphi}^4}{2M_P^2M^4}+\frac{3\lambda\varphi^4}{4M_P^2}\right)& \frac{\varphi \dot{\varphi}^2}{2M^4} =0.\end{aligned}$$ We see that a positive term $\frac{3}{2\widetilde{M}^4}\dot{\varphi}^2$, which comes from the higher-order kinetic term, is added to the coefficient of $\ddot{\varphi}$. One expects that this term makes the coefficient positive throughout the reheating phase. The same type of term comes in when we add an arbitrary power of $X$ to the Lagrangian, and thus those terms can also be used. The sound speed squared of the fluctuations is also modified as $$c_s^2 = \frac{1 + \frac{1}{2\widetilde{M}^4}\dot{\varphi}^2 - (4H\dot{\varphi}+2\ddot{\varphi})\frac{\varphi}{M^4} - \frac{\varphi^2 \dot{\varphi}^4}{2M_{Pl}^2M^8} } {1 + \frac{3}{2\widetilde{M}^4}\dot{\varphi}^2 - 6H\frac{\dot{\varphi}\varphi}{M^4} + 2 \frac{\dot{\varphi}^2}{M^4} + \frac{3\varphi^2 \dot{\varphi}^4}{2M_{Pl}^2M^8} }. \label{cs2}$$ One finds a positive contribution both in the numerator and in the denominator, which is expected to help avoid negative values of $c_s^2$. We carried out numerical calculations of the background evolution, using the equation of motion of the scalar field and the gravitational evolution equation \[the explicit form shown in Eq. (\[back12\]) with $\ell=2$, $m=1$, $n=0$\]. The parameters $M$ and $\widetilde{M}$ were determined by requiring the amplitude of the curvature fluctuations to be $2.2\times 10^{-9}$ at $k=0.05\mathrm{Mpc}^{-1}$ [@Ade:2013zuv], which we have tentatively identified with the comoving Hubble scale 60 e-folds before the end of inflation. The case in which this scale corresponds to different e-folds will be shown later. The perturbation values were determined using the general formulas of Generalized G-inflation [@Kobayashi:2011nu]. We have two parameters $M$ and $\widetilde{M}$ to tune the power spectrum amplitude, so we obtain a one parameter family of possible values for the parameters. Figure \[fig:mtildem\] shows the relation between $M$ and ${\widetilde M}$ that reproduces the correct amplitude of the power spectrum of the primordial scalar perturbation. ![The relation between $M$ and ${\widetilde M}$ that reproduces the correct amplitude of the power spectrum of the primordial scalar perturbation. For $M<1.6 \times 10^{-5} M_{P}$, the Galileon term induces instability at small scales during the oscillating stage, and the present universe will not be realized. []{data-label="fig:mtildem"}](M_tildeM2.eps){width="11cm"} Instabilities are avoided when $$M>1.6\times 10^{-5}M_P.$$ $\widetilde{M}$ is determined by the value of $M$, which for the smallest value of $M$ is $$\widetilde{M} \simeq 4.8 \times 10^{-6} M_P.$$ For smaller values of $M$, the effect of the Galileon term after inflation becomes too strong compared to the kinetic term, leading to instabilities. Slow-roll inflation can be realized even when $M \rightarrow \infty$, which corresponds to $$\widetilde{M} = 2.7 \times 10^{-6}M_P. \label{kindom}$$ For the smallest values of $M$, or $(M,{\widetilde M})=(1.6 \times 10^{-5} M_P, 4.8 \times 10^{-6} M_P)$, the spectral index becomes $$n_s \simeq 0.964,$$ and the tensor-to-scalar ratio is $$r \simeq 0.155.$$ Raising the value of $M$, both the tensor-to-scalar ratio and the scalar spectral index become smaller. For $M\rightarrow \infty$, we find $$n_s \simeq 0.958,$$ $$r \simeq 0.116.$$ Although the Lagrangian leads to a consistent cosmology, there is no reason for the specific $X^2$ term to dominate over the other higher-order kinetic terms possibly present. In the next section, we generalize the above model and explore its predictions. Before we proceed, we would like to comment on another possible term that could be added, which is of the form $$\varphi^2 X.$$ This term also gives a positive contribution to both the coefficient of $\ddot{\varphi}$ in the equation of motion, and the sound speed squared. In the case in which this term dominates, this model becomes identical to running kinetic inflation [@Nakayama:2010kt] (see [@Cai:2012va] for another possible role of this type of term in the bouncing scenario). Here, we do not pursue this possibility and concentrate on the higher-order kinetic term. Generalized Lagrangian and its predictions ------------------------------------------ Here we further analyze the predictions the model entails by using a generalized version of the above model. The full action we consider is of the form $$S= \int d^4 x \sqrt{-g}\left[ \frac{M_{P}^2}{2}R +k(X) - V(\varphi) +g(\varphi) h(X) \Box \varphi \right].$$ The formulas for the background equations of motion and the perturbations are presented in Appendix A. Let us consider the inflationary stage in which the energy density is dominated by the potential term, and in which the slow-roll parameters defined as $$\begin{aligned} \epsilon:=-\frac{\dot H}{H^2}, \quad \eta:=-\frac{\ddot\varphi}{H\dot\varphi}, \quad \alpha:=\frac{\dot g}{H g},\end{aligned}$$ satisfy $$\begin{aligned} \epsilon,\;|\eta|,\;|\alpha|\ll 1.\end{aligned}$$ The Friedmann equation (\[back1\]) reduces to $$\begin{aligned} 3M_P^2H^2\simeq V,\end{aligned}$$ while Eq. (\[back2\]) implies that $$\begin{aligned} k, \; Xk_X,\; Hgh\dot\varphi \lesssim M_P^2H^2\times {\cal O}(\epsilon).\end{aligned}$$ Eq. (\[back31\]) in the slow-roll approximation is given by $$\begin{aligned} 3k_X H\dot\varphi +V' -18H^2ghm \simeq 0,\end{aligned}$$ where we have used $$\begin{aligned} \frac{2X g''}{H^2g}=\left(-\epsilon + \eta + \alpha+\frac{\dot\alpha}{H\alpha}\right)\alpha\ll 1.\end{aligned}$$ and introduced $m$ defined as $$\begin{aligned} m(X):=\frac{Xh_X}{h}.\end{aligned}$$ Similarly, from Eqs. (\[pertFs\]) and (\[pertGs\]) we have $$\begin{aligned} {\cal F}_S&\simeq \frac{1}{M_P^2 H^2}\left( Xk_X-4Hgh\dot\varphi m \right), \\ {\cal G}_S&\simeq\frac{1}{M_P^2H^2}\left[ Xk_X+2X^2k_{XX}-6 Hgh\dot\varphi \left(m^2+Xm_X\right) \right].\end{aligned}$$ We now consider two extreme cases analytically, one in which the kinetic term dominates the dynamics of the Higgs field, and the other in which the Galileon term dominates the dynamics of the Higgs field. For $Xk_X, X^2k_{XX}\gg Hgh\dot\varphi$, which is the case where the kinetic term dominates, the known result of k-inflation [@ArmendarizPicon:1999rj; @Garriga:1999vw] is reproduced, $$\begin{aligned} {\cal F}_S\simeq \epsilon,\quad c_s^2\simeq \frac{k_X}{k_X+2Xk_{XX}},\end{aligned}$$ where the background equation was used to derive the first equation. The power spectrum of $\zeta$ is given by $$\begin{aligned} {\cal P}_\zeta\simeq\frac{1}{8\pi^2 c_s\epsilon}\left(\frac{H}{M_P}\right)^2, \quad n_s-1\simeq -2\epsilon -\frac{\dot {c_s}}{Hc_s}-\frac{\dot\epsilon}{H\epsilon} \label{ns_kin},\end{aligned}$$ while the tensor-to-scalar ratio is $r\simeq 16c_s\epsilon$. Note in passing that $\dot \epsilon/H\epsilon\simeq 2\epsilon-\eta - \eta/c_s^2$. In the opposite limit, $Xk_X, X^2k_{XX}\ll Hgh\dot\varphi$, we have $$\begin{aligned} {\cal F}_S\simeq \frac{4}{3}\epsilon,\quad c_s^2\simeq \frac{2}{3(m+Xm_X/m)}. \label{cs_m}\end{aligned}$$ We are in particular interested in the model with $h\propto X^m$ where $m={\mathrm{const.}}$ In this case, $c_s^2\simeq 2/(3m)$ and thus $$\begin{aligned} {\cal P}_\zeta\simeq\frac{1}{8\pi^2 \epsilon}\frac{3\sqrt{6m}}{8} \left(\frac{H}{M_P}\right)^2, \quad n_s-1\simeq -2\epsilon -\frac{\dot\epsilon}{H\epsilon}, \quad r\simeq \frac{64}{9}\sqrt{\frac{6}{m}}\epsilon.\end{aligned}$$ Since $\dot \epsilon/H\epsilon \simeq \epsilon+\alpha-(2m+1)\eta$, we find $$\begin{aligned} n_s-1\simeq -3\epsilon-\alpha+(2m+1)\eta. \label{ns_gal}\end{aligned}$$ For $m=1$ the result of the original Higgs G-inflation [@Kamada:2010qe] is reproduced.\ More concretely, we concentrate on the model with $g\propto \varphi^{2n+1}$ and $h\propto X^m$. We take the following ansatz during inflation: $$\begin{aligned} \epsilon=\frac{b}{N+b}, \quad \eta=\frac{c}{N+b}, \quad \alpha=\frac{d}{N+b}, \label{slow_def}\end{aligned}$$ with $b,c$, and $d$ being constant, which will be justified later. Here, $N$ is the number of e-folds defined by $dN:=-Hdt$ and inflation ends at $N=0$. Then, we find $$\begin{aligned} H\sim (N+b)^b,\quad X\sim (N+b)^{2c}, \quad \varphi\sim (N+b)^{c-b+1}, \quad g(\varphi)\sim (N+b)^{-d},\end{aligned}$$ so that $$\begin{aligned} b=2(c-b+1),\quad (2n+1)(c-b+1)=-d \quad\Rightarrow\quad c=\frac{3}{2}b-1,\quad d=-\frac{2n+1}{2}b.\end{aligned}$$ If $Xk_X \gg Hgh\dot\varphi$ and a single term $X^\ell$ dominates over the other terms in $k$, then, we find from the slow-roll equation of motion $$\begin{aligned} 3(c-b+1)=(2\ell -1)c+b.\end{aligned}$$ We therefore arrive at $$\begin{aligned} b=\frac{2\ell-1}{3\ell-2}, \ c=\frac{1}{2(3\ell -2)}, \ d=-\frac{(2n+1)(2\ell-1)}{2(3\ell-2)}. \label{eq:bcd}\end{aligned}$$ Thus, from Eqs. (\[ns\_kin\]) and (\[slow\_def\]), we obtain $$n_s-1 = -\frac{7\ell-4}{(3\ell-2)N+2\ell-1},$$ where we used $$c_s^2 \simeq \frac{1}{2\ell-1}\simeq \mathrm{const.} \label{cs_l}$$ The tensor-to-scalar ratio is $$r=16c_s\epsilon \simeq \frac{16\sqrt{2\ell-1}}{(3\ell -2)N +2\ell -1}.$$ If $Xk_X \ll Hgh\dot\varphi$, again, we find from the slow-roll equation of motion $$\begin{aligned} 3(c-b+1)=2b+(2n+1)(c-b+1)+2mc,\end{aligned}$$ leading to $$\begin{aligned} b=\frac{2m}{3m+n+1}, \ \ c=-\frac{n+1}{3m+n+1}, \ \ d=-\frac{m(2n+1)}{3m+n+1}. \label{eq:bcd2}\end{aligned}$$ Using Eqs. (\[ns\_gal\]) and (\[slow\_def\]), we obtain $$n_s-1 \simeq -\frac{7m+n+1}{(3m+n+1)N+2m},$$ and $$r\simeq\frac{64}{9}\sqrt{\frac{6}{m}}\epsilon \simeq \frac{128\sqrt{6m}}{9\left[(3m+n+1)N + 2m\right]}.\label{analyticr}$$ Note that Eqs. (\[eq:bcd\]) and (\[eq:bcd2\]) show that the constants $b, c,$ and $d$ do not depend on $N$ and hence justify the assumption (\[slow\_def\]). The analytical predictions for these models are presented in Fig. \[fig:prediction\]. Calculations using explicit Lagrangians are presented in Appendix B. The model parameters $\ell, m,$ and $n$ intoduced here are in general arbitrary, but since the sound speed during inflation becomes smaller the larger these parameters are, there is an upper bound on the parameters due to the non-Gaussianities produced. From the Planck constraints on non-Gaussianity [@Ade:2013ydc], we have a lower bound on the sound speed during inflation, $$c_s {\protect\raisebox{-0.7ex}{$\:\stackrel{\textstyle >}{\sim}\:$}}0.02.$$ Comparing this with (\[cs\_m\]) and (\[cs\_l\]), we obtain an upper bound on the model parameters $\ell$ and $m$: $$\ell, \ m {\protect\raisebox{-0.7ex}{$\:\stackrel{\textstyle <}{\sim}\:$}}\mathcal{O}(10^3).$$ ![Analytical predictions of the Galileon term dominating models and the higher-order kinetic term dominating models, compared with the Planck results [@Ade:2013uln]. The double line shows the case where the higher-order kinetic term is dominant, with varying $\ell$, while the other lines show the predictions of the Galileon term dominating case, fixing $n$ and varying $m$ or vice versa. []{data-label="fig:prediction"}](analytic2.eps){width="13cm"} Numerical Analysis ------------------ We now resort to numerical calculations to see which term dominates during inflation, while maintaining stability. For the Higgs field action, we used $$S = \int d^4x\sqrt{-g}\left[\frac{M_P^2}{2} +X + \frac{X^\ell}{\ell \widetilde{M}^{4(\ell-1)}} - \frac{1}{4}\lambda \varphi^4 + \frac{\varphi^{2n+1}X^m}{M^{2n+4m}}\Box \varphi\right].$$ We obtained the sets of $M$ and $\widetilde{M}$ with $\mathcal{P}_\zeta=2.2\times 10^{-9}$ at $N=60$, down to the smallest value of $M$ possible without encountering instabilities. ![Numerical results for $\ell=2$ and $m=1$ with several values of $n$. The continuous lines show the predictions for different combinations of $M$ and $\widetilde{M}$ that lead to $\mathcal{P}_\zeta=2.2\times10^{-9}$ at $N=60$ while avoiding instabilities. The black line is the analytical prediction of the Galileon term dominant case.[]{data-label="fig:prediction1"}](nsr_L222.eps){width="12cm"} ![Results for $\ell=2, m=2$ with several values of $n$, varying the value of $M$.[]{data-label="fig:prediction2"}](nsr_M222.eps){width="12cm"} ![Results for $\ell=3,\ m=1$, and $\ell=5,\ m=5$ for $N=60$ (top), and $\ell=3,\ m=1,\ n=5$ with different numbers of e-folds (bottom). The value of $M$ for each point is shown in the lower figure. []{data-label="fig:prediction3"}](nsr_L332.eps "fig:"){width="12cm"} ![Results for $\ell=3,\ m=1$, and $\ell=5,\ m=5$ for $N=60$ (top), and $\ell=3,\ m=1,\ n=5$ with different numbers of e-folds (bottom). The value of $M$ for each point is shown in the lower figure. []{data-label="fig:prediction3"}](nsr_L3332.eps "fig:"){width="12cm"} For $\ell=2$ and $m=1$, the results are shown in Fig. \[fig:prediction1\]. The focal point at $(n_s, r) = (0.958, 0.116)$ corresponds to the case where the dynamics is dominated by the higher-order kinetic term with no Galileon effect. This corresponds to the upper end of the double line in Fig. \[fig:prediction\], but the actual numerical values are slightly deviated due to the crudeness of the analytic estimate used to draw Fig. \[fig:prediction\]. The fact that the $n=0$ line corresponding to the original Higgs G-inflation is discontinued before reaching the black solid line in Fig. \[fig:prediction1\] indicates that the Galileon dominant case beyond the end point is not viable due to instabilities after inflation. We see that for large values of $n$, the Galileon term can dominate the Higgs field dynamics during inflation, recovering the analytical prediction of the Galileon dominant case, while for $n=0$, the higher-order kinetic term comes into effect whenever instabilities are avoided. We also show several numerical results for other parameters in Figs. \[fig:prediction2\] and \[fig:prediction3\]. These models typically predict a spectral index consistent with the Planc\]k results, and large values for the tensor-to-scalar ratio. In conclusion, our numerical calculation shows that introducing a higher-order kinetic term as well as a higher-order Galileon term helps avoid the unwanted instability during the reheating stage of Higgs G-inflation, while its predictions come to parameter regions that can be tested by the CMB observations in the near future. Conclusion ========== In this paper, we extended the Lagrangian of the Higgs G-inflation model by adding a higher-order kinetic term and modifying the form of the Galileon term. By adding the higher-order kinetic term, both the coefficient of $\ddot{\varphi}$ in the equation of motion and the sound speed squared receive a positive contribution, thereby avoiding the instabilities reported in Ref. [@Ohashi:2012wf]. Modifying the Galileon term enables the term to dominate over the higher-order kinetic term in the equation of motion during inflation, leading to a new class of models. As shown in Figs. \[fig:prediction1\] - \[fig:prediction3\], our model matches the observed range of the spectral index quite well, while the tensor-to-scalar ratio takes values between the 1$\sigma$ and 2$\sigma$ bounds for the parameter range we probed, unlike the original Higgs inflation model. This is good news for those working in B-mode polarization experiments. To suppress the amplitude of $r$, one should take either large $m$ or large $n$. In the limit of large $n$, the spectral index converges to $n_s=1-1/N\simeq 0.983$ for $N=60$, which is outside the 2$\sigma$ range. On the other hand, in the large $m$ limit, $n_s$ approaches $ n_s = 1- \frac{7}{3(N+2/3)}\simeq 0.962$ for $N=60$, which is in the preferred range. Hence, to suppress $r$ in this model, we can introduce higher-order interactions with large exponents. Our model thus accommodates a large region of the $n_s$ - $r$ plane, most of which is consistent with current observations. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported in part by the JSPS Grant-in-Aid for Young Scientists (B) No. 24740161 (T.Ko.), the Grant-in-Aid for Scientific Research on Innovative Areas No. 24111706 (M.Y.) and No. 21111006 (J.Y.), and the Grant-in-Aid for Scientific Research No. 25287054 (M.Y.) and No. 23340058 (J.Y.), and the Program for Leading Graduate Schools, MEXT, Japan (T.Ku.).\ General formulas ================ In this appendix, we give the general formulas for the equations of motion and primordial curvature perturbations. The general cosmological background and perturbation equations for the Lagrangian $$\begin{aligned} {\cal L}=\frac{M_P^2}{2}R+K(\varphi, X)-G(\varphi, X)\Box\varphi\end{aligned}$$ are found in Refs. [@Deffayet:2010qz; @Kobayashi:2010cm; @Kobayashi:2011nu]. The background equation of motion for the scalar field $\varphi$ is given by $$\begin{aligned} &\lefteqn{\left(K_X+2XK_{XX}+6H\dot{\varphi}G_X+6H\dot{\varphi}XG_{XX}-2G_{\varphi} -2XG_{X\varphi} \right)\ddot{\varphi}} \notag \\ &+3H\dot{\varphi}K_X - K_{\varphi} + 2(9H^2+3\dot{H})XG_X - 6 H \dot{\varphi}G_\varphi -2XG_{\varphi\varphi}+6H\dot{\varphi} XG_{X\varphi} + 2X K_{\varphi X}=0, \label{eqn:scalar}\end{aligned}$$ where the subscripts on $K$ and $G$ denote the derivative with respect to those variables. The gravitational field equations are given by $$\begin{aligned} 3M_{P}^2H^2 &= 2XK_X-K+6H\dot{\varphi}XG_{X}-2XG_{\varphi}, \label{eqn:grav1} \\ M_{P}^2\dot{H} &= -XK_X-3H\dot{\varphi}XG_{X}+2XG_{\varphi}+\ddot{\varphi}XG_X. \label{eqn:grav2}\end{aligned}$$ In the main text we focus on inflation models for which the functions $K$ and $G$ are of the form $$\begin{aligned} K=k(X)-V(\varphi),\quad G=-g(\varphi)h(X).\label{kgh}\end{aligned}$$ To study the potential-driven inflation in the above theory, it is convenient to introduce $$\begin{aligned} \epsilon:=-\frac{\dot H}{H^2}, \quad \eta:=-\frac{\ddot\varphi}{H\dot\varphi}, \quad \alpha:=\frac{\dot g}{H g}.\label{slowpar}\end{aligned}$$ The background gravitational field equations are now given by $$\begin{aligned} 3M_{P}^2 H^2&=2Xk_X-k+V-Hgh\dot\varphi\left(6m-\alpha\right),\label{back1} \\ M_{P}^2\dot H&= -Xk_X+Hgh\dot\varphi\left(3m+m\eta -\alpha\right),\label{back2}\end{aligned}$$ where we also introduced $$\begin{aligned} m(X):=\frac{Xh_X}{h}.\end{aligned}$$ The background scalar-field equation is $$\begin{aligned} 0=&H\dot{\varphi}\bigg{[}k_X(3-\eta) +H\dot{\varphi}gh_X\left(-9-3\alpha +6\eta +3\epsilon - \alpha\eta\right) \notag\\ &+2Hg\frac{h}{\dot{\varphi}}(3-\eta)\alpha-2Xk_{XX}\eta +6H\dot{\varphi}Xgh_{XX}\eta \bigg{]} +V'\left(1+\frac{2Xg''h}{V'}\right) \label{back31}\\ =&A\ddot\varphi+3k_X H\dot\varphi +V'\notag\\ &+H^2gh\left[6m\left(-3+\frac{Xk_X}{M_P^2H^2}- \frac{3mgh\dot{\varphi}}{M_P^2H}\right)+6\alpha\left(\frac{mgh\dot{\varphi}}{M_P^2H}-m+1\right)+\frac{2X g''}{H^2g}\right],\label{back3}\end{aligned}$$ where $$\begin{aligned} A=k_X+2Xk_{XX}-6\frac{Hgh\dot\varphi}{X}\left(m^2+Xm_X\right)+\frac{Hgh\dot\varphi}{X}(m+1)\alpha + \frac{6m^2g^2h^2}{M_P^2}.\end{aligned}$$ Here, we have eliminated $\dot{H}$ in Eq. (\[back3\]) using Eq. (\[back2\]). If $A(t_*)=0$ at some $t=t_*$, we cannot solve the time evolution of $\varphi$ for $t>t_*$. Therefore, we require that $A$ never crosses the zero. Note here that we have not employed any slow-roll approximations; the above equations can be used in the reheating stage as well as during inflation. The second-order action $S_2$ for the primordial curvature perturbation $\zeta$ is given by [@Kobayashi:2011nu] $$S_2 = M_P^2 \int d^4 x \,a^3\left[\mathcal{G}_S \dot{\zeta}^2 -\frac{\mathcal{F}_S}{a^2}(\vec{\nabla}\zeta)^2\right],$$ where $$\begin{aligned} \mathcal{F}_S &= \frac{M_P^2X}{\Theta^2}\bigg{[}K_X - 2 G_\varphi + 4 H \dot{\varphi} G_X + 2 \ddot{\varphi} G_X +2X \ddot{\varphi} G_{XX} + 2X G_{X\varphi} - \frac{2}{M_P^2}X^2 G_X^2\bigg{]}\\ &=\frac{M_P^2X}{\Theta^2}\left\{ k_X+\frac{Hgh\dot\varphi}{X}\left[-4m-(m-1)\alpha+2(m^2+Xm_X)\eta\right] -\frac{2}{M_P^2}g^2h^2m^2 \right\}, \\ {\cal G}_S&= \frac{M_P^2X}{\Theta^2}\bigg{[}K_X+2XK_{XX} - 2 G_\varphi + 6 H \dot{\varphi} G_X +6 H \dot{\varphi}X G_{XX} - 2X G_{X\varphi} + \frac{6}{M_P^2}X^2 G_X^2\bigg{]}\\ &=\frac{M_P^2X}{\Theta^2}\left\{ k_X+2Xk_{XX}+\frac{Hgh\dot\varphi}{X}\left[ -6\left(m^2+Xm_X\right)+(m+1)\alpha\right] +\frac{6}{M_P^2}g^2h^2m^2 \right\},\end{aligned}$$ and $$\begin{aligned} \Theta = M_P^2H\left(1+m\frac{Hgh\dot\varphi}{M_P^2H^2}\right).\end{aligned}$$ The sound speed is given by $c_s^2={\cal F}_S/{\cal G}_S$. The curvature perturbation shows a stable evolution provided that ${\cal F}_S>0$ and ${\cal G}_S>0$. Note again that no slow-roll approximations are made in deriving the above expressions, and hence the same conditions ${\cal F}_S>0$ and ${\cal G}_S>0$ can be used to judge the stability during the reheating stage. Analytical calculations {#sec4} ======================= In this section, we give analytic formulas using an explicit action of the form $$S = \int d^4x\sqrt{-g}\left[\frac{M_P^2}{2}R + X + \frac{X^\ell}{\ell \widetilde{M}^{4(\ell-1)}} - \frac{1}{4}\lambda \varphi^4 + \frac{\varphi^{2n+1}X^m}{M^{2n+4m}}\Box \varphi\right]$$ for the two extreme cases; one in which the higher-order kinetic term dominates the equation of motion, and the other in which the Galileon term dominates the equation of motion. This action corresponds to taking $$g(\varphi) = \varphi\left(\frac{\varphi^2}{M^2}\right)^n,\ \ h(X) = \left(\frac{X}{M^4}\right)^m,$$ and the term proportional to $X^\ell$ is to be understood as the dominant higher-order kinetic term during inflation. Evolution equations ------------------- From Eqs. (\[back1\]) - (\[back3\]) the evolution equations for this action are $$\begin{aligned} 3M_{P}^2 H^2&=X + \frac{(2\ell-1)X^\ell}{\ell\widetilde{M}^{4(\ell-1)}}+\frac{1}{4}\lambda \varphi^4-\frac{H\varphi^{2n+1}\dot{\varphi}X^m}{M^{2n+4m}}\left(6m-\frac{(2n+1)\dot{\varphi}}{H\varphi}\right),\label{back11} \\ M_{P}^2\dot{H}&= -X -\frac{X^\ell}{\widetilde{M}^{4(\ell - 1)}}+\frac{H\varphi^{2n+1}\dot{\varphi}X^m}{M^{2n+4m}}\left(3m-\frac{m\ddot{\varphi}}{H\dot{\varphi}} -\frac{(2n+1)\dot{\varphi}}{H\varphi}\right),\label{back12}\end{aligned}$$ $$\begin{aligned} &\left(1+\frac{(2\ell-1)X^{\ell-1}}{\widetilde{M}^{4(\ell-1)}}-\frac{6m^2H\varphi^{2n+1}\dot{\varphi}X^{m-1}}{M^{2n+4m}}+\frac{2(2n+1)(m+1)\varphi^{2n}X^{m}}{M^{2n+4m}} + \frac{6m^2\varphi^{4n+2}X^{2m}}{M_P^2M^{4n+8m}}\right)\ddot\varphi\notag\\ &+3H\dot\varphi\left(1+ \frac{X^{\ell-1}}{\widetilde{M}^{4(\ell-1)}}\right) +\lambda \varphi^3 +\frac{H^2\varphi^{2n+1}X^{m}}{M^{2n+4m}}\left[-9m+\frac{3mX}{M_P^2H^2}+ \frac{3mX^\ell}{\ell M_P^2\widetilde{M}^{4(\ell-1)}H^2}\right.\notag\\ &\left.+\frac{6m(2n+1)\varphi^{2n}X^{m+1}}{M_P^2M^{2n+4m}H^2}-\frac{6(2n+1)(m-1)\dot{\varphi}}{H\varphi}-\frac{3m\lambda\varphi^4}{4M_P^2H^2}+\frac{4n(2n+1)X}{H^2\varphi^2}\right] =0,\label{back13}\end{aligned}$$ The coefficients in $S_2$ become $$\mathcal{F}_S=\frac{M_P^2X}{\Theta^2}\left[1 + \frac{X^{\ell-1}}{\widetilde{M}^{4(\ell-1)}} -\frac{H\varphi^{2n+1}\dot{\varphi}X^{m-1}}{M^{2n+4m}}\left(4m+\frac{(m-1)(2n+1)\dot{\varphi}}{H\varphi}+\frac{2m^2\ddot{\varphi}}{H\dot{\varphi}}\right)-\frac{2m^2\varphi^{4n+2}X^{2m}}{M_P^2M^{4n+8m}}\right], \label{pertFs}$$ $$\mathcal{G}_S=\frac{M_P^2X}{\Theta^2}\left[1 + \frac{(2\ell-1)X^{\ell-1}}{\widetilde{M}^{4(\ell-1)}}- \frac{H\varphi^{2n+1}\dot{\varphi}X^{m-1}}{M^{2n+4m}}\left(6m^2-\frac{(m+1)(2n+1)\dot{\varphi}}{H\varphi}\right)+\frac{6m^2\varphi^{4n+2}X^{2m}}{M_P^2M^{4n+8m}}\right], \label{pertGs}$$ with $$\Theta=M_P^2H + \frac{m\varphi^{2n+1}\dot{\varphi}X^m}{M^{2n+4m}}.$$ In the following, we analytically investigate the system introduced above for the two extreme cases. $X^\ell$ domination ------------------- First, we investigate the case in which the higher-order kinetic term dominates the inflation dynamics. This corresponds to the condition $$k_X \gg |H\dot{\varphi}gh_X|,\ 1\label{galileonkinX}$$ throughout inflation. Assuming the slow-roll conditions, $$|\epsilon|,\ |\eta|,\ |\alpha| \ll 1, \label{slow-rollX}$$ Eqs. (\[back11\])-(\[back13\]) can be approximated as $$3M_P^2H^2 \simeq V = \frac14 \lambda \varphi^4, \label{friedmann}$$ $$M_P^2\dot{H} \simeq -\frac{X^\ell}{\widetilde{M}^{4(\ell-1)}}, \label{friedmannX}$$ $$3H\dot{\varphi}\frac{X^{\ell-1}}{\widetilde{M}^{4(\ell-1)}}+ \lambda \varphi^3 \simeq 0.$$ Solving the last equation for $\dot{\varphi}$, we obtain $$\dot{\varphi} \simeq -\bigg{(}\sqrt{\frac{\lambda}{3}}2^\ell \widetilde{M}^{4(\ell-1)}M_P\varphi\bigg{)}^\frac{1}{2\ell-1}, \label{varphidotX}$$ where use has been made of Eq. (\[friedmann\]) to eliminate $H$. The number of e-folds $N$ can be calculated as $$\begin{aligned} N &= \int Hdt = \int \frac{H}{\dot{\varphi}} d\varphi \notag \\ &=-\int\sqrt{\frac{\lambda}{12}}\frac{\varphi^2}{M_{P}}\bigg{|}\sqrt{\frac{\lambda}{3}}2^\ell\widetilde{M}^{4(\ell-1)} {M_{P}}\varphi\bigg{|}^{-\frac{1}{2\ell-1}}d\varphi \notag \\ &= -\frac{2\ell-1}{3\ell-1}2^{-\frac{5\ell-2}{2\ell-1}}3^{-\frac{\ell-1}{2\ell-1}}\lambda^{\frac{\ell-1}{2\ell-1}}\widetilde{M}^{-\frac{{4(\ell-1)}}{2\ell-1}}M_{P}^{-\frac{2\ell}{2\ell-1}}\left[\varphi^{\frac{6\ell-4}{2\ell-1}}\right]^{\varphi_\mathrm{end}}_\varphi.\end{aligned}$$ $\varphi_{\mathrm{end}}$ is given by the condition $\epsilon=1$, which reads $$\varphi_{\mathrm{end}} = 2^{\frac{5\ell-2}{6\ell-4}}3^{\frac{\ell-1}{6\ell-4}}\lambda^{-\frac{\ell-1}{6\ell-4}}\widetilde{M}^{\frac{{4(\ell-1)}}{6\ell-4}}{M_{P}^{\frac{\ell}{3\ell-2}}}.$$ This leads to $$N = -\frac{2\ell-1}{3\ell-2}+\frac{2\ell-1}{3\ell-2}2^{-\frac{5\ell-2}{2\ell-1}}3^{-\frac{\ell-1}{2\ell-1}}\lambda^{\frac{\ell-1}{2\ell-1}}\widetilde{M}^{-\frac{{4(\ell-1)}}{2\ell-1}}{M_{P}^{-\frac{2\ell}{2\ell-1}}}\varphi^{\frac{6\ell-4}{2\ell-1}},$$ and hence we express the value of $\varphi$ in terms of $N$: $$\varphi = \left(\frac{3\ell-2}{2\ell-1}\right)^{\frac{2\ell-1}{6\ell-4}}2^{\frac{5\ell-2}{6\ell-4}} 3^{\frac{\ell-1}{6\ell-4}}\lambda^{-\frac{\ell-1}{6\ell-4}}\widetilde{M}^\frac{{4(\ell-1)}}{6\ell-4}{M_{P}^{\frac{\ell}{3\ell-2}}}\left(N +\frac{2\ell-1}{3\ell-2}\right)^{\frac{2\ell-1}{6\ell-4}}.$$ Then, the slow-roll parameters can be expressed in terms of $N$ as $$\begin{aligned} &\epsilon = -\frac{\dot{H}}{H^2}=-\frac{1}{2H}\frac{d}{dt}\ln H^2 \simeq -\frac{2\dot{\varphi}}{H\varphi} \simeq \frac{2\ell-1}{3\ell-2} \left(N+ \frac{2\ell-1}{3\ell-2}\right)^{-1}, \\ &\eta = -\frac{\ddot{\varphi}}{H\dot{\varphi}} =-\frac{1}{H}\frac{d}{dt}\ln|\dot{\varphi}| \simeq -\frac{\dot{\varphi}}{(2\ell-1)H\varphi} \simeq \frac{1}{2(3\ell-2)} \left(N+ \frac{2\ell-1}{3\ell-2}\right)^{-1}. \label{eq:X2slow}\end{aligned}$$ Thus, by an explicit calculation we have established the validity of the ansatz (\[slow\_def\]); $b=(2\ell-1)/(3\ell-2), c=1/2(3\ell-2)$. The parameter $\widetilde{M}$ is determined by calculating the curvature perturbations generated from this model. In the present case, from Eqs. (\[pertFs\]) and (\[pertGs\]), the coefficients in the perturbation action $S_2$ read $$\begin{aligned} \mathcal{F}_S &\simeq \frac{X^\ell}{H^2M_P^2\widetilde{M}^{4(\ell-1)}}, \nonumber \\ \mathcal{G}_S &\simeq \frac{(2\ell-1)X^\ell}{H^2M_P^2\widetilde{M}^{4(\ell-1)}}.\end{aligned}$$ This leads to the sound speed $$c_s^2 := \frac{\mathcal{F}_S}{\mathcal{G}_S} \simeq \frac{1}{2\ell-1}.$$ The curvature perturbations in the present case are expressed in terms of $N$ as $$\begin{aligned} \mathcal{P}_\zeta &= \frac{1}{8\pi^2 c_s {\mathcal{F}_S}} \left(\frac{H}{M_P}\right)^2 \notag\\ &= (3\ell-2)^{\frac{7\ell-4}{3\ell-2}}(2\ell-1)^{-\frac{11\ell-6}{6\ell-4}}2^{-\frac{5\ell-6}{3\ell-2}} 3^{-\frac{\ell}{3\ell-2}} \pi^{-2} \lambda^{\frac{\ell}{3\ell-2}} \left(\frac{\widetilde{M}}{M_P}\right)^{\frac{8(\ell-1)}{3\ell-2}} \left(N+\frac{2\ell-1}{3\ell-2}\right)^{\frac{7\ell-4}{3\ell-2}}.\end{aligned}$$ Solving this for $\widetilde{M}$, we have $$\widetilde{M}=\left(\mathcal{P}_\zeta \pi^2\right)^\frac{3\ell-2}{8(\ell-1)} (3\ell-2)^{-\frac{7\ell-4}{8(\ell-1)}}(2\ell-1)^{\frac{11\ell-6}{16(\ell-1)}} 2^{\frac{5\ell-6}{8(\ell-1)}}3^{\frac{\ell}{8(\ell-1)}}\lambda^{-\frac{\ell}{8(\ell-1)}}\left(N+\frac{2\ell-1}{3\ell-2}\right)^{-\frac{7\ell-4}{8(\ell-1)}}M_P.$$ For $\ell=2$, $N=60$ and $\lambda=0.01$, we obtain $$\widetilde{M} \simeq 2.7\times 10^{-6}M_P$$ to explain the present universe with $\mathcal{P}_\zeta = 2.2 \times 10^{-9}$. The spectral index and tensor-to-scalar ratio can also be calculated analytically. The spectral index is calculated using the expression for the power spectrum: $$n_s -1 = \frac{d\ln \mathcal{P}_\zeta }{d\ln k} \simeq \frac{1}{H}\frac{d}{dt}\ln \left(H^4 \dot{\varphi}^{-2\ell}\right) =-4\epsilon + 2\ell\eta.$$ For the tensor power spectrum, we obtain the same expression as in the usual canonical potential driven inflation: $$\mathcal{P}_t = \frac{8}{M_P^2}\left(\frac{H}{2\pi}\right)^2.$$ Thus, the tensor-to-scalar ratio is calculated as $$r:= \frac{\mathcal{P}_t}{\mathcal{P}_\zeta} = 16 c_s \mathcal{F}_S \simeq \frac{16\sqrt{2\ell-1}}{(3\ell-2)N+2\ell-1}.$$ The tensor-to-scalar ratio becomes smaller for larger $\ell$. Galileon domination ------------------- Next, we investigate the case in which the Galileon term dominates over the newly introduced higher-order kinetic term, namely $$|H\dot{\varphi}gh_X| =\left|H\frac{m \varphi^{2n+1} \dot{\varphi}^{2m-1}}{2^{m-1}M^{2n+4m}}\right|\gg k_X\label{galileonkin}$$ during inflation. This happens when we take large $n\gg1$. In this case, the analytic formulas in Ref. [@Kamada:2010qe] can be used with slight modifications. Assuming the slow-roll conditions, $$|\epsilon|,\ |\eta|,\ |\alpha| \ll 1, \label{slow-roll}$$ Eqs. (\[back11\])-(\[back13\]) can be approximated as $$3M_P^2H^2 \simeq V = \frac14 \lambda \varphi^4, \label{friedmannG}$$ $$M_P^2\dot{H} \simeq \frac{3mH\varphi^{2n+1}\dot{\varphi}^{2m+1}}{2^{m}M^{2n+4m}},$$ $$\frac{9mH^2 \varphi^{2n+1} \dot{\varphi}^{2m}}{2^{m-1}M^{2n+4m}}+ \lambda \varphi^3 \simeq 0.\label{sloweomG}$$ Solving the last equation for $\dot{\varphi}$, we obtain $$\dot{\varphi} \simeq -\bigg{(}M_P \frac{2^{\frac{m+1}{2}}M^{n+2m}}{\sqrt{3m}\varphi^{n+1}}\bigg{)}^\frac{1}{m}, \label{varphidot}$$ where we used Eq. (\[friedmannG\]) and took the negative sign for $\dot{\varphi}$. From the numerical calculations in the main text, we see that in order to avoid instabilities after inflation, the higher-order kinetic term has to dominate over the Galileon term by the end of inflation. This situation makes analytical calculations substantially difficult, so here we focus on the parameter regions where we can ignore the number of e-folds of inflation during which the equation of motion is dominated by the higher-order kinetic term. This is realized when we have a large value of $n$, as seen from the numerical calculations. In this case, we can approximately calculate the number of e-folds by assuming that the Galileon term dominated until the end of inflation: $$\begin{aligned} N &= \int Hdt = \int \frac{H}{\dot{\varphi}} d\varphi \notag \\ &=-\int\sqrt{\frac{\lambda}{12}}\frac{\varphi^2}{M_P}\bigg{(}\frac{\sqrt{3m}\varphi^{n+1}}{M_P2^{\frac{m+1}{2}}M^{n+2m}}\bigg{)}^\frac{1}{m}d\varphi \notag \\ &= -\sqrt{\frac{\lambda}{12}}\frac{1}{M_P}\bigg{(}\frac{\sqrt{3m}}{M_P 2^{\frac{m+1}{2}}M^{n+2m}}\bigg{)}^\frac{1}{m}\frac{m}{3m + n+ 1}\bigg{[}\varphi^{\frac{3m+n+1}{m}}\bigg{]}^{\varphi_{\mathrm{end}}}_\varphi ,\label{efoldg}\end{aligned}$$ where we used Eq. (\[varphidot\]). $\varphi_{\mathrm{end}}$ is determined by the condition that $$\epsilon = - \frac{\dot{H}}{H^2} \simeq \frac{9XH\dot{\varphi}G_X}{V} =1,$$ which gives $$\varphi_{\mathrm{end}}\simeq \bigg{[} 2^\frac{5m+1}{2m} 3^\frac{m-1}{2m} m^{-\frac{1}{2m}} \lambda^{-\frac{1}{2}} M_P^\frac{m+1}{m} M^\frac{2m+n}{m} \bigg{]}^{\frac{m}{3m+n+1}}. \label{end}$$ Inserting this value into Eq. (\[efoldg\]) leads to $$N = \frac{m}{3m+n+1} \left( 2^{-\frac{3m+1}{2m}} 3^{-\frac{m-1}{2m}} m^{\frac{1}{2m}} \lambda^{\frac{1}{2}} M_P^{-\frac{m+1}{m}} M^{-\frac{2m+n}{m}} \varphi^{\frac{3m+n+1}{m}} -2 \right),$$ which yields the relation between the value of the Higgs field and e-foldings $N$ before the end of inflation, $$\varphi = \left[ 2^{3m+1} 3^{m-1} m^{-1} \lambda^{-m} M_P^{2(m+1)} M^{2(2m+n)} \left( \frac{3m+n+1}{m}N+2 \right)^{2m} \right]^{\frac{1}{2(3m+n+1)}}. \label{varphiend}$$ Then, the slow-roll parameters can be expressed in terms of $N$ as $$\begin{aligned} &\epsilon = -\frac{\dot{H}}{H^2}=-\frac{1}{2H}\frac{d}{dt}\ln H^2 \simeq -\frac{2\dot{\varphi}}{H\varphi} \simeq \frac{2m}{N(3m+n+1)+2m}, \\ &\eta = -\frac{\ddot{\varphi}}{H\dot{\varphi}} =-\frac{1}{H}\frac{d}{dt}\ln|\dot{\varphi}| \simeq \frac{n+1}{mH}\frac{d}{dt}\ln \varphi \simeq -\frac{n+1}{N(3m+n+1)+2m}, \\ &\alpha = \frac{{(2n+1)}\dot{\varphi}}{H\varphi} \simeq {-}\frac{{(2n+1)}m}{N(3m+n+1)+2m}. \label{eq:gslow}\end{aligned}$$ The parameter $M$ is determined by calculating the curvature perturbations generated from this model. In the present case, the coefficients $\mathcal{F}_S$ and $\mathcal{G}_S$ are given, from Eqs. (\[pertFs\]) and (\[pertGs\]), by $$\begin{aligned} \mathcal{F}_S &\simeq - \frac{2m \varphi^{2n+1} \dot{\varphi}^{2m+1}}{2^{m-1}HM_{Pl}^2M^{2n+4m}}, \nonumber \\ \mathcal{G}_S &\simeq - \frac{3m^2 \varphi^{2n+1} \dot{\varphi}^{2m+1}}{2^{m-1}HM_{Pl}^2M^{2n+4m}}.\end{aligned}$$ This leads to the sound speed $$c_s^2 := \frac{\mathcal{F}_S}{\mathcal{G}_S} \simeq \frac{2}{3m}.$$ The curvature perturbation in the present case is expressed in terms of $N$ as $$\begin{aligned} \mathcal{P}_\zeta &= \frac{1}{8\pi^2 c_s \mathcal{F}_S} \left(\frac{H}{M_P}\right)^2 = - \frac{2^{m-5} H^3 M^{4m+2n}}{m \pi^2 c_s \varphi^{2n+1} \dot{\varphi}^{2m+1}} \nonumber \\ &= \pi^{-2} \left[ 2^{-\frac{39m+17n+13}{2}} 3^{\frac{7m+n-3}{2}} m^{\frac{3m+n-3}{2}} \lambda^{m+n+1} \right. \nonumber \\ & \qquad \left. M_P^{-4(2m+n)} M^{4(2m+n)} \left( \frac{3m+n+1}{m}N + 2 \right)^{7m+n+1} \right]^{\frac{1}{3m+n+1}}. \label{Gpower}\end{aligned}$$ By solving this equation for $M$, we obtain the expression for the model parameter $M$: $$\begin{aligned} M = \left[ \left(\mathcal{P}_\zeta \pi^2 \right)^{3m+n+1} 2^{\frac{39m+17n+13}{2}} 3^{-\frac{7m+n-3}{2}} m^{-\frac{3m+n-3}{2}} \lambda^{-(m+n+1)} \left( \frac{3m+n+1}{m}N + 2 \right)^{-(7m+n+1)} \right]^{\frac{1}{4(2m+n)}}M_P.\end{aligned}$$ Substituting this back into Eq. (\[varphiend\]), we acquire the expression for the value of the Higgs field: $$\varphi \simeq [\mathcal{P}_\zeta \pi^2]^\frac{1}{4} 2^{\frac{17}{8}} 3^{-\frac18} m^{-\frac{1}{8}} \lambda^{-\frac{1}{4}} \left( \frac{3m+n+1}{m}N+2 \right)^{-\frac{1}{4}} M_{P}. \label{gphi}$$ The spectral index can be calculated using the expression for the power spectrum Eq. (\[Gpower\]): $$\begin{aligned} n_s -1 = \frac{d\ln \mathcal{P}_\zeta }{d\ln k}&= 3\frac{d\ln H}{d\ln k} - (2n+1)\frac{d\ln \varphi}{d\ln k} - (2m+1) \frac{d\ln \dot{\varphi}}{d\ln k} \notag\\ &=-3\epsilon {- \alpha} +(2m+1)\eta. $$ Inserting the expressions for the slow-roll parameters Eq. (\[eq:gslow\]) gives $$n_s - 1 \simeq -\frac{7m+n+1}{N(3m+n+1)+2m},$$ which agrees with the calculation in the main text. The tensor power spectrum is of the same expression as before, $$\mathcal{P}_t = \frac{8}{M_P^2}\left(\frac{H}{2\pi}\right)^2.$$ Thus, the tensor-to-scalar ratio becomes $$r:= \frac{\mathcal{P}_t}{\mathcal{P}_\zeta} = 16c_s \mathcal{F}_S = - \frac{c_s {m}\varphi^{2n+1} \dot{\varphi}^{2m+1}} {2^{m-6}HM_P^2M^{4m+2n}} \simeq \frac{2^{\frac{15}{2}} 3^{-\frac32} m^{\frac12} }{N(3m+n+1)+2m}.$$ The tensor-to-scalar ratio becomes smaller for larger $n$, which is the region where the analysis here is valid. [99]{} A. A. Starobinsky, Phys. Lett. B [**91**]{}, 99 (1980). K. Sato, Mon. Not. Roy. Astron. Soc.  [**195**]{}, 467 (1981). A. H. Guth, Phys. Rev. D [**23**]{}, 347 (1981). G. Aad [*et al.*]{} \[ATLAS Collaboration\], Phys. Lett. B [**716**]{}, 1 (2012) \[arXiv:1207.7214 \[hep-ex\]\]. S. Chatrchyan [*et al.*]{} \[CMS Collaboration\], Phys. Lett. B [**716**]{}, 30 (2012) \[arXiv:1207.7235 \[hep-ex\]\]. Y. Fukuda [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Rev. Lett.  [**81**]{}, 1562 (1998) \[hep-ex/9807003\]. M. C. Gonzalez-Garcia and M. Maltoni, Phys. Rept.  [**460**]{}, 1 (2008) \[arXiv:0704.1800 \[hep-ph\]\]. J. L. Cervantes-Cota and H. Dehnen, Nucl. Phys. B [**442**]{}, 391 (1995) \[astro-ph/9505069\]. F. L. Bezrukov and M. Shaposhnikov, Phys. Lett. B [**659**]{}, 703 (2008) \[arXiv:0710.3755 \[hep-th\]\]. A. O. Barvinsky, A. Y. .Kamenshchik and A. A. Starobinsky, JCAP [**0811**]{}, 021 (2008) \[arXiv:0809.2104 \[hep-ph\]\]. C. Germani and A. Kehagias, Phys. Rev. Lett.  [**105**]{}, 011302 (2010) \[arXiv:1003.2635 \[hep-ph\]\]. K. Kamada, T. Kobayashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. D [**83**]{}, 083515 (2011) \[arXiv:1012.4238 \[astro-ph.CO\]\]. K. Nakayama and F. Takahashi, JCAP [**1011**]{}, 009 (2010) \[arXiv:1008.2956 \[hep-ph\]\]. T. Kobayashi, M. Yamaguchi and J. Yokoyama, Prog. Theor. Phys.  [**126**]{}, 511 (2011) \[arXiv:1105.5723 \[hep-th\]\]. K. Kamada, T. Kobayashi, T. Takahashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. D [**86**]{}, 023504 (2012) \[arXiv:1203.4059 \[hep-ph\]\]. Y. Hamada, H. Kawai and K. -y. Oda, arXiv:1308.6651 \[hep-ph\]. C. Deffayet, O. Pujolas, I. Sawicki and A. Vikman, JCAP [**1010**]{}, 026 (2010) \[arXiv:1008.0048 \[hep-th\]\]. T. Kobayashi, M. Yamaguchi and J. Yokoyama, Phys. Rev. Lett.  [**105**]{}, 231302 (2010) \[arXiv:1008.0603 \[hep-th\]\]. J. Ohashi and S. Tsujikawa, JCAP [**1210**]{}, 035 (2012) \[arXiv:1207.4879 \[gr-qc\]\]. J. A. Casas, J. R. Espinosa and M. Quiros, Phys. Lett. B [**342**]{}, 171 (1995) \[hep-ph/9409458\]. J. Elias-Miro, J. R. Espinosa, G. F. Giudice, G. Isidori, A. Riotto and A. Strumia, Phys. Lett. B [**709**]{}, 222 (2012) \[arXiv:1112.3022 \[hep-ph\]\]. P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], arXiv:1303.5076 \[astro-ph.CO\]. Y. -F. Cai, D. A. Easson and R. Brandenberger, JCAP [**1208**]{}, 020 (2012) \[arXiv:1206.2382 \[hep-th\]\]. C. Armendariz-Picon, T. Damour and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 209 (1999) \[hep-th/9904075\]. J. Garriga and V. F. Mukhanov, Phys. Lett. B [**458**]{}, 219 (1999) \[hep-th/9904176\]. P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], arXiv:1303.5084 \[astro-ph.CO\]. P. A. R. Ade [*et al.*]{} \[Planck Collaboration\], arXiv:1303.5082 \[astro-ph.CO\]. [^1]: [email protected] [^2]: [email protected] [^3]: [email protected] [^4]: [email protected] [^5]: [email protected] [^6]: Another Higgs inflation model without any higher derivative couplings was also proposed recently [@Hamada:2013mya]. [^7]: One may wonder if $\lambda$ can be (accidentally) of the order of $10^{-13}$ at the inflationary scale, which explains the correct magnitude of the primordial curvature perturbations with robust oscillation after inflation. However, even in this case, the tensor-to-scalar ratio is too large to be compatible with the Planck results as long as the logarithmic correction is not so significant. [^8]: Here we assume that the new physics does not affect the inflation dynamics except for modification of the equation of motion due to the Galileon term. [^9]: We did not reproduce the result from [@Ohashi:2012wf], which claims the upper bound of the Higgs quartic coupling $\lambda$ as $\lambda<2.7 \times 10^{-8}$.

ArXiv

--- abstract: 'We prove several types of scaling results for Wigner distributions of spectral projections of the isotropic Harmonic oscillator on ${{\mathbb R}}^d$. In prior work, we studied Wigner distributions $W_{\hbar, E_N(\hbar)}(x, \xi)$ of individual eigenspace projections. In this continuation, we study Weyl sums of such Wigner distributions as the eigenvalue $E_N(\hbar)$ ranges over spectral intervals $[E - \delta(\hbar), E + \delta(\hbar)]$ of various widths $\delta(\hbar)$ and as $(x, \xi) \in T^*{{\mathbb R}}^d$ ranges over tubes of various widths around the classical energy surface $\Sigma_E \subset T^*{{\mathbb R}}^d$. The main results pertain to interface Airy scaling asymptotics around $\Sigma_E$, which divides phase space into an allowed and a forbidden region. The first result pertains to $\delta(\hbar) = \hbar$ widths and generalizes our earlier results on Wigner distributions of individual eigenspace projections. Our second result pertains to $\delta(\hbar) = \hbar^{2/3}$ spectral widths and Airy asymptotics of the Wigner distributions in $\hbar^{2/3}$-tubes around $\Sigma_E$. Our third result pertains to bulk spectral intervals of fixed width and the behavior of the Wigner distributions inside the energy surface, outside the energy surface and in a thin neighborhood of the energy surface.' address: - 'Department of Mathematics, Texas A&M and Facebook AI Research, NYC' - 'Department of Mathematics, Northwestern University, Evanston, IL 60208, USA' author: - Boris Hanin and Steve Zelditch title: 'Interface asymptotics of Wigner-Weyl distributions for the Harmonic Oscillator' --- [^1] \[INTRO\] Introduction ====================== This article is part of a series [@HZ19] studying the scaling asymptotics of Wigner distributions $$\sum_{N \in {\mathbb N}} f_{\hbar, E} (E_N(\hbar)) W_{\hbar, E_N(\hbar)}(x, \xi),\;\; (x, \xi) \in T^*{{\mathbb R}}^d$$ of various spectral projection kernels for the isotropic Harmonic Oscillator, $$\label{Hh} \widehat{H}_{h} = \sum_{j = 1}^d \left(- \frac{\hbar^2}{2} \frac{\partial^2 }{\partial x_j^2} + \frac{ x_j^2}{2} \right) \quad\mathrm{on}\quad L^2({{\mathbb R}}^d,dx).$$ As is well-known, the spectrum of $\widehat{H}_{\hbar}$ consists of the eigenvalues, $$\label{ENh} E_N(\hbar)=\hbar{\ensuremath{\left(N+d/2\right)}},\qquad N = 0, 1, 2, \dots.$$ We denote the corresponding eigenspaces by $$\label{VhE} V_{\hbar, E_N(\hbar)}: = \{\psi \in L^2({{\mathbb R}}^d): \hat{H}_{\hbar} \psi = E_N(\hbar) \psi \},\;\; (\dim V_{\hbar, E_N(\hbar)} \simeq N^{d-1}),$$ and the eigenspace projections by $$\label{PiDEF} \Pi_{\hbar, E_N(\hbar)}: L^2({{\mathbb R}}^d) \to V_{\hbar,E_N(\hbar)}.$$ The semi-classical Wigner distributions of the projections are defined by $$\label{WIGNERDEF1} W_{\hbar, E_N(\hbar)}(x, \xi) := \int_{{{\mathbb R}}^d} \Pi_{\hbar, E_N(\hbar)} \left( x+\frac{v}{2}, x-\frac{v}{2} \right) e^{-\frac{i}{\hbar} v \cdot \xi} \frac{dv}{(2\pi h)^d} .$$ When $E_N(\hbar) = E + o(1)$ as $\hbar \to 0$, $W_{\hbar, E_N(\hbar)}$ is thought of as the ‘quantization’ of the energy surface, $$\label{SIGMAEDEF} \Sigma_E =\{(x, \xi) \in T^*{{\mathbb R}}^d: H(x, \xi): = {{\textstyle \frac 12}}(||x||^2 + ||\xi||^2) = E\},$$ and is thought of as an approximate $\delta$-function on . This is true in the weak\* sense, but the pointwise behavior is quite a bit more complicated and is studied in [@HZ19]. The purpose of this article is to study ‘Weyl sums’ of Wigner distributions (the ‘Wigner-Weyl distributions’ of the title) of the form $$\label{Whfdelta} W_{\hbar, f, \delta(\hbar)}(x, \xi): = \sum_{N} f{\ensuremath{\left(\delta(\hbar)^{-1}(E- E_N(\hbar))\right)}}W_{ \hbar, E_N}(x, \xi),$$ where $\delta(h)$ controls the width and $f$ controls the smoothness of the localization. Thus, the essential point is to simultaneously study two types of ‘localization’: - Spectral localization, determined by the choice of scale $\delta(\hbar)$. For instance, if $\delta(\hbar)=\hbar^\gamma$ and $f$ is the indicator function of the interval $[-a,b]$, then $W_{\hbar, f, \delta(\hbar)}(x, \xi)$ the is the sum of eigenspace Wigner distributions $W_{\hbar, E_N(\hbar)}$ for energies $E_N(\hbar)$ in the spectral interval $[E-b\hbar^\gamma, E+a\hbar^\gamma]$; - Phase space localization, determined by a choice of region $U \subset T^* {{\mathbb R}}^d$ over which $(x,\xi)$ can vary. For example, in different results we will consider $U$ to be the $\hbar^{2/3}$ tube around the energy surface $\Sigma_E$ as well as regions deeper in the interior or exterior of $\{H \leq E\}$. In [@HZ19] we studied the extreme case of spectral localization, where the window contained just one distinct eigenvalue $E_N(\hbar)$ and it was assumed that $E_N(\hbar) = E$ for a fixed $E \in {{\mathbb R}}_+$. In particular, we studied the asymptotics of $W_{\hbar, E_N(\hbar)}(x,\xi)$ in different regions around $\Sigma_E$. Roughly, we found with $H(x,\xi)=({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2)/2$ denoting the classical Hamiltonian that for fixed $E$, $$(2\pi)^d W_{\hbar, E}(x,\xi) ~\simeq~ \begin{cases} \hbar^{-d+1/2}H(x,\xi)^{-d/2}{\ensuremath{\times}}\mathrm{~oscillatory},&\quad 0<H(x,\xi)< E\\ \hbar^{-d+1/3}\mathrm{Ai}{\ensuremath{\left(\hbar^{-2/3}{\ensuremath{\left(H(x,\xi)-E\right)}}\right)}},&\quad H(x,\xi)-E\approx \hbar^{2/3}\\ O(\hbar^\infty),&\quad H(x,\xi)> E \end{cases}.$$ ![A plot of $(2\pi \hbar)^d H(x,\xi)^{-d/2} W_{\hbar, E}(x,\xi)$ as a function of the radial variable $\rho=\sqrt{2H(x,\xi)}$ in phase space when $E=1/2$ and $\hbar\approx 0.01$. The Airy behavior with peak $\hbar^{1/3}\approx .22$ near the energy surface $\rho=1=\sqrt{2E}$ and the oscillatory behavior with amplitude $\hbar^{1/2}\approx .1$ in the ‘bulk’ $H(x,\xi)<2E=1$ are clearly visible.[]{data-label="fig:individual-wigner"}](Wigner_HO_indiv){width=".5\textwidth"} Here, $\mathrm{Ai}$ is the Airy function (see Section \[AIRYAPP\]). These behaviors are illustrated in Figure \[fig:individual-wigner\] (see also Section \[INDIVIDUALSECT\]). In this continuation, we study ‘Weyl sums’ of Wigner distributions over a variety of spectral windows and describe the impact of the different spectral localizations on the scaling asymptotics of the corresponding Wigner distributions in different phase space regions. Let us begin by introducing the three types of spectral localization we are studying and the interfaces in each type. - \(i) $\hbar$-localized Weyl sums over eigenvalues in an $\hbar$-window $E_N(\hbar) \in [E - a \hbar, E + b \hbar] $ of width $O(\hbar)$. More generally we consider smoothed Weyl sums $W_{\hbar, E, f} $ with weights $f(\hbar^{-1}(E_N(\hbar) - E))$; see for such $\hbar$-energy localization. This is the scale of individual spectral projections but is substantially more general than the results of [@HZ19]. The scaling and asymptotics are in Theorem \[ELEVELLOC\]. For general [Schrödinger ]{}operators, $\hbar$- localization around a single energy level leads to expansions in terms of periodic orbits. Since all orbits of the classical isotropic oscillator are periodic, the asymptotics may be stated without reference to them. The generalization to all [Schrödinger ]{}operators will be studied in a future article. - \(ii) Airy-type $\hbar^{2/3}$-spectrally localized Weyl sums $W_{\hbar, f, 2/3}(x, \xi)$ over eigenvalues in a window $[E - a \hbar^{2/3}, E + a \hbar^{2/3}]$ of width $O(\hbar^{2/3})$. See Definition \[INTERDEF\] for the precise definition. The levelset $\Sigma_E$ is viewed as the interface. The scaling asymptotics of its Wigner distribution across the interface are given in Theorems \[RSCOR\] and \[SHARPh23INTER\]. To our knowledge, this scaling has not previously been considered in spectral asymptotics. - \(iii) Bulk Weyl sums $\sum_{N: \hbar(N +\frac{d}{2}) \in [E_1,E_2]} W_{\hbar, E_N(\hbar)}(x, \xi)$ over energies in an $\hbar$-independent ‘window’ $[E_1, E_2]$ of eigenvalues; this ‘bulk’ Weyl sum runs over $\simeq \hbar^{-1}$ distinct eigenvalues; See Definition \[BULKDEF\]. We are mainly interested in its scaling asymptotics around the interface $\Sigma_{E_2}$ (see Theorem \[BULKSCALINGCOR\]). However, we also prove that the Wigner distribution approximates the indicator function of the shell $\{E_1 \leq H \leq E_2\} \subset T^* {{\mathbb R}}^d$ (see Proposition \[pp:PBK-leading\]). As far as we know, this is also a new result and many details are rather subtle because of oscillations inside the energy shell. Indeed, the results of [@HZ19] show that the indvidual terms in the sum grow like $W_{\hbar, E_N(\hbar)}(x,\xi)\simeq \hbar^{-d+1/2}$ when $H(x,\xi)\in (E_1,E_2).$ Proposition \[pp:PBK-leading\], in contrast, shows although the bulk Weyl sums have $\simeq \hbar^{-1}$ such terms, their sum has size $\hbar^{-d}$, implying significant cancellation. We are particularly interested in ‘interface asymptotics’ of the Wigner-Weyl distributions $W_{\hbar, f, \delta(\hbar)}$ (see ) around the edge (i.e. boundary) of the spectral interval when $(x, \xi)$ is near the corresponding classical energy surface $\Sigma_E$. Such edges occur when $f$ is discontinuous, e.g. the indicator function of an interval. In other words, we integrate the empirical measures below over an interval rather than against a Schwartz test function. At the interface, there is an abrupt change in the asymptotics with a conjecturally universal shape. Theorem \[ELEVELLOC\] gives the shape of the interface for $\hbar$-localized sums, Theorem \[RSCOR\] gives the shape for $\hbar^{2/3}$ localized sums and Theorem \[BULKSCALINGCOR\] gives results on the bulk sums. Our results concern asymptotics of integrals of various types of test functions against the weighted empirical measures, $$\label{EMPDEF} d\mu_{\hbar}^{(x, \xi)}(\tau): = \sum_{N =0}^{\infty} W_{\hbar, E_N(\hbar)}(x, \xi) \delta_{E_N(\hbar)}(\tau),$$ and of recentered and rescaled versions of these measures (see below). A key property of Wigner distributions of eigenspace projections is that the measures are signed, reflecting the fact that Wigner distributions take both positive and negative values, and are of infinite mass: \[INFINITEPROP\] The signed measures are of infinite mass (total variation norm). On the other hand, the mass of is finite on any one-sided interval of the form, $[-\infty, \tau]$. Also, $\int_{{{\mathbb R}}} d\mu_{\hbar}^{(x, \xi)} = 1$ for all $(x, \xi). $ Moreover, the $L^2$ norms of the terms $W_{\hbar, E_N(\hbar)}$ grows in $N$ like $N^{\frac{d-1}{2}}$ (see Section \[NORMSECT\]). Hence, the measures are highly oscillatory and the summands can be very large. This makes it difficult to study integrals of over intervals using Tauberian arguments. Instead we rely on the special spectral properties of , encapsulated in Lemma \[NICEFORM\] below. While the results of [@HZ19] are very special to the isotropic oscillator, many of the results of this article should be universal among [Schrödinger ]{}operators, at least for energy levels that are not critical values of the classical Hamiltonian. We plan to study general [Schrödinger ]{}operators, and the universality of scaling asymptotics, in a future article. We begin with the case of the isotropic oscillator to determine the form of the laws, which would be much more complicated to prove in the general case. \[hSCSECT\] Interior asymptotics for $\hbar$-localized Weyl sums ---------------------------------------------------------------- The first result we present pertains to the $\hbar$-spectrally localized Weyl sums of type (i), defined by taking $\delta(\hbar)=\hbar$ in : $$W_{\hbar, E, f}(x, \xi) :=\sum_N f(\hbar^{-1}(E-E_N(\hbar))) W_{\hbar, E_N(\hbar)}(x,\xi),\qquad f \in {\mathcal{S}}({{\mathbb R}}).\label{UhfDEF}$$ \[ELEVELLOC\] Fix $E>0$, and let $W_{\hbar, E, f} $ be the Wigner distribution as in with $f$ an even Schwartz function. If $H(x,\xi)>E,$ then $W_{\hbar, E,f}(x,\xi)=O(\hbar^\infty).$ In contrast, when $0<H(x, \xi) < E$, set $H_E:=H(x,\xi)/E$ and define $$\begin{aligned} t_{+,\pm,k}&:= 4\pi k \pm 2\cos^{-1}{\ensuremath{\left(H_E^{1/2}\right)}},\quad t_{-,\pm,k}:= 4\pi {\ensuremath{\left(k+\frac{1}{2}\right)}} \pm 2\cos^{-1}{\ensuremath{\left(H_E^{1/2}\right)}},\qquad k \in {{\mathbb Z}}. $$ Fix any $\delta>0.$ Then $$W_{\hbar, f, E}(x,\xi)=\frac{\hbar^{-d+1}{\ensuremath{\left(1+O_\delta(\hbar^{1-\delta})\right)}}}{(2E)^{1/2}(2\pi)^dH_E^{d/2}(H_E^{-1}-1)^{1/4}}~\sum_{\pm_1,\pm_2\in {\left\{+,-\right\}}} \frac{e^{\pm_2 i{\ensuremath{\left(\frac{\pi}{4}-\frac{4E}{\hbar}\right)}}}}{{\ensuremath{\left(\pm_1\right)}}^d}\sum_{k\in {{\mathbb Z}}} \widehat{f}(t_{\pm_1,\pm_2,k})e^{\frac{iE}{\hbar}t_{\pm_1,\pm_2,k}},$$ where the notation $O_\delta$ means the implicit constant depends on $\delta.$ Note that there are potentially an infinite number of ‘critical points’ in the support of $\hat{f}$. We prove Theorem \[ELEVELLOC\] in Section \[WIGNERINTRO\]. \[SECT23\] Interface asymptotics for smooth $\hbar^{2/3}$-localized Weyl sums ----------------------------------------------------------------------------- We now consider less spectrally localized Wigner distributions but that are both spectrally localized and phase-space localized on the scale $\delta(\hbar)= \hbar^{2/3}$. They are mainly relevant when we study interface behavior around $\Sigma_E$ of Weyl sums. \[INTERDEF\] Let $H(x, \xi) = ({\left\lVertx\right\rVert}^2 + {\left\lVert\xi\right\rVert}^2)/2$, and assume that $(x, \xi)$ satisfies $$H(x,\xi) = E + u{\ensuremath{\left(\hbar/2E\right)}}^{2/3}.\label{E:near-shell}$$ Let $\delta(h) = \hbar^{2/3}$ in and define the interface-localized Wigner distributions by $$\begin{array}{lll} W_{\hbar, f, 2/3}(x, \xi): & = & \sum_{N} f(h^{-2/3} (E-E_N(\hbar)))W_{ \hbar, E_N}(x, \xi) \;\; \end{array}$$ \[RSCOR\] Assume that $(x, \xi)$ satisfies with ${\left\lvert u \right\rvert}<\hbar^{-2/3}.$ Fix a Schwartz function $f \in {\mathcal{S}}({{\mathbb R}})$ with compactly supported Fourier transform. Then $$W_{\hbar, f, 2/3}(x, \xi) = (2\pi \hbar)^{-d} I_0(u;f,E)~+~O((1+{\left\lvert u \right\rvert})\hbar^{-d+2/3}),$$ where $$I_0(u; f, E) = \int_{{{\mathbb R}}} f(-\lambda/C_E) \mathrm{Ai}{\ensuremath{\left(\lambda + \frac{u}{E}\right)}} d\lambda,\qquad C_E=(E/4)^{1/3}.$$ More generally, there is an asymptotic expansion $$W_{\hbar, f, 2/3}(x, \xi) ~\simeq~(2\pi\hbar)^d\sum_{m\geq 0}\hbar^{2m/3}I_m{\ensuremath{\left(u;f,E\right)}}$$ in ascending powers of $\hbar^{2/3}$ where $I_m(u;f,E)$ are uniformly bounded when $u$ stays in a compact subset of ${{\mathbb R}}.$ The calculations show that the results are valid with far less stringent conditions on $f$ than $f \in {\mathcal{S}}({{\mathbb R}})$ and $\widehat{f}\in C_0^\infty$. To obtain a finite expansion and remainder it is sufficient that $\int_{{{\mathbb R}}} |\widehat{f}(t)| |t|^k dt < \infty$ for all $k.$ It is not necessary that $\hat{f} \in C^k$ for any $k >0$. Theorem \[RSCOR\] is proved in Section \[RSCOR-pf\]. Sharp $\hbar^{2/3}$-localized Weyl sums --------------------------------------- Next we consider the sums of Definition \[INTERDEF\] when $f$ is the indicator function of a spectral interval, $$f = {\bf 1}_{[ \lambda_- ,\lambda_+]}.$$ Equivalently, we fix positive integers $n_{_\pm}$ such that $$\lambda_{\pm}=\hbar^{1/3}n_{\pm}~~\text{are bounded},$$ and consider the corresponding Wigner-Weyl sums $ W_{\hbar, f, 2/3}(x, \xi)$ of Definition \[INTERDEF\]: $$\label{W23EDEF} W_{2/3,E,\lambda_{\pm}}(x,\xi):=\sum_{N:\, \lambda_- \hbar^{2/3}\leq E_N(\hbar)-E< \lambda_+\hbar^{2/3}}W_{\hbar, E_N(\hbar)}(x,\xi)=\sum_{N=N(E,\hbar)+n_-}^{N(E,\hbar)+n_+-1}W_{\hbar, E_N(\hbar)}(x,\xi) ,$$ where $N(E,\hbar)=E/\hbar - d/2$. Thus, the sums run over spectral intervals of size $\simeq \hbar^{2/3}$ centered at a fix $E>0$ and consist of a sum of $\simeq \hbar^{-1/3}$ Wigner functions for spectral projections of individual eigenspaces. The following extends Theorem \[RSCOR\] to sharp Weyl sums at the cost of only giving a 1-term expansion plus remainder. \[SHARPh23INTER\] Assume that $(x, \xi)$ satisfies ${\ensuremath{\left({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2\right)}}/2 =E+ u{\ensuremath{\left(\hbar/2E\right)}}^{2/3}$ with ${\left\lvert u \right\rvert}<\hbar^{-2/3}.$ Then, $$W_{2/3,E,\lambda_{\pm}}(x,\xi)=(2\pi\hbar)^{-d}C_E\int_{-\lambda_+}^{-\lambda_-}\operatorname{Ai}{\ensuremath{\left(\frac{u}{E}+\lambda C_E\right)}} d\lambda + O{\ensuremath{\left(\hbar^{-d+1/3-\delta}+(1+{\left\lvert u \right\rvert})\hbar^{-d+2/3-\delta}\right)}},$$ where $ C_E = (E/4)^{1/3}.$ Theorem \[SHARPh23INTER\] is proved in Section \[2/3sumSect\]. It can be rephrased in terms of weighted empirical measures $$\label{mu2/3} d \mu^{u, E, \frac{2}{3}}_{\hbar} := \hbar^d\; \sum_{N} W_{ \hbar, E_N(\hbar)}{\ensuremath{\left( E + u{\ensuremath{\left(\hbar/2E\right)}}^{2/3}\right)}} \delta_{[ \hbar^{-2/3} (E- E_N(\hbar) )]}.$$ obtained by centering and scaling the family . Thus, for $(x, \xi)$ satisfying ${\ensuremath{\left({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2\right)}}/2 = E + u{\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3},$ and for $f \in {\mathcal{S}}({{\mathbb R}})$, $$W_{\hbar, f, 2/3}(x, \xi): = \hbar^{-d} \int_{{{\mathbb R}}} f(\tau) d \mu^{u, E, \frac{2}{3}}_{\hbar}(\tau), \;\;\; W_{2/3,E,\lambda_{\pm}}(x,\xi) = \hbar^{-d} \int_{\lambda_-}^{\lambda^+} d \mu^{u, E, \frac{2}{3}}_{\hbar}(\tau).$$ \[BULKSECT\] Bulk sums ---------------------- We next consider Weyl sums of eigenspace projections corresponding to an energy shell (or window) $[E_1, E_2]$. We consider both sharp and smoothed sums. \[BULKDEF\] Define the ‘bulk’ Wigner distributions for an $\hbar$-independent energy window $[E_1, E_2]$ by $$\label{WE} W_{\hbar,[E_1, E_2]}(x, \xi): \sum_{N: E_N(\hbar) \in [E_1,E_2]} W_{\hbar, E_N(\hbar)}(x, \xi).$$ More generally for $f \in C_b({{\mathbb R}})$ define $$\label{WEf} W_{\hbar, f}(x, \xi):= \sum_{N=1}^{\infty} f(\hbar(N +d/2)) \; W_{\hbar, E_N(\hbar)}(x, \xi).$$ Our first result about the bulk Weyl sums concerns the smoothed Weyl sums $W_{\hbar, f}.$ \[smoothedbulk\]For $f \in {\mathcal{S}}({{\mathbb R}})$ with $\hat{f} \in C_0^{\infty}$, $W_{\hbar, f}(x, \xi)$ admits a complete asymptotic expansion as $\hbar \to 0$ of the form, $$\left\{ \begin{array}{lll} W_{\hbar, f}(x,\xi) & \simeq & (\pi \hbar)^{-d} \sum_{j = 0}^{\infty} c_{j, f, H}(x, \xi) \hbar^j,\; \rm{with} \\&&\\c_{0, f, H} (x, \xi) & = & f(H(x, \xi)) = \int_{{{\mathbb R}}} \hat{f}(t) e^{i t H(x,\xi)} dt. \end{array}\right.$$ In general $c_{k, f, H}(x, \xi)$ is a distribution of finite order on $f$ supported at the point $(x, \xi)$. The proof is given in Section \[BULKfSECT\] and merely involves Taylor expansion of the phase. Interior/exterior asymptotics for bulk Weyl sums of Definition \[BULKDEF\] -------------------------------------------------------------------------- From Proposition \[smoothedbulk\], it is evident that the behavior of $W_{\hbar, [E_1,E_2]}(x,\xi)$ depends on whether $H(x, \xi) \in (E_1, E_2)$ or $H(x, \xi) \notin [E_1, E_2]$. Some of this dependence is captured in the following result. \[pp:PBK-leading\] We have, $$W_{\hbar, [E_1, E_2]}(x, \xi) = \left\{ \begin{array}{ll} (i)\; (2\pi\hbar)^{-d} (1+O(\hbar^{1/2})), & H(x, \xi) \in (E_1, E_2),\\ &\\ (ii) \; O(\hbar^{-d +1/2}), & H(x, \xi) < E_1, \\&\\ (iii) \; O(\hbar^{\infty}), & H(x, \xi) > E_2 \end{array} \right.$$ The two ‘sides’ $0 < H(x, \xi) < E_1$ and $H(x, \xi) > E_2$ behave differently because the Wigner distributions have slowly decaying tails inside an energy ball but are exponentially decaying outside of it (see Section \[INDIVIDUALSECT\]). If we write $W_{\hbar, [E_1, E_2]}(x, \xi) = W_{\hbar, [0, E_2]}(x, \xi) -W_{\hbar, [0, E_1]}(x, \xi) $, we see that the two cases with $H(x, \xi) > E_1$ are covered by results for $W_{\hbar, [0, E]}$ with $E = E_1 $ or $E = E_2$. When $H(x, \xi) < E_1$, then both terms of $W_{\hbar, [0, E_2]}(x, \xi) -W_{\hbar, [0, E_1]}(x, \xi) $ have the order of magnitude $\hbar^{-d}$ and the asymptotics reflect the cancellation between the terms. The boundary case where $H(x, \xi) = E_1,$ or $H(x, \xi)= E_2$ is special and is given in Theorem \[RSCOR\]. Interface asymptotics for bulk Weyl sums of Definition \[BULKDEF\] ------------------------------------------------------------------ Our final result concerns the asymptotics of $W_{\hbar, [E_1, E_2]}(x, \xi)$ in $\hbar^{2/3}$-tubes around the ‘interface’ $H(x, \xi) = E_2$. Again, it is sufficient to consider intervals $[0, E]$. It is at least intuitively clear that the interface asymptotics will depend only on the individual eigenspace projections with eigenvalues in an $\hbar^{2/3}$-interval around the energy level $E$, and since they add to $1$ away from the boundary point, one may expect the asymptotics to be similar to the interface asymptotics for individual eigenspace projections in [@HZ19]. \[BULKSCALINGCOR\] Assume that $(x, \xi)$ satisfies ${\ensuremath{\left({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2\right)}}/2 = E + u{\ensuremath{\left(\hbar/2E\right)}}^{2/3}$ with ${\left\lvert u \right\rvert}<\hbar^{-2/3}.$ Then, for any ${\varepsilon}>0$ $$W_{\hbar, [0, E]}(x, \xi) = {\ensuremath{\left(2\pi \hbar\right)}}^{-d} \left[\int_0^{\infty} \operatorname{Ai}{\ensuremath{\left(\frac{u}{E}+\lambda\right)}} d\lambda +O(\hbar^{1/3-{\varepsilon}}{\left\lvert u \right\rvert}^{1/2})+ O({\left\lvert u \right\rvert}^{5/2}\hbar^{2/3-{\varepsilon}})\right],$$ where the implicit constant depends only on $d,{\varepsilon}.$ The Airy scaling the Wigner function is illustrated in Figure \[fig:bulk-Wigner-Airy\]. \[HEURISTICSECT\] Heuristics ---------------------------- In Sections \[NORMSECT\] - \[INDIVIDUALSECT\], we review the results on $L^2$ norms and pointwise asymptotics of Wigner distribution of eigenspace projections . Wigner distributions are normalized so that the Wigner distribution of an $L^2$ normalized eigenfunction has $L^2$ norm $1$ in $T^*{{\mathbb R}}^d$. Due to the multiplicity $N^{d-1}$ of eigenspaces , the $L^2$ norm of $W_{\hbar, E_N(\hbar)}$ is of order $N^{\frac{d-1}{2}}$. In the main results, we sum over windows of eigenvalues, e.g. $\lambda_- \hbar^{2/3}\leq E-E_N(\hbar)< \lambda_+ \hbar^{2/3}$ , resp. $E_N(\hbar) \in [0,E]$ in . Inevitably, the asymptotics are joint in $(\hbar, N)$. As $\hbar \downarrow 0$, the number of $N$ contributing to the sum grows at the rate $\hbar^{-\frac{1}{3}}$, resp. $\hbar^{-1}$. Due to the $N$-dependence of the $L^2$ norm, terms with higher $N$ have norms of higher weight in $N$ than those of small $N$ but the precise size of the contribution depends on the position of $(x, \xi)$ relative to the interface $\{H = E\}$ and of course the relation . As reviewed in Section \[INDIVIDUALSECT\], $W_{\hbar, E_N(\hbar)}(x, \xi)$ peaks when $H(x, \xi) = E_N(\hbar)$, exponentially decays in $\hbar$ when $H(x, \xi) > E_N(\hbar)$ and has slowly decaying tails inside the energy ball $\{H < E_N(\hbar)\}$, which fall into three regimes: (i) Bessel near $0$, (ii) oscillatory or trigonometric in the bulk, and (iii) Airy near $\{H = E\}.$ In terms of $N$, when holds, and $H(x, \xi) < E_N(\hbar)$, then $W_{\hbar, E_N(\hbar)}(x, \xi) \simeq \hbar^{-d + 1/2} \simeq N^{d -1/2}$. Near the peak point, when $H(x,\xi)-E_N(\hbar)\approx \hbar^{2/3}$, we have in contrast $W_{\hbar, E_N(\hbar)}(x, \xi) \simeq \hbar^{-d + 1/3} \simeq_E N^{d - 1/3}.$ It follows that the terms with a high value of $N$ and with $ E_N(\hbar) \geq H(x, \xi)$ in contribute high weights. There are an infinite number of such terms, and so is a signed measure of infinite mass (as stated in Proposition \[INFINITEPROP\].) This is why we mainly consider the restriction of the measures to compact intervals. See Section \[EMPSECT\] for more on these measures. Discussion of Results --------------------- Proposition \[pp:PBK-leading\] shows that when properly normalized (by multiplying by $\hbar^{d}$), the Wigner distribution is asymptotically equal to $1$ in the allowed region $\{E_1 \leq H \leq E_2\}$, and asymptotically equal to $0$ in the complementary forbidden region, where $H(x, \xi) = {{\textstyle \frac 12}}(||\xi||^2 + ||x||^2)$ is the classical Hamiltonian; this proves a conjecture in [@JZ]. The main question is the shape of the interface interpolating between the values $0$ and $1$. The result (Theorem \[RSCOR\]) is that the scaled interface shape around $\{H = E_2\}$ is the graph of the exponentially decaying side of the Airy function from the region around its global maximum (slightly to the left of $0$) to the right to $+\infty$. The addition of many terms in the Weyl sum fills in the region $\{E_1 < H < E_2\}$ to an asymptotically constant function (see Figure \[fig:bulk-Wigner-Airy\]). As the figure shows, there is an asymmetry between the upper endpoint $E_2$ and the lower endpoint $E_1$, reflecting the asymmetric behavior of the Airy function on positive/negative axis. On the positive axis, $\mathrm{Ai}(x) \sim C x^{-1/4} e^{- 2x^{3/2}/3}, \;\; x \in {{\mathbb R}}_+, x\to \infty$, i.e. the decay is almost as fast as in the complex domain [@ZZ17]. However, $\mathrm{Ai}(-x) \sim C x^{-1/4} \sin\left(2 x^{3/2}/3 + \pi/4\right)$ when $x \to \infty$. As a result, there is a slowly decaying oscillating tail of $W_{\hbar, E_N(\hbar))}(x, \xi)$ with $E_N(\hbar) \simeq E$ for $(x, \xi)$ inside the energy surface $\Sigma_E$ but a fast exponential decay outside of it. The oscillatory behavior inside the energy surface is in marked contrast to the Erf interface behavior in the complex domain studied in [@ZZ17] (see Section \[PRIOR\]). It is interesting to compare the results of Theorems \[RSCOR\]-\[SHARPh23INTER\] and Theorem \[BULKSCALINGCOR\]. In both theorems, we assume that $(x, \xi)$ satisfies ${\ensuremath{\left({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2\right)}}/2= E + u{\ensuremath{\left(\hbar/2E\right)}}^{2/3}.$ For sharp $\hbar^{2/3}$ localization to an interval $[E+\lambda_-\hbar^{2/3},E+\lambda_+\hbar^{2/3}],$ we get $$W_{2/3,E,\lambda_{\pm}}(x,\xi)=(2\pi\hbar)^{-d}C_E\int_{-\lambda_+}^{-\lambda_-}\operatorname{Ai}{\ensuremath{\left(\frac{u}{E}+\lambda C_E\right)}} d\lambda + O(\hbar^{-d+1/3-\delta}),\qquad C_E = (E/4)^{1/3},$$ while at the interface of the bulk sums we get essentially the same asymptotics, $$W_{\hbar, [0, E]}(x, \xi) = {\ensuremath{\left(2\pi \hbar\right)}}^{-d} \int_0^{\infty} \operatorname{Ai}{\ensuremath{\left(\frac{u}{E}+\tau\right)}} d\tau + O((1+{\left\lvert u \right\rvert})\hbar^{-d+1/3-{\varepsilon}}).$$ One may understand this coincidence by decomposing the bulk Weyl sum, $$\label{WEDECintro} W_{\hbar, [0,E]}(x, \xi)= \sum_{N: \hbar(N +\frac{d}{2}) \in [0, E - \delta(\hbar)]} W_{\hbar, E_N(\hbar)}(x, \xi) +\sum_{N: \hbar(N +\frac{d}{2}) \in [E - \delta(\hbar), E]} W_{\hbar, E_N(\hbar)}(x, \xi),$$ with $\delta(\hbar) \simeq \hbar^{2/3}$, into terms with energy levels either sufficiently close or sufficiently well-separated from the boundary level $E$. One then shows that the terms far from the interface do not make a principal order contribution. Intuitively, the reason is that the summands $W_{\hbar, E_N(\hbar)}$ with $E_N(\hbar) \simeq a < E$ are exponentially decaying to the right of $H(x,\xi) = a$, or better to the right of the maximum point at $\hbar^{2/3} (a - \zeta_0)$ where $\zeta_0$ is the first critical point of $\rm{Ai}(x)$, hence contribute little to the asymptotics for the given $(x, \xi)$. The further that $E_N(\hbar) \simeq a$ moves left of $E$, the more exponentially decaying $W_{\hbar, E_N(\hbar)}(x, \xi)$ becomes. Hence it is only at the edge of the interval $[0, E - C \hbar^{2/3}]$ that the terms of the first sum can contribute non-trivially to the asymptotics. Each of the terms with $E_N(\hbar) \simeq E + C \hbar^{2/3} $ is in fact positive since the largest zero of $\rm{Ai}(x)$ is approximately $ - 2.3381.$ Each is of size $h^{-d+1/3}$ (see Section \[INDIVIDUALSECT\]) and there are $\hbar^{-1/3}$ such terms, so the sum over the interval $[E - C \hbar^{2/3}, E]$ does account for the order of magnitude. This heuristic discussion ignores the fact that the terms are signed and that cancellation is needed for the terms with $E_N(\hbar) \leq E - C \hbar^{2/3}$ to obtain the correct order of magnitude. \[Outline\]Outline of proofs ----------------------------- As in [@HZ19], we express Wigner functions $W_{\hbar, E_N(\hbar)}$ of projections $\Pi_{\hbar, E_N(\hbar)}$ onto the $N^{th}$ eigenspace of $\widehat{H}_\hbar$ as Fourier coefficients of the Wigner function ${\mathcal{U}}_{\hbar}(t)$ of the Propagator $U_{\hbar}(t)= e^{- i \frac{t}{\hbar} \hat{H}_{\hbar}}$. To make this precise, notice that on the level of Schwartz kernels, $$\label{PROJUT} \Pi_{\hbar, E_N(\hbar)}(x,y)=\int_{-\pi}^{\pi} U_\hbar(t-i{\varepsilon},x,y) e^{-{{\frac{i}{h}}}(t-i{\varepsilon}) E_N(\hbar)} \frac{dt}{2\pi} $$ is a Fourier coefficient of the boundary value of the function $U_\hbar(t,x,\xi)$ of $t$, which is holomorphic for $t$ in the lower half-plane. The Wigner function of the propagator ${\mathcal{U}}_\hbar(t,x,\xi)$ is defined as $$\label{WIGU} {\mathcal{U}}_\hbar(t, x,\xi) := \int e^{- {{\frac{i}{h}}}2 \pi \xi p} U_{\hbar}{\ensuremath{\left(t,x + \frac{p}{2}, x - \frac{p}{2}\right)}}dp.$$ Since the map from a kernel to its Wigner trasnform si linear, we can take the Wigner transform of the identities above between $\Pi_{\hbar, E_N(\hbar)}$ and $U_\hbar(t,x,y)$ to see that $W_{\hbar, E_N(\hbar)}$ is essentially a Fourier coefficient of ${\mathcal{U}}_\hbar(t):$ $$W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{-\pi}^\pi e^{itE_N(\hbar)/\hbar} {\mathcal{U}}_\hbar(t,x,\xi)\frac{dt}{2\pi}.\label{E:Fourier-Wigner}$$ This expression is profitably combined with the following exact formula for the Wigner distribution ${\mathcal{U}}_{\hbar}(t, x, \xi) $ of the propagator. \[P:Wigner-prop\] We have, $${\mathcal{U}}_{\hbar}(t, x, \xi) =(2\pi\hbar\cos((t+i0)/2))^{-d} e^{-2iH\tan(t/2)/\hbar}, \quad H=H(x,\xi)=\frac{1}{2}{\ensuremath{\left({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2\right)}}.$$ The Wigner distribution ${\mathcal{U}}_\hbar(t,x,\xi)$ is well-defined as a distribution (see [@HZ19]). For any $f\in \mathcal S({{\mathbb R}})$ and any $\delta(\hbar)>0$ we may use Fourier inversion to write $$\begin{aligned} \notag W_{\hbar, f, E}(x,\xi)&~=~\sum_{N\geq 0} f{\ensuremath{\left(\delta(\hbar)^{-1}{\ensuremath{\left(E-E_N(\hbar)\right)}}\right)}}W_{\hbar, E_N(\hbar)}(x,\xi)\\ &~=~\int_{{{\mathbb R}}}\widehat{f}(t) e^{itE\delta(\hbar)^{-1}}{\mathcal{U}}_\hbar(t\hbar\delta(\hbar)^{-1}, x,\xi)\frac{dt}{2\pi}.\label{E:smoothed-sum}\end{aligned}$$ To obtain the results on smoothed Wigner-Weyl sums (Theorems \[ELEVELLOC\] and \[RSCOR\] as well as Proposition \[smoothedbulk\]), where $f$ is Schwartz, we use this formula with $\delta(\hbar)=\hbar,\hbar,1$ and simply Taylor expand the phase and amplitude from Proposition \[P:Wigner-prop\]. A key point is that ${\mathcal{U}}_\hbar(t,x,\xi)$ is a smooth function away from $t\in \pi {{\mathbb Z}}$ and that the contribution to the integral in from a neighborhood of any of these points is $O(\hbar^\infty).$ Since ${\mathcal{U}}_\hbar(t)$ is periodic, it suffices to check this near $t=\pi,$ which is the content of a result from [@HZ19]. To state it, fix $ \delta>0$ and define a sequence of smooth of cut-off functions $$\chi_{\hbar}:{{\mathbb R}}{\ensuremath{\rightarrow}}[0,1],\quad \chi_{\hbar}(t)= \begin{cases} 1,&\quad {\left\lvert t-\pi \right\rvert}<\hbar^\delta\\ 0, &\quad {\left\lvert t-\pi \right\rvert} > 2\hbar^\delta \end{cases} .$$ \[LOCLEM\] Fix ${\varepsilon}>0$. Uniformly over $(x,\xi)$ with $H(x,\xi)>{\varepsilon},$ the localized version $$\label{E:smoothed-sum-2} \int_{{{\mathbb R}}} e^{-itE/\hbar} \chi_\hbar(t)\widehat{f}(t) {\mathcal{U}}_\hbar(t,x,\xi)\frac{dt}{2\pi}$$ of is $O(\hbar^\infty).$ The idea behind this lemma is that ${\mathcal{U}}_\hbar(t)$ is not locally $L^1$ near $t\in \pi{{\mathbb Z}},$ but is nonetheless well-defined as a tempered distribution in the sense that the integral $$\label{ucalhf} {\mathcal{U}}_{\hbar, f}(x, \xi) : =\int_{{{\mathbb R}}} \hat{f}(t) {\mathcal{U}}_{\hbar}(t, x, \xi)dt.$$ is well-defined for $f \in {\mathcal{S}}({{\mathbb R}}).$ In fact, ${\mathcal{U}}_{\hbar}(t+\pi, x, \xi)$ resembles the position space propagator with $y = \xi$ except that its phase lacks the term $- \frac{2}{\cos t/2} \langle x, \xi \rangle $. Note that for small $t$ the phase of is essentially $\frac{x^2 + y^2}{2 t} - \frac{1}{t}\langle x, y \rangle = |x - y|^2/2t$ and that as $t \to 0$ this kernel weakly approaches $\delta(x - y)$. Similarly, ${\mathcal{U}}_{\hbar} (t, x, \xi) \to \delta_0(x, \xi)$ as $t \to \pi$, the Dirac distribution at the point $(0,0) \in T^* {{\mathbb R}}^n$. Indeed, at $t=\pi$ it is the Wigner distribution of $\delta(x + y)$, which is $\int_{{{\mathbb R}}^n} \delta(x - \frac{v}{2} + x + \frac{v}{2}) e^{- i \langle v, \xi \rangle} dv = \delta(2x) \delta(\xi). $ Since ${\mathcal{U}}_\hbar$ is a locally $L^1$ function at times $t \notin \pi {{\mathbb Z}}$ and is a measure when $t \in \pi {{\mathbb Z}}$ it is a measure for all $t$. To prove our sharp results when $f$ is the indicator function of an interval (Theorems \[SHARPh23INTER\] and \[BULKSCALINGCOR\] as well as Proposition \[pp:PBK-leading\]), we make key use of the special properties of the isotropic oscillator. Namely, we use the special nature of its spectrum, which after rescaling by $\hbar$ and subtracting $d/2$ forms an arithmetic progression. A simple identity is useful when dealing with sums over spectral intervals rather than smooth sums. Since we use it several times, we record it in advance: \[NICEFORM\] We have, $$\label{EXACT} \sum_{N: E_1\leq E_N(\hbar) < E_2} W_{\hbar, E_N(\hbar)}(x, \xi) = \int_{S^1} \frac{e^{i N_2 t} -e^{iN_1t}}{e^{it} -1}e^{itd/2}{\mathcal{U}}_{\hbar}(t, x, \xi) \frac{dt}{2\pi},$$ where $N_i=E_i/\hbar-d/2,\, i=1,2.$ The identity follows directly from by noting that $E_N(\hbar)/\hbar = N+d/2$ and summing a partial geometric series. The expression ${\ensuremath{\left(e^{i N_2 t} -e^{iN_1t}\right)}}/{\ensuremath{\left(e^{it} -1\right)}}$ when $N_1=0$ is the Dirichlet kernel, and its analysis depends on the scale of $N_2-N_1$ (see the discussion in Section \[PROOFBULK1\] for an explanation of which intervals in $t$ are important for analyzing ). In several cases, we will find that the dominant contribution comes from $t\approx 0$. \[PRIOR\]Prior results and further questions -------------------------------------------- As mentioned above, this article is a continuation of [@HZ19], which contains further references to the vast literature on Wigner distributions and the isotropic Harmonic Oscillator. Most of that literature pertains to dimension 1. As also mentioned, the problems studied here can be posed for general [Schrödinger ]{}operators, and we expect some of the interface scaling results to be universal. The weights of the empirical measures are given by $W_{\hbar, E_N(\hbar)}(x, \xi)$, whose size depends on the pair $(\hbar, N)$. A potentially interesting option is to weight the terms in with functions $w_N$ of $N$, e.g. $N^{-\alpha}$ or the weights $(E_N(\hbar) - E)_+^{\alpha}$ of a Riesz mean, to counter-act the growth of the Wigner weights as $N \to \infty$ and change the infinite measures to finite ones. The interface asymptotics of Theorem \[RSCOR\] and Theorem \[BULKSCALINGCOR\] may seem reminiscent of the Gibbs phenomenon for Fourier series of indicator functions, where one obtains a universal $\frac{\sin t}{t}$-integral interface in configuration space (the circle) at a jump. Generalizations to manifolds are given in [@PT97]. They are of a different nature than the interface results here, although there is some common ground. Here, the interface is due to a sharp cutoff ${\bf 1}_{[E_1, E_2]}$ in the spectral window, while the Gibbs phenomenon pertains to convergence of eigenfunction expansions of discontinuous functions of the physical space variable $x$. In particular, the Gibbs phenomenon is essentially a physical space phenomenon while the interface phenomena studied here are essentially phase space phenomena. To clarify the analogy with the Gibbs phenomenon, consider the [Schrödinger ]{}operator $\hbar D$ on the circle $S^1$, where $D =\frac{d}{i dx}$. Its eigenvalues (resp. eigenfunctions) are $E_n(\hbar) = \hbar n$ (resp. $e^{i n x}$). For each $x$ we define a (complex) empirical measure by $\mu_h^x : = \sum_{n =-\infty}^{\infty} e^{i n x} \delta_{\hbar n} $. This is not really analogous to because the weights are physical space functions rather than phase space functions. Then $$\mu_h^x[-E, E] = D_{\hbar,E}(x):=\frac{\sin (N(\hbar, E) + {{\textstyle \frac 12}})x}{\sin \frac{x}{2}},$$ where $N(\hbar, E) =\max\{n: E_n(\hbar) \leq E\} \iff n \leq \hbar^{-1} E\}$, and where $D_{\hbar, E}$ is the Dirichlet kernel. This kernel arises in the proof Lemma \[NICEFORM\] and in , suggesting a connection to the Gibbs phenomenon. In the proof of the Gibbs phenomenon for step functions with a jump at $x = 0$, one lets $x_n = y (n + {{\textstyle \frac 12}})^{-1}$ for some $y >0$. Then, $D_{\hbar, E}(x_{N(\hbar, E)} ) = \frac{\sin y}{y} + O(\hbar)$. The indefinite integral $Si(y) = \int_0^{y} \frac{\sin t}{t} dt$ gives the interface asymptotics of the Gibbs phenomenon when the jump occurs at $x = 0$. In contrast, the problem we study for the Harmonic oscillator is the behavior of Wigner-Weyl sums over $\{n: \hbar n \leq E\}$ at the boundary $|\xi| = E$ of an energy ball $\{|\xi| \leq E\} \subset T^*S^1$. Since the energy curve projects over all of $S^1$ without singularities, there is no analogue of allowed or forbidden regions; and since $\mu_{\hbar}^x$ does not depend on the momentum $\xi$, there is no interface behavior to consider at $\{|\xi|= E\}$. There could exist an analogous interface problem if one defined phase space measures on $T^*S^1$ analogous to Wigner distributions. However, the most natural analogues are $\delta$-functions on energy levels, about which little is left to be known. In this direction, we note that Guillemin-Sternberg [@GS81] studied Weyl asymptotics in many realizations of the metaplectic representation, e.g. the [Schrödinger ]{}representation studied here, the Bargmann-Fock representation studied in [@ZZ17] and in a largely unexplored context of the Lagrangian Grassmannian, a generalization of $S^1$. More precisely, they studied intertwining Fourier integral operators between Harmonic Oscillators (and other quadratic Hamiltonians) from $L^2({{\mathbb R}}^d)$ to $L^2(\Lambda_d)$ where $\Lambda_d$ is the Lagrangian Grassmannian. When $d =1$, $\Lambda_d = S^1/{{\mathbb Z}}_2$. It might be interesting to study the induced map from Wigner distributions to phase space distributions in the setting of $\Lambda_d$. In view of the analogies between Harmonic oscillators and Fourier series, there should exist a direct analogue of the Gibbs phenomenon for ‘Fourier-Wigner’ series of radial functions on $T^* {{\mathbb R}}^d$. Namely, the Wigner distributions form an orthonormal basis of radial functions on $T^*{{\mathbb R}}^d$ (see [@HZ19] for background). Hence the ‘Fourier-Wigner series’ expansion of the indicator function ${\bf 1}_{\{H \leq E\}}$ of an energy ball in terms of must diverge across the energy surface $\Sigma_E$ in a universal way. Almost certainly, it also diverges at the origin, giving a Wigner analogue of the ‘Pinsky phenomenon.’ Background ========== In this section we give a rapid review of the results on Wigner distributions of individual eigenspace projections in [@HZ19], so that the reader may compare them with the results on Weyl sums of projections in this article. We also review the relation to Laguerre polynomials and prove Proposition \[INFINITEPROP\]. \[NORMSECT\]Norms and normalizations ------------------------------------ Recalling the notation in Section \[INTRO\], the isotropic oscillator admits the spectral decomposition $$L^2({{\mathbb R}}^d, dx) = \bigoplus_{N\in \mathbb N } V_{\hbar,E_N(\hbar)}, \qquad \widehat{H}_\hbar|_{V_{h,E_N(\hbar)}} = \hbar (N + d/2).$$ The Wigner transform, which is typically viewed as a mapping $${\mathcal{W}}_{\hbar}: L^2({{\mathbb R}}^d \times {{\mathbb R}}^d) \to L^2(T^*{{\mathbb R}}^d),$$ in fact acts on Schwartz kernel distributions $K_{\hbar} \in {\mathcal{D}}'({{\mathbb R}}^d \times {{\mathbb R}}^d)$ as follows: $${\mathcal{W}}_{\hbar}(K_{\hbar})(x, \xi): = (2 \pi\hbar)^{-d} \int_{{{\mathbb R}}^d} K_{\hbar} \left( x+\frac{v}{2}, x-\frac{v}{2} \right) e^{-\frac{i}{\hbar} v \xi} \frac{dv}{(2\pi h)^d}. \label{E:WignerK}$$ Taking $K_\hbar$ to be an eigenspace projection $\Pi_{\hbar, E_N(\hbar)}(x,y)$ gives the Wigner distributions $ W_{\hbar, E_N(\hbar)}(x, \xi)$ of individual eigenspace projections . In [@HZ19] it is proved that, $$\label{INTEGRALS}\left\{ \begin{array}{ll}(i) & \int_{T^* {{\mathbb R}}^d} W_{\hbar, E_N(\hbar)}(x, \xi) dx d \xi = \rm{Tr} \Pi_{\hbar, E_N(\hbar)} = \dim V_{\hbar, E_N(\hbar)} = \binom{N+d-1}{d-1} \\ & \\(ii) & \int_{T^* {{\mathbb R}}^d}\left| W_{\hbar, E_N(\hbar)}(x, \xi) \right|^2 dx d \xi = \rm{Tr} \Pi^2_{\hbar, E_N(\hbar)} = \dim V_{\hbar, E_N(\hbar)}=\binom{N+d-1}{d-1}\\ & \\ (iii)& \int_{T^* {{\mathbb R}}^d} W_{\hbar, E_N(\hbar)}(x, \xi) \overline{W_{\hbar, E_M(\hbar)}(x, \xi)}dx d \xi = \mathrm{Tr} \Pi_{\hbar, E_N(\hbar)} \Pi_{\hbar, E_M(\hbar)}= 0, \; \rm{for}\; M \not=N. \end{array} \right.,$$ In these equations, $N=\frac{E}{\hbar}-\frac{d}{2},$ and $\binom{N+d-1}{d-1}$ is the composition function of $(N,d)$ (i.e. the number of ways to write $N$ as an ordered us of $d$ non-negative integers). Thus, the sequence, $$\{ {\ensuremath{\left(\dim V_{\hbar, E_N(\hbar)}\right)}}^{-1/2} W_{\hbar, E_N(\hbar)}\}_{N=1}^{\infty}\subset L^2({{\mathbb R}}^{2n})$$ is orthonormal. \[INDIVIDUALSECT\]Asymptotics of individual eigenspace projection Wigner distributions -------------------------------------------------------------------------------------- Next we review the results of [@HZ19] on pointwise asymptotics of Wigner distributions of individual eigenspace projections, which are necessary to understand the behavior of sums. In each result we fix one term by requiring that $E_N(\hbar) = E$, and then give pointwise results depending on the position of $(x, \xi)$ with respect to . The first result gives Airy scaling asymptotics around the energy surface: \[SCALINGCOR-old\] Fix $E>0,d\geq 1$. Assume $E_N(\hbar) = E$. Suppose $(x,\xi)\in T^*{{\mathbb R}}^d$ satisfies $$\label{udef} H(x,\xi)=E + u {\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3},\qquad u \in {{\mathbb R}},\, H(x,\xi)=\frac{{\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2}{2}$$ with ${\left\lvert u \right\rvert}<\hbar^{-1/3}$. [^2] Then, $$\label{E:W-scaling} W_{\hbar, E_N(\hbar)}(x, \xi)= \begin{cases} \frac{2}{(2\pi \hbar)^d} {\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{1/3} {\ensuremath{\left(\operatorname{Ai}(u/E) + O{\ensuremath{\left((1+{\left\lvert u \right\rvert})^{1/4}u^2\hbar^{2/3}\right)}}\right)}},&\qquad u<0\\ \frac{2}{(2\pi \hbar)^d} {\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{1/3} \operatorname{Ai}(u/E){\ensuremath{\left(1 + O{\ensuremath{\left((1+{\left\lvert u \right\rvert})^{3/2}u\hbar^{2/3}\right)}}\right)}},&\qquad u>0 \end{cases}$$ The second gives pointwise asymptotics when $(x, \xi)$ lies in the interior of the energy ball (the ‘bulk’): \[BESSELPROP\] Fix $E>0$ and suppose $E_N(\hbar)=E.$ For each $(x,\xi)\in T^*{{\mathbb R}}^d$ write $$H_E := \frac{H(x,\xi)}{E}=\frac{{\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2}{2E},\qquad \nu_E:=\frac{4E}{\hbar}.$$ Fix $0<a < 1/2.$ Uniformly over $a\leq H_E \leq 1-a$, there is an asymptotic expansion, $$W_{\hbar, E_N(\hbar)}(x, \xi)=\frac{2}{(2\pi\hbar)^d}\left[\frac{J_{d-1}(\nu_E A(H_E))}{A(H_E)^{d-1}}\alpha_0(H_E)+O{\ensuremath{\left( \nu_E^{-1}{\left\lvert \frac{J_{d}(\nu_E A(H_E))}{A(H_E)^{d}} \right\rvert}\right)}}\right].$$ In particular, uniformly over $H_E$ in a compact subset of $(0,1),$ we find $$W_{\hbar, E_N(\hbar)}(x, \xi) = (2\pi\hbar)^{-d+1/2} P_{H,E}\cos{\ensuremath{\left(\xi_{\hbar, E,H}\right)}}+O{\ensuremath{\left(\hbar^{-d+3/2}\right)}},\label{E:interior-cosine}$$ where we’ve set $$\xi_{\hbar, E,H} =-\frac{\pi}{4} -\frac{2H}{\hbar}{\ensuremath{\left(H_E^{-1}-1\right)}}^{1/2}+\frac{2E}{\hbar}\cos^{-1}{\ensuremath{\left(H_E^{1/2}\right)}}$$ and $$P_{E,H}:={\ensuremath{\left(\pi E^{1/2}{\ensuremath{\left(H_E^{-1}-1\right)}}^{1/4}{\ensuremath{\left(H_E\right)}}^{d/2}\right)}}^{-1}.$$ The third gives asymptotics in the ‘forbidden region’ where $(x, \xi)$ lies in the exterior of the energy surface. \[EXTDECAY\] Suppose that $H_E=H(x, \xi)/E>1$ and let $E_N(\hbar) =E$. Then, there exists $C_1>0$ so that $$|W_{\hbar, E_N(\hbar)}(x, \xi)| \leq C_1 \hbar^{-d + \frac{1}{2}} e^{- \frac{2 E}{\hbar}[\sqrt{H_E^2 -H_E} - \cosh^{-1}\sqrt{H_E} ]}.$$ Moreover, as $H(x,\xi) \to \infty$, there exists $C_2>0$ so that $$|W_{\hbar, E_N(\hbar)}(x, \xi)| \leq C_2 \hbar^{-d + \frac{1}{2}} e^{- \frac{2H(x,\xi)}{\hbar}}.$$ \[WLSECT\] Wigner-Laguerre --------------------------- In [@HZ19] (see also [@T; @T12]), it is shown that $$\label{E:Wigner-sp} W_{\hbar, E_N(\hbar)}(x, \xi) = \frac{(-1)^N}{(\pi \hbar)^d} e^{- 2H/\hbar} L^{(d-1)}_N(4H/\hbar),\qquad H=H(x,\xi)=\frac{{\left\lvert x \right\rvert}^2+{\left\lvert \xi \right\rvert}^2}{2},$$ where $L_N^{(d-1)}$ is the associated Laguerre polynomial of degree $N$ and type $d-1$. Classical asymptotics of Laguerre functions (see [@AT15; @AT15b] for recent results and references) show that for $x$ small relative to $N$, $e^{-x/2} L^{(d-1)}_N(x)$ can be asymptotically approximated by Bessel functions. To the right, in the bulk, the oscillatory behavior is modeled by trigonometric functions, and in a neighborhood of the extreme right by Airy functions. In this article, we are concerned with sums over $N$ of such Laguerre functions, and use semi-classical methods different from the classical ones, but this description is obviously visible in the main results of this article and in [@HZ19]. \[EMPSECT\] Empirical measures: Proof of Proposition \[INFINITEPROP\] --------------------------------------------------------------------- First let us consider the (conditionally convergent) integral of . Letting $K_{\hbar} = \Pi_{\hbar, E_N(\hbar)}$ and noting that $\sum_{N = 0}^{\infty}\Pi_{\hbar, E_N(\hbar)}(x,y) = \delta(x - y)$, we have $$\begin{aligned} \hbar^d\sum_N W_{ \hbar, E_N}(x, \xi) &=\hbar^d {\mathcal{W}}_{\hbar} (\sum_{N =1}^{\infty} \Pi_{\hbar, E_N(\hbar)})= \hbar^d {\mathcal{W}}_{\hbar} \delta(x - y)\\ &= \int_{{{\mathbb R}}^d} \delta(x - v/2, x + v/2) e^{\frac{i}{\hbar} \xi \cdot v } dv \equiv 1. \end{aligned}$$ On the other hand, the absolute sum is infinite. To see this, we use the identity and classical asymptotics of associated Laguerre functions. The sums in are weighted by $W_{\hbar, E_N(\hbar)}(x, \xi)$. We now show that the measures have infinite mass even if we $L^2$-normalize the Laguerre functions to form an orthonormal basis $$\psi_N^{\alpha} = \left(\frac{2^{-\alpha} \Gamma(N +1)}{\Gamma(N + \alpha +1)}\right)^{{{\textstyle \frac 12}}} L_N^{\alpha}({{\textstyle \frac 12}}r^2) e^{-\frac{1}{4} r^2},$$ of $L^2({{\mathbb R}}_+, r^{2 \alpha +1} dr)$. By , we want to show that $$\sum_{N = 1}^{\infty} \left(\frac{2^{-\alpha} \Gamma(N +1)}{\Gamma(N + \alpha +1)}\right)^{{{\textstyle \frac 12}}} | e^{- 2H/\hbar} L^{(d-1)}_N(4H/\hbar)| = \infty.$$ Since the exponential factor is independent of $N$, it may be removed from the sum. Moreover, the argument is independent of $N$ so it suffices to show that $$\sum_{N =1}^{\infty} \left(\frac{2^{-\alpha} \Gamma(N +1)}{\Gamma(N + \alpha +1)}\right)^{{{\textstyle \frac 12}}} | L^{(d-1)}_N(x) | = \infty, \forall x > 0.$$ The relevant asymptotics are the Hilb asymptotics of the Bessel regime, where $x$ is small relative to $N$, $$e^{-\frac{x}{2}} x^{\alpha/2} L_n^{(\alpha)}(x) = \frac{(n + \alpha)!}{n!} (N x)^{-\alpha/2} J_{\alpha}(2 \sqrt{Nx}) + {\varepsilon}(x, n),$$ where $$N = n + \frac{\alpha +1}{2}, \;\;{\varepsilon}(x, n) = \left\{\begin{array}{ll} x^{\alpha/2 + 2} O(n^{\alpha}), & 0 < x < \frac{c}{n}, \\ &\\ x^{5/4} O(n^{\alpha/2 - 3/4}), & \frac{c}{n} < x < C. \end{array} \right.$$ Note that $\frac{\Gamma(n + \alpha +1)}{\Gamma(n)} \simeq n^{\alpha}$, so $\frac{(n + \alpha)!}{n!} (N )^{-\alpha/2} \simeq n^{\alpha/2}.$ The $L^2$-normalizing constant puts in a factor asymptotic to $N^{-\alpha/2}$, cancelling the power in the Hilb asymptotics. In addition, we have the Bessel asymptotics of fixed order and large argument. When $0 < x < \sqrt{\alpha +1}$, $J_{\alpha}(x) \sim \frac{1}{\Gamma(\alpha +1)} (\frac{x}{2})^{\alpha}).$ When $x >> |\alpha^2 - \frac{1}{4}|$ then $J_{\alpha}(x) \sim \sqrt{\frac{2}{\pi x}} \cos(x - \frac{\alpha \pi}{2} - \frac{\pi}{4}). $ Clearly, the latter is the relevant form, and it puts in an additional factor of $N^{-{{\textstyle \frac 12}}}$. It follows that the summands are of asymptotic order $N^{-{{\textstyle \frac 12}}},$ and therefore the mass is infinite. A fortiori the mass of the measures is infinite. To complete the proof of Proposition \[INFINITEPROP\] we need to show that the mass of is finite on one-sided infinite intervals $[-\infty, \tau]$. This is due to the ‘two-sidedness’ of the Airy function. The main point (see Section \[HEURISTICSECT\]) is that terms with $E_N(\hbar) < H(x, \xi)$ are exponentially small and may be neglected. Therefore, $$\left| \mu_h^{(x, \xi)} \right| [-\infty, \tau] = \sum_{N: E_N(\hbar) \in [-\infty, \tau]} |W_{\hbar, E_N(\hbar)}(x, \xi)| \simeq \sum_{N: H(x, \xi) \leq E_N(\hbar) \leq \tau} |W_{\hbar, E_N(\hbar)}(x, \xi) | < \infty$$ for any $\hbar > 0$. Wigner distributions of the propagator -------------------------------------- Taking to be the propagator $U_\hbar(t)$ of , its Wigner transform may be calculated explicity by using the Mehler formula for $U_\hbar(t,x,y)$ in the position representation (see e.g. [@F]) $$\label{E:Mehler} U_h(t, x,y) = e^{-{{\frac{i}{h}}}t H_\hbar}(x,y)= \frac{1}{(2\pi i \hbar \sin t)^{d/2}} \exp\left[ \frac{i}{\hbar}\left( \frac{{\left\lvert x \right\rvert}^2 + {\left\lvert y \right\rvert}^2}{2} \frac{\cos t}{\sin t} - \frac{x\cdot y}{\sin t} \right) \right], $$ where $t \in {{\mathbb R}}$ and $x,y \in {{\mathbb R}}^d$. The right hand side is singular at $t=0.$ It is well-defined as a distribution, however, with $t$ understood as $t-i0$. Indeed, since $\widehat{H}_\hbar$ has a positive spectrum the propagator $U_\hbar$ is holomorphic in the lower half-plane and $U_\hbar(t, x, y)$ is the boundary value of a holomorphic function in $\{{{\operatorname{Im}\,}}t < 0\}$. \[WIGNERINTRO\] Uniform asymptotics of Wigner functions of spectral projections: Proof of Theorem \[ELEVELLOC\] ================================================================================================================ In this section we prove Theorem \[ELEVELLOC\] about asymptotics for the Wigner-Weyl sums $$W_{\hbar, f, E}(x,\xi)= \sum_N f(\hbar^{-1}(E-E_N(\hbar)) W_{\hbar, E_N(\hbar)}(x,\xi)$$ defined in , where $f$ is a Schwartz function. Our starting point is to combine with $\delta(\hbar)=\hbar$ with Proposition \[P:Wigner-prop\], which together read $$W_{\hbar, f, E}(x,\xi)= \int_{{\mathbb R}}e^{-itE/\hbar} \widehat{f}(t) A_\hbar(t)\exp{\ensuremath{\left(-\frac{i}{\hbar}\Psi(t,x,\xi)\right)}}\frac{dt}{2\pi}.\label{E:elevelloc-1}$$ Here, $$A_\hbar(t)={\ensuremath{\left(2\pi \hbar \cos(t/2)\right)}}^{-d},\qquad \Psi(t,x,\xi)=2H\tan(t/2),\quad H=H(x,\xi)=\frac{{\left\lvert \xi \right\rvert}^2+{\left\lvert \xi \right\rvert}^2}{2}.$$ Note that this distribution is actually a smooth function away from $t\in {\left\{\pi, 3\pi, \ldots\right\}}.$ By Lemma \[LOCLEM\], it remains to evaluate $$\label{E:elevelloc-3} \int_{{\mathbb R}}e^{-itE/\hbar} \widehat{f}(t)\eta_\hbar(t) e^{-itE/\hbar}{\mathcal{U}}_{\hbar}(t,\rho)\frac{dt}{2\pi}~=~ \int_{{\mathbb R}}\widehat{f}(t)\eta_\hbar(t) A(t)e^{-\frac{i}{\hbar}{\ensuremath{\left(\Psi(t,E,\rho)-tE\right)}}}\frac{dt}{2\pi},$$ where $\eta_\hbar:{{\mathbb R}}{\ensuremath{\rightarrow}}[0,1]$ is a smooth, periodic cut-off function with $$\eta_\hbar(t)= \begin{cases} 1,&\quad {\left\lvert t-(2k+1)\pi \right\rvert}>2\hbar^\delta,\,\,\forall k\in {{\mathbb Z}}\\ 0,&\quad {\left\lvert t-(2k+1)\pi \right\rvert}<\hbar^\delta,\,\,\text{for some }k\in {{\mathbb Z}}\\ \end{cases}.$$ We do this by stationary phase. The critical point equation for $\Psi(t, E, \rho)-tE$ as a function of $t$ is $$\label{CPE1} H_E=\cos^2{\ensuremath{\left(\frac{t}{2}\right)}},\qquad H_E:=\frac{\rho}{2E}.$$ If $H_E>1$ (i.e. the point $(x,\xi)$ at which we evaluate is outside the energy surface), there are no real critical points for $\Psi$ and the lemma of non-stationary phase shows that is $O(\hbar^\infty)$ as desired. If $H_E\in (0,1),$ we use the method of stationary phase (Lemma \[L:SP LO\]). The critical points of the phase $\Psi(t,E,\rho)-tE$ in are $$\begin{aligned} t_{+,\pm,k}&= 4\pi k \pm 2\cos^{-1}{\ensuremath{\left(H_E^{1/2}\right)}},\quad t_{-,\pm,k}= 4\pi {\ensuremath{\left(k+1/2\right)}} \pm 2\cos^{-1}{\ensuremath{\left(H_E^{1/2}\right)}},\qquad k \in {{\mathbb Z}}.\end{aligned}$$ where $\cos^{-1}:[-1,1]{\ensuremath{\rightarrow}}[0,\pi]$ and $$\cos(t_{\pm_1,\pm_2, k})=\pm_1 H_E^{1/2},\qquad \sin(t_{\pm_1,\pm_2, k})=\pm_1\pm_2 {\ensuremath{\left(1-H_E\right)}}^{1/2}.$$ Notice that since $H_E>0,$ all the critical points are a bounded distance away from ${\left\{(2k+1)\pi,\, k\in {{\mathbb Z}}\right\}}$ and hence are inside the support of $\eta_\hbar$ for all $\hbar$ sufficiently small. The stationary phase analysis of around $t_{\pm,\pm, k}$ is the same for all $k,$ so let us focus on $k=0.$ Since $$\rho = 4E H_E,$$ we have for every $k$ that $$\Psi(t_{\pm_1,\pm_2,k})=4EH_E\frac{\pm_1\pm_2{\ensuremath{\left(1-H_E\right)}}^{1/2}}{\pm_1H_E^{1/2}}= \pm_24E{\ensuremath{\left(H_E(1-H_E)\right)}}^{1/2}$$ and $$\Psi''(t_{\pm_1,\pm_2, k})=\frac{\rho}{2}\frac{\sin(t_{\pm_1,\pm_2,k}/2)}{\cos^3(t_{\pm_1,\pm_2,k}/2)}=\pm_2 2E{\ensuremath{\left(H_E^{-1}-1\right)}}^{1/2}$$ Note that $(2\pi \hbar)^{d-1/2}$ times the intergral in becomes $$\begin{aligned} {\ensuremath{\left(2\pi \hbar\right)}}^{-1/2} \int_{{\mathbb R}}\frac{\widehat{f}(t)\eta_\hbar(t)}{\cos(t/2)^d} e^{\frac{i}{\hbar}{\ensuremath{\left(-\Psi(t,E,\rho)+tE\right)}}}\frac{dt}{2\pi}.\end{aligned}$$ Applying the stationary phase expansion (Lemma \[L:SP LO\]), completes the proof (we obtain an error $\hbar^{1-2\delta}$ from the fact that after expanding to order $k$ in stationary phase, the remainder is a universal constant times the supremum of the ${\ensuremath{\left(k+1\right)}}^{st}$ derivative of the amplitude, and each derivative of $\eta_\delta$ give a power of $\hbar^{-\delta}.$). $\square$ Interface asymptotics for smooth $\hbar^{2/3}$-localized Weyl sums: Proof of Theorem \[RSCOR\] {#RSCOR-pf} =============================================================================================== We now prove Theorem \[RSCOR\]. Our starting point is with $\delta(\hbar)=\hbar:$ $$W_{f,\hbar, 2/3}(x,\xi)=\sum_{N\geq 0} f{\ensuremath{\left(\hbar^{-2/3}{\ensuremath{\left(E-E_N(\hbar)\right)}}\right)}}W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{{{\mathbb R}}} \widehat{f}(t) e^{\frac{i}{\hbar^{2/3}}tE} {\mathcal{U}}(t\hbar^{1/3},x,\xi)\frac{dt}{2\pi}.$$ Taylor expanding the result of Proposition \[P:Wigner-prop\] we obtain the following pointwise asymptotics $$\label{u13} e^{\frac{i}{\hbar^{2/3}}tE}{\mathcal{U}}(t\hbar^{1/3},x,\xi) = (2\pi \hbar)^{-d}e^{i\left[t\frac{E-H}{\hbar^{2/3}}-\frac{t^3}{12}H\right]}{\ensuremath{\left(1+O((1+{\left\lvert t \right\rvert})^5\hbar^{2/3})\right)}},$$ and, moreover, that the left hand side has a complete asymptotic expansion in ascending powers of $\hbar^{2/3}.$ Since $\widehat{f}$ is Schwartz, touine Taylor expansion shows that $W_{f,\hbar, 2/3}(x,\xi)$ has a complete asymptotic expansion in ascending powers of $\hbar^{2/3}$ as well. To compute the leading term, recall that $$H=H(x,\xi)=E + u {\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3}.$$ This assumption and imply that $$\label{WfFT} W_{f,\hbar, 2/3}(x,\xi)=(2\pi \hbar)^{-d}\int_{{{\mathbb R}}} \widehat{f}(t)e^{-i\left[ \frac{tu}{(2E)^{2/3}} + \frac{t^3}{12}E\right]}\frac{dt}{2\pi} {\ensuremath{\left(1+O((1+{\left\lvert u \right\rvert})\hbar^{2/3})\right)}},$$ Changing variables $t\mapsto -C_E^{-1} t$, with $C_E= (4/E)^{1/3}$, we arrive at the representation $$W_{f,\hbar, 2/3}(x,\xi)=-(2\pi \hbar)^{-d}\int_{{{\mathbb R}}} C_E\widehat{f}(-t)(C_E t)e^{i\left[ \frac{tu}{E} + \frac{t^3}{3}\right]}\frac{dt}{2\pi}{\ensuremath{\left(1+O((1+{\left\lvert u \right\rvert})\hbar^{2/3})\right)}}.$$ Using that that inverse Fourier transform of $-C_E\widehat{f}(-tC_E)$ in the $t{\ensuremath{\rightarrow}}\lambda$ variables evaluated at $\lambda$ is $f(-\lambda/C_E)$, that the Airy function is the inverse Fourier transform of $e^{it^3/3}$, and that the Fourier transform is an isometry on $L^2$, we therefore find $$W_{f,\hbar, 2/3}(x,\xi)=(2\pi \hbar)^{-d}\int_{{{\mathbb R}}} f( -\lambda/C_E ) \operatorname{Ai}{\ensuremath{\left(\lambda + \frac{u}{E}\right)}} d\lambda+O((1+{\left\lvert u \right\rvert})\hbar^{-d+2/3}),\qquad C_E=(4/E)^{1/3},$$ completing the proof of Theorem \[RSCOR\]. Interface asymptotics for sharp $\hbar^{2/3}$-localized Weyl sums: Proof of Theorem \[SHARPh23INTER\] \[2/3sumSect\] ==================================================================================================================== In this section we consider the sharp $\hbar^{2/3}$ localized Weyl sums around an energy level $E$ and consider the interface scaling around the energy surface . Recall that $$W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{S^1}e^{it(E_N(\hbar)-E)/\hbar} e^{itE/\hbar}{\mathcal{U}}_{\hbar}(t,x,\xi)\frac{dt}{2\pi}.$$ We claim that in the regime $$H(x,\xi) = E + u{\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3},\qquad E_N(\hbar)-E=O(\hbar^{1-{\varepsilon}})\label{E:surface-scaling}$$ we also have for any $\delta>0$ $$W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{S^1}\chi(\hbar^{-2/3+\delta}t)e^{it(E_N(\hbar)-E)/\hbar} e^{itE/\hbar}{\mathcal{U}}_{\hbar}(t,x,\xi)\frac{dt}{2\pi}+O(\hbar^\infty)$$ where $\chi$ is a smooth cutoff the is identically $1$ near $0$ and supported in $[-1,1].$ To see this, we begin with the exact distributional formula $$W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{S^1}e^{itE_N(\hbar)/\hbar}{\mathcal{U}}_{\hbar}(t,x,\xi)\frac{dt}{2\pi}=\int_{S^1}e^{it(E_N(\hbar)-E)/\hbar}e^{itE/\hbar}{\mathcal{U}}_{\hbar}(t,x,\xi)\frac{dt}{2\pi}.$$ As in Lemma \[LOCLEM\], we can localize away from the singular $t=\pi$ for the amplitude of ${\mathcal{U}}_\hbar(t,x,\xi)$ at the cost of an error of size $O(\hbar^\infty).$ Then, on the support of $1-\chi(\hbar^{-2/3+\delta}t)$ and away from $t=\pi$, there are no critical points for the combined phase $itE-i\tan(t/2)\rho$, and hence, using the assumption from that $E-E_N(\hbar)=O(\hbar^{1-{\varepsilon}})$ we may apply the lemma of non-stationary phase. Thus, to prove Theorem \[SHARPh23INTER\], it is enough to prove if for the localized integral $$W_{2/3,E,\lambda_{\pm}}^{\chi}(x,\xi):=\int_{{{\mathbb R}}} \chi(\hbar^{-2/3+\delta}t)\frac{e^{itn_+}-e^{itn_-}}{e^{it}-1} e^{itE/\hbar} {\mathcal{U}}_\hbar(t,x,\xi)\frac{dt}{2\pi}+O(\hbar^{\infty}),\label{E:loc-interface}$$ where the notation $n_{\pm}$ is from the definition of the localized Wigner function and we have $n_+-n_- \simeq \hbar^{-1/3}.$ Using that for any $a,b$ $$e^{ita}-e^{itb}=it\int_a^b e^{it\lambda}d\lambda,$$ rescaling $t\mapsto t\hbar^{1/3}$, and Taylor expanding the phase and amplitude in $e^{itE/\hbar}{\mathcal{U}}_\hbar(t,x,\xi)$, we find that that $W_{2/3,E,\lambda_{\pm}}^{\chi}(x,\xi)$ equals $$(2\pi\hbar)^{-d}\int_{{{\mathbb R}}}{\bf 1}_{{\left\{s\in [\lambda_-,\lambda_+]\right\}}} \int_{{{\mathbb R}}} \chi(\hbar^{\delta}t) e^{-it (\hbar^{2/3}{\ensuremath{\left(E-H\right)}}-\lambda) + i (t^3/3)(E/4)}\frac{dtd\lambda}{2\pi}+O(\hbar^{-d+1/3-6\delta}).$$ Using that $H=H(x,\xi)=E+u{\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3}$ and changing variables $$(\lambda,t)\mapsto (-\lambda,-tC_E^{-1})\qquad C_E={\ensuremath{\left(4/E\right)}}^{1/3}$$ shows that $W_{2/3,E,\lambda_{\pm}}^{\chi}(x,\xi)$ equals $$(2\pi\hbar)^{-d}C_E\int_{{{\mathbb R}}}{\bf 1}_{{\left\{\lambda\in [-\lambda_+,-\lambda_-]\right\}}} \int_{{{\mathbb R}}} \chi(\hbar^{\delta}C_Et) e^{i{\ensuremath{\left(t (\lambda C_E+u/E) + t^3/3\right)}}}\frac{dtd\lambda}{2\pi}+O(\hbar^{-d+1/3-6\delta}+(1+{\left\lvert u \right\rvert})\hbar^{-d+2/3-6\delta}).$$ Since (see [@HZZ16 Lem. 3.1]) for any $\delta>0$ $$\int_{{\left\lvert t \right\rvert}<\hbar^{-\delta}}e^{itu+it^3/3}\frac{dt}{2\pi}=\operatorname{Ai}(u)+O(\hbar^\infty),$$ we find because $\chi(\hbar^{\delta}t)=1$ for all ${\left\lvert t \right\rvert}<\hbar^{-\delta},$ $$W_{2/3,E,\lambda_{\pm}}^{\chi}(x,\xi)=(2\pi\hbar)^{-d}C_E\int_{-\lambda_+}^{-\lambda_-}\operatorname{Ai}{\ensuremath{\left(\frac{u}{E}+C_E\lambda\right)}} d\lambda + O{\ensuremath{\left(\hbar^{-d+1/3-6\delta}+(1+{\left\lvert u \right\rvert})\hbar^{-d+2/3-6\delta}\right)}}.$$ This completes the proof. Bulk asymptotics: Proofs of Propositions \[smoothedbulk\] and \[pp:PBK-leading\] ================================================================================ Asymptotics for $\hat{f} \in C_0({{\mathbb R}})$: Proof of Proposition \[smoothedbulk\] {#BULKfSECT} --------------------------------------------------------------------------------------- Fix $f\in \mathcal S({{\mathbb R}})$ with $\hat{f}\in C_0({{\mathbb R}}).$ Relation with $\delta(\hbar)=1$ and $E=0$ reads $$[W_{\hbar, f}(x,\xi)=\sum_{N\geq 0} f(E_N(\hbar)) W_{\hbar, E_N(\hbar)}(x,\xi)=\int_{{{\mathbb R}}}\widehat{f}(t){\mathcal{U}}_{\hbar}(-t\hbar,x,\xi) \frac{dt}{2\pi},$$ where we recall that ${\mathcal{U}}_{\hbar}$ is the Wigner function of the propagator. Proposition \[P:Wigner-prop\] therefore yields $$W_{\hbar, f}(x,\xi)= (2\pi \hbar)^{-d}\int_{{{\mathbb R}}} \widehat{f}(t){\ensuremath{\left(\cos(t\hbar/2)+i0\right)}}^{-d}e^{-\frac{2iH}{\hbar}\frac{\sin(-t\hbar/2)}{\cos(-t\hbar/2)}}\frac{dt}{2\pi},\quad H=H(x,\xi)=\frac{{\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2}{2}.$$ For $\hbar$ sufficiently small, ${\mathcal{U}}_{\hbar}(-t\hbar, x,\xi)$ has no singularities on the support of $\widehat{f}.$ Taylor expanding in $\hbar$ therefore yields a complete asymptotic expansion for $W_{\hbar, f},$ as claimed. To complete the proof of Proposition \[smoothedbulk\], note that $${\mathcal{U}}_\hbar(-t\hbar, x,\xi)=(2\pi \hbar)^{-d}e^{itH}{\ensuremath{\left(1+O(\hbar^2)\right)}}.$$ Hence, the leading term in the expansion is $$(2\pi\hbar)^{-d}\int_{{{\mathbb R}}}\hat{f}(t) e^{itH}\frac{dt}{2\pi} =(2\pi\hbar)^{-d}f(H(x,\xi)),$$ as desired. As discussed above, if $|D^{\ell} \widehat{f}| \in L^1({{\mathbb R}})$ for all $\ell \leq d$, then after integrating by parts $d$ times the integral becomes absolutely convergent. We may then calculate the limit $\hbar \to 0$ by dominanted convergence. The limit is $$H^{-d} \int_{{{\mathbb R}}} D^d \widehat{f}(t) e^{i tH} \frac{dt}{2\pi}= \int_{{{\mathbb R}}} \widehat{f}(t) e^{itH}\frac{dt}{2\pi} = f(H(x,\xi)),$$ as in the Schwartz case. Proof of Proposition \[pp:PBK-leading\] (i) {#PROOFBULK1} ------------------------------------------- Proposition \[pp:PBK-leading\] is formally Proposition \[smoothedbulk\] in the case $f = {\bf 1}_{[E_1, E_2]} (E_N(\hbar))$. It is sufficient to study the intervals $[0, E]$ (where we assume $H(x, \xi) > 0$), and so we assume henceforth that the interval has this form. Writing $f(E_N(\hbar)) := {\bf 1}_{[0, E]}(E_N(\hbar)),$ we are interested in $$W_{\hbar, [0,E]}(x,\xi)=\sum_{N: E_N(\hbar) \leq E} W_{\hbar, E_N(\hbar)}(x, \xi).$$ The endpoint of the sum is the largest value $N(\hbar, E) \in {\mathbb N}$ for which $E_N(\hbar)=\hbar(N + \frac{d}{2}) \leq E$, i.e. for which $N \leq \hbar^{-1} E - \frac{d}{2}$. Lemma \[NICEFORM\] with $N_1=0$ and $N_2=N(\hbar, E)$ reads $$\label{EXACT-new}W_{\hbar, [0,E]}= \int_{S^1} {\mathcal{U}}_{\hbar}(t, x, \xi) e^{itd/2} D_{\hbar, E}(t) \frac{dt}{2\pi}, \qquad D_{\hbar,E}(t):=\frac{e^{ i N(\hbar, E)t} -1}{e^{it} -1},$$ where $D_{\hbar, E}$ is the Dirichlet kernel. Before turing to the detailed analysis of the expression in the previous line, note that three intervals in $t$ are important to this oscillatory integral expression for the pointwise Weyl sum $W_{\hbar, [0,E]}(x,\xi)$ as $\hbar{\ensuremath{\rightarrow}}0.$ The first contribution comes from the large peak of the Dirichlet kernel around $t=0$: $${\left\lvert t \right\rvert}\leq N(\hbar,E)^{-1}\simeq \hbar\quad \Rightarrow \quad D_{\hbar, E}(t)\simeq N(\hbar, E)\simeq \hbar^{-1}.$$ Using that the expression for ${\mathcal{U}}_\hbar$ has a prefactor of size $\hbar^{-d},$ we will see that this interval gives an $\hbar^{-d}(1+O(\hbar))$ contribution to $W_{\hbar, [0,E]}$ as $\hbar{\ensuremath{\rightarrow}}0$ when $H(x,\xi)<E$ and is $O(\hbar^\infty)$ when $H(x,\xi)>E.$ The second contribution comes from the stationary phase points of ${\mathcal{U}}_\hbar.$ As we will see, when $H(x,\xi)<E,$ these points are well-separated from both $t=0$ and $t=\pi.$ A simple stationary phase argument will show that the resulting contribution is of size $O(\hbar^{-d+1/2}).$ In contrast, if $H(x,\xi)>E$, then there are no stationary phase points and this integral will also contribute $O(\hbar^\infty).$ Finally, we appeal to the argument from Lemma \[LOCLEM\] to show that that the apparent singularity $t=\pi$ of the amplitude $\cos(t/2)^{-d}$ in the expression for ${\mathcal{U}}_\hbar$ in fact is negligible (i.e. has size $O(\hbar^\infty)$) in the semi-classical limit $\hbar{\ensuremath{\rightarrow}}0.$ We now turn to the details. Fix a parameter $M>0$, which we will later take to be $\hbar^{-{\varepsilon}}$ for ${\varepsilon}>0$ sufficiently small, and let $\chi:{{\mathbb R}}{\ensuremath{\rightarrow}}[0,1]$ be a smooth bump function: $$\chi(t):= \begin{cases} 1,\qquad t \leq 1 \\ 0,\qquad t > 2 \end{cases}.$$ We first compute the contribution to $W_{\hbar, [0,E]}(x,\xi)$ from the peak of the Dirichlet kernel. Specifically, let us check that with ${\varepsilon}$ sufficiently small and positive for $M= \hbar^{-{\varepsilon}},$ we have $$\int_{-\pi}^\pi \chi((M\hbar)^{-1} t){\mathcal{U}}_{\hbar}(t, x, \xi) e^{itd/2} D_{\hbar, E}(t) \frac{dt}{2\pi} = \begin{cases} (2\pi\hbar)^{-d}{\ensuremath{\left(1 + O(M\hbar)\right)}},&~~ H(x,\xi)<E\\ O(\hbar^\infty),&~~ H(x,\xi)>E \end{cases}.\label{E:Bulk-Wig-Dir}$$ This will turn out the be the dominant contribution to $W_{\hbar, [0,E]}(x,\xi)$ in the case $H(x,\xi)<E.$ The integrand in is supported on ${\left\lvert t \right\rvert}\leq 2 M\hbar$. In this regime, Taylor expansion gives $${\mathcal{U}}_\hbar(t,x,\xi)=(2\pi\hbar)^{-d}e^{-\frac{i}{\hbar}tH}(1+O((M\hbar)^2)+O(M^3\hbar^2)),\qquad H = H(x,\xi) =\frac{{\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2}{2}.$$ Next, note that for any $\beta\in {{\mathbb R}}$ $$e^{i\beta t}-1=i\beta t \int_0^1 e^{i\beta t (1-s)}ds.$$ Hence, on the support of the integrand in , $$e^{itd/2}D_{\hbar, E}(t)= Ne^{\frac{i}{\hbar}tE}\int_0^1 e^{-iNts}ds (1+O(M\hbar)),\qquad N = N(\hbar, E)=\frac{E}{\hbar}-\frac{d}{2}.$$ Thus, using that ${\left\lvert D_{\hbar, E}(t) \right\rvert}=O(\hbar^{-1})$ on the support of $\chi((M\hbar)^{-1}t)$, which has size $O(M\hbar),$ the integral in equals $$(2\pi\hbar)^{-d}\int_{-\infty}^\infty \int_0^N e^{\frac{i}{\hbar}t\left[(E-H)-s\right]} \chi{\ensuremath{\left((M\hbar)^{-1}t\right)}} \frac{dsdt}{2\pi}$$ plus an error of size $O(M^2\hbar^{-d+1})+O(M^3\hbar^{-d+2}).$ Changing variables $t\mapsto t (\hbar M)$ and performing the $dt$ integral, the expresion in the previous line becomes $$(2\pi\hbar)^{-d}(M\hbar)\int_0^N \widecheck{\chi}{\ensuremath{\left(M\hbar (E-H - s)\right)}} ds,$$ where $\widecheck{\chi}$ is the inverse Fourier transform. Finally, changing variables $s\mapsto \hbar M(E-H - s)$, this integral becomes $$(2\pi\hbar)^{-d}\int_{M(E-H-N)}^{M(E-H)} \widecheck{\chi}{\ensuremath{\left(s\right)}} ds=(2\pi \hbar)^d{\ensuremath{\left(\chi(0) + O(\hbar^\infty)\right)}}=(2\pi \hbar)^d{\ensuremath{\left(1 + O(\hbar^\infty)\right)}},\label{E:inverse-fourier}$$ provided $M=\hbar^{\varepsilon}$ for any ${\varepsilon}>0$ and assuming that $H<E$. In contrast, when $H>E,$ this integral is $O(\hbar^\infty)$ since $\widecheck{\chi}$ is rapidly decaying at infinity. This proves . The same argument as in Lemma \[LOCLEM\] combined with shows that $$W_{\hbar, [0,E]}= (2\pi \hbar)^{-d}(1+O(\hbar^{1-{\varepsilon}}))+ \int_{-\pi}^\pi (1-\chi((M\hbar)^{-1} t))\psi_\delta(t){\mathcal{U}}_{\hbar}(t, x, \xi) e^{itd/2} D_{\hbar, E}(t) \frac{dt}{2\pi} +O(\hbar^\infty),$$ where $\psi_\delta$ is a smooth cut-off function that localizes the integral away from a $\hbar^{\delta}$ neighborhood of $t=\pi.$ Thus, to complete the proof of Proposition \[pp:PBK-leading\], it remains to check $$\int_{-\pi}^\pi (1-\chi((M\hbar)^{-1} t))\psi_\delta(t){\mathcal{U}}_{\hbar}(t, x, \xi) e^{itd/2} D_{\hbar, E}(t) \frac{dt}{2\pi} =O(\hbar^{-d+1/2})\label{E:Bulk-Wig-SP}$$ where $M=\hbar^{{\varepsilon}}$ and ${\varepsilon}, \delta$ are sufficiently small. Our argument uses the method of stationary phase. We write $${\mathcal{U}}_{\hbar}(t, x, \xi) e^{itd/2} D_{\hbar, E}(t)= \frac{e^{-\frac{i}{\hbar}(\Psi(t,x,\xi)-tE)}-e^{-\frac{i}{\hbar}\Psi(t,x,\xi)}}{(\cos(t/2)+i0)^{d}(e^{it}-1)},\quad \Psi(t,x,\xi)=\tan(t/2) ({\left\lVertx\right\rVert}^2+{\left\lVert\xi\right\rVert}^2).$$ On the support of $(1-\chi((M\hbar)^{-1}t))\psi(t)$, the phase function $\Psi$ has no critical points, while the amplitude has its $k^{th}$ derivative bounded by $\hbar^{k\max{\left\{\delta, {\varepsilon}\right\}}}.$ Thus, the integral in corresponding to term with $e^{-\frac{i}{\hbar}\Psi(t,x,\xi)}$ gives contribution $O(\hbar^\infty).$ In contrast, the critical point equation for $\Psi(t,x,\xi)-tE$ is $$\cos^2(t/2)=H(x,\xi)/E.\label{E:crits-again}$$ This has two real solutions bounded away from $t=0,\pi$ when $0<H(x,\xi)<E$ and has no real solutions if $H(x,\xi)>E$. In the former case, these critical points are easily seen to be non-degenerate and the usual method of stationary phase show that the contribution to $W_{hbar,[0,E]}(x,\xi)$ is of order $\hbar^{-d+1/2}$ when ${\varepsilon},\delta$ are sufficiently close to $0.$ In the latter case, the integral is $O(\hbar^\infty).$ This completes the proof of Proposition \[pp:PBK-leading\].$\square$ Interface asymptotics for bulk Wigner distributions: Proof of Theorem \[BULKSCALINGCOR\] ======================================================================================== Throughout this section we fix an energy level $E$ and consider the energy window or shell $[0, E]$. We are interested in the interface asymptotics behavior of bulk Wigner functions at points $(x, \xi)$ which satisfy $$\label{ASSUME}H(x, \xi)= \frac{{\left\lvert x \right\rvert}^2+{\left\lvert \xi \right\rvert}^2}{2} = E + u{\ensuremath{\left(\frac{\hbar}{2E}\right)}}^{2/3}.$$ Combining Lemma \[NICEFORM\] gives $$W_{\hbar, [0,E]}(x,\xi)=\int_{-\pi}^\pi \chi(t) D_{\hbar, E}(t)e^{itd/2} {\mathcal{U}}_\hbar(t,x,\xi)\frac{dt}{2\pi}+O(\hbar^\infty)$$ where $$D_{\hbar, E}(t)=\frac{e^{itN_E}-1}{e^{it}-1},\qquad N_E=E/\hbar - d/2$$ and $\chi$ is a smooth cutoff that is identically $0$ near $t=\pm \pi.$ Proposition therefore shows that, up to an error of size $O(\hbar^\infty),$ $$W_{\hbar, [0,E]}(x,\xi)=(2\pi \hbar)^{-d}N_E \int_{-\pi}^\pi \chi_1(t)D_{\hbar, E}(t)e^{itd/2}\cos(t/2)^{-d} e^{-\frac{2i H}{\hbar}\tan(t/2)}\frac{dt}{2\pi},$$ where as before $H=H(x,\xi).$ Note that both phase functions $2H\tan(t/2)$ and $2H\tan(t/2)-tE$ have no critical points outside ${\left\lvert t \right\rvert}\leq C\hbar^{1/3}{\left\lvert u \right\rvert}^{1/2},$ where $C$ is a fixed sufficiently large constant. Thus, we may further localize: $$W_{\hbar, [0,E]}(x,\xi)=(2\pi \hbar)^{-d}\int_{-\pi}^\pi \chi(t\hbar^{-1/3+{\varepsilon}}{\left\lvert u \right\rvert}^{-1/2})D_{\hbar, E}(t)e^{itd/2}\cos(t/2)^{-d} e^{-\frac{2i H}{\hbar}\tan(t/2)}\frac{dt}{2\pi}+O(\hbar^\infty)$$ As above we use that $$D_{\hbar, E}(t)=-\frac{it}{e^{it}-1} N_E\int_0^1 e^{itN_E(1-s)}ds=-(1+O(t))e^{itE/\hbar}\int_0^{N_E} e^{-its}ds$$ to find that, up to errors of size $O(\hbar^\infty)$ the normalized Wigner function $(2\pi\hbar)^d W_{\hbar, E}(x,\xi)$ equals $$N_E\int_0^{1}\int_{-\pi}^\pi \chi(t\hbar^{-1/3}{\left\lvert u \right\rvert}^{-1/2})(1+O(t))e^{\frac{i }{\hbar}{\ensuremath{\left(2H\tan(t/2)+t(s-E)\right)}}}\frac{dt}{2\pi}ds.$$ Taylor expanding yields $$2H\tan(t/2)+t(s-E)=t(H-E+s) ~+~Ht^3/12 + O(t^5).$$ Thus, $(2\pi \hbar)^d W_{\hbar, [0,E]}(x,\xi)$ is $$N_E\int_0^1 \int_{-\infty}^\infty \chi(t\hbar^{-1/3+{\varepsilon}}{\left\lvert u \right\rvert}^{-1/2})e^{\frac{i}{\hbar}{\ensuremath{\left(t(H-E+s)+Et^3/12\right)}}}{\ensuremath{\left(1+R(t,s)\right)}}\frac{dt}{2\pi}ds,$$ where the remainder $R$ satisfies $$R(t,s)=O(t)+O(t^5/\hbar)+O(t^3\hbar^{-1/3}{\left\lvert u \right\rvert}).$$ Since the integrand vanishes when ${\left\lvert t \right\rvert}\geq \hbar^{1/3-{\varepsilon}}{\left\lvert u \right\rvert}^{1/2},$ we find that $(2\pi \hbar)^d W_{\hbar, [0,E]}(x,\xi)$ is $$N_E\left[\int_0^1 \int_{-\infty}^\infty \chi(t\hbar^{-1/3+{\varepsilon}}{\left\lvert u \right\rvert}^{-1/2})e^{\frac{i}{\hbar}{\ensuremath{\left(t(H-E+s)+Et^3/12\right)}}}\frac{dt}{2\pi}ds+O(\hbar^{1/3}{\left\lvert u \right\rvert}^{1/2})+O(\hbar^{2/3}{\left\lvert u \right\rvert}^{5/2})\right].$$ Changing variables to $$T=(4\hbar/E)^{1/3}t,\qquad\sigma = s(\hbar/2E)^{2/3}$$ and using that $\hbar N_E = E + O(\hbar),$ we obtain the following expression for $(2\pi \hbar)^d W_{\hbar, [0,E]}(x,\xi):$ $$\int_{u/E}^{u/E+(2E/\hbar)^{2/3}} \int_{-\infty}^\infty \chi(T\hbar^{{\varepsilon}}{\left\lvert u \right\rvert}^{-1/2})e^{\frac{i}{\hbar}{\ensuremath{\left(T\sigma+T^3/3\right)}}}\frac{dt}{2\pi}d\sigma= \int_0^{\infty}\operatorname{Ai}(u/E+t)dt$$ plus an error of size $ O(\hbar^{1/3}{\left\lvert u \right\rvert}^{1/2})+O(\hbar^{2/3}{\left\lvert u \right\rvert}^{5/2}).$ This completes the proof of Theorem \[BULKSCALINGCOR\]. \[APPENDIX\] Appendix ===================== Stationary Phase Expansion {#SO} -------------------------- We recall here the following simple version of the stationary expansion, which we use in several proofs. \[L:SP LO\] Suppose $a,S\in \mathcal S({{\mathbb R}})$ and $S$ is a complex-valued phase function such that ${{\operatorname{Im}\,}}S|_{{{\operatorname{Supp\,}}}(a)} \geq 0$ with a unique non-degenerate critical point at $t_0\in {{\operatorname{Supp\,}}}(a)$ satisfying ${{\operatorname{Im}\,}}S(t_0)=0.$ Define $$I(h)=(2\pi h)^{-1/2}\int_{{{\mathbb R}}}e^{iS(x)/h}a(x)dx.$$ Then $$\begin{aligned} \label{E:Stationary Phase 2} I(h) = C(S) \left[a(t_0)+O(h)\right],\qquad C (S)= e^{i \frac{\pi}{4} \text{sgn} S''(0)} {\ensuremath{\left(\frac{2\pi h}{{\left\lvert S''(t_0) \right\rvert}}\right)}}^{1/2}.\end{aligned}$$ \[AIRYAPP\] Appendix on the Airy function ----------------------------------------- The Airy function is defined by, $$Ai(z) = \frac{1}{2 \pi i} \int_L e^{v^3/3 - z v} dv,$$ where $L$ is any contour that beings at a point at infinity in the sector $- \pi/2 \leq \arg (v) \leq - \pi/6$ and ends at infinity in the sector $\pi/6 \leq \arg(v) \leq \pi/2$. In the region $|\arg z| \leq (1 - \delta) \pi$ in ${{\mathbb C}}- \{{{\mathbb R}}_-\}$ write $v = z^{{{\textstyle \frac 12}}} + i t ^{{{\textstyle \frac 12}}}$ on the upper half of L and $v = z^{{{\textstyle \frac 12}}} - i t^{{{\textstyle \frac 12}}}$ in the lower half. Then $$\label{AIRYASYM} \operatorname{Ai}(z) = \Psi(z) e^{- \frac{2}{3} z^{3/2}}, \; \rm{ with}\; \Psi(z) \sim z^{-1/4} \sum_{j = 0}^{\infty} a_j z^{- 3j/2}, \;\; a_0 = \frac{1}{4} \pi^{-3/2}.$$ [HHHH]{} Aptekarev, A. I.; Tulyakov, D. N. Asymptotic behavior of the Lp-norms of Laguerre polynomials. (Russian) Uspekhi Mat. Nauk 70 (2015), no. 5(425), 177–178; translation in Russian Math. Surveys 70 (2015), no. 5, 955–957. \[A. I. Aptekarev and D. N. Tulyakov, "Asymptotics of Lp-norms of Laguerre polynomials and entropic moments of D-dimensional oscillator”, Preprint No. 41, Keldysh Inst. Appl. Math., 2015\]. L. Baez-Duarte, Central limit theorem for complex measures. J. Theoret. Probab. 6 (1993), no. 1, 33-56. Berry, M. V. Semi-classical mechanics in phase space: a study of Wigner’s function. Philos. Trans. Roy. Soc. London Ser. A 287 (1977), no. 1343, 237–271. M.V. Berry, Some quantum-to-classical asymptotics. [*Chaos et physique quantique*]{} (Les Houches, 1989), 251-304, North-Holland, Amsterdam, 1991 G. Folland, Harmonic Analysis in Phase Space, Ann. of Math. Stud., vol. 122, Princeton University Press, 1989 C.L. Frenzen and R. Wong, Uniform asymptotic expansions of Laguerre polynomials. SIAM J. Math. Anal. 19 (1988), no. 5, 1232–1248. V. Guillemin and S. Sternberg, The metaplectic representation, Weyl operators and spectral theory. J. Funct. Anal. 42 (1981), no. 2, 128-225. Boris Hanin, Steve Zelditch, Peng Zhou Nodal Sets of Random Eigenfunctions for the Isotropic Harmonic Oscillator, International Mathematics Research Notices, Vol. 2015, No. 13, pp. 4813–4839, (2015) (arXiv:1310.4532) Boris Hanin, Steve Zelditch and Peng Zhou, Scaling of harmonic oscillator eigenfunctions and their nodal sets around the caustic. Comm. Math. Phys. 350 (2017), no. 3, 1147-1183 (arXiv:1602.06848). B. Hanin and S. Zelditch, Interface Asymptotics of Eigenspace Wigner distributions for the Harmonic Oscillator, arXiv:1901.06438. Hörmander, Lars [*The analysis of linear partial differential operators. III.* ]{} Classics in Mathematics. Springer-Verlag, Berlin, 2003. A. J. E. M. Janssen and S. Zelditch, Szegö limit theorems for the harmonic oscillator. Trans. Amer. Math. Soc. 280 (1983), no. 2, 563–587. F. W. J. Olver, [*Asymptotics and special functions.* ]{} Academic Press, New York. M.A. Pinsky, M. Taylor, Pointwise Fourier inversion: A wave equation approach, J. Fourier Anal. Appl. 3 (6) (1997) 647-703. MR1481629 S. I. Resnick, [*A probability path*]{}. Modern Birkhauser Classics. Birkauser/Springer, New York, 2014. D. Robert,[*Autour de l’approximation semi-classique.* ]{} Progress in Mathematics, 68. Birkhäuser Boston, Inc., Boston, MA, 1987. S. Thangavelu, Lectures on Hermite and Laguerre expansions. With a preface by Robert S. Strichartz. Mathematical Notes, 42. Princeton University Press, Princeton, NJ, 1993. S. Thangavelu, Hermite and Laguerre semigroups: some recent developments. Orthogonal families and semigroups in analysis and probability, 251-284, Semin. Congr., 25, Soc. Math. France, Paris, 2012 E.P. Wigner, “On the quantum correction for thermodynamic equilibrium”, Phys. Rev. 40 (June 1932) 749-759. S. Zelditch and P. Zhou, Interface asymptotics of partial Bergman kernels on $S^1$-symmetric Kaehler manifolds, to appear in J. Symp. Geom. (arXiv:1604.06655). S. Zelditch and P. Zhou, Central Limit theorem for spectral Partial Bergman kernels, to appear in Geom. Topl. arXiv:1708.09267. [^1]: Research partially supported by NSF grant DMS-1810747. [^2]: The errors blow up when $u=\hbar^{-1/3}$.

ArXiv

epsf =1200 plus 4pt minus 4pt 1.0in **ABSTRACT** Most of the present understanding of the S=1 quantum spin chains displaying the Haldane gap is coming from the so-called valence-bond-solid (VBS) Hamiltonian which has an exactly known ground state. We show that this point is characterized by the onset of short-range incommensurate spin correlations in the one-parameter family of Hamiltonians $ H_\theta = \cos\theta \sum_{i} {\bf S}_i \cdot {\bf S}_{i+1} + \sin\theta \sum_i ({\bf S}_i \cdot{\bf S}_{i+1} )^2 $. This gives a physical meaning to this special point. We establish precise values for the gaps, correlations, the string order parameter, and identify the VBS point as a disorder point in the sense of classical statistical mechanics. It is a quantum remnant of the classical transition between a ground state with long-range Néel order and a ground state with incommensurate long-range order. [**I. INTRODUCTION**]{} It has been conjectured by Haldane$^1$ that antiferromagnetic quantum spin chains have a disordered ground state with a gap to spin excitation when the spins are integer. This phenomenon has been studied extensively over the years and a simple physical picture has emerged through the consideration of the so-called valence-bond solid (VBS) Hamiltonian$^2$. This peculiar Hamiltonian contains, in addition to the simplest isotropic bilinear nearest-neighbor exchange, a biquadratic term in the case of the spin S=1 chain. This modification transforms the Hamiltonian in a sum of projection operators and, as a consequence, the ground state is known exactly and has a simple structure. This is different from the exactly integrable models: here nothing is exactly known about the excited states. It is believed that this model is smoothly connected to the usual nearest-neighbor antiferromagnet: they share the same physics. More precisely, there is a hidden topological long-range order$^3$ that is common to both Hamiltonians$^{4,5}$ and which is revealed clearly in the VBS Hamiltonian. It is interesting to note that a similar situation happens also in the fractional quantum Hall effect$^6$: here the Laughlin wave function which is the exact ground state of an approximate Hamiltonian does possess the hidden order which is revealed in the anyonic gauge. The VBS nature of the ground state of the S=1 spin chain has led to the curious consequence of effective spins S=1/2 at the end of open chains: this has been observed theoretically$^7$ by numerical means and experimentally$^{8,9}$. If we concentrate on the VBS model in the $S=1$ case, it can be written as the sum of a bilinear and a biquadratic spin-spin interaction between nearest neighbors. It is thus natural to study it as a special case of the general bilinear-biquadratic isotropic quantum $S=1$ chain: $$H_\theta = \cos\theta \sum_{i} {\bf S}_i \cdot {\bf S}_{i+1} + \sin\theta \sum_i ({\bf S}_i \cdot{\bf S}_{i+1} )^2 , \eqno(I.1)$$ with $\theta$ varying between 0 and $2\pi$. All energies are measured in units of the global exchange coupling which is omitted everywhere in this paper. The VBS Hamiltonian corresponds to the value $\theta_{vbs}$ with $\tan \theta_{vbs} =1/3$. In this case, each term in the sum in Eq.(I.1) ${\bf S}_i \cdot {\bf S}_{i+1} +({\bf S}_i \cdot{\bf S}_{i+1} )^2 /3$ is the projector on the spin S=2 state of the two neighboring spins $i, i+1$. This fact leads to a simple ground state wavefunction$^2$. The behaviour of this model as a function of $\theta$ has been studied by numerous authors$^{10-30}$. If one increases $\theta$ starting from the bilinear Hamiltonian $\theta =0$ which is known to possess a Haldane gap there is no phase transition till $\theta =\pi/4$ and thus the VBS Hamiltonian ($\theta_{vbs}={\rm atan} (1/3)<\pi/4$) is smoothly connected to the usual bilinear Heisenberg model. However its precise physical meaning has remained so far unexplained. In this paper, we clarify the physical meaning of the VBS point in the phase diagram of the family (I.1) of models. Consider first the classical limit $S\rightarrow\infty$ of Eq.(I.1). For $\theta =0$ the ground state is antiferromagnetically long-range ordered with ordering wavevector $q=\pi$ (Néel state). When $\theta$ is increased, the order becomes incommensurate when $\theta >\theta_c$ with $\tan \theta_c =1/2$: the wavevector shifts from $q=\pi$. As a consequence, the static structure factor $S(q)$ has a delta peak at $q=\pi$ when $\theta <\theta_c$ and a delta peak at $q<\pi$ when $\theta >\theta_c$. Now, in the quantum case $S<\infty$ with S [*integer*]{}, fluctuations wash out long-range order and we are left with short-range order below a characteristic correlation length $\xi$, according to Haldane’s conjecture. The delta peak of $S(q)$ is thus smeared and acquire a finite width given by $\xi^{-1}$. It is important to note that, due to this finite width, the incommensurate behaviour cannot be seen immediately in the quantity $S(q)$ when $q$ shifts away from the commensurate position. This is best understood by considering the analytic structure of $S(q)$ in the complex q-plane. Short-range order means that singularities (poles or branch points) are away from the real axis at a distance $\approx \xi^{-1}$. In the commensurate phase, the real part of the nearest singularity is $\pi$ and the peak of $S(q)$ for q real is also at $q=\pi$. If we increase the parameter $\theta$, at some value the real part of the leading singularity will move away from $q=\pi$: this means that real-space correlations oscillate with a new period. However, due to the width of the peak in the structure factor, the maximum of $S(q)$ remains at $q=\pi$ till the shift $\Delta q$ of the real part reaches a value $O(\xi^{-1})$. Then for larger values of $\theta$, the structure factor will exhibit an incommensurate peak. We have obtained evidence that, right at the VBS point, the correlations become incommensurate in real space: it is a “disorder” point in the language of classical statistical mechanics$^{31}$. For a [*larger*]{} value of $\theta$, the function $S(q)$ exhibits the signature of incommensurability$^{29}$: this point is properly called a Lifshitz point. This splitting of the classical phenomenon at $\theta_c$ is typical of systems with only short-range order. We proceed by first recalling briefly in section II the state of knowledge on the bilinear-biquadratic $S=1$ quantum spin chain. In section III we import the concepts of so-called “disorder” points from classical statistical mechanics. They are discussed by use of simple classical spin models. In section IV, using the Density Matrix Renormalization Group algorithm$^{32}$, we calculate energy gaps, correlation functions and correlation lengths, as well as the $S=1$ string order parameter in the neighborhood of the VBS point. We demonstrate that the spin correlations exhibit a change of behaviour in real space right at the VBS point and that this point is a quantum example of a “disorder” point. Section V contains our conclusions. [**II. PHASE DIAGRAM OF THE BILINEAR-BIQUADRATIC SPIN CHAIN**]{} Let us represent the phase diagram of models $H_\theta$ as in figure 1. For $\theta=0$ and $\theta=\pi$, one finds the isotropic (anti)ferromagnetic quantum Heisenberg model. Antiferromagnetic (resp. ferromagnetic) models corresponds to $-\pi/2<\theta<+\pi/2$ (resp. $+\pi/2<\theta<3\pi/2$). Some points in the phase diagram have been studied in detail and we summarize below the current knowledge: $\bullet$ $\theta=0$: Isotropic antiferromagnetic quantum Heisenberg model: this well-studied model has a non-degenerate disordered ground state with exponentially decaying antiferromagnetic correlations obeying a law $\langle {\bf S}_0 \cdot {\bf S}_n\rangle\approx (-)^n\exp(-n/\xi)/\sqrt{n}$ and a gapped spectrum (Haldane gap $\approx 0.41$). The static structure factor $S(q)$ is a square-root Lorentzian peaked at $q=\pi$. $\bullet$ $\theta=0.1024\pi$ ($\tan\theta = 1/3$): this is the VBS model with exact valence-bond-solid ground state. The spin correlations are [*purely*]{} exponential with a correlation length $\xi = (\ln 3)^{-1} \approx 0.91$. There is a gap in the spectrum ($\Delta = 0.664$). $\bullet$ $\theta=0.25\pi$: this is the Lai-Sutherland model, the Hamiltonian is a sum of permutation operators and exactly integrable by the Bethe ansatz$^{10,11}$. The ground state is unique and the model is critical. The corresponding conformal theory is $SU(3)_{k=1}$. There are zero-energy modes for $q=0, \pm 2\pi/3$. $\bullet$ $\theta=-0.25\pi$: The model is solvable exactly by the nested Bethe ansatz$^{12,13}$. One finds a critical system with a unique ground state. The conformal theory is $SU(2)_{k=2}$. There are zero-energy modes at $q=0,\pi$. $\bullet$ $\theta=-0.50\pi$: the physics is that of a dimerized state; the order parameter is given by the coefficient $c_2$ in the singlet-singlet correlation: $$\langle ({\bf S}_i {\bf S}_{i+1}) ({\bf S}_j {\bf S}_{j+1}) \rangle \rightarrow c_1 +(-1)^{i-j}c_2 , \eqno(II.1)$$ for $|i-j|\rightarrow\infty$. The ground state is twice degenerate in the thermodynamic limit and the spectrum is gapped ($\Delta=0.17$). The correlation length is given as $\xi=42.2$: these are exact results$^{18-21}$. $\bullet$ $\theta=-0.75\pi$: possible location of a continuous phase transition from a ferromagnetic to a dimerized phase$^{23,25,26}$. $\bullet$ $\theta=\pi$: This is the isotropic ferromagnetic Heisenberg model. There is ferromagnetic order with gapless excitations. The ground state is the ferromagnetic state for $\pi/2<\theta <5\pi /4$. With these points, one constructs the following phase diagram$^{14,17}$: Starting at $\theta=\pi$, one finds an ordered ferromagnetic state without gap. The ferromagnetic phase terminates at $\theta=-0.75\pi$. A continuous phase transition leads to a dimerized state. A prediction by Chubukov$^{22}$ of a non-dimerized nematic phase seems refuted by Fath and Solyom$^{25,26}$. In the dimerized phase, the ground state is a singlet with a double degeneracy due to a $Z_2$ symmetry breaking. The order parameter is given by $c_2$ in the correlation function (II.1). A continuous phase transition at $\theta=-0.25\pi$ leads to a Haldane phase, with a unique disordered ground state, exponentially decaying correlations and a gapped spectrum. This gapped phase ends at the Lai-Sutherland point $\theta =+0.25\pi$ where a continuous transition leads to a phase which is possibly trimerized (see refs. 25,26 for a detailed discussion). One is back to ferromagnetic phase for $\theta =+0.5\pi$. This phase diagram is displayed in figure 1. Up to now, the VBS point appears to be generic in the Haldane phase. Recently, Bursill, Xiang and Gehring$^{29}$ considered, using the DMRG, the Fourier transform of the spin-spin correlations, i.e. the static structure factor: $$S(q)=\sum_n {\rm e}^{iqn}\langle {\bf S}_n \cdot {\bf S}_0 \rangle , \eqno(II.2)$$ In the Haldane phase, at the isotropic point $\theta =0$, $S(q)$ is a square-root Lorentzian with a peak at $q=\pi$. Since parity is unbroken in the phases we discuss, we restrict the momenta to the interval $(0,\pi)$. It was found that the peak of the Fourier transform $S(q)$ starts to move away from $q=\pi$ towards $q=\pm 2\pi/3$. This happens at $\tan \tilde{\theta} = 0.43806(4)$, or $\tilde{\theta}=0.1314\pi$. They found also that when $\theta\rightarrow 0.25\pi$, the peak reaches $2\pi /3$ in agreement with the period-3 zero modes that are seen at the Lai-Sutherland point. Their conclusion is then that there are three regions between $\theta=0$ and $\theta=0.50\pi$: $\bullet$ $0 < \theta < \tilde{\theta}$: short-ranged antiferromagnetic correlations. $\bullet$ $\tilde{\theta} < \theta < 0.25\pi$): short-ranged [*spiral*]{} order; the peak of the Fourier transform shifts from $q=\pi$ to $q=2\pi/3$; the spectrum is still gapped. $\bullet$ $0.25\pi\leq \theta < 0.50\pi$: possible trimerized phase beyond the Lai-Sutherland phase transition. If one considers the classical limit of model (I.1), there is a related phenomenon. For $\theta$ smaller than $\theta_c=\arctan (1/2)=0.148\pi$, the ground state is the usual commensurate Néel order with wavevector $q=\pi$, while beyond this value of $\theta$ the ground state becomes an incommensurate spiral characterized by wavevector $q$ such that $\cos q=-{1\over 2} {\rm cotan} \theta$. Of course the classical ground states have long-range ordering and when going to finite spin values this order becomes short range. We will show in section IV that this short-range order naturally splits the commensurate-incommensurate transition in [*two*]{} distinct phenomena: one happens at $\theta_{vbs}$ where the spin oscillations becomes incommensurate in real space and one happens at $\tilde\theta$ where incommensurability becomes obvious in the structure factor. Before discussing our results, we now recall the corresponding concepts of classical statistical physics, first developed by Stephenson$^{31}$. [**III. SHORT-RANGE ORDER AND “DISORDER” POINTS**]{} If one starts from a classical model and considers finite integer spins then, according to Haldane’s conjecture, there is only short-range order and a finite correlation length. This is, roughly speaking, an example of “quantum paramagnetism”. Finite spin is in a sense equivalent to a finite temperature. In Haldane’s mapping$^1$ onto a nonlinear sigma model, the coupling constant is equal to the inverse of the spin while in the nonlinear sigma model describing classical two-dimensional systems the coupling is the temperature itself. This means that an integer spin chain has a physics which is related to that of a two-dimensional spin system at nonzero temperature. Since there is no long-range ordering in such a two-dimensional system according to the Mermin-Wagner theorem, the ground state of the spin chain is short-range ordered. We consider thus classical systems in their paramagnetic phase to understand the physics of finite-spin chains. Strictly speaking, one should consider classical two-dimensional systems but in fact, for our purposes, the physics is absolutely similar to that of three-dimensional systems above the critical temperature. Let us consider a magnetic Hamiltonian that exhibits two ordered low-temperature phases, one with commensurate correlations and the other with incommensurate correlations. One may think for example of a square lattice of classical spins with nearest-neighbor exchange $J_1$ and third-nearest-neighbor $J_3$: when $J_3/J_1>1/8$ one destabilizes the Néel order and obtains an incommensurate spiral whose pitch evolves continuously. We note $P$ any parameter that controls the zero-temperature phase transition (e.g. anisotropy, pressure, ratio of exchange couplings, etc.). A generic phase diagram is given in figure 2. We consider the case where these low-temperature phases are separated from the disordered paramagnetic high-temperature phase by [*continuous*]{} transitions (If the classical system is 2D then the $T_c$ is zero and the reasoning is unchanged). It is clear that the short-ranged correlations in the disordered phase will be of variable nature: “close” to the commensurate phase, they will be commensurate; “close” to the incommensurate phase, they will be incommensurate. One can guess that there will be a line in this phase diagram, where the correlations change their behaviour; this change will be linked to correlations of very short range, thus to a state with a minimum of short range order. Hence the name of disorder line. If one moves along path (A) in the paramagnetic phase in figure 2, there should be a change in the correlations. If one considers the real-space spin-spin correlations along path (A) they will develop incommensurate oscillations at some point $A_D$. If one considers now the correlations in Fourier space $S(q)$, one finds that the peaks of the Fourier transform still stays at the value for commensurate correlations, even though the real-space correlations are already incommensurate, due to the finite correlation length: the peak width is linked to $\xi^{-1}$. It is only “closer” to the incommensurate phase that the peak will start to shift. This will happen at a second point $A_L$ on path (A) in fig.2. This is easy to understand by taking a simplified form for $S(q)$: $$S(q)= {1\over \alpha (q-q_x)^2 +(q-q_x)^4 +\xi^{-2}}. \eqno(III.1)$$ For convenience, we shift the momenta to set $q_x =0$. It is only for $\alpha <0$ that $S(q)$ has a double-peak structure. When $\alpha >0$ is large enough, $\alpha^2 >4\xi^{-2}$, all the poles of (III.1) are on the imaginary axis in complex-q space and the real-space correlations do not oscillate. But when $\alpha^2 <4\xi^{-2}$ the poles have a real part and thus there are real-space oscillations. It is only when the real part of the poles is large enough that the structure factor itself displays a two-peak shape. This effect is entirely due to the finite correlation length $\xi$ i.e. short-range ordering and finite width of the peak in $S(q)$. It is clear from the simple example above that it is only the analytic structure of $S(q)$ that matters and not our peculiar eq.(III.1). As such, this is a general behaviour. The starting point of real-space oscillations is the disorder point. It extends in the plane of Fig.2 in a disorder line D. The starting point for the double peak structure in $S(q)$ is the Lifshitz point, extending in a line L in fig.2. In experiments, one normally measures the structure factor in reciprocal space and will thus observe this line. It is also clear that the two lines must end in the multicritical point where the three phases meet, which is thus necessary for their existence. To avoid unnecessary generalizations, we take the results from an example treated by an RPA method in Ref.33. It is an Ising spin chain with a ferromagnetic $J_1$ interaction between nearest neighbors and an antiferromagnetic $J_2$ between next-to-nearest neighbors. The RPA treatment shows that there are three regimes. The disorder temperature is: $$T_D = {J_1^2}/{4|J_2|} + 2|J_2|. \eqno(III.2)$$ One derives the following expressions for $x$ large: $T < T_D$: $\langle S(0)S(x) \rangle \simeq e^{-x/\xi_-(T)}$ . $T = T_D$: $\langle S(0)S(x) \rangle \simeq x {\rm e}^{-k_0x}$ with $\cosh k_0a = J_1/4|J_2|$. $T > T_D$: $\langle S(0)S(x) \rangle \simeq {\rm e}^{-x/\xi_+(T)} \cos (q(T)x)$ with $q(T) \sim (T-T_D)^{1/2}$. For $\xi_{\pm}(T)$ one finds that, on the commensurate side, the correlation length exhibits an infinite derivative at $T_D$; it will typically be very small, but not necessarily a minimum or zero. The derivative on the incommensurate side is finite (see figure 3). This characterizes a disorder line of the first kind. There are two more special properties: (i) The susceptibility shows a particularly simple form at the disorder line. (ii) If one considers the correlation functions and compares them to an Ornstein-Zernicke correlation function (for $x$ large) for a $d$-dimensional system: $$\langle S(0) S(x) \rangle \simeq e^{-x/\xi(T)}/r^{(d-1)/2} , \eqno(III.3)$$ one sees that the correlation functions are those for $d=1$, as expected for a chain, except at the disorder point: formally, they correspond to $d=-1$. Let us add that the incommensurate correlations are given by a wave vector $q$ which shifts continuously from the commensurate value $q=0$; the exponent $1/2$ is however non-universal. In the spin chain problem with S=1, we are always in the paramagnetic phase i.e. we are following a path like (A) in fig.2 when varying the parameter $\theta$ in the model (I.1). We thus expect to cross the disorder point and the Lifshitz point that are the quantum remnants of the classical transition at $\theta_c$. [**IV. THE NEIGHBORHOOD OF THE VBS POINT**]{} For our calculations, we use the DMRG: see Ref.32 for a detailed discussion of the algorithm. We apply it to chains of a length $L=96$ and keep $M=80$ states. This is sufficient to find truncation errors smaller than $10^{-12}$ in the considered region. It is therefore not necessary to extrapolate results in $M$, as they are extremely close to the exact results. For the VBS point, we recover the exact results within machine precision. Due to the very small correlation lengths (typically smaller than 3) a length $L=96$ is sufficient to obtain the results of the thermodynamic limit. From the DMRG viewpoint, this situation is ideal. There is however a problem with purely computational errors: the spin-spin correlations are, for a distance of 30 to 40 sites, of the order $10^{-13}$ or less. As they are obtained by summing small numbers below machine precision (in a REAL\*8 calculation), they must be rejected. To judge the importance of this effect, we have adopted the following strategy: as the system under study is isotropic and disordered in the region where the Haldane and the trimerized phase meet, the spin-spin correlations must obey the relation $$\langle S^+_iS^-_j \rangle = \langle S^x_iS^x_j + S^y_iS^y_j \rangle = 2 \langle S^z_i S^z_j \rangle. \eqno(IV.1)$$ These two quantities are calculated independently; we reject correlations that show a deviation of more than a thousandth from this relation. As a matter of fact, we find that the correlations $\langle S^z_i S^z_j \rangle$ reach a minimum value which oscillates randomly around $10^{-14}$, whereas $\langle S^+_i S^-_j \rangle$ continues to diminish regularly. We conclude that the values for the latter correlation are more precise; analyzing the calculation, we find that for the latter all the important contributions and weights show the same sign, whereas it changes for the former. The key of our analysis is not to analyze the Fourier transform of the correlations, but to analyze them directly in real space. We will thus show that the VBS point, so far without special role in the phase diagram, is effectively a disorder point. It shows all the characteristics of a disorder point of the first kind, as described in section III. Consider the real-space correlations (figure 4) for some values of $\theta$ between $0.10\pi$ and $0.125\pi$. This includes the VBS point ($\theta_{vbs}=0.1024\pi$). The correlations for $\theta<0.1024\pi$ are perfectly antiferromagnetic; this is most evident in a logarithmic plot of $|(-1)^{n}\sqrt{n} \langle {\bf S}_n \cdot {\bf S}_0\rangle|$. For antiferromagnetic correlations, the curve shows no modulations. Above $\theta_{vbs} = 0.1024\pi$, the logarithmic plots show oscillations with periods that become shorter for increasing $\theta$, to end at a period of 3 for $\theta\rightarrow 0.25\pi$. Thus the VBS point is a disorder point. The correlations are no longer antiferromagnetic, but already incommensurate: this point is missed if one considers only the Fourier transform as in Ref.29. These modulations can be understood from the classical law for an incommensurate high-temperature phase (adapted to $d=2$): $$\langle {\bf S}_n \cdot {\bf S}_0\rangle \approx \cos [q(\theta)n] {{{\rm e}^{-n/\xi(\theta)}}\over {\sqrt{n}} }= (-1)^n \cos [(\pi-q)n]{ {{\rm e}^{-n/\xi(\theta)}}\over {\sqrt{n}}}. \eqno(IV.2)$$ The modulations should thus show a period $\pi/(\pi-q)$, which is easier to see than the period originating directly from $\cos qn$. To show that the correlation functions can be well described by (IV.2), we have attempted a direct fit of our results. This fit is complicated by the fact that there are effectively three parameters to be controlled, the wave vector $q$, the correlation length $\xi$ and also a phase factor $\phi$, from replacing $n=i-j$ by $(n-\phi)$ in the argument of the cosine. We find that the fit is extremely sensitive to the parameter values, which allows for a good fit. As an example, we take $\theta=0.115\pi$, sensibly below the point given so far for the change of the correlations. We obtain the fit shown in figure 5, for $\phi=0.65$, $\xi=1.08$ and $\pi-q=0.198\pi$. The wave vector $q$ has already shifted by 20 percent from the antiferromagnetic value $q=\pi$. For $\theta$ closer to the VBS point, the fit is made more complicated by the errors in the correlation function due to the finite precision of the computer. To estimate the behaviour $q(\theta)$, we therefore consider the periodicity: For a period $p$, $\pi - q \approx \pi/p$. The discrete nature of the problem limits this approach. We find the behaviour shown in figure 6. Clearly, is is compatible with $q \propto (\theta-\theta_{vbs})^{\sigma}$ with $0<\sigma<1$, the behaviour of a disorder point of the first kind. The curve would be compatible with $\sigma \approx 1/2$; but we are in no position to give a precise estimate. We have also calculated the Fourier transform of the correlation function and find results in agreement with those given by Bursill et al.$^{29}$. The VBS point has no special significance for its behaviour; the point $\tilde{\theta}=0.1314\pi$ where the peak starts to shift, can now be identified as a Lifshitz point: see fig.7. As expected in a system with short-range order, it is distinct from the disorder (VBS) point. For the correlation lengths, we find that they show a minimum for the VBS point, with an infinite slope (numerically: very large) for $\xi$ in the commensurate regime ($\theta < 0.1024\pi$), and a slow increase in the incommensurate regime with a finite slope, given in figure 8. The correlation lengths have been found by different methods: in the regime $\theta < \theta_{vbs}$, we have compared the spin-spin correlations numerically and graphically to a law $\exp (-n/\xi) /\sqrt{n}$, which were in all cases in good agreement. We estimate the precision of the results of the order of 1 per cent: the truncation errors are of the order $10^{-13}$, a serious underestimation can therefore be excluded. For $L=96$ and $\xi \approx 1 - 2$, finite size effects are of no importance. The situation is more complicated for the regime $\theta > \theta_{vbs}$. As we have seen, the fit of the theoretically expected behaviour to the found curve is rather complex. If one considers a plot of $|(-1)^{n}\sqrt{n} \langle {\bf S}_i \cdot {\bf S}_j \rangle|$ (figure 4), one finds that a linear fit for the maxima is quite good. This is stable in the sense that a factor $\cos qx$ influences the logarithm least when it is close to 1. Very generously estimated, the error of the graphical evaluation should be below 5 per cent. We estimate that for $\theta > 0.15\pi$ the underestimation due to a non-negligible truncation error dominates. For $\theta_{vbs} < \theta \leq 0.11\pi$, we could not obtain the correlation length: on the one hand, the periods due to the incommensurability are too long to separate them well from the exponential behaviour. On the other hand, neither a fit $\exp -r/\xi$ nor $\exp (-r/\xi)/\sqrt{r}$ is satisfactory. We don’t know whether this observation is due to the crossover of the behaviour of the correlation function or simply due to problems of the numerical method. In any case, the results indicate strongly an extrapolation to the VBS point with a finite slope. The other important quantities are the gap and the string order parameter $O_{\pi}(i,j)=\langle S^z_i \exp (i\pi \sum_{k=i+1}^{j-1} S^z_k) S^z_j\rangle$. The gap shows a maximum for $\theta \approx 0.123\pi$, which is thus linked neither to the disorder nor the Lifshitz point: see figure 9. For the transition points $\theta=\pm 0.25\pi$ we have not obtained serious estimates: the critical fluctuations imply a greater $M$, and the vanishing gap is difficult to see. Our results are well compatible with a zero gap, but not precise enough to give a serious estimate. The point $\theta=-0.20\pi$ is sufficiently close to the transition to cause the same problem. For the VBS point, we find a gap value $\Delta=0.664$, in agreement with Ref.25. Since the gap is smooth at $\theta_{vbs}$, this implies that the spin wave velocity has a singularity at this point ($c=\Delta\xi$). The string order parameter in the thermodynamic limit $i-j\rightarrow\infty$ has its extremum for the VBS point: $|O_{\pi}| = 4/9$: see figure 10. One sees that the VBS point, though a point of minimal [*spin*]{} order, is a point of maximum hidden topological order. To complete the identification of the VBS point as a disorder point, we note that there is the equivalent of the particularly simple form of the susceptibility of the classical model of section III. The exact correlations at the VBS point obey a one dimensional Ornstein-Zernicke form: they are purely exponential (no prefactor), whereas the non-linear sigma model yields two-dimensional correlation laws. This “dimensional reduction” is accompanied by a particularly simple form of the Hamiltonian: it can be decomposed into a sum of local projection operators. The problem loses its quantum character and turns into a classical one-dimensional problem. [**V. CONCLUSION**]{} We have shown that the VBS point$^2$ is a disorder point in the sense of classical statistical mechanics. It is thus identified as a point which is not just by chance exactly solvable, but shows this property for more profound physical reasons. The results of Bursill et al. who have considered the Fourier transform of the spin-spin correlations fit naturally in the picture we have developed: the point they identify as the point where correlations change is simply the Lifshitz point following the definition given above. Quantum fluctuations in the integer spin chain wash out long-range order. As a consequence, the transition that happens at the classical level for $\theta_c$ is no longer a phase transition. However the change of short-range correlations still happens in the quantum system at the VBS point in real-space. Due to the finite correlation length this is not seen immediately in $S(q)$, hence the Lifshitz point which is distinct. Since $\xi\rightarrow\infty$ when the spin increases one expects that these two points should merge in the classical limit right at $\theta_c$. This is summarized in figure 11. This implies the following description of the phase diagram: There are three regions for $0 < \theta < 0.25\pi$: $0 < \theta < \theta_{vbs}$: Short-range antiferromagnetic correlations; Haldane phase with gapped spectrum. A generic description is given by the VBS model or also the isotropic antiferromagnetic Heisenberg model. $\theta_{vbs} < \theta < \tilde{\theta}$: Incommensurate short-range correlations with a wave vector $q<\pi$, that shifts away from $\pi$ as $\pi - q \propto (\theta-\theta_{vbs})^{\sigma}$, $\sigma \approx 1/2$. In the Fourier transform, the peak stays at $q = \pi$. The spectrum is gapped: we expect that the low-lying Haldane modes are now at the incommensurate wavevector. Part of the VBS physics remains valid: the gap, the hidden order, and the free spins 1/2 at the ends of an open chain. $\tilde{\theta} < \theta < 0.25\pi$: The physics is similar to that of the former region; the peak of the Fourier transform shifts from $q = \pi$ to $q=\pm 2\pi/3$ and the incommensurate correlations become visible. The spectrum is gapped. There is however no profound physical difference between this region and the former one. The above picture does not challenge conventional wisdom in the sense that the usual Heisenberg Hamiltonian $\theta =0$ shares the same physics with the VBS Hamiltonian and there are no additional phase transitions between the point$^{12,13}$ $\theta =-0.25\pi$ and the Lai-Sutherland point $\theta =+0.25\pi$. However the VBS point itself means that the physics has changed beyond $\theta_{vbs}$: due to the incommensurate correlations, it is no longer possible to capture the low-lying excitations by a nonlinear sigma model with symmetry breaking pattern $O(3)/O(2)$ as used originally by Haldane. In the incommensurate regime, the full rotation group is broken down. Appropriate nonlinear sigma models have been contemplated before$^{34}$. Since they involve a nonabelian symmetry, they are generically massive, consistent with the gapped nature of the Haldane phase. There is an interesting relationship$^{6}$ with the fractional Quantum Hall effect. The present understanding$^{35}$ of the physics of the FQHE at filling $\nu =1/m$, m odd, is based on Laughlin’s wavefunction $\psi_m$ which is the exact ground state of a truncated Hamiltonian. In addition, this function embodies the hidden long-range order that is revealed in the anyonic gauge. This is similar to the situation of the VBS wavefunction. Expectation values computed with the Laughlin wavefunction correspond to a classical statistical problem which is the two-dimensional one-component plasma (2D OCP): this is Laughlin’s plasma analogy. The case of the full Landau level $\psi_1$ corresponds to the special point$^{36}$ $\Gamma =2$ of the 2D OCP ($\Gamma$ being the ratio of the squared electric charge to the temperature) at which the density correlations begin to oscillate in real space, a precursor phenomenon of the crystallization that occurs at $\Gamma \approx 140$ when the plasma is dilute enough. This special point has also some of properties expected from a disorder point$^{36}$. This is similar to what we observe at the VBS point. [**Acknowledgements**]{} It is a pleasure to thank N. Elstner and O. Golinelli for fruitful discussions. **FIGURE CAPTIONS** [**Figure 1**]{}: Phase diagram of the bilinear-biquadratic $S=1$ isotropic quantum spin chain as a function of $\theta$. Solid lines: transition points; dashed lines: other special points. H(AFM): isotropic antiferromagnet; VBS: the valence-bond-solid model; L: the crossover point studied in Ref.29; H (FM): isotropic ferromagnet. [**Figure 2**]{}: Schematic phase diagram: a disordered high-$T$ phase is linked by two continuous transitions to two ordered low-$T$ phases. $P$ is a parameter which controls the nature of the ground state. The dashed lines represent the disorder (D) and Lifshitz (L) lines, where the behaviour of the correlations changes in real and in Fourier space respectively. [**Figure 3**]{}: Correlation length at a disorder point of the first kind (schematical from RPA): the commensurate phase is to the left. [**Figure 4**]{}: Real-space spin-spin correlations as a function of the distance, for several values of $\theta$ below and above the VBS point (at $\theta_{vbs} =0.1024\pi$). The modulations appear above the VBS point. Here $K_n ={\rm ln} \left[|(-)^n \sqrt{n}\langle {\bf S}_0\cdot {\bf S}_n\rangle |\right]$. [**Figure 5**]{}: Comparison between the spin-spin correlations predicted (solid line) and calculated numerically (squares + dashed line) for $\theta=0.115\pi$, above $\theta_{vbs}$, but below $\tilde{\theta}$. The dotted line is $(-)^n \sqrt{n} \exp (n/\xi) \langle S^z_0 S^z_n\rangle$ and the solid line is $\cos \left[(n-\phi)(\pi-q)\right]$. [**Figure 6**]{}: Wave vector $\pi - q$ characteristic of the spin correlations. There is a singularity at the VBS point consistent with the identification of a disorder point. [**Figure 7**]{}: The Fourier transform $S(q)$ as a function of momentum for various values of $\theta$. The two-peak structure appears only for $\theta>\tilde\theta =0.1314\pi$ while nothing is seen at $\theta_{vbs}$. [**Figure 8**]{}: Correlation lengths for various values $\theta$. The minimum is at the VBS point. [**Figure 9**]{}: Gaps in the Haldane phase as a function of $\theta$. The maximum value is between the VBS point and the Lifshitz point. [**Figure 10**]{}: String order parameter $|O_{\pi}|$ in the thermodynamic limit in the Haldane phase. The maximum is exactly at the VBS point. [**Figure 11**]{}: Schematic phase diagram of integer spin chains with biquadratic coupling: model (I.1). Here the inverse spin plays the role of a temperature: the S=1 case corresponds to short-range ordered phases. Of course this picture may be altered by other types of ordering such as dimerization. **REFERENCES** [\[1\]]{}F. D. M. Haldane, Phys. Rev. Lett. [**50**]{}, 1153 (1983); Phys. Lett. A[**93**]{}, 464 (1983). [\[2\]]{}I. Affleck, T. Kennedy, E.H. Lieb and H. Tasaki, Phys. Rev. Lett.[**59**]{}, 799 (1987); Comm. Math. Phys. [**115**]{}, 477 (1988). [\[3\]]{}M. den Nijs and K. Rommelse, Phys. Rev. [**B 40**]{}, 4709 (1989). [\[4\]]{}T. Kennedy, J. Phys. C: Cond. Matter, C[**2**]{}, 5737 (1990). [\[5\]]{}T. Kennedy and H. Tasaki, Phys. Rev. B[**45**]{}, 304 (1992). [\[6\]]{}S. M. Girvin and D. P. Arovas, Phys. Scr. T[**27**]{}, 156 (1989). [\[7\]]{}S. Yamamoto and S. Miyashita, Phys. Rev. B[**48**]{}, 9528 (1993). [\[8\]]{}M. Hagiwara, K. Katsumata, I. Affleck, B. I. Halperin and J. P. Renard, Phys. Rev. Lett. [**65**]{}, 3181 (1990); S. H. Glarum, S. Geshwind, K. M. Lee, M. L. Kaplan and J. Michel, [*ibid.*]{} [**67**]{}, 1614 (1991). [\[9\]]{}O. Avenel, J. Xu, J. S. Xia, M. F. Xu, B. Andraka, T. Lang, P. L. Moyand, W. Ni, P. J. C. Signore, C. M. van Woerkns, E. D. Adams, G. G. Ihas, M. W. Meisels, S. E. Nagler, N. S. Sullivan, Y. Takano, D. R. Talham, T. Goto and N. Fujiwara, Phys. Rev. B[**46**]{}, 8655 (1992). [\[10\]]{}G. V. Uimin, JETP Lett. [**12**]{}, 225 (1970). [\[11\]]{}C. K. Lai, J. Math. Phys. [**15**]{}, 1675 (1974); B. Sutherland, Phys. Rev. B[**12**]{}, 3795 (1975). [\[12\]]{}L. A. Takhtajan, Phys. Lett. [**87A**]{}, 479 (1982). [\[13\]]{}H. M. Babudjian, Phys. Lett. [**90A**]{}, 479 (1982); Nucl. Phys. B[**215**]{}, 317 (1983). [\[14\]]{}I. Affleck, Nucl. Phys. B[**265**]{}, 409 (1986). [\[15\]]{}J. Oitmaa, J. B. Parkinson and J. C. Bonner, J. Phys. C[**19**]{}, L595 (1986). [\[16\]]{}J. Sólyom, Phys. Rev. B[**36**]{}, 8642 (1987). [\[17\]]{}N. Papanicolaou, Nucl. Phys. B[**305**]{} \[FS23\], 367 (1988). [\[18\]]{}J. B. Parkinson, J. Phys. C[**20**]{}, L1029 (1987); [*ibid.*]{} C[**21**]{}, 3793 (1988) [\[19\]]{}M. N. Barber and M. T. Batchelor, Phys. Rev. B[**40**]{}, 4621 (1989). [\[20\]]{}A. Klümper, Europhys. Lett. [**9**]{}, 815 (1989); J. Phys. A[**23**]{}, 809 (1990). [\[21\]]{}E. S. Sørensen and A. P. Young, Phys. Rev. B[**42**]{}, 754 (1990). [\[22\]]{}A. V. Chubukov, J. Phys. Condens. Matter [**2**]{}, 1593 (1990); Phys. Rev. B[**43**]{}, 3337 (1991). [\[23\]]{}G. Fáth and J. Sólyom, Phys. Rev. B[**44**]{}, 11836 (1991). [\[24\]]{}R. F. Bishop, J. B. Parkinson and Y. Xian, J. Phys. Condens. Matter [**5**]{}, 9169 (1993). [\[25\]]{}G. Fáth and J. Sólyom, Phys. Rev. B[**47**]{}, 872 (1993). [\[26\]]{}G. Fáth and J. Sólyom, J. Phys. Condens. Matter [**5**]{}, 8983 (1993). [\[27\]]{}Y. Xian, J. Phys. C[**5**]{}, 7489 (1993). [\[28\]]{}T. Xiang and G. Gehring, Phys. Rev. B[**48**]{}, 303 (1993). [\[29\]]{}R. J. Bursill, T. Xiang and G. Gehring, J. Phys. A[**28**]{}, 2109 (1995). [\[30\]]{}G. Fáth and J. Sólyom, Phys. Rev. B[**51**]{}, 3620 (1995). [\[31\]]{} J. Stephenson, Can. J. Phys. [**47**]{}, 2621 (1969); ibid. [**48**]{}, 1724 (1970); ibid. [**48**]{}, 2118 (1970); J. Math. Phys. [**12**]{}, 420 (1970). [\[32\]]{} S. R. White, Phys. Rev. Lett. [**69**]{}, 2863 (1992); Phys. Rev. B[**48**]{}, 10345 (1993); S. R. White and D. A. Huse, Phys. Rev. B[**48**]{}, 3844 (1993). [\[33\]]{} T. Garel and J. M. Maillard, J. Phys. C: Solid State Phys. [**19**]{}, L505 (1986). [\[34\]]{} T. Dombre and N. Read, Phys. Rev. B[**39**]{}, 6797 (1989). [\[35\]]{} See the articles by R. B. Laughlin and S. M. Girvin in “The Quantum Hall Effect”, R. E. Prange and S. M. Girvin editors, Springer-verlag, New York, 1990. [\[36\]]{} B. Jancovici, Phys. Rev. Lett. [**46**]{}, 386 (1981). plus 2pt minus 2pt **PHYSICAL MEANING OF THE AFFLECK-KENNEDY-LIEB-TASAKI** **S=1 QUANTUM SPIN CHAIN** U. Schollwöck, Th. Jolicœur and T. Garel *Service de Physique Théorique* *C.E. Saclay* *F-91191 Gif-sur-Yvette, France* 1.0in Submitted to: [*Physical Review B*]{} 4.0in May 1995 PACS No: 75.40.Mg, 75.10.Jm. SPhT/95-054 = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm = 15.0truecm

ArXiv

--- author: - 'G. Cataldi' - 'A. Brandeker' - 'G. Olofsson' - 'C. H. Chen' - 'W. R. F. Dent' - 'I. Kamp' - 'A. Roberge' - 'B. Vandenbussche' bibliography: - 'bibliography.bib' date: 'Received; accepted' subtitle: 'Can gas-dust interactions explain the belt’s morphology?' title: 'Constraints on the gas content of the Fomalhaut debris belt[^1]' --- [The 440Myr old main-sequence A-star Fomalhaut is surrounded by an eccentric debris belt with sharp edges. This sort of a morphology is usually attributed to planetary perturbations, but the orbit of the only planetary candidate detected so far, Fomalhaut b, is too eccentric to efficiently shape the belt. Alternative models that could account for the morphology without invoking a planet are stellar encounters and gas-dust interactions.]{} [We aim to test the possibility of gas-dust interactions as the origin of the observed morphology by putting upper limits on the total gas content of the Fomalhaut belt.]{} [We derive upper limits on the 158$\mu$m and 63$\mu$m emission by using non-detections from the Photodetector Array Camera and Spectrometer (PACS) onboard the *Herschel Space Observatory*. Line fluxes are converted into total gas mass using the non-local thermodynamic equilibrium (non-LTE) code <span style="font-variant:small-caps;">radex</span>. We consider two different cases for the elemental abundances of the gas: solar abundances and abundances similar to those observed for the gas in the $\beta$ Pictoris debris disc.]{} [The gas mass is shown to be below the millimetre dust mass by a factor of at least $\sim$3 (for solar abundances) respectively $\sim$300 (for $\beta$ Pic-like abundances).]{} [The lack of gas co-spatial with the dust implies that gas-dust interactions cannot efficiently shape the Fomalhaut debris belt. The morphology is therefore more likely due to a yet unseen planet (Fomalhaut c) or stellar encounters.]{} Introduction ============ Fomalhaut is one of the best studied examples out of several hundred main-sequence stars known to be surrounded by dusty discs commonly known as *debris discs*. The presence of circumstellar material around this nearby [ 7.7pc; @vanLeeuwen_2007], $440\pm40$Myr old [@Mamajek_2012] A3V star was first inferred by the detection of an infrared excess above the stellar photosphere due to thermal emission from micron-sized dust grains [@Aumann_1985]. The dusty debris belt around Fomalhaut has been resolved at infrared wavelengths with *Spitzer* [@Stapelfeldt_etal_2004] and the *Herschel* Photodetector Array Camera and Spectrometer (PACS) [@Acke_etal_2012]. The dust is thought to be derived from continuous collisions of larger planetesimals or cometary objects [@Backman_Paresce_1993]. To trace these parent bodies, observations at longer wavelengths sensitive to millimetre-sized grains are used. Millimetre-sized grains are less affected by radiation pressure and thus more closely follow the distribution of the parent bodies. @Holland_etal_1998 were the first to resolve the Fomalhaut belt in the sub-millimetre. Recent data from the Atacama Large Millimeter/submillimeter Array (ALMA) show a remarkably narrow belt with a semi-major axis $a\sim140$AU [@Boley_etal_2012 hereafter ]. The belt has also been observed in scattered star light with the *Hubble Space Telescope* [@Kalas_etal_2005; @Kalas_etal_2013]. Two key morphological characteristics emerging from the @Kalas_etal_2005 observations are i) the ring’s eccentricity of $e\sim0.1$ and ii) its sharp inner edge. Both features have been seen as evidence for a planet orbiting the star just inside the belt [e.g. @Quillen_2006]. A candidate planetary body (Fomalhaut b) was detected [@Kalas_etal_2008], but subsequent observations showed that its orbit is highly eccentric ($e=0.8\pm0.1$), making it unlikely to be responsible for the observed morphology [@Kalas_etal_2013; @Beust_etal_2014; @Tamayo_2014]. There remains the possibility that a different, hereto unseen planet is shaping the belt; infrared surveys were only able to exclude planets with masses larger than a Jupiter mass [@Kalas_etal_2008; @Marengo_etal_2009; @Janson_etal_2012; @Janson_etal_2014]. Actually, the extreme orbit of Fomalhaut b might be a natural consequence of the presence of an additional planet in a moderately eccentric orbit [@Faramaz_etal_2015]. Alternatively, the observed belt eccentricity may be caused by stellar encounters [e.g.  @Larwood_Kalas_2001; @Jalali_Tremaine_2012]. In particular, Fomalhaut is part of a wide triple system. @Shannon_etal_2014 recently showed that secular interactions or close encounters with one of the companions could result in the observed belt eccentricity. We focus on yet another candidate mechanism to explain the observed morphology. It is known that gas-dust interactions can result in a clumping instability, organising the dust into narrow rings [@Klahr_Lin_2005; @Besla_Wu_2007]. @Lyra_Kuchner_2013 recently presented the first 2D simulations of this instability and found that some rings develop small eccentricities, resulting in a morphology similar to that observed for the Fomalhaut debris belt. The mechanism obviously requires the presence of gas beside the dust (typically, a dust-to-gas ratio of $\epsilon\lesssim1$ is required), but we know from objects such as $\beta$ Pictoris [e.g. @Olofsson_etal_2001] or 49 Ceti [e.g. @Hughes_etal_2008] that debris discs are not always devoid of gas. We test the applicability of the clumping instability to the case of Fomalhaut by searching for 158$\mu$m and 63$\mu$m gas emission using *Herschel* PACS. We do not detect any of the emission lines and use our data to put stringent upper limits on the gas content of the Fomalhaut debris belt. Observations and data reduction =============================== Fomalhaut was observed using PACS [@Poglitsch_etal_2010] onboard the *Herschel Space Observatory* [@Pilbratt_etal_2010]. The integral field unit PACS consists of 25 spatial pixels (spaxels), each covering $9.4\arcsec\times9.4\arcsec$ on the sky. We used PACS in line spectroscopy mode (PacsLineSpec) to observe the 158$\mu$m and the 63$\mu$m line regions (observation IDs 1342257220 and 1342210402 respectively). We observed in chop/nod mode. Figure \[PACS\_CII\_lineemission\] shows the placement of the PACS spaxels relative to the Fomalhaut debris belt in the case of the observations. We reduced the data using the “background normalisation” pipeline script within the *Herschel* Interactive Processing Environment (HIPE) version 12.0 [@Ott_2010]. The background normalisation pipeline is recommended for faint sources or long observations since it corrects more efficiently for detector drifts compared to the “calibration block” pipeline. The pipeline performs bad pixel flagging, cosmic ray detection, chop on/off subtraction, spectral flat fielding, and re-binning. We choose a re-binning with oversample=4 (thus gaining spectral resolution[^2] by making use of the redundancy in the data) and upsample=1 (thus keeping the individual data points independent). Finally, the pipeline averages the two nodding positions and calibrates the flux. The noise in the PACS spectra is lowest at the centre of the spectral range and increases versus the spectral edges. We discard the very noisy spectral edges. We then subtract the continuum by fitting linear polynomials to the spectra, where the line region is masked. The pipeline generated noise estimate “stddev” is used to give the highest weight to the data points at the centre of the spectral range[^3]. ![Placement of the 25 PACS spaxels relative to Fomalhaut (marked by the white star) in the case of the observations. The grey scale indicates the integrated emission for each spaxel within a window of width equal to 2.5\*FWHM of the line. The ellipse shows the location of the peak density of mm grains according to the best-fit model by . The corresponding figure for the data looks very similar. The 15 spaxels with a yellow bold edge were summed to produce Fig. \[upperlimits\].[]{data-label="PACS_CII_lineemission"}](lineemission_CII){width="\hsize"} Results and analysis ==================== Upper limits on the and emission from PACS {#flux_upperlimits} ------------------------------------------ Whereas both lines remain undetected, we do detect the dust continuum at both wavelengths in the spaxels covering the belt. At 63$\mu$m, we see additional continuum emission at the position of the star. Within the absolute flux calibration accuracy of PACS, the total detected continuum is consistent with the photometry presented by @Acke_etal_2012. We adopt a forward modelling strategy to compute upper limits on the line flux. We simulate PACS observations from a map of the line intensity on the sky. The PACS Spectrometer beams (version 3)[^4] are used to calculate the integrated line flux registered in each spaxel from the input map. The line is assumed to be unresolved. Thus the computed flux is assumed to be contained in a Gaussian-shaped line with a full width at half maximum (FWHM) determined by the PACS instrument[^5]. These simulated spectra can then be compared to the real observations. The input sky map is constructed under the assumption that the gas is co-spatial with the dust, since we are interested in constraining the possibility of gas-dust interactions. We take the best-fit model of the distribution of mm grains derived by as a description of the line luminosity density (energy from line emission per volume) and derive a sky map assuming optically thin emission. The only free parameter in the model is a global, positive scaling of the sky map, determining the total flux received from the belt. We adopt a simple Bayesian approach with a flat prior for the scaling factor. Assuming Gaussian noise for the PACS data points, a posterior probability distribution can be derived, which is used to compute an upper limit on the scaling factor and thus on the total line emission. The noise is estimated for each spaxel individually from two 2.5\*FWHM wide spectral windows, placed sufficiently far from the line centre so to avoid any potential line emission. However, the windows are also sufficiently far from the noisy edges of the spectral range. Thus we do not significantly overestimate the noise in the line region itself. Table \[upper\_limits\_flux\_table\] shows upper limits on the total line emission of gas co-spatial with the dust at 99% confidence level. The line gives a significantly stronger upper limit because of the lower noise level in these data. We also estimated upper limits using Monte Carlo simulations by repeatedly fitting our model to new realisations of the data with added noise. This approach gives upper limits that are smaller by $\sim$5% for and $\sim$20% for compared to the Bayesian calculation. We state the more conservative Bayesian values here. Compared to co-adding flux from the spaxels covering the belt (Fig. \[PACS\_CII\_lineemission\]), our approach using a model of the spatial distribution allows for a stronger upper limit (Fig. \[upperlimits\]). This is expected since more information is used to constrain the maximum flux compatible with the data. ----------- ---------------------- Line Flux (Wm$^{-2}$) 158$\mu$m $<2.2\times10^{-18}$ 63$\mu$m $<1.0\times10^{-17}$ ----------- ---------------------- : Upper limits (99% confidence level) on line fluxes.[]{data-label="upper_limits_flux_table"} ![Line fluxes ( and ) obtained when co-adding spaxels marked with bold edges in Fig. \[PACS\_CII\_lineemission\]. A 99% confidence level upper limit was derived by integrating over a 2\*FWHM wide wavelength range centred at the line wavelength and propagating the error on the integrated flux. The dashed line shows the Gaussian profile containing this upper limit flux. The plain line shows the profile obtained by co-adding the same spaxels containing our simulated spectra corresponding to the Bayesian upper limit, demonstrating the stronger limit obtained with our detailed model.[]{data-label="upperlimits"}](upperlimits){width="\hsize"} Upper limits on the total gas mass in the Fomalhaut debris belt --------------------------------------------------------------- We use the non-local thermodynamic equilibrium (non-LTE) excitation and radiative transfer code <span style="font-variant:small-caps;">radex</span> [@vanderTak_etal_2007] to convert the upper limits on the and line emission (Table \[upper\_limits\_flux\_table\]) into gas masses. The emission lines are assumed to be excited by collisions with atomic hydrogen and electrons. We do not attempt to calculate the thermal balance, but instead consider a range of kinetic gas temperatures between 48K (the dust temperature, ) and $10^4$K, which is the temperature where collisional ionisation of both and becomes important. The / ratio is fixed to solar. For the abundance of the other elements, we consider two cases: solar abundances [@Lodders_2003] and abundances as observed in the $\beta$ Pictoris gaseous debris disc, where the gas mass is dominated by and [@Cataldi_etal_2014; @Brandeker_etal_2015_subm]. The latter choice of abundances is further motivated by a recent study of the 49 Ceti debris disc, suggesting a C-rich gas similar to $\beta$ Pic [@Roberge_etal_2014]. For a given temperature, we search the amount of gas that is reproducing either of the upper limits ( or ). Usually, the gas mass is limited by the flux. The flux is the limiting factor only for kinetic temperatures in excess of 150K and solar abundances (see Fig. \[total\_gas\_mass\]). The densities of the collision partners ( and ) determine the excitation of the emission lines and therefore the total amount of gas necessary to reproduce a certain flux, but the total amount of gas also implies certain values for the and densities. Therefore, to construct a self-consistent model, an iterative approach is necessary. Starting with a guess of the and densities, we use <span style="font-variant:small-caps;">radex</span> to compute the amount of (or if is limiting the total gas mass) reproducing the corresponding flux upper limit. Next, the ionisation fraction of is estimated. We assume equilibrium between photoionisation and recombination. Ionising photons from two different sources are included: Fomalhaut’s photosphere from an ATLAS9 model [@Castelli_Kurucz_2004] with $T_\mathrm{eff}=8500$K, $\log g=4.0$ and $\log Z=0$, and the interstellar UV field [taken from @Draine_1978], where the latter is the dominant component. Ionisation by cosmic rays is negligible. A simplifying assumption is that all electrons are coming from the ionisation of , i.e. $n_{\element{C}}=n_{\element[-][]{e}}$. This should be an excellent approximation if the gas has $\beta$ Pic-like abundances. For solar abundances, we are likely underestimating the electron density because of the presence of highly ionised elements such as or . This only affects the ionisation calculation, but not the excitation of the lines, since for solar abundances, is dominating the collisional excitation[^6]. The ionisation fraction of is giving us the total amount of and in the model. The total gas mass (and in particular the density) can then be derived from the assumed elemental abundances, where elements up to atomic number $Z=38$ are considered (hydrogen through strontium). We iterate until the and input densities are equal to the output densities implied by the total gas mass. We assume to be completely neutral in our models because of its high ionisation potential. This is justified by computing the ionisation of for various electron densities and temperatures, showing that remains largely neutral for $n_\mathrm{e}\gtrsim10^{-2}$cm$^{-3}$. While <span style="font-variant:small-caps;">radex</span> assumes a homogeneous medium, we calculated the upper limits on the line emission using the dust profile as a description of the line luminosity density. We convert a gas mass output from <span style="font-variant:small-caps;">radex</span> to a density by distributing the mass according to the profile and taking the peak density. For example, the mass from <span style="font-variant:small-caps;">radex</span> implies an density via the peak density of the density profile. Changing this mass-to-density conversion by e.g. taking only half the peak density has only a minor effect on our results. Table \[upper\_limits\_gas\_table\] shows the upper limits on the gas mass and corresponding lower limits on the dust-to-gas ratio derived for the two elemental abundances considered. We also list upper limits on the and mass, the range of collision partner densities occurring and the maximum column densities, showing that the emission is optically thin. Figure \[total\_gas\_mass\] shows the gas mass as a function of the kinetic temperature. The upper limit on the gas mass comes from the lowest kinetic temperature (T=48K) and is below the millimetre dust mass derived by . For illustration, we also performed a simple calculation of the amount of necessary to reproduce the upper limit on the flux in the case of LTE for $\beta$ Pic-like abundances. It is seen that LTE is not a very accurate approximation, demonstrating the need for a non-LTE code like <span style="font-variant:small-caps;">radex</span> for the type of environment investigated here. ------------- ----------------------- ------------ ----------------------- ----------------------- ----- ----- --------------------- --------------------- ------------------ ------------------ abundances $M_{\mathrm{gas}}$ $\epsilon$ $M_{\element{C}}$ $M_{\element{O}}$ $N_\mathrm{max}$ () $N_\mathrm{max}$ () min max min max ($\mathrm{M}_\oplus$) ($\mathrm{M}_\oplus$) ($\mathrm{M}_\oplus$) (cm$^{-2}$) (cm$^{-2}$) solar $<5.0\times10^{-3}$ $>3.4$ $<1.2\times10^{-5}$ $<3.2\times10^{-5}$ 2 10 $5\times10^{3}$ $5\times10^{4}$ $8\times10^{14}$ $2\times10^{15}$ $\beta$ Pic $<5.7\times10^{-5}$ &gt;298 $<1.6\times10^{-5}$ $<4.1\times10^{-5}$ 6 12 $8\times10^{-2}$ $2\times10^{-1}$ $1\times10^{15}$ $3\times10^{15}$ ------------- ----------------------- ------------ ----------------------- ----------------------- ----- ----- --------------------- --------------------- ------------------ ------------------ ![Total gas mass as a function of the kinetic temperature, derived from models reproducing either the or flux limit from Table \[upper\_limits\_flux\_table\] (as indicated in the legend). The / ratio is fixed to solar. The gas masses are derived using the non-LTE code <span style="font-variant:small-caps;">radex</span> for both solar and $\beta$ Pic-like elemental abundances. In addition, the gas mass derived from a simple LTE calculation is shown for $\beta$ Pic-like abundances. The total gas mass remains below the mm dust mass (horizontal dashed line) over the entire temperature range. We take the gas masses at the lowest kinetic temperature as our upper limits (indicated by an arrow and shown in Table \[upper\_limits\_gas\_table\]).[]{data-label="total_gas_mass"}](total_gas_mass){width="\hsize"} Summary and conclusion ====================== Gas-dust interactions have been shown to concentrate dust into narrow, eccentric rings via a clumping instability. A necessary condition for the instability to develop is a dust-to-gas ratio $\epsilon\lesssim1$ [@Lyra_Kuchner_2013], although the instability is expected to be maximised for $\epsilon\approx0.2$ (W. Lyra 2014, private communication). We know from e.g. $\beta$ Pictoris or 49 Ceti that some debris discs contain gas such that $\epsilon$ is indeed of the order of unity. Prior to the present work, the only information on the gas content of the Fomalhaut debris belt came from non-detections of . @Dent_etal_1995 used a non-detection of the J=3-2 transition to put upper limits on the gas mass assuming LTE. Recently, @Matra_etal_2014 used an ALMA non-detection of the same transition to place a $3\sigma$ upper limit of $\sim$$5\times 10^{-4}$M$_\oplus$ on the CO content of the Fomalhaut belt (i.e. a dust-to-CO ratio $\gtrsim$30), taking the crucial importance of non-LTE effects into account. However, since is expected to be photo-dissociated by stellar or interstellar UV photons on a short timescale compared to the lifetime of the system [@Kamp_Bertoldi_2000; @Visser_etal_2009], the absence of does not strongly constrain the total gas mass. Constraints on the atomic gas mass are more useful. We derived upper limits on the atomic gas mass using non-detections of and emission and assuming either solar abundances or abundances as those observed in the $\beta$ Pic debris disc. The total gas mass is shown to be smaller than the dust mass. Because of the low gas content, gas-dust interactions are not expected to be working efficiently in the Fomalhaut belt. Therefore, the data indirectly suggest a second, yet unseen planet Fomalhaut c or stellar encounters as the cause for the morphology of the Fomalhaut debris belt. We would like to thank the referee, Jane Greaves, for useful and constructive comments that helped to clarify this manuscript. We also thank Elena Puga from the *Herschel* Science Centre Helpdesk for support with the *Herschel* beam products. This research has made use of the SIMBAD database (operated at CDS, Strasbourg, France), the NIST Atomic Spectra Database, the NORAD-Atomic-Data database and NASA’s Astrophysics Data System. PACS has been developed by a consortium of institutes led by MPE (Germany) and including UVIE (Austria); KU Leuven, CSL, IMEC (Belgium); CEA, LAM (France); MPIA (Germany); INAF-IFSI/OAA/OAP/OAT, LENS, SISSA (Italy); IAC (Spain). This development has been supported by the funding agencies BMVIT (Austria), ESA-PRODEX (Belgium), CEA/CNES (France), DLR (Germany), ASI/INAF (Italy), and CICYT/MCYT (Spain). [^1]: Based on *Herschel* observations. *Herschel* is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA. [^2]: With this choice of re-binning, the resolution is $\sim$60kms$^{-1}$ for the data and $\sim$22kms$^{-1}$ for the data. [^3]: Even if the noise estimate delivered by the pipeline probably underestimates the true uncertainty, it is still useful for fitting the continuum, since it describes the relative change of the noise over the spectral range. [^4]: <http://herschel.esac.esa.int/twiki/bin/view/Public/PacsCalibrationWeb#PACS_spectrometer_calibration> [^5]: 0.126$\mu$m for the 158$\mu$m line and 0.018$\mu$m for the 63$\mu$m line; see section 4.7.1 of the PACS Observer’s Manual, <http://herschel.esac.esa.int/Docs/PACS/pdf/pacs_om.pdf> [^6]: In the case of the -poor $\beta$ Pic-like abundances, collisional excitation is completely dominated by .

ArXiv

--- abstract: 'In a non-centrosymmetric crystal, the Zeeman interaction of the band electrons with an external magnetic field is highly anisotropic in the momentum space, vanishing along some high-symmetry planes. One of the consequences is that the paramagnetic susceptibility in superconductors without inversion symmetry, such as CePt$_3$Si, shows an unusual temperature dependence.' author: - 'K. V. Samokhin' title: | Paramagnetic properties of non-centrosymmetric superconductors:\ Application to CePt$_3$Si --- Electronic spin susceptibility measurements in the superconducting state provide one of the most useful tools for the identification of the pairing symmetry. Experimentally, the susceptibility $\chi(T)$ is measured by the Knight shift in the nuclear magnetic resonance frequency due to the hyperfine interaction of the conduction electrons with the nuclear magnetic moments. If the superconducting pairing occurs in the spin-singlet channel, then the Cooper pairs have spin $S=0$ and do not contribute to the magnetization $\bm{M}$ of the system, which is therefore entirely determined by the thermally excited quasiparticles. For a fully gapped order parameter one would see an exponentially decreasing $\chi(T)$ at low temperatures. In contrast, if the gap has zeros at the Fermi surface, then $\chi(T)\propto T^2$ for isolated point nodes, or $\chi(T)\propto T$ for line nodes. In the triplet case, when the Cooper pairs have spin $S=1$ and the order parameter is a vector $\bm{d}(\bk)$ in the spin space (see, e.g., Ref. [@Book]), the susceptibility depends on the mutual orientation of $\bm{d}$ and the external magnetic field $\bB$. If $\bB\parallel\bm{d}$, then the Cooper pairs do not contribute to $\bm{M}$, and $\chi(T)$ has the same temperature dependence as in the singlet case. If $\bB\perp\bm{d}$, then both the pairs and the excitations contribute to $\bm{M}$, so that $\chi(T)=\chi_n$ – the normal-state susceptibility. The observation of a flat Knight shift in Sr$_2$RuO$_4$ for the field in the basal $xy$ plane [@SRO-NMR], has been used as a proof of a spin-triplet pairing with $\bm{d}\parallel\hat z$. The theoretical picture described above is valid only if the superconducting crystal has an inversion center. Although this is the case in the majority of superconductors, some exceptions have been known since 1960’s [@early; @examples]. More recently, it was pointed out in Ref. [@GR01] that the surface superconductors, e.g. Na-doped WO$_3$ [@WO3], are intrinsically non-centrosymmetric simply because the two sides of the surface layer are manifestly non-equivalent. As for the bulk materials, the latest examples are CePt$_3$Si [@exp; @CePtSi] and UIr [@exp; @UIr]. Different models of superconductivity in CePt$_3$Si have been proposed in Refs. [@SZB04; @FAKS04]. In this Letter, we calculate the suppression of the critical temperature and the paramagnetic susceptibility in superconductors without an inversion center. We focus on the tetragonal symmetry relevant for CePt$_3$Si, and show in particular that the anisotropy of the temperature dependence of the susceptibility tensor $\chi_{ij}(T)$ is strikingly different from the centrosymmetric case. The spin susceptibility in two-dimensional non-centrosymmetric superconductors has been previously studied in Refs. [@BGR76; @Edelstein; @GR01; @Yip02], where the inversion symmetry breaking in the presence of a non-zero spin-orbit (SO) coupling was introduced using the Rashba model [@Rashba]. Very recently, a three-dimensional generalization of the Rashba model was applied to CePt$_3$Si in Ref. [@FAS04]. Treating the SO band splitting as a perturbation, it was found that the order parameter becomes a mixture of a spin-singlet (even in $\bk$) and a spin-triplet (odd in $\bk$) components, which gives rise to a non-zero residual susceptibility at $T=0$. In this Letter, we use a different approach based on the effective single-band Hamiltonian, which works for any crystal symmetry and arbitrary strength of the SO coupling. Guided by the fact that the SO band splitting is usually large compared to the superconducting energy scales, we consider both the Cooper pairing and the magnetic response independently in different bands. In contrast to the previous works, we construct the pairing interaction using the exact band states and explicitly take into account that all the pairing channels but one are suppressed [@And84]. The starting point of our analysis is the observation that in a crystal lacking an inversion center the electron bands are non-degenerate almost everywhere, except along some high-symmetry lines in the Brillouin zone. Indeed, without the inversion operation $I$, one cannot construct two orthogonal degenerate Bloch states at the same $\bk$. At zero SO coupling there is an additional symmetry in the system – the invariance with respect to arbitrary spin rotations, which restores the two-fold spin degeneracy of the bands. Here we assume that the SO coupling is sufficiently strong, so that the bands are well split. The results of Ref. [@SZB04] show that this is indeed the case in CePt$_3$Si, where the SO band splitting can be as large as 50-200 meV depending on the band. Assuming that there is no disorder in the crystal, the Bloch wave vector $\bk$ is a good quantum number in zero field. The free electron Hamiltonian for a non-degenerate band can be written as $H_0=\sum_{\bk}\epsilon(\bk)c^\dagger_{\bk}c_{\bk}$, where $\sum_{\bk}$ stands for the integration over the first Brillouin zone, and $\epsilon(\bk)$ is the quasiparticle dispersion, which takes into account the periodic lattice potential and the SO interaction. If the time-reversal symmetry is not broken in the normal phase, then the states $|\bk\rangle$ and $K|\bk\rangle\sim|-\bk\rangle$ are degenerate because of the Kramers theorem [@AFM; @comment]. It is the coupling between those states that leads to the formation of the Cooper pairs and the superconductivity in the system. The large band splitting strongly suppresses the pairing of electrons from different bands. Then, considering just one band (i.e. neglecting the inter-band pair scattering) and assuming that the interaction has a generalized Bardeen-Cooper-Schrieffer (BCS) form, we have $$\label{H BCS} H_{int}=\frac{1}{2}\sum\limits_{\bk,\bk'} V(\bk,\bk')c^\dagger_{\bk}c^\dagger_{-\bk}c_{-\bk'}c_{\bk'}.$$ The pairing potential can be written as $V(\bk,\bk')=\tilde V(\bk,\bk')t(\bk)t^*(\bk')$, where $\tilde V(\bk,\bk')=-V_\Gamma\sum_a\phi_a(\bk)\phi_a^*(\bk')$ is the part that transforms according to an irreducible representation $\Gamma$ of the normal-state point group $G$, $\phi_a(\bk)$ are the scalar basis functions of $\Gamma$, which are are nonzero only inside the energy shell of width $\epsilon_c$ near the Fermi surface, $V_\Gamma>0$ is the coupling constant, and $t(\bk)=-t(-\bk)$ are non-trivial phase factors in $K|\bk\rangle=t(\bk)|-\bk\rangle$ [@SC04]. Although anti-commutation of fermionic operators dictates that the mean-field order parameter $\Delta(\bk)=t(\bk)\sum_a\eta_a\phi_a(\bk)$ is odd in $\bk$ [@SZB04], its nodal structure is determined by the basis functions $\phi_a(\bk)$, which should be even because of the presence of $t(\bk)$ [@correct]. The focus of this article is on CePt$_3$Si, which has a non-centrosymmetric tetragonal crystal lattice described by the point group $G=\mathbf{C}_{4v}$. This group is generated by the rotations $C_{4z}$ about the $z$ axis by an angle $\pi/2$ and the reflections $\sigma_x$ in the vertical plane $(100)$, and has five irreducible representations: four one-dimensional (1D): $A_1$, $A_2$, $B_1$, $B_2$, and one two-dimensional $E$ [@LL3]. Here are the examples of the even basis functions: $\phi_{A_1}\propto k_x^2+k_y^2+ck_z^2$, $\phi_{A_2}\propto k_xk_y(k_x^2-k_y^2)$, $\phi_{B_1}\propto k_x^2-k_y^2$, $\phi_{B_2}\propto k_xk_y$, and $(\phi_{E,1},\phi_{E,2})\propto (k_xk_z,k_yk_z)$. We consider a small sample, of a dimension $d\leq\xi<\delta$, where $\xi$ is the superconducting correlation length and $\delta$ is the London penetration depth, which allows us to neglect the spatial variations of both the order parameter components $\eta_a$ and the magnetic field. Let us now turn on a uniform stationary magnetic field $\bB=\mathrm{curl}\,\bA$. Assuming that the pairing interaction is field-independent, $\bB$ can only affect the system through its coupling to the band states. At $\bB\neq 0$, the band dispersion function $\epsilon(\bk)$ is replaced by an effective band Hamiltonian in the momentum space, which can be represented as a power series in $\bB$: $\epsilon(\bk)\to{\cal E}(\bk,\bB)=\epsilon(\bK)+B_i\epsilon_{1,i}(\bK)+...$, where $\bK=\bk+(e/\hbar c)\bA(i\bm{\nabla}_{\bk})$ because of the requirements of gauge invariance [@H-eff]. The expansion coefficients must satisfy certain symmetry-imposed conditions, in particular the zero-field band dispersion $\epsilon(\bk)$ must be invariant under all operations from $G$. In addition, at $\bB\neq 0$ the Hamiltonian is invariant with respect to time reversal $K$ only if the sign of $\bB$ (and of $\bA$) is also changed, which imposes the following constraint on the function ${\cal E}$: $K^\dagger{\cal E}(-\bB)K={\cal E}(\bB)$. In the analysis of the “paramagnetic” properties of superconductors, the orbital effect of the field is neglected, which is achieved by putting $\bA=0$ in the effective band Hamiltonian. Then, for a two-fold degenerate band in the presence of inversion symmetry, ${\cal E}$ is a $2\times 2$ matrix, and the coupling to the magnetic field is described by a familiar Zeeman term: ${\cal E}_{\alpha\beta}(\bk,\bB)=\epsilon(\bk)\delta_{\alpha\beta}-B_i\mu_{ij}(\bk) \sigma_{j,\alpha\beta}$, with $\mu_{ij}(\bk)=\mu_{ij}(-\bk)$ being the tensor generalization of the Bohr magneton $\mu_B$ for the case of band electrons. The indices $\alpha\beta$ here are pseudospin indices [@UR85]. The Zeeman interaction splits the energies of the electrons forming the Cooper pairs and gives rise to the paramagnetic suppression of superconductivity [@CC62]. If the inversion symmetry is absent and the bands are non-degenerate, then the Zeeman term should be modified. The effective single-band Hamiltonian in the external field can be written as $$\label{H 0} H_0=\sum\limits_{\bk}\left[\epsilon(\bk)-\bB\blam(\bk)\right] c^\dagger_{\bk} c_{\bk},$$ which is markedly different from the centrosymmetric case. Here $\blam(\bk)$ is a real pseudovector, which satisfies the conditions $(g\blam)(g^{-1}\bk)=\blam(\bk)$, where $g$ is any operation from the point group $G$. Because of the time-reversal symmetry, we also have $\epsilon(-\bk)=\epsilon(\bk)$ and $\blam(-\bk)=-\blam(\bk)$, but ${\cal E}(-\bk,\bB)\neq{\cal E}(\bk,\bB)$. Explicit expressions for $\blam(\bk)$ can only be obtained in some simple models. For example, in an isotropic two-dimensional electron gas in the $xy$ plane with $G=\mathbf{C}_{\infty v}$, the combined effect of the SO coupling and the lack of inversion symmetry is described by an additional (Rashba) term in the single-particle Hamiltonian: $H_{SO}=\gamma\sum_{\bk}\bm{n}\cdot(\bm{\sigma}_{\sigma\sigma'} \times\bk)\,a^\dagger_{\bk\sigma}a_{\bk\sigma'}$, where $\bm{n}$ is the normal vector to the plane [@Rashba]. Diagonalization of the Hamiltonian in zero field gives two non-degenerate Rashba bands: $\epsilon_\pm(\bk)=\epsilon_0(\bk)\pm\gamma|\bk|$. At finite field, adding a usual Zeeman term $-\mu_B(\bB\cdot\bm{\sigma})$, and expanding the eigenlavues of the Hamiltonian in powers of $\bB$, we obtain ${\cal E}_\pm(\bk,\bB)=\epsilon_\pm(\bk)-\blam_\pm(\bk)\bB+O(B^2)$, where $\blam_\pm(\bk)=\pm\mu_B(\bk\times\bm{n})/|\bk|$. Thus the coupling of the Rashba bands with the field is highly anisotropic, in particular it vanishes for $\bB\parallel\bm{n}$. While a microscopic derivation of the effective single-band Hamiltonian (\[H 0\]) in more realistic systems can be done, at least in principle, using the procedures described in Refs. [@H-eff], it suffices for our purposes to work with a phenomenological expression for $\blam(\bk)$, which is compatible with all the symmetry constraints. We need an expression for $\blam(\bk)$, which satisfies (i) $\blam(-\bk)=-\blam(\bk)$, (ii) $(C_{4z}\blam)(C_{4z}^{-1}\bk)=\blam(\bk)$, and (iii) $(\sigma_x\blam)(\sigma_x^{-1}\bk)=\blam(\bk)$ (since $\blam$ is a pseudovector, we have $\sigma_x\blam\equiv IC_{2x}\blam=C_{2x}\blam$, where $C_{2x}$ is a rotation by an angle $\pi$ about the $x$ axis). It is straightforward to check that the general expression for $\blam(\bk)$ is given by $$\label{lambda CPS} \blam(\bk)=\tilde\phi_{E,2}(\bk)\hat x- \tilde\phi_{E,1}(\bk)\hat y+\tilde\phi_{A_2}(\bk)\hat z,$$ where $\tilde\phi_{E,1(2)}(\bk)$ and $\tilde\phi_{A_2}(\bk)$ are real odd functions, which transform according to the representations $E$ and $A_2$ respectively, e.g. $\tilde\phi_{A_2}\propto k_xk_yk_z(k_x^2-k_y^2)$ and $(\tilde\phi_{E,1},\tilde\phi_{E,2})\propto (k_x,k_y)$. We see that $\lambda_z=0$ along the five nodal planes of the $A_2$ representation, while $\lambda_x=\lambda_y=0$ along the $z$ axis. Now, we calculate the free energy ${\cal F}$ for the Hamiltonian $H=H_0+H_{int}$, defined by Eqs. (\[H 0\]) and (\[H BCS\]). We use the effective field theory in terms of the bosonic Matsubara fields $\eta_a(\tau)$, which can be introduced in a standard fashion by decoupling the pairing interaction (\[H BCS\]). The effective action for a uniform stationary order parameter in the mean-field approximation reads $$\label{S eff} S_{eff}=\frac{\beta{\cal V}}{2V_\Gamma}\sum\limits_a|\eta_a|^2 -\frac{1}{2}\mathrm{Tr}\ln G^{-1},$$ where $$\begin{aligned} \label{G inverse} &&G^{-1}(\bk,\omega_n)\nonumber\\ &&=\left(% \begin{array}{cc} i\omega_n-\epsilon(\bk)+\bB\blam(\bk) & -\Delta(\bk) \\ -\Delta^*(\bk) & i\omega_n+\epsilon(\bk)+\bB\blam(\bk) \\ \end{array}% \right)\qquad\end{aligned}$$ is the $2\times 2$ inverse Gor’kov Green’s function, $\omega_n=(2n+1)\pi T$ is the fermionic Matsubara frequency, and ${\cal V}$ is the system volume. The mean-field free energy is related to the saddle-point action (\[S eff\]): ${\cal F}=(1/\beta)S_{eff}$, and the magnetization density is $\bm{M}=-{\cal V}^{-1}(\partial{\cal F}/\partial\bB)$. The saddle point condition yields the self-consistency equation $$\label{gap eq} \frac{1}{V_\Gamma}\eta_a+T\sum\limits_n\sum\limits_{\bk}t^*(\bk)\phi_a^*(\bk)G_{12}(\bk,\omega_n)=0,$$ which determines the temperature and field dependence of the order parameter components. Substituting here the Green’s function (\[G inverse\]), one can see that the phase factors $t(\bk)$ drop out of the gap equation. Let us first find how the critical temperature is suppressed by the field. The equation for $T_c(\bB)$ is obtained by linearizing Eq. (\[gap eq\]) and can be written in the form $\det||K_{ab}||=0$, where $$\begin{aligned} \label{Kab} &&K_{ab}=\left[\ln\frac{T_{c0}}{T_c}+\psi\left(\frac{1}{2}\right)\right]\delta_{ab}\\ &&\qquad-\left\langle\phi_a^*(\bk)\phi_b(\bk)\re\psi\left(\frac{1}{2}- i\frac{\bB\blam(\bk)}{2\pi T_c}\right)\right\rangle_{FS},\nonumber\end{aligned}$$ where $\psi(x)$ is the digamma function, $T_{c0}\simeq 1.13\epsilon_c\exp(1/N_FV_\Gamma)$ is the critical temperature in zero field ($N_F$ is the density of states at the Fermi level), and the angular brackets denote the average over the Fermi surface. Next, differentiating the effective action (\[S eff\]) with respect to $\bB$, we find the magnetization density $$\label{M} \bm{M}=\frac{1}{2}T\sum\limits_n\sum\limits_{\bk}\blam(\bk)\left[ G_{11}(\bk,\omega_n)+G_{22}(\bk,\omega_n)\right].$$ The uniform susceptibility tensor is defined in the usual manner as $\chi_{ij}=\partial M_i/\partial B_j|_{\bB=0}$. One can easily check using Eq. (\[gap eq\]) that the corrections to the order parameter components $\eta_a$ in a weak magnetic field are quadratic in $\bB$, which means that in the calculation of $\chi_{ij}$ one can neglect the field dependence of $\Delta(\bk)$, to obtain $$\label{chi ij} \chi_{ij}(T)=\frac{1}{4T}\sum\limits_{\bk} \frac{\lambda_i(\bk)\lambda_j(\bk)}{\cosh^2[E(\bk)/2T]},$$ where $E(\bk)=\sqrt{\epsilon^2(\bk)+|\Delta(\bk)|^2}$. We now apply the general theory to CePt$_3$Si. The pseudovector $\blam(\bk)$ in this case is given by Eq. (\[lambda CPS\]). Since all three components of $\blam$ are in general non-zero, we expect the superconducting critical temperature to be suppressed for all orientations of the magnetic field. According to Eq. (\[Kab\]), the magnitude of the suppression depends on many factors: the shape of the Fermi surface, the symmetry of the order parameter, and also the explicit form of the functions $\tilde\phi$ in Eq. (\[lambda CPS\]). The Fermi surface of CePt$_3$Si is quite complicated and consists of six sheets [@SZB04]. It is not known which one (or ones) of them are superconducting. The order parameter symmetry is not known either, although the observation of a linear $T$-dependence of the specific heat [@exp; @CePtSi] probably indicates that the order parameter has lines of nodes. In view of all this uncertainty, it seems to be premature to discuss the paramagnetic suppression in CePt$_3$Si quantitatively. More interesting qualitative conclusions can be drawn from the analysis of the susceptibility $\chi_{ij}(T)$. It follows from Eq. (\[chi ij\]) that the main contribution to the susceptibility tensor at low temperatures comes from the thermally excited nodal quasiparticles. If the order parameter $\Delta(\bk)$ has no zeros at the Fermi surface (e.g. for the $A_1$ representation), the susceptibilities are exponentially small in all directions. Let us consider the 1D order parameters corresponding to the representations $A_2$, $B_1$, or $B_2$, for which the lines of nodes are symmetry-imposed. In this case, the components of the susceptibility tensor are given by $\chi_{xx}=\chi_{yy}=\chi_\parallel$ and $\chi_{zz}=\chi_\perp$, where $$\begin{aligned} \label{chi parall} &&\chi_\parallel(T)=\frac{1}{8T}\sum\limits_{\bk} \frac{\tilde\phi_{E,1}^2(\bk)+\tilde\phi_{E,2}^2(\bk)}{\cosh^2[E(\bk)/2T]},\\ \label{chi perp} &&\chi_\perp(T)=\frac{1}{4T}\sum\limits_{\bk} \frac{\tilde\phi_{A_2}^2(\bk)}{\cosh^2[E(\bk)/2T]}.\end{aligned}$$ It is straightforward to show that for the field in the basal plane, $\chi_\parallel(T)\propto T$, since $\tilde\phi_{E,1}^2(\bk)+\tilde\phi_{E,2}^2(\bk)$ is in general non-zero everywhere, except the poles of the Fermi surface. This behavior is characteristic of the systems with lines of nodes. In contrast, for the field orientation along the $z$ axis, we have from Eq. (\[chi perp\]) $\chi_\perp(T)\propto T^3$, since $\tilde\phi_{A_2}(\bk)$ vanishes at the nodal lines for all the 1D order parameters. Such temperature dependence is never seen in the centrosymmetric case. For a two-component order parameter $\Delta(\bk)\sim\eta_1k_xk_z+\eta_2k_yk_z$ with a horizontal line of zeros at $k_z=0$, one also has a $T^3$-behaviour of $\chi_\perp(T)$. On the other hand, the anisotropy and the temperature dependence of $\chi_\parallel(T)$ is determined by $(\eta_1,\eta_2)$. We would like to note that, in contrast to Refs. [@GR01; @FAS04], our approach does not yield a finite value of the susceptibility at $T=0$. The explanation is that the residual susceptibility in the systems described by the Rashba model comes from the interband transitions between the Rashba bands (Van Vleck susceptibility), which are not affected by the transition into the superconducting state [@Yip02]. In our model, the quasiparticles in each band respond to the external field independently of the other bands. If to take the interband transitions into account, then the observed susceptibility will be $\chi_{tot}(T)=\chi_{0}+\chi(T)$, where the first term is the temperature-independent background that comes, e.g., from the Van Vleck processes, and the second term is the single-band contribution (\[chi ij\]). Other mechanisms that might affect the residual susceptibility include the contributions from the unpaired sheets of the Fermi surface, or spin-reversing scattering at impurities or surface imperfections [@residual; @chi]. In conclusion, we have shown that the paramagnetic responses of superconductors with and without inversion center are qualitatively different. The most important feature is that, in the latter case, the coupling of the non-degenerate band electrons with the external field is strongly momentum-dependent and vanishes, for symmetry reasons, along some high-symmetry planes in the Brillouin zone. This results in a high anisotropy of the susceptibility in the superconducting state with lines of nodes, from $\chi_\parallel(T)\propto T$ to $\chi_\perp(T)\propto T^3$, which can be used as a clear-cut experimental test of our theory. This work was supported by the Natural Sciences and Engineering Research Council of Canada. [99]{} V. P. Mineev and K. V. Samokhin, [*Introduction to Unconventional Superconductivity*]{} (Gordon and Breach, London, 1999). K. Ishida, H. Mukuda, Y. Kitaoka, K. Asayama, Z. Q. Mao, Y. Mori, and Y. Maeno, Nature **396**, 658 (1998). P. W. Anderson and E. I. Blount, Phys. Rev. Lett. **14**, 217 (1965); J. F. Schooley, W. R. Hosler, and M. L. Cohen, Phys. Rev. Lett. **12**, 474 (1964); A. C. Lawson and W. H. Zachariasen, Phys. Lett. **38A**, 1 (1972); V. V. Bogatko and Yu. N. Venevtsev, Fiz. Tverd. Tela **25**, 1495 (1983) \[Sov. Phys. – Solid State **25**, 859 (1983)\]. L. P. Gor’kov and E. I. Rashba, Phys. Rev. Lett. **87**, 037004 (2001). S. Reich and Y. Tsabba, Eur. Phys. J. B **9**, 1 (1999). E. Bauer, G. Hilscher, H. Michor, Ch. Paul, E. W. Scheidt, A. Gribanov, Yu. Seropegin, H. Noël, M. Sigrist, and P. Rogl, Phys. Rev. Lett. **92**, 027003 (2004). T. Akazawa, H. Hidaka, T. Fujiwara, T. C. Kobayashi, E. Yamamoto, Y. Haga, R. Settai, and Y. Onuki, J. Phys.: Condens. Matter **16**, L29 (2004). K. V. Samokhin, E. S. Zijlstra, and S. K. Bose, Phys. Rev. B **69**, 094514 (2004). P. A. Frigeri, D. F. Agterberg, A. Koga, and M. Sigrist, Phys. Rev. Lett. **92**, 097001 (2004). L. N. Bulaevskii, A. A. Guseinov, and A. I. Rusinov, Zh. Eksp. Teor. Fiz. **71**, 2356 (1976) \[Sov. Phys. – JETP **44**, 1243 (1976)\]. V. M. Edelstein, Zh. Eksp. Teor. Fiz. **95**, 2151 (1989) \[Sov. Phys. – JETP **68**, 1244 (1989)\]; V. M. Edelstein, Phys. Rev. Lett. **75**, 2004 (1995). S. K. Yip, Phys. Rev. B **65**, 144508 (2002). E. I. Rashba, Fiz. Tverd. Tela **2**, 1224 (1960) \[Sov. Phys. – Solid State **2**, 1109 (1960)\]; Yu. A. Bychkov and E. I. Rashba, Pis’ma Zh. Eksp. Teor. Fiz. **39**, 66 (1984) \[JETP Letters **39**, 78 (1984)\]. P. A. Frigeri, D. F. Agterberg, and M. Sigrist, preprint cond-mat/0405179 (unpublished). P. W. Anderson, Phys. Rev. B **30**, 4000 (1984). One can show that the symmetry of the single-electron bands and of the superconducting order parameter is not affected by the fact that the normal state of CePt$_3$Si is antiferromagnetic, see Ref. [@exp; @CePtSi], and N. Metoki *et al*, J. Phys.: Condens. Matter **16**, L207 (2004). I. A. Sergienko and S. H. Curnoe, cond-mat/0406003 (unpublished). The gap symmetry was analyzed in Ref. [@SZB04] using the odd basis functions $\phi_a(\bk)$. This has been corrected in Ref. [@SC04]. L. D. Landau and E. M. Lifshitz, [*Quantum Mechanics*]{} (Butterworth-Heinemann, 2002). R. Peierls, Z. Phys. **80**, 763 (1933); J. M. Luttinger, Phys. Rev. **84**, 814 (1951); W. Kohn, Phys. Rev. **115**, 1460 (1959); G. Wannier, Rev. Mod. Phys. **34**, 645 (1962); E. I. Blount, Phys. Rev. **126**, 1636 (1962); L. M. Roth, J. Phys. Chem. Solids **23**, 443 (1962). K. Ueda and T. M. Rice, Phys. Rev. B **31**, 7114 (1985). B. S. Chandrasekhar, Appl. Phys. Lett. **1**, 7 (1962); A. M. Clogston, Phys. Rev. Lett. **9**, 266 (1962). R. A. Ferrell, Phys. Rev. Lett. **3**, 262 (1959); P. W. Anderson, Phys. Rev. Lett. **3**, 325 (1959); A. A. Abrikosov and L. P. Gor’kov, Zh. Eksp. Teor. Fiz. **42**, 1088 (1962) \[Sov. Phys. – JETP **15**, 752 (1962)\].

ArXiv

23.5cm 1ex -40pt = .5ex [**December 1992,** ]{}\ [**THE THEORY OF TURBULENCE IN TWO DIMENSIONS** ]{}\ \ Introduction ============ During the last thirty years we witnessed an amazing unification of physical ideas. Concepts and notions of the totally different regions of physics appeared to be almost isomorphic. Spontaneous symmetry breaking and renormalization group are the most notable examples. In this paper I extend slightly this set of isomorphic ideas. Namely, I apply methods of conformal field theory ( CFT ) to the problem of two-dimensional turbulence. Short version of this paper has been published earlier \[1\]. The major puzzle in the theory of turbulence is the following. It is commonly believed, that at large Reynolds numbers we are dealing with the stationary statistical regime, describing velocity distribution. In other words, there should exist a time-independent probability $ P=P[~v_{\alpha}\left( x \right)~]$ (where $ v_{\alpha}\left( x \right)$ ) is the velocity). This probability should commute with the equations of motion, or: $$\int{dx\left( {\delta P \over \delta v_{\alpha}\left( x \right)} \right)\dot{v}_{\alpha}\left( x \right)}=0$$ (where it is understood that we express $ \dot{v}_{\alpha}$ through $ v_{\alpha}$ by the use of Navier - Stokes equations). At large Reynolds numbers viscosity can be neglected and hydrodynamics becomes a hamiltonian theory with the hamiltonian: $$\label{p} H~=~\int{dx {1 \over 2}v_{\alpha}^{2}}$$ If so, the equation (1) means that P is the integral of motion for this system. Since it is believed that in general there are no non-trivial integrals for the system (2) (in two-dimensions there are, but these are irrelevant for our discussion) we could erroneously conclude that: $$\label{3} P~=P\left( H \right)~=~exp\left\{ -\beta H \right\}$$ (the last equality follows from the usual additivity assumption). This is the Gibbs distribution for the temperature $ \beta ^{-1}$ and it clearly does not describe the turbulent flow, in which we expect permanent energy flux. The resolution of this puzzle is that in fact there are extra integrals of motion which form $ P[~v_{\alpha}]$. They are, however, highly non-local and non-polynomial. It would be difficult and unnecessary to write them down explicitly, because there still be a problem of averaging with respect to very complicated distributions. In what follows we will mostly use the analogue of the equation (1) applied directly to the correlation functions. Nevertheless we will briefly comment on the origin of the extra integral of motions for the systems with weak coupling. In this case they are just naive field-theoretic constants of motion formed of “ in” and “out” operators. When expressed in terms of the Heisenberg fields they become non-local, and describe the distributions of weak turbulence. The central point of this paper is, however, quite different. We will consider the equations for the correlation functions: $$\label{4} <~\dot{v}_{\alpha_{1}}\left( x_{1} \right)v_{\alpha_{2}}\left( x_{2} \right)\cdots~>+<~v_{\alpha_{1}}\left( x_{1} \right)\dot{v}_{\alpha_{2}}\left( x_{2} \right)\cdots~>+\cdots=0$$ (where $ \dot{v}_{\alpha}$ is expressed in terms of v by the use of equations of motion). These are standard equations (The Hopf equations \[2\] ) which express N - point functions in terms of the N + 1 - point function. The existing procedures for dealing with these equations are based on some kind of closure hypothesis expressing, say the 4 - point function as a square of 2 - point function. We take a different approach. Namely, we try to satisfy (4) exactly, by the use of conformal field theory, assuming that the developed turbulence in the inertial range possesses infinite conformal symmetry. There is no proof that this must necessarily the case, and that our conformal solutions are the only possible ones. What we have done is only a successful ansatz which solves exactly eqs (4). Chains of equations solved by conformal field theories. ======================================================== In this section I will sum up some basic properties of CFT, needed below as a representative example, I will use the most familiar case of the Ising model. The Ising model is a $ \varphi^{4}$ -field theory\[3\]. That means that if we treat $ \varphi$ as a fluctuating variable, it satisfies the equation: $$\label{5} \partial^{2}\varphi\left( x \right)+\mu \varphi\left( x \right)=g\varphi^{3}\left( x \right)$$ implying the relations between N - point functions of $ \varphi$ and N + 3 - point functions, known as Schwinger - Dyson equations. The constant $ \mu_{0}$ must be fine - tuned in order to put the system to the critical point, or in order to allow (5) to have conformal invariant solutions. Now let us try to answer the basic question of this section : how CFT solves eqs (5)? In CFT we are dealing directly with the sets of correlation functions at the critical point. The basic assumptions \[4\] are that there exists a set of the so called primary operators $ \left\{ O_{k}\left( x \right)\right\} $ with anomalous dimensions $ \left( \Delta_{k}, \bar{\Delta} _{k}\right)$ (referring to the rescaling of variables $ z ~=~x_{1}+i x_{2}$ and $ \bar{z}~=~x_{1}-ix_{2}$. It is assumed that under arbitrary analytic transformation $ z {}~\Rightarrow~f\left( z \right)$  ,$ O_{k}$ transform as: $$\label{6} O_{k}\left( z,\bar{z} \right)\Longrightarrow\left( {\partial {f}\over \partial z} \right)^{\Delta_{k}}\overline{\left({\partial {f}\over \partial z}\right)^{\Delta_{k}}}~O_{{k}}\left( f,\bar{f} \right)$$ This transformation is generated by the analytic energy-momentum tensors T(z) and $ \bar{T}(z)$ (related to the tensor $ T _{\alpha\beta}\left( x \right) $ as $ T,\bar{T}~=~T_{11}-T_{22}\pm 2iT_ {12})$. Conformal invariance implies that the trace of the energy-momentum tensor is zero. The basic result of \[4\] was that together with primary operators, CFT is bound to contain the so-called secondary or descendent operators. They come from the conformal deformations of the primary ones. In order to describe them it is convenient to introduce the Virasoro algebra: $$\label{7} L_{n}= \oint{~dz~z^{n+1}T\left( z \right) }$$ $$\label{8} [~L_{n}, L_{m}~]~=~\left( n-m \right)L_{n+m}+{c \over 12}n\left( n^{2} -1\right)\delta_{n+m,0}$$ The secondary operators are given by: $ O_{k}^{n_{1}\cdots n_{l},m_{1}\cdots m_{n}}~=L_{-n_{1}}\ldots L_{-n_{l}}\overline{L_{-m_{1}}\ldots L_{-m_{n}}}O_{k}$ The correlation functions of the secondaries are expressed by the differential relations through the correlators of the primaries \[4\]. The second basic feature of CFT is the operator product expansion. It says that $$\label{} O_{k}\left( z \right)O_{l}\left( 0 \right)~=~\sum{f_{kl}^{m}\left( z\bar{z} \right)^{\Delta_{m}-\Delta_{k}-\Delta_{l}}O_{m}\left( 0 \right)+\mbox{secondaries}}$$ The structure constants satisfy simple (in principal) consistency relation coming from the associativity of the algebra. Thus, the classification problem of CFT is reminiscent of the classification problem of the Lie algebras, but is vastly more complicated. However, in \[4\] we have found a simplest set of CFT’s, the so called minimal models. In these models the number of the primaries is finite (but the total number of operators is, of course, infinite). The Ising model is the minimal model of the type (3,4) with only 3 primary operators. They are unit operator I , spin $ \sigma$ and energy density $\varepsilon$. The fusion rules (9) in this case are: $$\begin{aligned} %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR [\sigma]~[\sigma]~=~[I]+[\varepsilon]\\~[\sigma]~[\varepsilon]~=[\sigma]\\~[\varepsilon][\varepsilon]~=~[I]\end{aligned}$$ The brackets in these symbolic formulas mean the “conformal class”, namely the primary operator itself together with all it descendents. Now we are coming to the central point. Let us identify the field $ \varphi$ of the $ \varphi ^{4}$ theory with the field $ \sigma$ of the conformal field theory. We have to examine the quantities\ $ <\varphi^{3 }\left( x \right)~\varphi\left( y_{1} \right ) \cdots\varphi\left( y _{n}\right ) >$. However, the conformal field theory is applicable only when all points are well separated compared to the lattice spacing. The way out of this problem is to introduce the point - splitted definition of the $ \varphi^{3}\left( x \right)$. Let us examine the object: $$\label{} \overline{\lim_{a\rightarrow 0}{}}\sigma\left( x+{a \over 2} \right)\sigma\left( x-{a \over 2} \right)\sigma\left( x \right)\equiv"\sigma^{3}\left( x \right)"\stackrel{\rm def}{=}\Phi\left( x,a \right)$$ The symbol $\overline{lim} $ here means that we first of all average over directions of $a$ and then take $|a| $ to be much smaller then all other distances $ |x-y_{i}|$ and $| y_{i}-y_{j}|$. Keeping it at first much larger than the lattice spacing, we can apply the fusion rules (9). Since $$\label{} [\sigma]~[\sigma]~[\sigma]~=~[\sigma]$$ We obtain: $$\label{} \Phi\left( x,a \right)\approx|a|^{-4\Delta_{\sigma}}[c_{1}\sigma\left( x \right)+c_{2}a^{2}\partial^{2}\sigma\left( x \right)+\ldots]$$ Although this equation is correct only for $ |a|~>>~l$ (where l is a lattice spacing), with any natural definition of the theory it can be extrapolated to $ |a|\sim~l$. The reason is that with any isotropic regularization at the distance we will get instead of (15): $$\label{} \Phi\left( x,a \right)=l^{-4\Delta_{\sigma}}f\left( {|a| \over l} \right)\sigma\left( x \right)+\ldots$$ (where precise form of f depends on the way we regulate the theory). As a result, if we take $ a{\ \lower-1.2pt\vbox{\hbox{\rlap{$<$}\lower5pt\vbox{\hbox{$\sim$}}}}\ }l$ we still have the equation (15), and hence the basic field equation (5) is satisfied. All this means one simple thing -field equations of motion, when read from the right to the left are nothing but operator product expansions. This fact was well understood from the very beginning of the field theory for critical phenomena. Point-splitting and directional averaging over $ a$ is also nothing new - already in quantum electro-dynamics it has been necessary to introduce the isotropic cut-off in the momentum space, which is equivalent to the procedure we discuss. To summarize - in this section we have seen how the structure of the operator product expansion reflects equations, satisfied by the correlation functions. Solutions of the inviscid Hopf equations by the conformal field theory. ======================================================================= Let us turn now to the Hopf equations (4) and try to solve it in the spirit of the proceeding section. In two dimensions it is convenient to introduce vorticity $ \omega$ and the stream function $ \psi$ given by $$\begin{aligned} \label{} v_{\alpha}\left( x \right)~=~e_{\alpha\beta}\ \partial_{\beta}\psi~~~ \omega\left( x \right)~=~e_{\alpha\beta}\partial_{\alpha}v_{\beta}~=~\partial^{2}\psi\end{aligned}$$ They satisfy Navier - Stokes equations: $$\label{} \dot{\omega}~+~e_{\alpha\beta} \partial_{\alpha}\psi~ \partial_{\beta}\partial^{2}\psi~=~\nu \partial^{2}\omega$$ If the stirring force is present, it must be added to the RHS of (18). If we assume, that for the large Reynolds numbers there exist the inertial range of scales in which both viscosity and the stirring force are negligible, we can as a first step examine the inviscid Hopf equation: $$\begin{aligned} \label{} <\dot{\omega} \left( x _{1}\right) ~\omega\left( x_{2} \right)\cdots>~+~<\omega\left( x_{1} \right)~\dot{\omega}\left( x_{2} \right)~\cdots>~+~\cdots~=~0\\ \dot{\omega}\left( x \right)~=~-~"e_{\alpha\beta}\partial_{\alpha }\psi\partial_{\beta}\partial^{2}\psi" \end{aligned}$$ Here, just as in the previous section, we must be careful in defining the correlators at the coinciding points. The reason there have been tied to the existence of the lattice cut-off (where conformal theory fails). The reason here is analogous - we expect that in the momentum space only $ \omega\left( {k } \right)$ with $ {1 \over L}~<~| {k}|~<{1 \over a}$ should satisfy (19). Here a is the ultraviolet cut-off proportional, as we see later, to some power of the viscosity $ \nu$ while L is the infrared cut-off, dictated by the scale of the stirring force. The UV - cut - off in the momentum space means that we must use the point- splitted definition of the product in (20) $$"e_{\alpha\beta}\partial_{\alpha}\psi\left( x \right)\partial_{\beta}\partial^{2}\psi\left( x \right)"=\overline{\lim_{a\rightarrow 0}{}}e_{\alpha\beta}\partial_{\alpha}\psi\left( x+a \right)\partial_{\beta}\partial^{2}\psi\left( x \right)$$ Let us assume now, that $ \psi$ is a primary operator of some yet unknown conformal field theory and compute the RHS of (21) by the use of operator algebra. Suppose that we have the structure: $$\label{} \psi\left( x+a \right)\psi\left( x \right)~=~\left( a\bar{a} \right)^{\Delta_{\phi}-2\Delta_{\psi}}\left\{ \phi\left( x \right)+\mbox{descendents} \right\}$$ Here we introduced the operator $ \phi$ which is the minimal dimension operator in the operator product (22). In unitary theories (as in the Ising model) it is a unit operator. However, only Gibbs states (from which “nothing disappears”) are described by the unitary theories. In turbulence we have “flux states” from which conserved quantities leak away, and it is natural to expect that they will be described by the non-unitary theories. In this case most of operators have negative dimensions and $ \phi$ is non-trivial. It is clear that when the operation (21) is applied to (22), the leading term in the expansion gives zero. This happens because the result must be pseudoscalar and we can’t form it out of $ \phi$ and its derivatives. Hence we have to consider subleading descendents in (22). They have the general structure: $$\begin{aligned} \mbox{descendents}=\sum{c_{\left\{ n \right\}\left\{ m \right\}}}L_{-n_{1}}\ldots L_{-n_{k}}\overline{L_{-m_{1}}\ldots L_{-m_{l}}}\phi\left( x %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR \right)\nonumber\\({a\bar{a}})^{\Delta_{\phi}-2\Delta_{\psi}}{a}^{\sum{n}}{\bar{a}}^{\sum{m}}\end{aligned}$$ Differentiation and directional averaging in (21) the leading term in (23): $$\label{} "e_{\alpha\beta}\partial_{\alpha}\psi\left( x \right)\partial_{\beta}\partial^{2}\psi\left( x %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR \right)"=\mbox{const}({a\bar{a}})^{\Delta_{\phi}-2\Delta_{\psi}}~[L_{-2}\bar{L}_ {-1}^{2}-\bar{L}_{-2}L^{2}_{-1}]~\phi$$ This one of the basic formulas of this work. The reason, why this descendent appeared in (24) is quite simple. The LHS of (24) is pseudoscalar and hence it changes sign under complex conjugation. The lowest descendent which has the same property is the one we find in (24). In deriving (24) we accounted for the UV - divergency, but disregarded the IR one. We will return to the infrared side of the problem below. Let us assume for the moment that IR - divergencies are not present. If so, we derive a rather amazing conclusion from the eq.(24). Namely, there are two possibilities. The first one is $ \Delta_{\phi}~\leq~2\Delta_{\psi}$ (negative:defect of dimensions"). In this case the only way to satisfy (19) would be to assume that the operator $$\label{} \Omega~=~\left( L_{2}\bar{L}^{2}_{-1}~-~\bar{L}_{-2} L^{2}_{-1}\right)~\phi$$ is either zero or is a symmetry of the underlying CFT. The latter statement just stresses the fact that the relation (19) implies that the infinitesimal variation of $ \omega$ , $ \dot{\omega}~=~\Omega$ must not change the correlation functions, or as in the case of decaying turbulence rescale them. It is quite possible that non-trivial examples of such symmetries exist. However if we restrict ourselves with the simple minimal models, we will have to conclude that the RHS of (25) must be zero. This is easy to achieve if we require that the minimal dimension operator $ \phi$ is degenerate on the level two. That means that it satisfies the equation \[4\]: $$\label{} \left( L_{-2} ~-~{3 \over 2\left( 2\Delta_{\phi} +1\right)}L^{2}_{-1}\right) {}~\phi~=~0$$ If so, $ \Omega$ = 0 . The simplest example, in which this relation is satisfied is the minimal model (2 , 5) with $ \Delta_{\phi}=\Delta_{\psi}=-{1 \over 5}$ Even more surprising is the fact, that if $ \Delta_ {\phi}~>~2\Delta_{\psi}$ then no extra restrictions are needed because as $ a~\rightarrow~0$ the RHS of (25) is zero, and the inviscid Hoft equation is satisfied. So, any CFT with the positive defect of dimensions solves the Hopf equations. This is a “turbulent” counter part of the fact that there are continuosly many parasite solutions of the static Euler equations. Let us remember that if we look at the “laminary” solutions of $$\label{} e_{\alpha\beta}\partial_{\alpha}\psi\partial_{\beta}\partial^{2 }\psi~=~0$$ we find that if $ \partial^{2}\psi~=~f\left( \psi \right)$ with arbitrary $ f\left( \psi \right)$ then (27) is satisfied. It is well known, however,that matching with the viscous region by the boundary layer consideration removes this ambiguity. This analogy makes it clear that the crucial point for selecting correct solutions of the inviscid Hopf equations must be matching relations at the UV region of the cut-off momenta, where viscosity takes over. It is also clear, that since we are dealing with negative dimensions, we have to analyze the infrared diver gencies.We can say that in turbulence there are two boundary layers in the “momentum space”, each requiring a separate theory. In the following sections we will discuss some partial results in this direction. The infrared problem ===================== The CFT’s which are most interesting for the turbulence problem contain negative dimensions. That means that the formal correlation functions are positive powers of the distance. It is clear that we have to be very careful in the infrared region and treat it separately. Let us begin with the 2 - point function $ <\psi\left( 0 \right)\psi\left( x \right)>$. When the dimension $ \Delta_{\psi}<0$, CFT gives $$\label{} <\psi\left( 0 \right)\psi\left( x \right) \stackrel{\rm\left( conf \right)}{>}~=~-|x|^{4|\Delta_{\psi}|}$$ This is clearly unphysical. The correct prescription should take into account the fact, that only momenta in the inertial range are described by CFT. So, we should expect that, if we take the Fourier transform of (28) $$\label{} <\psi\left( k \right)~\psi\left( -k \right)>~=~const{1 \over |k|^{2+4|\Delta_{\psi}|}}$$ this result must be used only if $ L^{-1}<|k|<a^{-1}$. We do not know the contribution from the $ |k|\sim L^{-1}$. The only thing which can be said is that when returning to the x space it should give expressions, analytic in $ x^{2}$. So, we have: $$\label{} <~\psi\left( 0 \right)\psi\left( x \right)~>\sim\int_{|k|{\ \lower-1.2pt\vbox{\hbox{\rlap{$>$}\lower5pt\vbox{\hbox{$\sim$}}}}\ }L^{-1}}{e^{ikx}|k|^{-2-4|\Delta_{\psi}|}}\sim\left( c_{1}L^{4|\Delta_{\psi}|} +\ldots \right)-|x|^{4|\Delta_{\psi}|}$$ with the first bracket representing the contribution of the infrared modes. When we turn to the multi - point functions, the situation is essentially the same. When we examine their momentum space form: $$\label{} G\left( k_{1}\cdots k_{N} \right)~=~<\psi\left( k_{1} \right) \cdots\psi\left( k_{N} \right)>~~~ \left( \Sigma k_{j}~=~0 \right)$$ we will assume that the CFT formulas for this quantity work,provided that: $$\label{} L^{-1}<|k_{i}|<a^{-1}$$ and $$\label{} L^{-1}<|k_{i 1} +\cdots+k_{i l}|<a^{-1}$$ The (33) condition means that we have no small momenta transferred in any channel. Again, in the coordinate space all that means that the physical N - point function is equal to the conformal one plus the terms, analytic in some $ \left( x_{i}~-~x_{j} \right)$. Notice, that if we understand the momentum integrals as analytically continued in $ \left( \Delta_{\psi} \right)$ from the region, where they converge, we get precisely conformal answer in the x - space. In order to get the physical correlator within this analytic regularization, one has to add $ \delta$ - functions in the k - space, representing the infrared modes. For instance, within the analytic prescription, we have instead of (29) $$\begin{aligned} \label{} <\psi\left( k \right)\psi\left( -k \right)>~=~{const \over |k|^{2+4|\Delta_{\psi}|}}~+~ a_{1 }L^{4|\Delta_{\psi}|}\delta\left( k \right)+a_{2}L^{4|\Delta_{\psi} |-2} {\partial^{2} \over \partial k^{2}} \delta\left(k \right)~+~\cdots\end{aligned}$$ When Fourier - transformed, (34) gives again (30). So, the convenient way to summarize this situation is to say, that we have some sort of Bose - condensate in the momentum space, formed by the large scale motions. The physical correlators differ from the conformal ones by the condensate terms (we will call them $ \delta $- terms later). The $ \delta$ - terms occur whenever any sum of momenta in the correlator is equal to zero. In the coordinate space they are represented by the terms analytic in $ \left( x_{i}-x_{j} \right)$. Occurrence of the $ \delta $- terms is not surprising , because we have strirring forces acting at k=0. The main problem with them is that we do not know yet how to determine their form from the dynamics. The other related problem, which we will consider now, is whether they destroy our solution of the Hopf equations. The danger comes from the fact that in the sect.3 we dealt with the purely conformal correlators, and this deri vation has to be reconsidered. In order to do that, let us rewrite our basic equation in the momentum space: $$\label{} <\dot{\omega}\left( {q} \right)\omega\left( {f} _{1}\right)\cdots\omega\left( {f}_{n} \right)>~=~ -\int{d^{2}k[k , q]\left( \left( q-k \right)^{2} -{k}^{2}\right)}<~\psi \left( k \right)\psi \left( q-k \right)\omega \left( f_{1} \right)\cdots\omega \left( f_{n} \right)~>$$ Generally speaking, this integral has infrared divergency as $ k\rightarrow0$ (or $ k\rightarrow q $ ). The results of the proceeding section guarantee that if we treat this IR divergency by the analytic regularization, the RHS of (35) will be zero, as before. The problem now is to study what happens when we simply impose the IR cut off in (35). In order to do this, we need to know the behavior of the correlations functions when one of the momenta is much smaller then the others. Let us first evaluate the divergencies at $ k\rightarrow 0$. This limit in the coordinate space is dominated by the large R limit in the Fourier transform: $$\label{} <\psi\left( k \right)~\psi\left( q-k \right)\omega\left( f_{1} \right)\cdots\omega\left( f_{N} \right)>~=~ \int{d^{2}Re^{ikR}}~<\psi\left( R \right)\psi\left( 0 \right)\omega\left( f_{1} \right)\cdots\omega\left( f_{N} \right)>$$ The large R behavior is governed by the fusion rule, where we replace all operators except $ \psi\left( R \right)$ by $ \psi\left( 0 \right)$ .  Using notations of \[4\], we have the formula: $$<\psi\left( R \right)\psi\left( 0 \right)\omega\left( f_{1} \right)\ldots >~=~\sum_{N}{<0|\psi\left( R \right)|\Delta+N><\Delta +N|\psi\left( 0 \right)\omega\left( f_{1} \right)\ldots |0>}$$ where the states $ |\Delta + N>$ correspond to the secondary operators of the $ \psi$. Using the fact that $$\label{} <0|\psi\left( R \right)|\Delta+N>\infty{1 \over R^{2\Delta+N}}$$ we obtain the following leading terms in the asymptotic expansion $$\begin{aligned} \label{} <\psi\left( k \right)\psi\left( q-k \right)\omega\left( f_{1} \right)\cdots\omega\left( f_{N} \right)>\approx{c_{1} \over|k|^{6+\lambda}}~<\Delta| \psi\left( q \right)\omega{f_{1}} \cdots\omega\left( f_{N} \right)|0>\nonumber\\~+c_{2} \left\{ {k_{+} \over |k|^{6+\lambda}} <\Delta|L_{1}\psi \left( q \right)\omega\left( f_{1} \right)\cdots\omega\left( f_{N} \right)\ 0 > + c.c. \right\} +\cdots \end{aligned}$$ When we substitute (39) into (35), the first leading term drops out after integration over directions of k the second term does contribute. Simple computation gives: $$\label{} <\dot{\omega}\left( q \right)~\omega\left( f_{1} \right)\cdots\omega\left( f_{n} \right)>=const L^{2+\lambda} ~ \left\{ q_{+}<\Delta|L_{1}\omega\left( q \right)\cdots\omega\left( f_{n} \right)| >-c.c. \right\}~+~O\left( L^{\lambda} \right)$$ This leading term can be interpreted as fluctuating transport by the large scale eddies. Indeed, the relation (40) is equivalent to the operator equation: $$\label{} \dot{\omega}\left( x \right)=L^{2+\lambda}\hat{u}_{\alpha}\partial_{\alpha}\omega\left( x \right)$$ Here we introduced a new $ x $ - independent operators $ \hat{u}_{\pm}$by the rule: $$\label{} <0|\hat{u}_{+}\omega\left( x_{1} \right)\ldots \omega\left( x_{n} \right)|0>~=~i<\Delta_{\psi}|L_{1}\omega\left( x_{1} \right)\ldots \omega\left( x_{n} \right)|0>$$ and$ \hat{u}_{-}=\hat{u}^{*}_{+}$. It is clear, that although this infrared counterterm makes $ \dot{\omega}\neq 0 $, the Hopf equation is still satisfied, due to the translation invariance. As we go to the order of $ L^{\lambda}$counterterms we find more complicated forms. By the same technic it is possible to show, that the complete divergencies give the following equation: $$\label{} %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR %% FOLLOWING LINE CANNOT BE BROKEN BEFORE 80 CHAR \dot{\omega}=\hat{u}_{\alpha}\partial_{\alpha}\omega~+~\hat{h}_{\alpha\beta}\partial_{\alpha} \partial_{\beta}\psi$$ with the $$\label{} <0|\hat{h}_{++}\omega\left( x_{1} \right)\cdots\omega\left( x_{N} \right)|0>~ \sim ~<\Delta|\left( L_{2}+\lambda L^{2}_{1} \right)\omega\left( x_{1} \right) \cdots\omega\left( x_{N} \right)|0>$$ (where $ \hat{h}_{\alpha\beta}$- is a traceless tensor). This contribution does not cancel in the Hopf equation. There are also other infrared counterterm with this property, occurring from the region in the (45) where: $| k+f_{i1}+ \cdots +f_{il} |\rightarrow 0 $ (for some choice of $ \left\{ f_{i} \right\}$ ). All that means that if we had defined the physical correlation functions by simply cutting of the momentum integrals in the Fourier transform, the Hopf equation would be destroyed. On the other hand, if we had defined all integrals by analytic regularization it would mean the identification of the physical and conformal correlators. In this case the Hopf equations is satisfied, but the positivity is broken. The way out of this problem is to find the $ \delta$ - terms which do not spoil the Hopf equation but restore the positivity. The most blatant violation of positivity is easily cured, if we postulate that any correlation function contains the constant part, coming from the zero modes: $$\label{} <O_{n1} \left( x_{1} \right)O_{n2}\left( x_{2} \right) \cdots O_{nN}\left( x_{N} \right) >^{phys}~=~c_{n1} \cdots n_{N}\left( L^{-2\sum_ {j}\Delta_{nj}} \right)~+~<O_{n1}\cdots O_{nN}>^{conf}$$ (where C - are some constants). This type of $ \delta$- term does not contribute to the Hopf equations at non - zero momenta, and at the same time prevails over the negative conformal contributions, since $ |x_{i}~-~x_{j}|~\ll L$. That doesn’t mean that there could be no other IR counterterms. In fact it is possible to show that there are many types of the $ \delta$ - terms, consistent with the Hopf equations at $ q_{i}\neq0$. In order to determine them we need to include the stirring force terms (which arise at $ q_{i}=0$) and to use the matching relations. This is the problem for the future work. At present it is not quite clear to what extent the $ \delta$- terms are universal i.e. independent of the large scale structures. There is another related issue which we should discuss in this section.[^1] It is the question of the vacuum expectation values of different operators. When defining our conformal contributions, we have implicitly assumed that all these expectation values (VEV) are zero. In general, however we have: $$\label{} <O_{n}\left( x \right)>~=~C_ {n}\cdot{L^{-2\Delta_{n}}}$$ If $ \Delta_{n}~<~0$ , and $c_{n} \neq 0$ these VEV considerably modify correlation functions, as was noticed by Al. Zamolodchikov\[5 \]. The reason for that is simple. From the OPE we get: $$\label{} <\psi\left( r \right)\psi\left( 0 \right)>~=~r^{-4\Delta\psi} <I>~+~r^ {2\left( \Delta_{\phi} -2\Delta_{\psi}\right)}~<\phi>+ \cdots = r^{4|\Delta_{\psi}|}~\left\{ 1+<\phi>\cdot\left( {L \over r} \right)^{2|\Delta_{\phi}|} ~+~\cdots\right\}$$ The second term in (48) is the dominating one. The question, whether VEV are non-zero is determined by the infrared boundary conditions. Our assumption that $ c_{n}~=~0$ is not fundamental. Actually our consideration of the Hopf equations remains intact even if $ c_{n}\neq 0$. The only thing which becomes more complicated is the infrared divergency. Also, the UV - matching of the next section requires modified consideration. The question of the $ c_{n}$ values belongs to the same category of the IR problems which we left for the future work. At present the situation resembles quantum chromodynamics, where we have not solved the IR problem, but can, nevertheless, test its small distance behaviour. Let us summarize. The true physical correlators can be obtained from the conformal ones either by defining the momentum space integrals in the sense of analytic regularization or in the usual sense, with the IR cut - off. In both cases it is necessary to add to these expressions the $ \delta$-terms, containing $ \delta$ functions of certain external or transferred momenta. These terms in the coordinate space contain the analytic dependence on certain distances. Precise form of the $ \delta $ - terms requires for its definition the matching with the stirring force - the task not accomplished in this paper. The Flux States ================ In the above considerations we accounted for viscosity in a somewhat symbolic sense. Namely, we have used the inviscid Hopf equations, but assumed that at the very high momenta our power-like correlation functions start to drop rapidly because of the viscosity, which was assumed as the origin of the ultra- violet cut-off. It is clear that the precise nature of this cut-off must be important. For instance, we can imagine an “elastic” cut-off which preserves the hamiltonian structure and the “inelastic” one which introduces dissipation. It is clear that they should correspond to different physics. In this section we’ ll try to impose the dissipation condition on the theory. The basic idea is roughly the same as in the classical Kolmogorov treatment of turbulence. Namely, we will assume (self-consistently) that integrals of motion are dissipated at the UV cut-off, and in the inertial range of momenta they are just transferred. It implies that the flux in the momentum space of conserved integrals must be constant. One can say that while Gibbs distributions are uniform on the surfaces of fixed values of conserved quantities, the turbulent distributions are located on the surfaces of the constant fluxes of the corresponding quantities. The constant flux in the momentum space is also a familiar object in the other part of physics. It is responsible for the anomalies in the quantum field theory. When one considers, for example, massless fermions in the electromagnetic field the chirality (the number of left minus the number of right Dirac particles), is not conserved due to axial anomaly. This happens because the ultra-violet regularization breaks conservation of the axial current. After the chirality is injected at the UV cut-off, it propagates in the momentum space to the physical region. There is, therefore, a clear and useful analogy with what we are discussing. It has been noticed by Kraichnan \[6\], that in two dimensions the most important flux is that of enstrophy (we will comment on that later). Let us derive the constraint on the theory which follows from the enstrophy conservation. Consider first the enstrophy contained in the modes with definite momentum: $$\label{} h \left( \vec{k} \right) ~=~<\omega \left( \vec{k } \right) ~\omega\left( \vec{-k} \right)>$$ Without viscosity and external force, $ h \left( \vec{k} \right) $would satisfy the continuity equation in the momentum space. When these factors are accounted for, the equation becomes: $$\label{} \dot{h}\left( k \right)+{\partial \over \partial k_{i}}{J^{(h)}_{i}}=\nu k^{2}h\left( k \right)+\Phi\left( k \right)$$ Here $ \Phi\left( \vec{k} \right)$ is a contribution of the external forces which is non-zero only for $ |\vec{k}|\sim {1 \over L}$. Let us notice also, that in isotropic turbulence the only non-zero component of $ \vec{J}^{\left( h \right)}$ is the radial one $ J^{\left( h \right)}\left( k \right)$. Let us now chose the momentum q lying in the inertial range and integrate over $ |\vec{k}|~>~|\vec{q}|$. Since in the steady state $ \dot{h}\left( k \right)~=~0$ we get $$\label{} -J^{(h)}\left( q \right)=~\nu\int_{|k|>|q|}{k^{2}h\left( k \right)d^{2}k}$$ Now comes an important point - suppose that the RHS of (51) is UV- divergent and defined by $ |k|\sim {1 \over a}$. Then the q -dependence of RHS can be neglected. As a result, we obtain: $$\label{} J^{(h)}\left( q \right)\approx - \nu \int{k^{2}h\left( k \right)}d^{2}k=\mbox{const}$$ for $ {1 \over L}\ll|q|\ll{1 \over a}$. The meaning of this constant flux condition is transparent. It just says that enstrophy is dissipated only at $ |q| \sim {1 \over a}$ and hence, the conservation law implies that its flux must be constant in the inertial range. This is one of matching conditions we must impose on the inviscid solutions. If we express $ \dot{\omega\left( \vec{q} \right)}$ by means of eq. (35) we get: $$\label{} J^{(h)}\left( q \right)=~-\int_{|k|>q}{<\dot{\omega \left( k \right)}\omega\left( -k \right)>}=~\int_{|k|<q}{<\dot{\omega\left( k \right)}\omega\left( -k \right)>}$$ (we used here the fact, that (35) conserves total enstrophy). We see from here that all $ \vec{k}- s$ from the inertial range give zero contribution. At the same time, using our general conjecture concerning the infrared terms (46) we obtain: $$\label{} J^{(h)}\left( q \right)\propto {1 \over L^{\Delta_{\dot{\omega}}+\Delta_{\omega}}}$$ The right hand side of (51) is independent of L and thus we obtain the matching condition: $$\label{} \Delta_{\omega}~+~\Delta_{\dot{\omega}}~=~0$$ If we had used the first term in ( 52), where $ k \sim {1 \over L}$ are absent, the IR contribution would come from the integral (35) for $ \dot{\omega}$ , yielding the same result. If we recall, that: $ \Delta_{\omega}~=~\Delta_{\psi}~+~1$ and  $ \Delta_{\dot{\omega}}~=~\Delta_{\phi}+2$ we get a following condition for the enstrophy flux state: $$\begin{aligned} [\psi]~[\psi]~=~[\phi]+\ldots \nonumber\\ ~\Delta_{\psi}+\Delta_{\phi}+3=0\end{aligned}$$ This result heavily depends on our conjectures about IR - contributions, namely on the fact that $ <\omega\dot{\omega}>$ infrared part is non-zero. Had we considered the energy flux state, the formula ( 55) would be replaced by: $$\label{} \Delta_{\psi}~+~\Delta_{\phi}~+~2~=~0$$ Apart from the energy and the enstrophy there are also higher conserved integrals of the type: $$\label{} I_{n}~=~\int{\omega^{n}\left( x \right)~d^{2}x}$$ Before commenting on these integrals, let us explain why the formulas (56) and (57) do not contradict each other. (After all, energy and enstrophy are both conserved in the inviscid system). The apparent contradiction is removed by the fact that in the enstrophy flux state the integral for the energy dissipation, analogous to (50) is given by: $$\label{} J^{(\varepsilon)}\left( q \right)=~\nu \int_{|k|>q}{d^{2}k<\omega\left( k \right)\omega\left( -k \right)>}$$ Contrary to (50), (58) is not UV - divergent and the dissipation is scale - dependent in the inertial range. In this case we can’t use (51) anymore. No obvious matching condition arise in this case, since (58) tells us that as $ \nu\rightarrow 0 $ the energy flux tends to zero, while the enstrophy flux persists. This is essentially the picture, advocated by Kraichnan \[6\]. As we turn to the integrals $I_{n}$ we have to decide, whether they have nonzero flux and whether this gives us any new information. We are unable at this point to give complete analyses of this question and will present some non-rigorous estimates, leaving complete resolution for the future work. We have: $$\label{} \dot{I}_{n}~=~\nu \int{<\nabla ^{2}\omega\left( x \right)\omega^{n-1} \left(x \right)>d^{2}x}$$ The main problem is how to regularize this expression. We must use OPE in order to define $ \omega^{n-1}$ .If n is even, which is imposed by parity conservation, we get $$\label{} \omega^{n-1}\left( x \right)\infty~\Psi\left( x \right)~+~\cdots$$ and as a result: $$\label{} \dot{I_{n}}\propto \int{<~\triangle \omega~~\psi>}\propto J^{(\varepsilon)}$$ We picked up $ \psi$ operator only in(61) since due to orthogonality of primary operators, all others will not contribute. This implies that the $ I_{n}$- transfer is zero in the inviscid limit and there are no extra constraints on the theory coming from it (apart, may be from some inequalities). However, this line of arguments must be considered in more details (with careful point - splitted definition of all parties involved, before the last conclusion could be trusted). In particular,the other groupings of $ \omega$ -s may involve some other operators. Much work here is to be done. An interesting related question is that of the inverse cascades of higher integrals. [^2] Strong infrared divergency in their dissipation makes them very similar to the energy and opens this possibility for $ I_{n}$ . Whether this is the case is an unsolved question. Unfortunately, any kind of the naive pseudophysical arguments is misleading here. To conclude this section, let us discuss the physical picture of the flux state. It is not exactly the Kolmogorov’s one. We have constant flux of conserved quantity through the inertial range. In the Kolmogorov’s mechanism the transfer occurs by the interaction of three modes with all wave vectors lying in the inertial range. Such transfer is absent in our picture. Instead we have the triad interaction in which one of the modes is infrared. We can say that the infrared (large scale) modes serve as catalyzers for the flux. This is an important qualitative prediction of our theory, which can be tested. Possible spectra of conformal turbulence. ========================================= Let us discuss solutions of conformal turbulence with constant enstrophy flux. The analyses of the previous sections indicates that we must start from the CFT, satisfying certain constrains which follow from the Hopf equation and constant flux condition. The first constraint tells us that if we have the fusion rule: $$\label{} [\psi]~[\psi]~=~[\phi]~+~\cdots$$ where $ \phi$ is a minimal dimension operator in the product (62), then it must be either $ \Delta_{\phi}>2\Delta_{\psi}$ or $ \phi$ - operator must be degenerate at the level two. The second constraint depends on the assumption about vacuum expectation values. The simplest version of the theory, when these values are zero gives: $$\label{} \Delta_{\phi}~+~ \Delta_{\psi}~=~-3$$ More options arise if we assume that if $ <O_{n}>~\neq~0$ In this case the correlation functions $ \omega\omega$ and $ \dot{\omega}\omega$ are determined not by the unit operator but by the minimal dimension operators. As a result, if we look at OPE, we find: $$\begin{aligned} <\omega\left( z \right)\omega\left( 0 \right)>\propto \left( \bar{z}z \right)^{\Delta_{\phi}-2\Delta_{\omega}} ~<\phi> \nonumber\\ \Delta_{\omega}=\Delta_{\psi}+1\end{aligned}$$ and $$\label{} <\dot{\omega}~\omega>\propto\left( \bar{z}z \right)^{-\left( \Delta_{\omega}+\Delta_{\dot{\omega}} \right)}\left( \bar{z}z \right)^{\Delta_{\chi}}$$ where $ \chi$ is a minimal dimension operator in the product $ [\psi]~[\psi]~[\psi] .$ So, in this case the constant flux condition is replaced by: $$\label{} \Delta_{\phi}~+~\Delta_{\psi}~-~\Delta_{\chi}~=~-3$$ while the energy spectrum is given by: $$\label{} E\left( k \right)~ \sim~k^{4\Delta_{\psi}-2\Delta_{\phi}+1}$$ One can also consider intermediate possibilities when some of the operators have vacuum averages and some do not. At present we do not have clear idea how the choice should be made. It may depend on the external conditions and the way turbulence is generated. Below we will examine a simplest possible model, the “minimal” minimal model. By that we mean the theory which satisfies our requirements and has the smallest number of primaries. If we associate each primary operator with the certain type of motion then it is indeed the simplest (but by no means the only) type of turbulence. Let us look at the minimal models of the type (2,2N +1), and assume, that vacuum expectation values are absent. In this models we have the set of N primary operators $ [\psi_{s}]$, with $s=1, ~\cdots N$. The fusion rules are: $$\begin{aligned} \label{} [\psi_{s_{1}}]~[\psi_{s_{2}}]~=~[\psi_{s_{3}}] +[\psi_{s_{3}-2}]+ \ldots \nonumber\\ s_{3}=min\left( s_{1}+s_{2}-1,2N+1-\left( s_{1}+s_{2}-1 \right) \right)\end{aligned}$$ The dimensions are given by: $$\label{} \Delta_{s}~=~-{\left( 2N - s \right)\left( s-1 \right) \over 2\left( 2N +1 \right)}$$ If we identify the stream function $ \psi$ with the primary field $ \psi_{S}$ with some S , then the flux condition gives: $$\label{} \Delta_{s}+\Delta_{2s-1}~=~-3$$ (if $ 2\left( 2s - 1 \right)~<~2N+1). $ This diophantine equation has unique solution, s=4; N=10 . Thus, we conclude that the minimal turbulence is described by the (2 ,21) minimal model. Anomalous dimensions in this case are $$\begin{aligned} \label{} \Delta_{\psi}~=~\Delta_{4}~=~-{8 \over 7}\\ \Delta_{\phi}~=~\Delta_{7}~=~-{13 \over7}\end{aligned}$$ As a result we obtain in this case the energy spectrum $$\label{} E\left( k \right)~\infty~k^{-{25 \over7}}$$ It seems to be consistent with observations. \[7\] An interesting question if this model is parity conservation. There is no a priori reason why we should exclude parity breaking solutions from the consideration. Let us look what is the situation in our model. The stream function $ \psi$ is a pseudoscalar. That means that if parity is conserved, the 3 - point function must contain vector products. But in CFT this is impossible: all 3-point functions are simple powers of distances.[^3] Hence, in conformal turbulence the necessary condition for parity conservation is: $$\label{} <\psi~\psi~\psi>~=0$$ This is indeed the case, since: $$\label{} [\psi_4][\psi_4]=[\psi_7]+[\psi_5]+[\psi_3]+[\psi_1]$$ and there is no $ [\psi_4]$ in the RHS. As we look at higher correlations the situation is less clear. Indeed, the 5 - point functions of $ \psi$ are still zero (by the similar arguments) while the 7 -point functions are non-zero. If parity is conserved, they must contain vector products. It is not obvious, whether this can be done, or, in other words, whether we can combine conformal blocks in the 7 - point function in a antisymmetric (under reflections) way. This we leave for the future analyses, noticing only, that if we consider the standard symmetric combinations of the conformal blocks, we get a very peculiar picture in this model. namely, the parity violation, if present at all, is well hidden - one must go to the seven - point functions in order to notice it! Let us point out that parity is automaticaly conserved in the non-minimal models. In this case there exists infinite number of the degenerate operators, forming closed algebra preserving parity \[4\]. Conclusion. =========== There are infinitely many other solutions of the flux conditions, the fixed points with larger number of structures. They have been considered in \[8\]. It is hard to say, whether all matching conditions have been exploited and whether we indeed have infinite number of discrete fixed points. If so, the turbulent fluid has infinite discrete regimes and must be able to jump from one regime to the other, generating new quasistable structures. we are not in a position yet to prove this conjecture, but it is certainly within the scope of our methods. As far as observations are concerned, the best way to check conformal invariance is to study momentum representation of the correlation functions when all the momenta and all their partial sums belong to the inertial range. In this way we avoid the infrared region which is scarcely understood. The situation is analogous to that in quantum chromodynamics - there we know dynamics at high momenta and quite ignorant in the infrared. Nevertheless, QCD has been tested. The theory of conformal turbulence gives explicit and unambiguous predictions concerning these type of correlators. They are expressed in terms of hypergeometric functions, and, more importantly have specific “fusion rule” structure when one momentum is getting much larger than the others. The most direct way to test the theory is to check this fusion property. For example, one takes the 4 - point function and reduce it to the 3 - point function by taking one large momentum. This 3 - point function in the momentum space is a simple hypergeometric function, which becomes just a product of two powers if one does the fusion procedure again. This test is especially important, since we are not certain which solution of conformal turbulence is realized in a given experiment. Since the structure of fusion rules define the theory the test is capable to answer this question. On the other hand, may be the most interesting part of the theory is the infrared one, still hidden from us. With further work it might become possible to get equations, governing the behavior of the vacuum expectation values of $<O_{n}>$ in space and time. This is the problem analogous to the one of equation of state in critical phenomena. In any case, the theory presented here is but a first step in the long future investigations. Acknowledgements ================ I am grateful to A.Migdal, E.Siggia, V.Yakhot and A.Zamolodchikov for many important remarks on this work. I also thank D. Makogonenko for her help in preparation of this article. This work was partially supported by the National Science Foundation under contract PHYS-90-21984. [^1]: I am grateful to A. Zamolodchicov for drawing my attention to it [^2]: I acknowledge useful discussion of this point with G. Falcovich and A. Hanany [^3]: Some time ago A. Migdal conjectured that in 3d turbulence T- invariance forbids the 3-point functions. Whether it is true is an interesting open question. In our case the only thing which can be proved is the combined PT-invariance .

ArXiv

--- author: - Nathalie Bertrand - 'Patricia Bouyer[^1]' - Thomas Brihaye - 'Pierre Carlier$^*$' title: 'When are Stochastic Transition Systems Tameable?' --- [10]{} Parosh A. Abdulla, Noomene [Ben Henda]{}, and Richard Mayr. Decisive [M]{}arkov chains. , 3(4), 2007. Parosh A. Abdulla, Nathalie Bertrand, Alexander Rabinovich, and [ Ph]{}ilippe Schnoebelen. Verification of probabilistic systems with faulty communication. , 202(2):141–165, 2005. Manindra Agrawal, S. Akshay, Blaise Genest, and P. S. Thiagarajan. Approximate verification of the symbolic dynamics of [M]{}arkov chains. In [*Proc. 27th Annual Symposium on Logic in Computer Science (LICS’12)*]{}. [IEEE]{} Computer Society, 2012. Rajeev Alur. . PhD thesis, Stanford University, Stanford, CA, USA, 1991. Rajeev Alur and Mikhail Bernadsky. Bounded model checking for [GSMP]{} models of stochastic real-time systems. In [*Proc. 9th International Workshop on Hybrid Systems: Computation and Control (HSCC’06)*]{}, volume 3927 of [*Lecture Notes in Computer Science*]{}, pages 19–33. Springer, 2006. Rajeev Alur and David L. Dill. A theory of timed automata. , 126(2):183–235, 1994. Christel Baier, Nathalie Bertrand, Patricia Bouyer, [Th]{}omas Brihaye, and Marcus Gr[ö]{}[ß]{}er. Probabilistic and topological semantics for timed automata. In [*Proc. 27th Conference on Foundations of Software Technology and Theoretical Computer Science (FSTTCS’07)*]{}, volume 4855 of [*Lecture Notes in Computer Science*]{}, pages 179–191. Springer, 2007. Christel Baier, Nathalie Bertrand, Patricia Bouyer, [Th]{}omas Brihaye, and Marcus Gr[ö]{}[ß]{}er. Almost-sure model checking of infinite paths in one-clock timed automata. In [*Proc. 23rd Annual Symposium on Logic in Computer Science (LICS’08)*]{}, pages 217–226. IEEE Computer Society Press, 2008. Christel Baier, Boudewijn Haverkort, Holger Hermanns, and Joost-Pieter Katoen. Model-checking algorithms for continuous-time [M]{}arkov chains. , 29(7):524–541, 2003. Christel Baier and Joost-Pieter Katoen. . MIT Press, 2008. Christel Baier and Marta Z. Kwiatkowska. On the verification of qualitative properties of probabilistic processes under fairness constraints. , 66(2):71–79, 1998. Dani[è]{}le Beauquier, Alexander Moshe Rabinovich, and Anatol Slissenko. A logic of probability with decidable model-checking. In [*Proc. 16th International Workshop on Computer Science Logic (CSL’02)*]{}, volume 2471 of [*Lecture Notes in Computer Science*]{}, pages 306–321. Springer, 2002. Mikhail Bernadsky and Rajeev Alur. Symbolic analysis for [GSMP]{} models with one stateful clock. In [*Proc. 10th International Workshop on Hybrid Systems: Computation and Control (HSCC’07)*]{}, volume 4416 of [*Lecture Notes in Computer Science*]{}, pages 90–103. Springer, 2007. Nathalie Bertrand. PhD thesis, [É]{}cole Normale Sup[é]{}rieure de Cachan, Cachan, France, 2006. Nathalie Bertrand, Patricia Bouyer, [Th]{}omas Brihaye, and Pierre Carlier. Analysing decisive stochastic processes. In [*Proc. 43rd International Colloquium on Automata, Languages and Programming (ICALP’16)*]{}, volume 55 of [*LIPIcs*]{}, pages 101:1–101:14. Leibniz-Zentrum f[ü]{}r Informatik, 2016. Nathalie Bertrand, Patricia Bouyer, [Th]{}omas Brihaye, and Nicolas Markey. Quantitative model-checking of one-clock timed automata under probabilistic semantics. In [*Proc. 5th International Conference on Quantitative Evaluation of Systems (QEST’08)*]{}, pages 55–64. IEEE Computer Society Press, 2008. Nathalie Bertrand, Patricia Bouyer, Thomas Brihaye, Quentin Menet, Christel Baier, Marcus Gr[ö]{}[ß]{}er, and Marcin Jurdzi[ń]{}ski. Stochastic timed automata. , 10(4):1–73, 2014. Patricia Bouyer, Thomas Brihaye, Marcin Jurdzi[ń]{}ski, and Quentin Menet. Almost-sure model-checking of reactive timed automata. In [*Proc. 9th International Conference on Quantitative Evaluation of Systems (QEST’12)*]{}, pages 138–147. IEEE Computer Society Press, 2012. Tom[á]{}š Br[á]{}zdil, Jan Krč[á]{}l, Jan Křet[í]{}nsk[ý]{}, and Vojtěch [Ř]{}eh[á]{}k. Fixed-delay events in generalized semi-[M]{}arkov processes revisited. In [*Proc. 22nd International Conference on Concurrency Theory (CONCUR’11)*]{}, volume 6901 of [*Lecture Notes in Computer Science*]{}, pages 140–155. Springer, 2011. Pierre Brémaud. . Springer, 1999. Taolue Chen, Marco Diciolla, Marta Z. Kwiatkowska, and Alexandru Mereacre. Time-bounded verification of [CTMCs]{} against real-time specifications. In [*Proc. 9th International Conference on Formal Modeling and Analysis of Timed Systems (FORMATS’11)*]{}, volume 6919 of [*Lecture Notes in Computer Science*]{}, pages 26–42. Springer, 2011. Vincent Danos, Jos[é]{}e Desharnais, and Prakash Panangaden. Conditional expectation and the approximation of labelled [M]{}arkov processes. In [*Proc. 14th International Conference on Concurrency Theory (CONCUR’03)*]{}, volume 2761 of [*Lecture Notes in Computer Science*]{}, pages 468–482. Springer, 2003. Pedro R. D[’A]{}rgenio and Joost-Pieter Katoen. A theory of stochastic systems [P]{}art [I]{}: [S]{}tochastic automata. , 203(1):1–38, 2005. Jos[é]{}e Desharnais and Prakash Panangaden. Continuous stochastic logic characterizes bisimulation of continuous-time [M]{}arkov processes. , 56:99–115, 2003. Ernst-Erich Doberkat. . Chapman & Hall/CRC, 2007. Martin Fr[ä]{}nzle, Ernst Moritz Hahn, Holger Hermanns, Nicol[á]{}s Wolovick, and Lijun Zhang. Measurability and safety verification for stochastic hybrid systems. In [*Proc. 14th International Conference on Hybrid Systems: Computation and Control (HSCC’11)*]{}, pages 43–52, 2011. Martin Fr[ä]{}nzle, Holger Hermanns, and Tino Teige. Stochastic satisfiability modulo theory: [A]{} novel technique for the analysis of probabilistic hybrid systems. In [*Proc. 11th International Workshop on Hybrid Systems: Computation and Control (HSCC’08)*]{}, pages 172–186, 2008. Daniel Gburek, Christel Baier, and Sascha Kl[ü]{}ppelholz. Composition of stochastic transition systems. In [*Proc. 43rd International Colloquium on Automata, Languages and Programming (ICALP’16)*]{}, volume 55 of [*LIPIcs*]{}, pages 102:1–102:15. Leibniz-Zentrum f[ü]{}r Informatik, 2016. Peter W. Glynn. A [GSMP]{} formalism for discrete event systems. , 77(1):14–23, 1989. Erich Grädel, Wolfgang Thomas, and [Th]{}omas Wilke, editors. , volume 2500 of [*Lecture Notes in Computer Science*]{}. Springer, 2002. Thomas A. Henzinger. The theory of hybrid automata. In [*Proceedings, 11th Annual [IEEE]{} Symposium on Logic in Computer Science, New Brunswick, New Jersey, USA, July 27-30, 1996*]{}, pages 278–292. [IEEE]{} Computer Society, 1996. Andr[á]{}s Horv[á]{}th, Marco Paolieri, Lorenzo Ridi, and Enrico Vicario. Transient analysis of non-[M]{}arkovian models using stochastic state classes. , 69(7-8):315–335, 2012. S. Purushothaman Iyer and Murali Narasimha. Probabilistic lossy channel systems. In [*Proc. 7th International Joint Conference on Theory and Practice of Software Development (TAPSOFT’97)*]{}, volume 1214 of [*Lecture Notes in Computer Science*]{}, pages 667–681. Springer, 1997. Vijay Anand Korthikanti, Mahesh Viswanathan, Gul Agha, and YoungMin Kwon. Reasoning about [MDP]{}s as transformers of probability distributions. In [*Proc. 7th International Conference on Quantitative Evaluation of Systems (QEST’10)*]{}, pages 199–208. IEEE Computer Society Press, 2010. YoungMin Kwon and Gul Agha. Linear inequality [LTL]{} [(iLTL)]{}: [A]{} model checker for discrete time [M]{}arkov chains. In [*Proc. 6th International Conference on Formal Engineering Methods (ICFEM’04)*]{}, volume 3308 of [*Lecture Notes in Computer Science*]{}, pages 194–208. Springer, 2004. Fran[ç]{}ois Laroussinie, Nicolas Markey, and [Ph]{}ilippe Schnoebelen. Model checking timed automata with one or two clocks. In [*Proc. 15th International Conference on Concurrency Theory (CONCUR’04)*]{}, volume 3170 of [*Lecture Notes in Computer Science*]{}, pages 387–401. Springer, 2004. Marco Ajmone Marsan, Gianni Conte, and Gianfranco Balbo. A class of generalized stochastic [P]{}etri nets for the performance evaluation of multiprocessor systems. , 2(2):93–122, 1984. Prakash Panangaden. Measure and probability for concurrency theorists. , 253(2):287–309, 2001. Prakash Panangaden. . Imperial College Press, 2009. Marco Paolieri, Andr[á]{}s Horv[á]{}th, and Enrico Vicario. Probabilistic model checking of regenerative concurrent systems. , 42(2):153–169, 2016. Amir Pnueli. On the extremely fair treatment of probabilistic algorithms. In [*Proc. 15th Ann. Symp. Theory of Computing (STOC’83)*]{}, pages 278–290. ACM Press, 1983. Amir Pnueli and Lenore D. Zuck. Probabilistic verification. , 103(1):1–29, 1993. Alexander Rabinovich. Quantitative analysis of probablistic lossy channel systems. , 204(5):713–740, 2006. Sadegh Soudjani, Rupak Majumdar, and Alessandro Abate. Safety verification of continuous-space pure jump [M]{}arkov processes. In [*Proc. 22nd International Conference on Tools and Algorithms for the Construction and Analysis of Systems (TACAS’16)*]{}, volume 9636 of [ *Lecture Notes in Computer Science*]{}, pages 147–163. Springer, 2016. Moshe Y. Vardi. Automatic verification of probabilistic concurrent finite-state programs. In [*Proc. 26th Annual Symposium on Foundations of Computer Science (FOCS’85)*]{}, pages 327–338. IEEE Computer Society Press, 1985. Moshe Y. Vardi and Pierre Wolper. Reasoning about infinite computations. , 115(1):1–37, 1994. [**– Appendix –**]{} Former results already stated in the core of the paper are put in a box. New results are normally stated without box. [^1]: Supported by ERC project EQualIS

ArXiv

--- abstract: 'Keeping the stability can be counted as the essential ability of a humanoid robot to step out of the laboratory to work in our real environment. Since humanoid robots have similar kinematic to a human, humans expect these robots to be robustly capable of stabilizing even in a challenging situation like while a severe push is applied. This paper presents a robust walking framework which not only takes into account the traditional push recovery approaches (e.g., ankle, hip and step strategies) but also uses the concept of Divergent Component of the Motion (DCM) to adjust next step timing and location. The control core of the proposed framework is composed of a Linear-Quadratic-Gaussian (LQG) controller and two proportional controllers. In this framework, the LQG controller tries to track the reference trajectories and the proportional controllers are designed to adjust the next step timing and location that allow the robot to recover from a severe push. The robustness and the performance of the proposed framework has been validated by performing a set of simulations, including walking and push recovery using . The simulation results verified that the proposed framework is capable of providing a robust walking even in very challenging situations.' author: - | Mohammadreza Kasaei, Nuno Lau and Artur Pereira\ IEETA / DETI University of Aveiro 3810-193 Aveiro, Portugal\ {mohammadreza, nunolau, artur}@ua.pt bibliography: - 'IROS2019.bib' title: '**A Hierarchical Framework to Generate Robust Biped Locomotion Based on Divergent Component of Motion**' --- **Keywords:** Biped locomotion, Robust walk engine, Divergent component of motion, Linear inverted pendulum, Linear-Quadratic-Gaussian (LQG). Introduction {#sec:introduction} ============ Unsafeness of humanoid robots is the main reason for preventing these robots from stepping out of the laboratory and adapting to our environment. Since some decades ago, many types of research have been conducted and the capability of humanoid robots has been much improved to perform stable walking but it does not still satisfy human expectations. The question is *how is it that a human is adept at performing walking even in very challenging conditions but humanoid robots are only capable of performing slow walking?* Generating a fast and stable walking for humanoid robots is a multidisciplinary and complex subject due to the naturally unstable dynamics of these types of robots. To reduce the complexity of this subject, it is decomposed into several small independent modules generally. Then these modules will be connected hierarchically to generate a walking system. A common decomposition approach is using a simplified physical model of the robot’s dynamics and developing a walking system based on this model. In this approach, the overall structure of the walking system is decoupled into four hierarchy levels which are footstep planner, reference generators, push recovery and low-level controller. Using this structure has gained great attention because of reducing the complexity, increasing the flexibility and also the portability of the walking system [@kasaei2017reliable]. Linear Inverted Pendulum Model (LIPM) [@kajita2003biped] is one of the successful simplified physical models among the others which is capable of generating a feasible reference trajectory of the Center Of Mass (COM). This model restricts the overall dynamics into COM and it is able to generate a fast, efficient and feasible trajectory for the COM movement according to a set of pre-planned footsteps. Using this model, the desired trajectory of COM can be generated based on one of the analytical solution, preview control or Differential Dynamic Program (DDP), which are appropriate for real-time implementation [@kajita2003biped]. After generating the reference trajectories, the major task of the low-level controller is tracking the reference trajectories and compensating the tracking error to keep the stability of the robot. Indeed, the controller tries to control either the Zero Moment Point (ZMP) or Divergent Component of Motion (DCM) [@englsberger2015three]. The low-level controller has been formulated successfully using classical feedback controllers, Linear Quadratic Regulator (LQR)-based methods and also Model Predictive Control (MPC). ZMP-based controllers try to keep the robot’s stability by keeping the ZMP inside the support polygon. DCM-based approaches split the LIPM dynamics into stable and unstable parts and just by controlling the unstable part, keep the stability of the robot [@takenaka2009real]. Although both methods can handle small and normal tracking errors, in case of large disturbance, ZMP-based methods are not able to handle the large tracking error due to the limited size of the support polygon. DCM-based methods try to handle such situations by changing the landing location of the swing leg. In this paper, we tackle the problem of designing a robust walking framework based on DCM which takes into account adjusting the next step timing and location. The fundamental component of the proposed framework is a Linear Quadratic Gaussian (LQG) which not only can optimally track the reference trajectory in the presence of noise and disturbances but also can illuminate the steady-state error. Furthermore, by measuring the DCM error at each control cycle, two proportional controllers are designed to adjust the landing time and location of the swing leg to increase the withstanding level of the robot. The remainder of this paper is structured as follows: Section \[sec:related\_works\] gives an overview of related work. In Section \[sec:arct\], the overall architecture of the proposed system is presented and the functionality of each module is explained and verified. In Section \[sec:simulation\], two simulation scenarios are designed to verify the performance of the proposed system. Based on the simulation results, discussion and comparison will be given in Section \[sec:DISCUSSIONS\]. Finally, conclusions and future research are presented in Section \[sec:CONCLUSION\]. Related Work {#sec:related_works} ============ Many types of research have been conducted to realize the full potential of biped robots and they showed a robust balance recovery is an essential part of a biped walking system. DCM-based approaches are one of the successful and appropriate methods to develop a robust walking controller that can react against unpredictable external pushes. These approaches decouple the dynamics of COM into divergent and convergent components and, just by controlling the divergent component try to keep the stability of the robot [@takenaka2009real; @englsberger2015three; @englsberger2017smooth; @kamioka2018simultaneous]. Some of these researches will be reviewed briefly in the remainder of this section. Pratt et al. [@pratt2006capture] proposed an extended version of LIPM, which is composed of a flywheel to consider the momentum around the COM. According to this dynamics model, they defined the Capture Point (CP) concept that is a point on the ground where the robot should step to keep its stability. Using their model, Stephens [@stephens2007humanoid] determined a decision surface to responses to external disturbances. The decision surface breaks the recovery reactions into three particular strategies which were ankle, push and step strategy. Based on this decision surface, in case of small disturbances, ankle strategy is used to compensate for the error using shifting the Center of Pressure (COP) to apply more ankle torques. However, the COP is limited by the size of support polygon, the ankle strategy is not useful in the case of larger disturbances. In the case of a large disturbance, robot should use the hip and waist joints to generate angular momentum around the COM for stabilizing the states of the system (hip strategy). Similar to the ankle strategy, the maximum feasible angular momentum is limited and in the case of very large disturbances, robot should take steps to keep its stability. Morisawa et al. [@morisawa2014biped] proposed online walking generators with push recovery by utilizing the CP concept which was based on a PID controller. The performance of their method has been verified using performing a set of simulations with the  humanoid robot. Simulation results showed that the robot could perform walking on the uneven terrain while keeping its stability. Hopkins et al. [@hopkins2014humanoid] released the height constraint of LIPM and considered a dynamics model based on time-varying DCM and showed how a generic COM height trajectories could be generated by modifying the natural frequency of the DCM during stepping. They designed a walking on an uneven terrain scenario in a simulation environment to validate the performance of their method. The simulation results verified the capability of their method. Englsberger et al. [@englsberger2015three] proposed the 3D version of CP and introduced the Enhanced Centroidal Moment Pivot point (eCMP) and the Virtual Repellent Point (VRP). They showed how eCMP and VRP could encode the magnitude, direction of the external impact. They designed a closed-loop motion tracking control verified the robustness of the controller with regarding the several uncertainties using performing a set of simulations and also real experiments. Kryczka et al. [@kryczka2015online] proposed an online biped locomotion planner based on a nonlinear optimization technique that is capable of modifying the next step position and timing to keep the stability during walking. Their approach has been validated using performing some real experiments on the real humanoid platform COMAN. The results showed that online modifying the step position and timing could increase the recovery level of the robot in the face with external disturbances. Khadiv et al. [@6930e2c2955c4eb688fa0507e3bc2032] combined step location and timing adjustment together to generate robust gaits. Their approach is composed of two main stages. The first stage is responsible for specifying the nominal step location and step duration at the beginning of each step. In the second stage, the landing location and time of the swing leg will be modified at each control cycle according to the measured DCM using an optimization method. The performance of their method has been verified using different simulation scenarios. The simulation results showed that step time adjustment improved the robustness. Griffin et al. [@griffin2016model] designed an MPC to generate a stable dynamics walking based on a time-varying DCM. They used step positions and rotations as the control inputs and considered some reachability constraints to ensure that the generated step positions are kinematically reachable. They demonstrated the performance of their method by performing fast and stable walking with footstep adjustment using the simulated ESCHER humanoid robot. Shafiee-Ashtiani et al. [@shafiee2017robust] also showed a robust walking could be formulated as an MPC based on time-varying DCM concept. [c]{} \[fig:overall\] Kamioka et al. [@kamioka2017dynamic] proposed a dynamic gait re-planing method by applying the cyclic gait criterion based on the DCM. Their method does not only take footsteps positions and its timing into account but also it considers the locomotion mode (e.g., walking, running, and hopping). They used an approximated gradient descent algorithm to modify of footsteps and timing. Their method has been validated using real robot experiments. Later, in [@kamioka2018simultaneous], the authors proposed a new quadratic programming (QP) approach based on an analytical solution of DCM. By performing several real experiments, they showed that their method is capable of compensating disturbances in a hierarchical strategy. Jeong et al. [@jeong2017biped] introduced a closed-loop foot placement controller based on CP, which is capable of generating the desired ZMP according to the current CP error. They try to stabilize walking using developing disturbance adapting walking pattern generator based on CP, an ankle torque reference generator and also by adjusting the next footstep position and timing. Their method has been validated using performing a set of simulations in the Choreonoid simulator with a model of DRC-HUBO+. Architecture {#sec:arct} ============ In this section, we summarize the components of our walking system presented in Fig. \[fig:overall\]. Particularly, the proposed system is composed of four main modules which are explained in the remainder of this section. Walking State Machine --------------------- Due to the periodic nature of walking, it can be modeled by a state machine that state transitions are occurred based on an associated timer and also the conditions of the states. This state machine controls the overall process of the framework and composes of four main states which are depicted in Fig. \[fig:overall\]. In the *Idle state*, the robot stops and waiting for a start walking command. Once a walking command received, the state transits to the *Initialize state* and the robot tries to shift its COM to the first support foot to be ready for taking the first step. In *Single Support State* and *Double Support State*, walking trajectories are generated. It should be mentioned that the associated timer will be reset at the end of the double support state. Dynamic Planners ---------------- This module is responsible for generating all the walking reference trajectories. This process is started by planning a set of footprints according to some constraints (e.g., maximum step size, the distance between feet) and the input command which can be either walking velocity or given step info parameters. After planning the footprints, the step time planner, plans the time of each step according to the input velocity and also the step time adjuster module. This planner determines the total time of each step, including the time of single and double support phases. According to the generated footprints and step times, the trajectories of ZMP and swing leg will be determined by the ZMP planner and the swing leg planner, respectively. The ZMP planning procedure is determined using the following formulation: $$p= \begin{cases} f_{i} \qquad\qquad\qquad\qquad\qquad 0 \leq t < T_{ss} \\ f_{i}+ \frac{SL \times (t-T_{ss})}{T_{ds}} \qquad\qquad T_{ss} \leq t < T_{ss}+T_{ds} \end{cases} , \label{eq:zmpEquation}$$ where $p = [p_x \quad p_y]^\top$ is the generated ZMP, $t, T_{ss}, T_{ds}$ represent the time, duration of single and double support phases, respectively. $SL$ is the length of the step, $f_{i} = [f_{i,x} \quad f_{i,y}] \quad i \in \mathbb{N}.$ represents the planned foot prints on a 2D surface. As is mentioned before, $t$ will be reset at the end of each step . In order to generate the swing trajectory, a Bezier curve is used to move the swing leg smoothly during lifting and landing based on a set of input parameters which are the maximum swing height parameter ($z_s$), the generated foot prints and the step times. In our target framework, LIPM is used as our template dynamics model for generating the trajectory of COM and DCM. In the remainder of this section, it will be briefly reviewed and we will explain how the analytical solution of LIPM will be used to generate the reference trajectories of COM and DCM. LIPM considers some assumptions like restricting the vertical motion of COM to simplify and approximate the dynamics model of a humanoid robot using a first-order stable dynamics model as follows: $$\ddot{c} = \omega^2 ( c - p) \quad , \label{eq:lipm}$$ where $c = [x_c \quad y_c \quad z_c]^\top$ is the position of COM, represents the natural frequency of the pendulum, $p$ is the ZMP position. As is explained before, the reference trajectory of ZMP is already generated and it can be used to determine the position of the COM at the start and end of each step. Accordingly, by considering these positions as the boundary conditions for the (\[eq:lipm\]), it can be solved analytically as a boundary value problem and the trajectory of COM is determined as follows: $$\label{eq:com_traj_x0xf} \resizebox{0.95\linewidth}{!} {$ c(t) = p + \frac{ (p-c_f) \sinh\bigl((t - t_0)\omega\bigl)+ (c_0 - p) \sinh\bigl((t - t_f)\omega\bigl)}{\sinh((t_0 - t_f)\omega)}$ },$$ where $t_0$ and $t_f$ represent the starting and the ending time of a step, $c_0$, $c_f$ are the corresponding positions of COM at these times, respectively. After determining the COM trajectory, the DCM reference trajectory should be generated. Conceptually DCM defines a point that the robot should step to rest over the support foot [@pratt2006capture] and it’s dynamics is defined as follows: $$\zeta = c + \frac{\dot{c}}{\omega} \quad , \label{eq:dcm}$$ where $\zeta = [\zeta_x \quad \zeta_y]^\top$ represents the 2D DCM, $\dot{c}$ is the velocity of COM and $\omega$ is the natural frequency of the pendulum. To validate the performance of the planners in this module, an exemplary 6-steps walking has been planned. In this example, a simulated robot should walk diagonally with a step length ($SL$) of 0.5$m$ (equal in both X and Y directions), duration of single support ($T_{ss}$) and double support ($T_{ds}$) are considered to be 0.8$s$ and 0.2$s$, respectively. The maximum swing leg height ($z_s$) is assumed to be 0.025$m$ and the height of COM ($z_c$) is 1$m$. The results of the walking scenario are shown in Fig. \[fig:exPlanner\]. As results showed, this module is capable of generating the walking reference trajectories according to the input parameters. \[exPlanner\] [c c]{} &\ & \[fig:exPlanner\] Low Level Controller -------------------- The task of this module is to keep tracking the generated references even in the presence of noise and uncertainties. In order to develop this controller, LIPM and DCM are used to design a state-space system. According to the (\[eq:dcm\]), by taking derivative from both sides of this equation and then by substituting (\[eq:lipm\]) into the obtained result, a linear dynamics system will be obtained which can be represented in a state-space system as follows: $$\frac{d}{dt} \begin{bmatrix} c \\ \zeta \end{bmatrix} = \begin{bmatrix} -\omega & \omega \\ 0 & \omega \\ \end{bmatrix} \begin{bmatrix} c \\ \zeta \end{bmatrix} + \begin{bmatrix} 0 \\ -\omega \end{bmatrix} p \quad , \label{eq:statespace_zeta}$$ as this system shows, the COM does not need to be controlled and it is always converged to the DCM. Hence, just by controlling the DCM, the system can always be stable. In this dynamics system, COM and DCM are the states of the system and they are considered to be measurable at each control cycle. Based on this system, a Linear-Quadratic-Gaussian (LQG) controller is designed which is able to track the references robustly. As shown in the low-level controller module of Fig. \[fig:overall\], the controller is composed of a Kalman Filter (KF) to estimate the state of the system and eliminate the effect of measurement noise. Moreover, using the measurement of the system’s state, an integrator is used to cancel the steady-state error. The fundamental component of this controller is an optimal state-feedback gain which is designed as follows: $$u = -K \begin{bmatrix} \tilde{X} - X_{des}\\ X_i \end{bmatrix} \quad ,$$ where $\tilde{X}$ is the output of the KF, $X_{des}, X_i$ represent the desired trajectories and the output of the integrator, respectively. $K$ is a gain matrix which defines the optimal action according to the measured and the desired state. Indeed, this gain should be designed such that the controller is capable of tracking the reference with minimizing the following cost function: $$J(u) = \int_{0}^{\infty} \{ \phi^\intercal Q \phi + u^\intercal R u \} dt \quad ,$$ where $\phi = [\tilde{X} \quad X_i]^\intercal $, $Q$ and $R$ are selected based on trial and error commonly in order to balance the tracking performance and cost of control effort. It should be noted that a direct solution exists to find the $K$ matrix according to the $Q$ and $R$. To verified the performance and the robustness of the proposed controller, a simulation has been carried out. In this simulation, a six-step forward walking trajectories is generated ($SL_x$ = 0.5$m$, $SL_y$ = 0.0$m$, $T_ss = 1$, $T_ds =0$) and the controller should track the reference trajectories in presence of measurement noise. To examine the performance of the system, the state of the system and also real ZMP have been recorded while the simulated robot is walking (sampling rate = 500 Hz). The simulation results are depicted in Fig. \[fig:excontrol\]. The results showed that the controller could track the references even in the presence of noise. \[excontrol\] [c c]{} &\ & \[fig:excontrol\] Next Step Adjusters ------------------- In some situations like when a strong push is applied to the robot, the ZMP goes outside the support polygon and the low-level controller can not regain it back just by applying the compensating torques at the ankle and at the hip due to the saturation of these torques. Human combines two main strategies to cope with these situations: *changing the landing location of swing leg* and *adjusting the step time*. In our structure, these strategies have been developed in the *Next Step Adjusters* module. To implement these strategies, the current measurement of DCM is used as an initial condition for (\[eq:dcm\]) and this equation can be solved as an initial value problem to predict the landing position of the swing leg ($f_p$). The difference between the prediction and the next planned foot position is used as an error variable for the controllers of this module and it is defined as follows: $$\Delta f = \underbrace{(\zeta_t - f_i)e^{w(T_{ss}-t)}}_{f_p}- f_{i+1} \quad . \label{eq:dcm_delta_f}$$ A proportional controller is designed based on $\Delta f$ in order to adjust the next footstep location: $$\delta p = -k_{sa}\Delta f \quad, \label{eq:controller}$$ where $k_{sa}$ is the controller gain, $\delta p$ is the next step adjustment output which should be added to the output of the footstep planner in *Dynamic Planners* module. It should be mentioned that a compliance margin is defined to avoid unnecessary adjustments. Moreover, the output of the footstep planner is saturated in order to prevent planning a footstep outside the kinematically reachable area for the robot. Due to the speed limitation of the swing leg, this strategy is useful if the robot has enough time for reaching to the new landing location. In some cases, like while the robot is in a constrained environment or when it does not have enough time to change the landing location, the step time adjustment can be used as a recovery strategy. In such situations, human either decreases the step time to rapidly put its foot down or increases it to regain its stability. This strategy can be combined with a step adjusting strategy to improve recovery performance. According to the (\[eq:dcm\_delta\_f\]), step time and DCM are exponentially related together. Therefore, if we consider $T_{ss} = T_{ss}+\Delta t$, the step time adjustment ($\Delta t$) can be calculated using the following equation: $$\Delta t = \frac{1}{\omega} \log_e (\frac{\Delta f - f_{i+1}}{\zeta_t - f_i}) + t-T_{ss} \quad, \label{eq:steptime}$$ it should be noted that a first-order lag filter is used to avoid supper quick change in the step time adjustment because it can cause discontinuities in the generated swing leg position: $$T_{ss} = T_{ss}(1-k_f) + (T_{ss}+\Delta t)k_f\quad, \label{eq:lagfilter}$$ where $k_f$ specifies the effect of the step time changing in the current step time and it is determined by trials and error generally. \[dcmscheme\] [c]{} \[fig:dcmscheme\] Simulation {#sec:simulation} ========== In this section, a set of simulation scenarios has been designed and carried out to validate the performance of the proposed framework. The simulations have been performed using a simulated robot which is developed in . The physical property of the simulated robot is summarized in the following table: mass height of COM foot length foot width -------- --------------- ------------- ------------ 30$kg$ 1$m$ 0.15$m$ 0.075$m$ : Physical parameters of the simulated robot. \[tb:pushParams\] It should be noted that since the equations in sagittal and frontal planes are equivalent, the simulation results will be shown just in the sagittal plane. Scenario1: Keeping the stability while standing in single foot: --------------------------------------------------------------- The goal of this scenario is examining the performance of the low-level controller in regaining the stability of the robot during the single support phase. In this scenario, the simulated robot is considered to be stand in the single foot with a specified initial condition ($c_{x_0} $, $\dot{c}_{x_0}$) and the controller should control the states to return into ($0$,$0$) in maximum two seconds. Each simulation is started by selecting an initial condition over the range of \[-0.2 0.2\] at interval 0.02 $m$ for the position ($c_{x_0}$) and \[-1 1\] at interval 0.1 $m/s$ for the velocity ($\dot{c}_{x_0}$). According to these ranges, 441 simulations were performed and the results are graphically depicted in Fig. \[fig:push\_hip\_ankle\]. The left plot of this figure represents an overall result of the simulations. As is shown in this plot, if the initial state is selected from the green region, the controller can keep the stability; otherwise, the robot falls down. The right plots represent a successful simulation (green curve) and an unsuccessful simulation (red curve). \[push\_hip\_ankle\] [c c]{} &\ (a)&(b) \[fig:push\_hip\_ankle\] Scenario2: Keeping the stability while walking in place: -------------------------------------------------------- [c c c c]{} & & &\ & & &\ &&&\ [c c c c]{} & & &\ & & &\ &&&\ \[fig:robust\_ext2\] This scenario is focused on verifying the performance of the *Next Step Adjusters* module. In this simulation, while the simulated robot is walking in place, a severe push with impact duration $10ms$ is applied to the robot’s COM at . The simulation is started by performing some trial and error to find the maximum amplitude of impact that the robot could keep the stability just by applying compensating torques. The simulation results showed that at $F_x=325N$ robot could not regain its stability and fall down (see the first column of Fig. \[fig:robust\_ext\]). After determining this value, firstly, the step time adjuster has been disabled, and the controller should keep the stability just by adjusting the next step location. This simulation has been repeated three times with different amplitudes of the impacts to find the maximum level of withstanding. It should be noted that during the simulations, the reference and measured ZMP, COM, DCM and the output of the controller have been recorded to analyze the behavior of the controller. The simulation results of this scenario are depicted in Fig. \[fig:robust\_ext\]. The plots in each column represent all the actual and the reference trajectories. As is shown in the second column of this figure, after applying $F_x=350N$, the proposed controller changes the landing location of the swing leg to $0.71m$ to keep the stability. The simulation has been repeated with more severe impact ($F_x=400N$) and the results are depicted in the third column. As the results showed, the controller modified the landing location of the swing leg to $0.92m$ and could regain stability. According to the kinematic limitation of the simulated robot, the maximum step length that the simulated robot can take is $0.95m$. Hence to find the maximum impact that can be handled by step adjustment, the amplitude of the impact has been increased while the robot falls down. The results showed that $F_x=411N$ is the maximum level of impact that the simulated robot could handle using the proposed step adjustment strategy. According to the simulation results, the next step adjustment strategy improves the withstanding level of the robot up to 26%. After evaluating the capability of the next step adjustment strategy, the step time adjustment has been enabled and the simulation has been repeated four times with more severe impacts. In these simulations, the output of the step time adjuster is considered to be saturated at $\pm 0.2s$ which means the step time adjuster can increase or decrease the step time up to $0.2s$. The simulation results are shown in Fig. \[fig:robust\_ext2\]. The last simulation has been repeated ($F_x=412N$) to check the effectiveness of the step time adjustment. The simulation results are shown in the plots of the first column. The results showed that the step time adjuster decreased the step time by $0.12s$ and could keep the stability. The amplitude of the impact has been increased and the simulation has been repeated ($F_x=460N$). According to the simulation results, which are depicted in the plots of the second column, the simulated robot could regain its stability by decreasing $0.15s$ the step time. The amplitude of the impact has been increased and the simulation has been repeated until the output of the time step adjuster was not saturated. The simulation results of $F_x=515N$ and $F_x=530N$ are depicted in third and fourth columns, respectively. The simulation results showed that at $F_x=532N$, the output of the time step adjuster is saturated and robot could not keep its stability. The simulation results showed that adjusting step time improves the withstanding level of the simulated robot up to 29%. According to the simulation results, the *Next Step Adjuster* module improved the overall withstanding level of the is improved up to 63%. DISCUSSIONS {#sec:DISCUSSIONS} =========== Most of the presented works in the literature are based on online optimization methods (e.g., QP, MPC) and due to iterative nature of these algorithms, their performance is sensitive to the computation power of the resources. In comparison with those works, the most important property of the proposed system is removing the online optimization of MPC and QP without reducing the adaptiveness level of the controller. Based on the presented simulation results in previous sections, the proposed system is capable of generating walking which is not only adaptive but also robust against uncertainties and disturbances. Besides, the proposed system is computationally supper fast; hence, it does not require specific resources. Unlike [@morisawa2014biped; @englsberger2015three] which are based on PID control, the proposed system is designed based on optimal control approach which is more effective, moreover, the proposed controller uses the step time adjustment strategy to improve the stability of the robot. Conclusion {#sec:CONCLUSION} ========== In this paper, we have designed and developed a biped locomotion framework which was composed of four main modules that are organized in a hierarchical structure to fade the complexity of the walking problem and increase the flexibility of the framework. We explained the functionality of each module and illustrated how these modules interact with each other to generate stable locomotion. The fundamental modules of this framework are the *low-level controller* and *next step adjustment* which are responsible for motion tracking and keeping the stability. Particularly, the LIPM and DCM concepts were used to represent the overall dynamics of the system as a linear state-space system and using that, we formulated the walking problem as an LQG optimal controller to generate a robust control solution based on an offline optimization. The next step adjustment module was combined with the low-level controller module to improve the withstanding level of the framework by online modification of step time and location according to the measured DCM in each control cycle. The functionality of each module has been tested independently and the overall performance and robustness of the proposed framework were validated using performing several simulations. The simulation results showed that the next step adjustment module improves the overall withstanding level of the framework up to 63%. Our future work includes examining the performance of the proposed framework while generating walking on uneven terrain. Additionally, we would like to port our framework to several real humanoid platforms to show the portability and flexibility of the proposed framework. Acknowledgment {#acknowledgment .unnumbered} ============== This research is supported by Portuguese National Funds through Foundation for Science and Technology (FCT) through FCT scholarship SFRH/BD/118438/2016.

ArXiv

--- abstract: 'For a virtual knot $K$ and an integer $r\geq 0$, the $r$-covering $K^{(r)}$ is defined by using the indices of chords on a Gauss diagram of $K$. In this paper, we prove that for any finite set of virtual knots $J_0,J_2,J_3,\dots,J_m$, there is a virtual knot $K$ such that $K^{(r)}=J_r$ $(r=0\mbox{ and }2\leq r\leq m)$, $K^{(1)}=K$, and otherwise $K^{(r)}=J_0$.' address: - 'Department of Engineering Science, Osaka Electro-Communication University, Hatsu-cho 18-8, Neyagawa, Osaka 572-8530, Japan' - 'Department of Mathematics, Kobe University, Rokkodai-cho 1-1, Nada-ku, Kobe 657-8501, Japan' - 'Department of Mathematics, Kobe University, Rokkodai-cho 1-1, Nada-ku, Kobe 657-8501, Japan' author: - Takuji NAKAMURA - Yasutaka NAKANISHI - Shin SATOH title: A note on coverings of virtual knots --- Introduction {#sec1} ============ Odd crossings are first introduced for constructing a simple invariant called the odd writhe of a virtual knot by Kauffman [@Kau2]. By using odd crossings, Manturov defines a map from the set of virtual knots to itself by replacing the odd crossings with virtual crossings [@Man]. Later the notion of index is introduced in [@CG; @Hen; @ILL; @ST] which assigns an integer to each real crossing such that the parity of the index coincides with the original parity. The $n$-writhe is defined as a refinement of the odd writhe. Jeong defines an invariant called the zero polynomial of a virtual knot by using real crossings of index $0$ [@Jeo]. In fact, Im and Kim prove that the zero polynomial is coincident with the writhe polynomial of the virtual knot obtained by replacing the real crossings whose indices are non-zero with virtual crossings [@IK]. They also study the operation replacing the real crossings whose indices are not divisible by $r$ for a positive integer $r$. This operation is originally considered for flat virtual knots by Turaev [@Tur] where he calls the obtained knot the [*$r$-covering*]{}. The writhe polynomial $W_K(t)$ of a virtual knot $K$ is the polynomial such that the coefficient of $t^n$ is equal to the $n$-writhe of $K$. A characterization of $W_K(t)$ is given as follows. \[thm11\] For a Laurent polynomial $f(t)\in{{\mathbb{Z}}}[t,t^{-1}]$, the following are equivalent. - There is a virtual knot $K$ with $W_K(t)=f(t)$. - $f(1)=f'(1)=0$. For an integer $r\geq 0$, we denote by $K^{(r)}$ the $r$-covering of a virtual knot $K$. By definition, we have $K^{(1)}=K$ and $K^{(r)}=K^{(0)}$ for a sufficiently large $r$. The aim of this note is to prove the following. \[thm12\] Let $m\geq 1$ be an integer, $J_n$ $(0\leq n\leq m, n\ne 1)$ $m$ virtual knots, and $f(t)$ a Laurent polynomial with $f(1)=f'(1)=0$. Then there is a virtual knot $K$ such that $$K^{(r)}= \left\{ \begin{array}{ll} J_0 & \mbox{for }r=0 \mbox{ and }r\geq m+1, \\ K & \mbox{for } r=1, \\ J_r & \mbox{otherwise}, \\ \end{array} \right.$$ and $W_K(t)=f(t)$. This paper is organized as follows. In Section \[sec2\], we define the $r$-covering $K^{(r)}$ of a (long) virtual knot $K$. We also introduce an anklet of a chord in a Gauss diagram which will be used in the consecutive sections. In Sections \[sec3\] and \[sec4\], we study the $0$-covering $K^{(0)}$ and $r$-covering $K^{(r)}$ for $r\geq 2$ of a long virtual knot, respectively. In Section \[sec5\], we review the writhe polynomial of a virtual knot, and prove Theorem \[thm11\]. Gauss diagrams {#sec2} ============== A [*circular*]{} or [*linear Gauss diagram*]{} is an oriented circle or line equipped with a finite number of oriented and signed chords spanning the circle or line, respectively. The [*closure*]{} of a linear Gauss diagram is the circular Gauss diagram obtained by taking the one-point compactification of the line. Let $c$ be a chord of a Gauss diagram $G$ with sign $\varepsilon=\varepsilon(c)$. We give signs $-\varepsilon$ and $\varepsilon$ to the initial and terminal endpoints of $c$, respectively. We consider the case that $G$ is circular. The endpoints of $c$ divide the circle into two arcs. Let $\alpha$ be the arc oriented from the initial endpoint of $c$ to the terminal. See Figure \[fig01\]. The [*index*]{} of $c$ is the sum of signs of endpoints of chords on $\alpha$, and denoted by ${\rm Ind}_G(c)$ (cf. [@Che; @Kau; @ST]). In the case that $G$ is linear, the index of $c$ is defined as that of $c$ in the closure of $G$. ![The orientation and signs of a chord and its endpoints.[]{data-label="fig01"}](fig01.pdf) Let $G$ be a circular or linear Gauss diagram. For a positive integer $r$, we denote by $G^{(r)}$ the Gauss diagram obtained from $G$ by removing all the chords $c$ with ${\rm Ind}_G(c)\not\equiv 0~({\rm mod}\ r)$ (cf. [@Tur]). In particular, we have $G^{(1)}=G$. For $r=0$, we denote by $G^{(0)}$ the Gauss diagram obtained from $G$ by removing all the chords $c$ with ${\rm Ind}_G(c)\ne 0$. Since the number of chords of $G$ is finite, we have $G^{(r)}=G^{(0)}$ for sufficiently large $r$. Two circular Gauss diagrams $G$ and $H$ are [*equivalent*]{}, denoted by $G\sim H$, if $G$ is related to $H$ by a finite sequence of [*Reidemeister moves*]{} I–III as shown in Figure \[fig02\]. A [*virtual knot*]{} is an equivalence class of circular Gauss diagrams up to this equivalence relation (cf. [@GPV; @Kau2]). Similarly, the equivalence relation among linear Gauss diagrams are defined, and an equivalence class is called a [*long virtual knot*]{}. The [*trivial*]{} (long) virtual knot is presented by a Gauss diagram with no chord. ![Reidemeister moves on Gauss diagrams.[]{data-label="fig02"}](fig02.pdf) \[lem21\] Let $G$ and $H$ be circular or linear Gauss diagrams such that $G\sim H$. Then it holds that $G^{(r)}\sim H^{(r)}$ for any integer $r\geq 0$. $\Box$ Although only the case of a circular Gauss diagram is studied in [@IK; @Tur], Lemma \[lem21\] for a liner Gauss diagram can be proved similarly. \[def22\] Let $K$ be a (long) virtual knot, and $r\geq 0$ an integer. The [*$r$-covering*]{} of $K$ is the (long) virtual knot presented by $G^{(r)}$ for some Gauss diagram $G$ of $K$. We denote it by $K^{(r)}$. The well-definedness of $K^{(r)}$ follows from Lemma \[lem21\]. We have $K^{(1)}=K$ and $K^{(r)}=K^{(0)}$ for sufficiently large $r$. Let $c(G)$ denote the number of chords of a Gauss diagram $G$. The [*real crossing number*]{} of a (long) virtual knot $K$ is the minimal number of $c(G)$ for all Gauss diagrams $G$ of $K$, and denoted by ${\rm c}(K)$. \[lem23\] Let $K$ be a $($long$)$ virtual knot, and $r\geq 0$ an integer. Then it holds that ${\rm c}(K^{(r)})\leq {\rm c}(K)$. In particular, ${\rm c}(K^{(r)})={\rm c}(K)$ holds if and only if $K^{(r)}=K$. Let $G$ be a Gauss diagram of $K$ with $c(G)={\rm c}(K)$. Since $G^{(r)}$ is obtained from $G$ by removing some chords, we have $${\rm c}(K^{(r)})\leq c(G^{(r)})\leq c(G)={\rm c}(K).$$ In particular, if the equality holds, then $G^{(r)}=G$ and $K^{(r)}=K$. Let $c_1,\dots,c_n$ be chords of a Gauss diagram $G$. We add several parallel chords near an endpoint of $c_i$ $(i=1,\dots,n)$ to obtain a Gauss diagram $G'$. Here, the orientations and signs of the added chords are chosen arbitrarily. See Figure \[fig03\](i). The parallel chords added to $c_i$ are called [*anklets*]{} of $c_i$. We remark that the index of an anklet is equal to $\pm 1$. ![Anklets.[]{data-label="fig03"}](fig03.pdf) Let $c_1,\dots,c_n$ be chords of a Gauss diagram $G$. For any integers $a_1,\dots,a_n$, by adding several anklets to each $c_i$ near its initial endpoint suitably, we obtain a Gauss diagram $G'$ such that ${\rm Ind}_{G'}(c_i)=a_i$ for any $i=1,\dots,n$, and ${\rm Ind}_{G'}(c)={\rm Ind}_{G}(c)$ for any $c\ne c_1,\dots,c_n$. Put $d_i=a_i-{\rm Ind}_G(c_i)$ for $i=1,\dots,n$. We add $|d_i|$ anklets to $c_i$ near its initial endpoint such that the signs of right endpoints of the anklets are equal to $\varepsilon_i$, where $\varepsilon_i$ is the sign of $d_i$. See Figure \[fig03\](ii). Let $G'$ be the obtained Gauss diagram. Then we have $${\rm Ind}_{G'}(c_i)= {\rm Ind}_G(c_i)+\varepsilon_i|d_i|= {\rm Ind}_G(c_i)+d_i=a_i.$$ Furthermore the index of a chord other than $c_1,\dots,c_n$ does not change. The $0$-covering $K^{(0)}$ {#sec3} ========================== For an integer $n\geq 2$, we define a map $f_n:\{2,3,\dots,n\}\rightarrow{{\mathbb{Z}}}$ which satisfies $$\sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} f_n(i)=-1$$ for any integer $r$ with $2\leq r\leq n$. The map $f_n$ exists uniquely. Put $$P_n=\{i \mid 2\leq i\leq n, f_n(i)\ne 0\}.$$ For $n=10$, we have $$\left\{ \begin{array}{lrl} f_{10}(2)+f_{10}(4)+f_{10}(6)+f_{10}(8)+f_{10}(10)&=-1 & \mbox{for } r=2, \\ f_{10}(3)+f_{10}(6)+f_{10}(9)&=-1 & \mbox{for } r=3, \\ f_{10}(4)+f_{10}(8)&=-1 & \mbox{for } r=4, \\ f_{10}(5)+f_{10}(10)&=-1 & \mbox{for }r=5, \\ f_{10}(6)&=-1 & \mbox{for } r=6, \\ f_{10}(7)&=-1 & \mbox{for } r=7, \\ f_{10}(8)&=-1 & \mbox{for } r=8, \\ f_{10}(9)&=-1 & \mbox{for } r=9, \\ f_{10}(10)&=-1 & \mbox{for } r=10. \end{array}\right.$$ Therefore we have $$f_{10}(i)= \left\{ \begin{array}{rl} 2 & (i=2), \\ 1 & (i=3), \\ 0 & (i=4,5), \\ -1 & (6\leq i\leq 10) \end{array}\right. \mbox{ and } P_{10}=\{2,3,6,7,8,9,10\}.$$ \[thm32\] For any integer $n\geq 1$ and long virtual knot $J$, there is a long virtual knot $K$ such that $$K^{(r)}= \left\{ \begin{array}{ll} J & \mbox{for }r=0 \mbox{ and }r\geq n+1, \\ K & \mbox{for } r=1, \ and \\ O & \mbox{otherwise}. \\ \end{array} \right.$$ Here, $O$ denotes the trivial long virtual knot. Let $H$ be a linear Gauss diagram of $J$. We construct a linear Gauss diagram $G$ as follows: First, we replace each chord $c$ of $H$ by $1+\sum_{i\in P_n}|f_n(i)|$ parallel chords labeled $c_0$ and $c_{ij}$ for $i\in P_n$ and $1\leq j\leq |f_n(i)|$. The orientations of $c_0$ and $c_{ij}$’s are the same as that of $c$. The signs of them are given such that - $\varepsilon(c_0)=\varepsilon(c)$, and - $\varepsilon(c_{ij})=\varepsilon(c)\delta_i$ for $i\in P_n$, where $\delta_i$ is the sign of $f_n(i)\ne 0$. Next, we add several anklets to each of $c_0$ and $c_{ij}$’s such that - ${\rm Ind}_G(c_0)=0$, and - ${\rm Ind}_G(c_{ij})=i$ for $i\in P_n$. The obtained Gauss diagram is denoted by $G$. Figure \[fig04\] shows the case $n=10$ replacing each chord $c$ of $H$ with nine chords $c_0,c_{21},c_{22}, c_{31},\dots,c_{10,1}$ and several anklets. The boxed numbers of the chords indicate their indices. ![The case $n=10$.[]{data-label="fig04"}](fig04.pdf) By the conditions (i)–(iv), we have $G^{(0)}=H$; in fact, we remove the chords whose indices are non-zero from $G$ to obtain $G^{(0)}$. Similarly, we have $G^{(r)}=H$ for any $r\geq n+1$. Therefore it holds that $K^{(0)}=K^{(r)}=J$ for $r\geq n+1$. For $2\leq r\leq n$, $G^{(r)}$ is obtained from $G$ by removing the chords whose indices are not divisible by $r$. In particular, all the anklets are removed. Among the chords $c_0$ and $c_{ij}$’s, the sum of signs of chords whose indices are divisible by $r$ is equal to $$\varepsilon(c)+\varepsilon(c) \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} \delta_i|f_n(i)|= \varepsilon(c)+\varepsilon(c) \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} f_n(i)=0.$$ Therefore all the chords of $G^{(r)}$ can be canceled by Reidemesiter moves II so that $K^{(r)}$ is the trivial long virtual knot. The $r$-covering $K^{(r)}$ for $r\geq 2$ {#sec4} ======================================== For an integer $n\geq 2$, we define a map $g_n:\{2,3,\dots,n\}\rightarrow{{\mathbb{Z}}}$ which satisfies $$g_n(n)=1 \mbox{ and } \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} g_n(i)=0$$ for any integer $r$ with $2\leq r< n$. The map $g_n$ exists uniquely. Put $$Q_n=\{i\mid 2\leq i\leq n, g_n(i)\ne 0\}.$$ \(i) For $n=10$, it holds that $$\left\{ \begin{array}{lrl} g_{10}(2)+g_{10}(4)+g_{10}(6)+g_{10}(8)+g_{10}(10)&=0 & \mbox{for } r=2, \\ g_{10}(3)+g_{10}(6)+g_{10}(9)&=0 & \mbox{for } r=3, \\ g_{10}(4)+g_{10}(8)&=0 & \mbox{for } r=4, \\ g_{10}(5)+g_{10}(10)&=0 & \mbox{for }r=5, \\ g_{10}(6)&=0 & \mbox{for } r=6, \\ g_{10}(7)&=0 & \mbox{for } r=7, \\ g_{10}(8)&=0 & \mbox{for } r=8, \\ g_{10}(9)&=0 & \mbox{for } r=9, \\ g_{10}(10)&=1 & \mbox{for } r=10. \end{array}\right.$$ Therefore we have $$g_{10}(i)= \left\{ \begin{array}{rl} 1 & (i=10), \\ -1 & (i=2,5), \\ 0 & ({\rm otherwise}), \\ \end{array}\right. \mbox{ and } Q_{10}=\{2,5,10\}.$$ \(ii) For $n=12$, we have $$g_{12}(i)= \left\{ \begin{array}{rl} 1 & (i=2,12), \\ -1 & (i=4,6), \\ 0 & ({\rm otherwise}), \\ \end{array}\right. \mbox{ and } Q_{12}=\{2,4,6,12\}.$$ \[thm42\] For any integer $n\geq 2$ and long virtual knot $J$, there is a long virtual knot $K$ such that $$K^{(r)}= \left\{ \begin{array}{ll} J & \mbox{for }r=n, \\ K & \mbox{for } r=1, \ and \\ O & \mbox{otherwise}. \\ \end{array} \right.$$ The proof is similar to that of Theorem \[thm32\] by using $g_n$ instead of $f_n$. Let $H$ be a linear Gauss diagram of $J$. We construct a linear Gauss diagram $G$ of $K$ as follows: First, we replace each chord $c$ of $H$ by $\sum_{i\in Q_n} |g_n(i)|$ parallel chords labeled $c_i$ for $i\in Q_n$. The orientations of $c_i$’s are the same as that of $c$. The signs of them are given such that $\varepsilon(c_i)=\varepsilon(c)\delta_i$, where $\delta_i$ is the sign of $g_n(i)\ne 0$. Next, we add several anklets to each of $c_i$’s such that ${\rm Ind}_G(c_i)=i$ for $i\in Q_n$. The obtained Gauss diagram is denoted by $G$. In the left of Figure \[fig05\], we shows the case $n=10$ replacing each chord $c$ of $H$ with three chords $c_2,c_5,c_{10}$ and several anklets. In the right figure, the case of $n=12$ is given. ![The cases $n=10$ and $12$.[]{data-label="fig05"}](fig05.pdf) Since any chord $c$ of $G$ satisfies $1\leq |{\rm Ind}_G(c)|\leq n$, we obtain $G^{(0)}$ and $G^{(r)}$ for $r\geq n+1$ by removing all the chords from $G$. For $2\leq r<n$, $G^{(r)}$ is obtained from $G$ by removing the chords whose indices are not divisible by $r$. Among the chords $c_i$’s, the sum of signs of chords whose indices are divisible by $r$ is equal to $$\varepsilon(c) \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} \delta_i |g_n(i)|= \varepsilon(c) \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} g_n(i) =0.$$ Therefore all the chords of $G^{(r)}$ can be canceled by Reidemeister moves II so that $K^{(r)}$ is the trivial long virtual knot. Finally, for $r=n$, we have $G^{(n)}=H$ by definition immediately. We see that $g_n$ is coincident with a famous function as follows. \[prop43\] Let $\mu$ be the Möbius function. Then we have $$g_n(i)= \left\{ \begin{array}{cl} \mu(\frac{n}{i}) & \mbox{if $n$ is divisible by $i$, and}\\ 0 & \mbox{otherwise}. \end{array}\right.$$ Let $h_n(i)$ be the right hand side of the equation in the proposition. Since $h_n(n)=\mu(1)=1=g_n(n)$, it is sufficient to prove that $$\sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} h_n(i)=0$$ for any integer $r$ with $2\leq r< n$. Assume that $n$ is not divisible by $r$. Then $n$ is not divisible by any $i$ such that $r\leq i\leq n$ and $i\equiv 0$ (mod $r$). Therefore we have $$\sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} h_n(i)= \sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} 0=0.$$ Assume that $n$ is divisible by $r$. By the property of the Möbius function, it holds that $$\sum_{{r\leq i\leq n}\atop{i\equiv 0\ ({\rm mod}\ r)}} h_n(i)= \sum_{r|i|n} \mu(\frac{n}{i}) =\sum_{d|\frac{n}{r}}\mu(d)=0.$$ Therefore we have $g_n=h_n$. The writhe polynomial {#sec5} ===================== For an integer $n\ne 0$ and a sign $\varepsilon=\pm 1$, the [*$(n,\varepsilon)$-snail*]{} is a linear Gauss diagram consisting of a chord $c$ with $\varepsilon(c)=\varepsilon$ and $|n|$ anklets such that the indices of $c$ and each anklet are equal to $n$ and $1$, respectively. See Figure \[fig06\]. ![The snail $S(n,\varepsilon)$.[]{data-label="fig06"}](fig06.pdf) Let $G$ be a Gauss diagram of a (long) virtual knot $K$. For an integer $n\ne 0$, we denote by $w_n(G)$ the sum of signs of all chords $c$ of $G$ with ${\rm Ind}_G(c)=n$. Then $w_n(G)$ does not depend on a particular choice of $G$ of $K$; that is, $w_n(G)$ is an invariant of $K$. In [@ST], the proof is given for a virtual knot, and the case of a long virtual knot is similarly proved. It is called the [*$n$-writhe*]{} of $K$ and denoted by $w_n(K)$. The [*writhe polynomial*]{} of $K$ is defined by $$W_K(t)=\sum_{n\ne 0}w_n(K)t^n- \sum_{n\ne 0}w_n(K)\in{{\mathbb{Z}}}[t,t^{-1}].$$ This invariant was introduced in several papers [@CG; @Kau; @ST] independently. \[thm51\] Let $J$ be a long virtual knot, and $f(t)\in{{\mathbb{Z}}}[t,t^{-1}]$ a Laurent polynomial with $f(1)=f'(1)=0$. Then there is a long virtual knot $K$ such that - $K^{(r)}=J^{(r)}$ for any integer $r=0$ and $r\geq 2$, and - $W_K(t)=f(t)$. Put $g(t)=f(t)-W_J(t)=\sum_{n\in{{\mathbb{Z}}}}a_nt^n$. By Theorem \[thm11\], we have $g(1)=g'(1)=0$. Therefore it holds that $a_0=\sum_{n\ne 0,1}(n-1)a_n$ and $a_1=-\sum_{n\ne 0,1} na_n$. Let $H$ be a linear Gauss diagram of $J$. We construct a linear Gauss diagram $G$ by juxtaposing $H$ and $|a_n|$ copies of $S(n,\varepsilon_n)$ for every integer $n$ with $n\ne 0,1$ and $a_n\ne 0$. Here, $\varepsilon_n$ is the sign of $a_n\ne 0$. Let $K$ be the long virtual knot presented by $G$. The contribution of each snail $S(n,\varepsilon_n)$ to the writhe polynomial $W_K(t)$ is equal to $\varepsilon_nt^n-\varepsilon_nnt+\varepsilon_n(n-1)$. Therefore it holds that $$\begin{aligned} W_K(t)&=& W_J(t)+ \sum_{n\ne 0,1} |a_n|\bigr(\varepsilon_nt^n-\varepsilon_nnt+\varepsilon_n(n-1)\bigr)\\ &=& W_J(t)+\sum_{n\ne 0,1} a_n\bigr(t^n-nt+(n-1)\bigr)\\ &=& W_J(t)+g(t)=f(t). \end{aligned}$$ By definition, $S(n,\varepsilon_n)^{(r)}$ has the only chord $c$ if $n$ is divisible by $r$. Otherwise it has no chord. Therefore $G^{(r)}$ is equivalent to $H^{(r)}$, and hence $K^{(r)}=J^{(r)}$ for any integer $r=0$ and $r\geq 2$. \[thm52\] Let $m\geq 1$ be an integer, $J_n$ $(0\leq n\leq m, n\ne 1)$ $m$ long virtual knots, and $f(t)$ a Laurent polynomial with $f(1)=f'(1)=0$. Then there is a long virtual knot $K$ such that $$K^{(r)}= \left\{ \begin{array}{ll} J_0 & \mbox{for }r=0 \mbox{ and }r\geq m+1, \\ K & \mbox{for } r=1, \\ J_r & \mbox{otherwise}, \\ \end{array} \right.$$ and $W_K(t)=f(t)$. Let $K_0$ be a long virtual knot obtained by applying Theorem \[thm32\] to the pair of $m$ and $J_0$. Let $K_n$ be a long virtual knot obtained by applying Theorem  \[thm42\] to each pair of $n$ and $J_n$ $(2\leq n\leq m)$. We juxtapose $K_0,K_2,\dots,K_m$ to have a long virtual knot $K'$. Let $K$ be a long virtual knot obtained by applying Theorem \[thm51\] to the pair of $K'$ and $f(t)$. Then we see that $K$ is a desired long virtual knot. Let $J_n^{\circ}$ be a long virtual knot whose closure is $J_n$ $(0\leq n\leq m, n\ne 1)$. Let $K^{\circ}$ be a long virtual knot obtained by applying Theorem \[thm52\]. Then we see that the closure of $K^{\circ}$ is a desired virtual knot. [99]{} Z. Chen, [*A polynomial invariant of virtual knots*]{}, Proc. Amer. Math. Soc. [**142**]{} (2014), no. 2, 713–725. Z. Cheng and H. Gao, [*A polynomial invariant of virtual links*]{}, J. Knot Theory Ramifications [**22**]{} (2013), no. 12, 1341002, 33 pp. M. Goussarov, M. Polyak, and O. Viro, [*Finite-type invariants of classical and virtual knots*]{}, Topology [**39**]{} (2000), no. 5, 1045–1068. A. Henrich, [*A sequence of degree one Vassiliev invariants for virtual knots*]{}, J. Knot Theory Ramifications [**19**]{} (2010), no. 4, 461–487. Y. H. Im and S. Kim, [*A sequence of polynomial invariants for Gauss diagrams*]{}, J. Knot Theory Ramifications [**26**]{} (2017), no. 7, 1750039, 9 pp. Y. H. Im, K. Lee, and Y. Lee, [*Index polynomial invariant of virtual links*]{}, J. Knot Theory Ramifications [**19**]{} (2010), no. 5, 709–725. M.-J. Jeong, [*A zero polynomial of virtual knots*]{}, J. Knot Theory Ramifications [**25**]{} (2016), no. 1, 1550078, 19 pp. L. H. Kauffman, [*Virtual knot theory*]{}, European J. Combin. [**20**]{} (1999), no. 7, 663–690. L. H. Kauffman, [*An affine index polynomial invariant of virtual knots*]{}, J. Knot Theory Ramifications [**22**]{} (2013), no. 4, 1340007, 30 pp. V. O. Manturov, [*Parity and projection from virtual knots to classical knots*]{}, J. Knot Theory Ramifications [**22**]{} (2013), no. 9, 1350044, 20 pp. S. Satoh and K. Taniguchi, [*The writhes of a virtual knot*]{}, Fund. Math. [**225**]{} (2014), no. 1, 327–342. V. Turaev, [*Virtual strings*]{}, Ann. Inst. Fourier (Grenoble) [**54**]{} (2004), no. 7, 2455–2525 (2005).

ArXiv

--- abstract: 'While non-action-generated, but identically conserved, abelian/YM gauge vectors exist, they are unsuitable for building alternate field equations, because they have no stress-tensor, hence do not permit Poincare generators and, most physically, cannot consistently couple to gravity. Separately, their geometric analogues, covariantly conserved non-Lagrangian symmetric tensors, probably do not even exist, but their weak field, abelian, counterparts do, and share the vector fields’ absence of generators.' --- BRX TH-6645\ CALT-TH 2019-01 [**Non-Lagrangian Gauge Field Models are Physically Excluded**]{} [S. Deser]{} [ *Walter Burke Institute for Theoretical Physics,\ California Institute of Technology, Pasadena, CA 91125;\ Physics Department, Brandeis University, Waltham, MA 02454\ [[email protected]]{}* ]{} Introduction ============ This work answers an open question regarding massless gauge (abelian and non-) vector, and (less strongly) geometric tensor fields: Are there viable models with identically conserved field equations’ “left-hand-side" terms that are not derivable from actions? It is a non-trivial one, both formally and physically, as neither existence of such terms nor the proper physical grounds to exclude them are obvious; indeed it is still not known if non-singular geometrical terms even exist \[1\]. Vector terms do, but have no corresponding stress-tensors, hence no Poincare generators can even be defined there. More physically, they cannot consistently couple to gravity if they couple to any normal matter — or merely if their Lagrangian counterparts are also present. Geometric tensors’ (if any!) weak field versions also exist; the latter are excluded on the more formal, absence of Poincare generators, grounds. These no-go results preclude a large class of speculative models. Vectors ======= A sufficiently general set of abelian vector field equations is $$M^\mu = \partial_\nu \left[ X(F^2, \,^*F F) F^{\mu\nu}(A)\right] = j^\mu$$ where $\,^*F$ is the usual dual of $F$ and the arbitrary scalar $X$ depends only on the two simplest, algebraic, invariants. The divergence identities $\partial_\mu M^\mu=0$ are manifest from the antisymmetry of $F^{[\mu\nu]}$ contracted with the symmetric $\partial^2_{\mu\nu}$, irrespective of $X$. However, not all such $M^\mu$ are variations of an action: they must obey the Helmholz integrability conditions, which set stringent limits on $X$. So identical conservation does NOT require an action, already in these simple examples of vectors $V^\mu=\dd_\nu H^{[\mu\nu]}$. Perhaps surprisingly, this is not a purely abelian property, but holds also for non-abelian fields: there, we replace $\partial_\mu$ by the usual covariant color derivatives $D_\mu$ whose commutator is now the non-abelian field strength, $[D_\mu, D_\nu] = F_{\mu\nu}$. Yet the generalization of (1) remains transverse, owing to the antisymmetric structure constants, since $f_{abc} F^b_{\, \, \mu\nu} F^{c \mu\nu} = 0$ (the arguments of $X$ are now the color-singlet traces of $F^2$ and $\,^*FF$). Again, only algebraic symmetry properties are relevant. Indeed, even in curved space, ordinary conservation of (1) holds, because the divergence of the contravariant tensor density $\sqrt{-g} X F^{\mu\nu}$ is still a partial derivative and so in turn is its divergence, being that of a contravariant vector density. Are there any physically permitted models exploiting the above conservation properties, either stand-alone or by adding terms like (1) to Maxwell- or YM- like equations? Clearly, charge conservation is not affected, since both sides of (1) are conserved. To be sure, the expression for the charge does becomes a bit byzantine, involving both the longitudinal AND transverse electric fields, $$\oint d^2 {\bf S} \cdot X {\bf E} = \int d^3 x j^0 = Q.$$ Instead, the real obstruction is due first to the loss of Poincare generators caused by the absence of an action for the $\partial (XF)$ term: no action means no conserved $T_{\mu\nu}$. For example, the divergence of a would-be $T_{\mu\nu} = X T_{\mu\nu}(\hbox{Maxwell})$ is $\sim F^2 \dd_\mu X \ne 0$; the general proof is obvious since the only possible terms are $A F_{\mu\alpha} F^\alpha_{\, \, \, \nu} + g_{\mu\nu} B F^2$. Since adding non-action terms forbids stress-tensors, there are no Poincare generators; mass and spin cannot even be defined (the generators are as essential at classical as at quantum level). However, the more striking — and physical–contradiction comes when attempting (unavoidably, if these fields are to interact with any normal ones) to couple to gravity: the added terms (while still conserved, as we saw) depend on the metric, hence are acted on but do not react on, gravity, absent a properly conserved $T_{\mu\nu}$ contribution to gravity’s equations. This seeming violation of Newton’s third law is not immediately inconsistent — rather, the non-Lagrangian gauge field equation represents a sort of “test-field": the (source-free) gravitational and gauge field equations are separate. However, if there is also a normal, say Maxwell, part — its $T_{\mu\nu}$ is no longer conserved, and consistency is lost. Generally, if any normal matter interacts with the gauge field, its stress tensor will also no longer be conserved (on its shell) since it effectively contains the $A$-field as an “external", rather than (normal) dynamical, parameter. Note the contrast with Chern-Simons (CS) electrodynamics (or YM) in this respect: the CS term’s stress-tensor vanishes identically, yet the original Maxwell/YM stress-tensor stays conserved on full CS shell. A large class of speculative Maxwell and Yang-Mills extensions can thus be neglected. Gravity ======= Assume the (unlikely \[1\]) existence of identically conserved non-Lagrangian symmetric geometric tensors $S_{\mu\nu}$ ($D^n$ curvature; $g_{\alpha\beta}$) and consider the physical effects of adding them to normal gravitational field equations, $$G_{\mu\nu} + S_{\mu\nu} = T_{\mu\nu}(\hbox{matter;} g),$$ where $G_{\mu\nu}$ denotes any Lagrangian-based tensor (or $0$) and the (normal) matter source is covariantly conserved on its shell, independent of the metric’s dynamics. At linearized curvature level, where all explicit metrics as well as derivatives are flat-space, these models are similar to the abelian vector case: There are again identically conserved projection operators, generalizing $\partial_\nu H^{[\mu\nu]}$, namely the so-called superpotentials[^1] $V^{\mu\nu}=\partial^2_{\alpha\beta} H^{[\mu\alpha] [\nu\beta]}$, where $H$ has the algebraic symmetries of the Riemann tensor. For example, in $D=2$, $H$ degenerates to ${\epsilon}^{\mu\alpha} {\epsilon}^{\nu\beta} S$, so $V_{\mu\nu}$ becomes the transverse projector $(\partial^2_{\mu\nu} -\eta_{\mu\nu} \Box) S$, where $S$ is any scalar. Any non-Lagrangian linearized $S_{\mu\nu}$ is allowed, but as in the vector case, it has no associated stress tensor, hence loss of Poincare generators at this linearized level — corresponding to the Maxwell limit of the vector case. But this destroys all non-linear would-be models as well, since they would all have an abelian limit, just as YM contains Maxwell. Separately, we know \[1\] that any $S_{\mu\nu}$, were it to exist, starts (at least) at fourth derivative order on the curvatures, with obvious negative implications for ghost-like, and external non-Schwarzchild (if there are terms solely involving the Weyl tensor), solutions of (3). Comments ======== We have seen that while infinitely many non-Lagrangian conserved vector gauge terms exist, they are forbidden in flat space model-building owing to their obstruction to defining Poincare generators. This failure is compounded by the direct physical contradiction that they cannot consistently couple to (any) gravity, because they cannot affect the geometry as legitimate (on-shell) conserved sources, being only acted on by the metric without reacting on the latter’s dynamics, not having conserved stress tensors. Yet if they are to couple to any normal matter or even if a normal, “Maxwell", part is included, they would have to — but cannot — contribute in order to insure consistency, as we have seen. Separately, while existence of conserved symmetric non-Lagrangian geometric tensors is not (yet) excluded, we noted that even if they do exist, their abelian limit encounters the corresponding vector problems. Finally, a referee-induced comment on the use of Lagrange multipliers, the usual last resort. We could add a new vector field $B_\mu$, with a Lagrangian $L= B_\mu M^\mu$, or equivalently $L=B_{\mu\nu} X F^{\mu\nu}$, with a conserved $T_{\mu\nu}$ on $(A+B)$ shell, but it course vanishes if we set the multiplier $B=0$. The pitfall here is that spurious degrees of freedom are introduced, as is clear when $X=1$ there $L$ describes a $2$-photon system. Acknowledgements {#acknowledgements .unnumbered} ================ This work was supported by the U.S. Department of Energy, Office of Science, Office of High Energy Physics, under Award Number de-sc0011632. Long-term collaboration with Y. Pang, and with A. Waldron, on a complex of related problems is happily acknowledged. I thank J. Franklin for tech help. [99]{} S. Deser and Y. Pang arXiv:1811.03124. [^1]: In the GR literature, quantities of this type are used to represent harmless ambiguities of flat space stress tensors because they cannot contribute to any generators, whereas we use them as putative field equation contributions.

ArXiv

--- abstract: 'Machine learning, with a dramatic breakthrough in recent years, is showing great potential to upgrade the power system optimization toolbox. Understanding the strength and limitation of machine learning approaches is crucial to answer when and how to integrate them in various power system optimization tasks. This paper pays special attention to the coordination between machine learning approaches and optimization models, and carefully evaluates to what extent such data-driven analysis may benefit the rule-based optimization. A series of typical references are selected and categorized into four kinds: the boundary parameter improvement, the optimization option selection, the surrogate model and the hybrid model. This taxonomy provides a novel perspective to understand the latest research progress and achievements. We further discuss several key challenges and provide an in-depth comparison on the features and designs of different categories. Deep integration of machine learning approaches and optimization models is expected to become the most promising technical trend.' author: - 'Guangchun Ruan,  Haiwang Zhong,  Guanglun Zhang,  Yiliu He,  Xuan Wang,  Tianjiao Pu, \' bibliography: - 'refs.bib' title: 'Review of Learning-Assisted Power System Optimization [^1] [^2] ' --- [Shell : arXiv Preprint]{} smart grid, machine learning, deep learning, neural network, data-driven, artificial intelligence Introduction {#SEC-INTRO} ============ Background ---------- the advance of computing hardwares, artificial intelligence (AI) has entered into a booming period in recent years [@RV5]. Machines are proven to outperform conventional man-make algorithms in more and more applications, and there is a wide consensus that both AI and machine learning have not reached their peaks yet. Optimization is popular in tremendous applications in power system, e.g., optimal dispatch, planning, market clearing and security assessment. A efficient and reliable optimization performance is of great essence. However, conventional optimization methods have shown their limitations in some complicated and volatile environments, such as future power grids with a high penetration of renewable energy [@RN13]. These models tend to repeatedly solve some similar problems without accumulating any experience. Machine learning, in this aspect, is powerful in gaining experience from historical data and past decisions [@RN26]. Existing practices have shown that the integration of machine learning and power system optimization may bring some significant benefits [@RN19]. Bibliometric Analysis --------------------- A bibliometric analysis is conducted using the well-known database, Web of Science, to provide an overlook on the research trend. The searching query for Web of Science is designed as follows: The number of publications concerning the learning-assisted power system optimization, and the proportion in all power system optimization publications, indexed by Web of Science in recent 8 years, are showed in Fig. \[fig\_overview\]. The proportion in all power system optimization publications is calculated as the number of publications on learning-assisted power system optimization divided by the number of all power system optimization publications. When searching for the total number, a part of query expressions behind the second keyword “AND” are accordingly dropped. ![Research trends of learning-assisted power system optimization in recent years. The number of publications is shown with a bar chart, and the associated proportion of all publications is plotted by a line chart. A significant increase can be observed in the recent two years, indicating a fast-developing research hotspot. Data Source: Web of Science.[]{data-label="fig_overview"}](fig/Overview2.pdf){width="0.98\linewidth"} It can be observed from Fig. \[fig\_overview\] that, from 2012 to 2017, the proportions in all power system optimization publications are relatively small. In 2018 and 2019, this percentage reaches 7.23% and 8.84% respectively, much higher than before. One can infer from Fig. \[fig\_overview\] that the idea of applying machine learning approaches in power system optimization has attracted much more attention than before. The prospect and potential of learning-assisted power system optimization are getting recognized by more researchers. We carefully select a series of articles that are most representative among all. To grasp the latest developments, most selected references are published in the latest three years—$24$ articles are published in the first half of 2020, and $38$ articles are published in either 2018 or 2019. The reviewed articles are mainly chosen from the most popular journals in power system domain, such as *IEEE Transactions on Power System, IEEE Transactions on Smart Grid, Applied Energy, Journal of Power and Energy Systems*. Comparison with Related Review Articles --------------------------------------- There are two essential differences between this paper and related review articles. First, a brand-new taxonomy is proposed in this paper to better reveal the methodological features of how machine learning may improve optimization. Second, instead of providing a broad view of machine learning applications, this paper concentrates on the power system optimization and provides more in-depth technical discussions. Currently, most of existing review articles on machine learning and power system are quite similar in the structure of paper. Most of the reviews tend to categorize articles according to different applications [@RV3; @RV5; @RV6; @RV10; @RV11; @RV12; @RV13], e.g., the optimal power flow, economic dispatch, energy management, forecasting. The categorization like this is simple and clear for readers who are not familiar with power systems, such as researchers in computer science. Another kind of categorization is based on different learning methods. In [@RV1], machine learning approaches are summarized into seven categories: reinforcement learning, deep learning, transfer learning, parallel learning, hybrid learning, adversarial learning and ensemble learning. Another limitation is that the existing review articles fail to provide enough in-depth discussions because their topics are often very broad. Reference [@RV2] limited the topic to only dispatch and control, but it was still focused on the application level, similar to the reviews mentioned above. It is still unclear what role machine learning may play in power system optimization. Therefore, this paper intends to provide a different taxonomy to understand how machine learning approaches may help the power system optimization. The selected references will be divided into different categories based on their methodological features. We make further efforts to discuss about the key challenges in practical applications, and recommend some hopeful solutions. Contributions and Paper Structure --------------------------------- The major contributions of this review paper are summarized as follows. 1. We propose a novel and well-designed taxonomy according to different coordination pattern between machine learning approaches and power system optimization models. For each category, key technologies are studied and summarized from a series of typical and recent publications. 2. Major challenges in the learning-assisted power system optimization are fully discussed, and the latest research explorations are summarized to provide some possible solutions. In the rest of this paper, Section \[SEC-TAX\] proposes a well-designed taxonomy to categorize different researches. For each category, the key technologies are studied and summarized in Section \[SEC-TECH\]. We further discuss the major challenges and future opportunities in Section \[SEC-CHL\]. Section \[SEC-CONCL\] concludes the paper.  Taxonomy {#SEC-TAX} ======== This section provides a novel and well-designed taxonomy according to the different modeling patterns, which is completely different from the existing review papers that are focused on the applications or machine learning approaches. We pay special attention to a key question: in what aspect and to what extent can the machine learning approaches make a difference in some power system optimization tasks? Both the potential benefits and possible risks should be fully considered. Specifically, we focus on the coordination between the machine learning approaches and optimization models, and formulate four specific categories to extract valuable insights from existing literature. Fig. \[fig-framework\] provides an overall framework of the proposed taxonomy with the following four categories: a) **boundary parameter improvement**, b) **optimization option selection**, c) **surrogate model** and d) **hybrid model**. Let us consider the full-cycle configuration of a power system optimization model. Each optimization model is formulated by an objective function and several constraints, and the coefficients, or boundary parameters, of these formulas are important to guarantee a high-quality solution. Many classical optimization methods can be applied to solve an optimization model, and some optimization options, e.g., initial value or iterative step size, need to be defined in advance. Other alternatives are also powerful to get the final solution, including the surrogate machine learning model and well-designed hybrid model. Different from the classical optimization methods, these alternatives are able to improve the optimization process by the experience learned from the historical data. Based on the above discussion, the primary ideas of the proposed four categories are quite straightforward. The first category, boundary parameter improvement, intends to apply machine learning approaches for more accurate coefficients estimations. The second category is then focused on selecting better optimization options. The reinforcement learning and other specially designed machine learning approaches are included in the surrogate models or the third category. The last category, the hybrid models, covers a series of analytic and data-driven hybrid frameworks. Note that the last two categories are technically different from the first two—the coordination between the machine learning and optimization models change from a tandem structure to some iterative, coupled or other complex structures. Usually, these frameworks can offset the weakness inherent to using each part alone. Beyond forecasting, machine learning approaches are showing broad prospect for potential applications in power system optimization tasks. The taxonomy of this paper clearly points out several possibilities which is helpful in classification of existing researches as well as extending the current ideas to create new-style models. {width="0.8\linewidth"} Key Technologies {#SEC-TECH} ================ Based on the proposed taxonomy, this section summarizes the latest research progresses and key technologies in learning-assisted power system optimization. A list containing all reference details can be downloaded from [@Ref]. Fig. \[fig-Ref\] gives an overview of the selected references. Different colors are assigned according to the application scenarios, and the height of each slice is proportional to the number of selected references. We will next dig into the technical details of each category. Category 1. Boundary Parameter Improvement ------------------------------------------ In this category, the efficient machine learning approaches are combined to improve the accuracy of coefficients estimations, which can confine more accurate feasible region and further improve the quality of the optimum solution. Boundary parameter obtained from a traditional method may commonly not be accurate enough due to many reasons, including uncertainty caused by natural variability, randomness caused by human behaviors, and partial observability of power systems. **Uncertainty of natural variability.** This kind of uncertainty may lead to the inaccurate boundary parameter in power system optimization. With the high penetration of renewable energy, the power systems are faced with the challenges of increasing uncertainty. In  [@RN1], the authors applied the Generative Adversarial Networks to generate scenarios of renewable energy, which can be expanded easily to systems with an enormous number of uncertainties. Conditional Generative Adversarial Networks are then verified that it achieves better function accounting for uncertainties of wind power in contrast to classical methods [@RN2]. For fault type identification of transmission line with the large-scale renewable energy integration, a deep learning approach can accomplish automatic characteristic learning [@RN7]. At the time of decision making, it is uncertain what the parameters of energy price and available wind energy are. Reference [@RN9] combined the multivariate clustering technique and the recurrent neural network to deal with this uncertainty, of which a wind and storage power plant takes part in the pool market. Considering the socio-technical complexities appearing during the process of planning ,  [@RN15] developed a data-driven interdisciplinary modelling framework to analyze the distributed energy resources. **Randomness of human behaviors.** Previous literature has used the machine learning approaches to cut down the randomness resulting from a variety of human behaviors and obtain more accurate optimum solution. Reference [@RN3] proposed a long short-term memory neural network for decision-making problems to adequately handle the price uncertainty in electricity markets. The neural network plays a significant role in striking a balance between the comfort and energy in buildings [@RN10], achieving optimal dispatch in ancillary services market [@RN11] and reducing load curtailments [@RN12]. In [@RN16], a data-driven model was used to estimate parameters of price responses and try best to reduce prediction errors. A Stackelberg game-based market strategy was designed to increase profit of each market participant as much as possible, which guarantees operational security. **Partial observability.** Partial observability of power system may result in information loss and error, which makes it difficult to solve the optimum problems. Authors adopted deep learning approaches to identify time-varying parameter for composite load model [@RN4], phasor measurement unit (PMU) data manipulation attacks  [@RN5], phase in power distribution systems [@RN8] and built feature extraction framework for security rules [@RN17]. To enhance the observability of distribution systems,  [@RN6] presented a data-driven method to determine the daily consumption patterns of customers without smart meters. In addition, under any monitoring missing conditions, a data-driven Generative Adversarial Networks can be used to handle the dynamic security assessment without complex computation [@RN18]. Category 2. Optimization Option Selection ----------------------------------------- This category intends to increase the overall optimization efficiency by selecting better optimization options, e.g., initial values. Practical experiences show that some options have significant impacts on the convergence feature and speed, and the default settings might be far from optimality in some cases. Machine learning approaches, in this aspect, can provide effective guidance from the past experiences. A large proportion of researches adopt machine learning approaches to estimate a good initial value, which is beneficial to implement a warm starting. Reference [@RN41] proposed a “learn to initialize’’ strategy to improve the Gauss-Newton algorithm. The authors used a neural network to achieve this learning task, and designed a special loss function (only penalizing the maximized errors) to improve the overall performance. Reference [@RN48] established a predict-and-reconstruct approach to learn to predict the generation states (part of the decision variables), and reconstructed the phase angles (other part of the decision variables) using power flow equations. A deep neural network was trained for above tasks, and the network size was properly tuned according to the approximation accuracy. In [@RN39], a data-driven approach to reconstruct and mimic the solution of a centralized optimal power flow was proposed. The idea was that local controllers could achieve a near-optimal performance by learning the limited locally available data. Some extensions about the approximated initial values were discussed in [@RN43; @RN44; @RN45]. The supervised and transfer learning were applied in reference [@RN43] to estimate the Pareto front that could be regarded as a series of single initial values. This task was more difficult than the above tasks in [@RN41; @RN48; @RN39]. The numerical tests indicated that such estimation might cause large errors under some conditions, so careful validation and further fine-tuning were extremely important in this situation. Reference [@RN44] proposed a linear power flow model to quickly derive some approximated power flow states. This work was further extended in [@RN45] to tackle the challenges of the hidden measurement noises. The authors formulated three quadratic programming problems with Jacobian matrix guided constraints to make the approximated power flow model more robust to the data noises. The benefits in discrete optimization are more significant than expected. The early work [@RN49] introduced a combined approach for the unit commitment problem. The proposed approach first used a neural network to determine the discrete variables, and after that, applied a simulated annealing method to generate the continuous variables. According to the case study, the neural network found near optimal commitment results except for a minor difference, but it could help to achieve a roughly 50x computing speedup. A recent work [@RN47] made some further progresses in the security-constrained unit commitment. Many machine learning techniques have been adopted to learn from previously solved instances, and accelerate the computation by predicting the redundant constraints, good initial feasible solutions and affine subspaces where the optimal solution was likely to locate on. The authors achieved an averaged 4.3x speedup with optimality quality and 10.2x speedup without optimality guarantees but with no discernible solution difference. The insights were valuable that predicting warm starts or valid hyperplanes were significantly harder than the redundant constraints. Machine learning approaches were also useful for other optimization options. Reference [@RN46] created effective algorithm selectors by machine learning approaches and found less overloads and curtailment in the power flow management. Similar idea was also adopted in the unit commitment problem in [@RN42] where a learning model was trained to assign the weights for several given heuristic rules. Reference [@RN40] set up a three-stage framework (mid-term, short-term and real-time) for outage scheduling, and a nearest neighboring classifier was trained as a proxy to approximate the intermediate results to accelerate the mid-term decision. ![The basic structure of reinforcement learning. A reinforcement learning agent can sense the states of its environment and is able to take actions that affect the environment. The environment gives a reward according to the actions and the agent tries to maximize the total reward during a period.[]{data-label="fig-RL"}](rl_image2.pdf){width="0.95\linewidth"} Category 3. Surrogate Model --------------------------- This category seeks to completely replace traditional optimization models by some data-driven models. These surrogate models intend to deal with situations where analytic models are unavailable or too computationally expensive. They are also useful in performing flexible and dynamic operations to handle the increasing uncertainty in power systems. Reinforcement learning is the most prevalent method, and other methods based on supervised learning are also studied. As a major branch of machine learning, reinforcement learning has been specially developed for sequential decision-making problems that can be formulated as Markov decision process. It has been increasingly popular as a surrogate method to solve time-coupled optimization problems in power system operations. The basic idea of reinforcement learning is to construct an agent, which observes the states of the environment, acts according to its policy, receives the reward signal and upgrades its policy in order to maximize the expected long-term profit (represented by Q value). Various reinforcement learning algorithms have been developed with diverse methods of Q value estimation and policy representation. Combining reinforcement learning with deep learning, deep reinforcement learning further strengthens the agents’ ability to perceive the environment. Fig. \[fig-RL\] shows the basic structure of (deep) reinforcement learning. One typical advantage of reinforcement learning methods is that a model-free agent can develop an approximate model of the environment based on observations without any prior knowledge of the environment. When the response profiles of the entities in a distribution system were unknown, [@RN26] used deep neural networks to learn the entities’ behaviors, based on which an agent decided the optimal price signal of the distribution system. Reference [@RN25] applied fitted Q-iteration to the direct control of a heterogeneous cluster of electro-thermal loads with unknown characteristics. Historical data were added to the observations to obtain more information on the thermodynamic process, and a convolutional neural network was constructed to handle the high-dimensional inputs and capture the hidden patterns. Reference [@RN27] proposed the use of deep reinforcement learning to provide navigation for electric vehicles in need of recharging, so that the total travel time to the charging station and the charging cost were minimized. The model took advantage of the data from both the smart grid and the intelligent transportation system. Another advantage of reinforcement learning is its ability to achieve on-line control of flexible loads and storage devices with immediate response to the fluctuation of stochastic exogenous factors such as the renewable outputs and the real-time electricity price. For example,  [@Rh1] performed on-line building load optimization to minimize energy cost and users’ dissatisfaction with regard to the electricity price and PV output at every time step. Reference [@RN28] used an artificial neural network to perform hour-ahead price prediction, which was combined with reinforcement learning to optimize home demand response operation. Besides, similar methods have also been applied to microgrid scheduling scenarios, e.g., [@Rh2; @Rh3]. Distributed operation and incomplete information is also a motivation for applying reinforcement learning methods. With the deregulation of the electric power industry, every self-interested participant makes his own decisions with limited knowledge of the rest. Reference [@RN34] designed a multi-agent methodology based on deep policy gradient method, with every agent representing a self-interested generation company that explored its offering strategy through interactions with the market. For the purpose of current and voltage control, [@RN22] combined the consensus method and deep reinforcement learning to coordinate distributed generations in an islanded DC microgrid. To minimize the operation costs and the carbon emission, [@RN21] performed distributed reactive power optimization based on collaborative equilibrium Q-learning. Reference [@RN20] proposeed a multi-agent architecture to optimize the scheduling of EV charging, based on Q-learning and W-learning. The strategies of both selfish agents and collaborative agents were covered. Despite the popular applications of reinforcement learning, its limitations and defects should not be neglected. Many studies have applied model-free methods without paying attention to the physical characteristics and constraints. As a black box model, reinforcement learning’s performance heavily depends on the training data, and its behaviors are unknown under extreme circumstances that have not been learned before. Without the guarantee for safety and robustness, model-free methods are not suitable for situations allowing a narrow margin of error. In order to mitigate the above problem, safe reinforcement learning have been developed to guarantee that certain security constraints are met. The core idea of safe reinforcement learning is to introduce a penalty term corresponding to the security constraints, and minimizing the penalty term is prioritized during the learning process. Reference [@RN23] adopted safe deep reinforcement learning in the scheduling of EV charging in consideration of the charging constraints of EV batteries. Reference [@RN24] performed voltage and reactive power optimization of distribution network using a safe off-policy deep reinforcement learning algorithm to avoid voltage violations. Some studies have also considered the exploitation of physical information within deep learning methods. A model-based deep learning approach was proposed in [@RN36] for the calculation of probabilistic power flow. The training process of the neural network was guided by the physical characteristics of the grid, and the case study showed a great improvement in calculation speed. A graph convolutional network was trained in [@RN37] to capture the topology information of power system, based on which the optimal load-shedding under contingency was calculated. Reference [@RN38] used a convex neural network to approximate the electrothermal characteristics of a building, and then applied convex optimization methods to minimize the electricity consumption. In order to provide on-line control of plug-in electric bus, [@Rh5] chose length ratio as the representation of trip information, based on which a neural network was designed. Reference [@RN35] studied the real-time control of working fluid pump speed to optimize internal-combustion engine waste heat recovery. Dynamic programming and supervised learning methods were used to discover patterns, and the model inputs were specially designed according to the differential equations that describe the physical characteristics of the engine. ![Detailed structural explanations for the existing researches. An iterative structure takes learning steps and optimization steps alternately, while a coupled structure replaces some inaccurate parts in the optimization model with machine learning.[]{data-label="fig-cat4"}](cat4-subclass-combine.pdf){width="1\linewidth"} Category 4. Hybrid Model ------------------------ This category combines the machine learning and optimization models together to boost the overall performance. As shown in Fig. \[fig-framework\](d), there are two typical hybrid types, and more supplementary structural details are provided in Fig. \[fig-cat4\]. Specifically, one type is an iterative structure with machine learning and optimization steps, and the other is a coupled structure that embeds the machine learning models to replace some inaccurate parts in the optimization models. Above combinations reflect a deep integration and coordination between these two models, and is thus able to fully explore the hidden potentials. Some well designs can achieve higher modeling accuracy and optimality property at the same time, e.g., [@RN62; @RN64]. **Iterative structure.** This structure is very common in many optimization and optimal control tasks, and the typical flowchart is shown in Fig. \[fig-cat4\]. Reference [@RN62] proposed an accelerated distributed demand response algorithm, which applied a neural network to iteratively predict the consumers’ price responses. After that, the authors further developed a transformation model to search for better step sizes to enhance the performance of the optimization step. The most promising feature of this algorithm was that it can cut off 60-80% iterations and still guarantee optimality at the same time. In [@RN63], in addition to selecting step sizes, neural network was also beneficial to calculate the searching directions in a sequential linear programming model. This paper considered an information asymmetric situation where a retailer company needed to make decisions with limited knowledge of the consumers’ models. The dual neural networks were specially designed and one of them is working to transform and derive the searching directions. Some intelligent optimization algorithms were also implemented in some researches. Reference [@RN51; @RN50] formulated a similar optimal dynamic pricing model for retailers but used the genetic algorithm and mean-variance mapping optimization for final solutions respectively. In their models, the consumer responsive features were learned and coordinated with the intelligent optimization algorithms later. Reference [@RN59] established a three-stage model predictive control procedure where neural networks were continuously predicting the energy demand and renewable energy supply to support the optimal decisions in the next stage. In [@RN65], a hybrid model and data-driven simulation platform was operating in real-time for selecting the power system security features. Within this platform, the analytical models were establishing the samples, and these samples were then learned and analyzed to extract some fine-tuned security rules. **Coupled structure.** This structure is more complex than the above iterative structure. As shown in Fig. \[fig-cat4\], current researches have made efforts on embedding the machine learning models in the objective function or constraints, and the technical difficulties are mainly on how to design the optimization algorithms and transformation process. Reference [@RN68] extracted the mapping function from a neural network and incorporated this function as a constraint in an optimal power flow model. This hybrid model was then solved by a nonlinear programming optimizer. Reference [@RN52; @RN60] integrated the learning models in the constraints of the household or building energy optimization model. As for the optimization algorithm, [@RN52] applied particle swarm optimization, while [@RN53] applied a hybrid approach of exhaustive search method and subsequent quadratic programming. Similarly, the particle swarm optimization was also conducted in [@RN54] to solve the optimization with the constraints modeled by a radial basis function neural network. In [@RN56], a Bayesian neural network that learned to predict the steady-states was embedded in a preventive control model. This model was later solved by the derivation-free Bayesian optimization. Several measures are designed to transform the machine learning models so that the optimization model with these embedded parts could be easily solved. The piecewise linear equations were used as a transformation process in [@RN57] to replicate the neural network. After integrating these equations, the final model turned out to be a mixed integer programming. Reference [@RN64; @RN61] tried to choose different machine learning approaches as an alternative. In [@RN64], the sparse oblique decision tree was applied to learn some accurate, understandable and linear security rules for economic dispatch. These rules could be embedded in the optimization as several mixed integer linear constraints. In [@RN61], the authors conducted an extreme learning machine to enhance the hydrostatic tidal turbine control. This extreme learning machine was basically a linear model and can be easily transformed to some linear constraints. Other special technique included [@RN67], where the authors designed a sequential approximation method with dynamically trained neural networks. Such method might be more suitable for small networks because the dynamic training was often very time-consuming. Other researchers have formulated the objective functions with machine learning element. Reference [@RN58] trained a neural network based upon the combined heat and power simulation results from a physics-based model, and formulated an economic dispatch model with this neural network integrated in the objective function. The dispatch schedule was optimized by adopting a genetic algorithm. Reference [@RN55] designed a convolutional neural network classifier for the faulted line localization. The placement problem of phasor measurement units was therefore establishes as a hybrid model whose objective function was composed of the loss function of the convolutional neural network. This optimization was essentially a special hyperparameter optimization, and was later solved by a greedy algorithm. Comparison and Comments ----------------------- Overall, the selected references have very diverse ideas and features. We intend to briefly compare all above categories and make some further comments on their applications. **Difficulty of the learning tasks.** Generally, with the same data dimension and precision requirement, the Category 3 is likely to include the most difficult learning tasks. Here, the so-called “difficulty’’ can be roughly measured by how large a machine learning model is needed to finish the task. The Category 2 usually takes the second place, but the situations for the remaining two categories are uncertain. Although there are some exceptions, the surrogate models (Category 3), especially those reinforcement learning models, are usually most complicated to calibrate, and thus a large amount of data and more computing resources are needed. These models might meet great difficulty if the simulation data are not available or effective. Further, the boundary parameter improvement (Category 1) are often easier to achieve, but its potential benefits for the optimization model are strongly related to the specific cases. As for the remaining two categories, some physical knowledge may be relatively helpful to boost the overall performance. **Difficulty of the model designs.** Generally, for a similar task, one has to make more efforts, or even case-by-case designs, when applying the Category 4 or Category 2 methods. In contrast, the Category 3 is more likely to have some off-the-shelf tools, and little processing design is further needed. We highlight two significant features in the power system optimization tasks: highly sensitive to the decision errors, and dependent on extensive physical knowledge. Leveraging these physical knowledge to design a dedicated framework or upgrade the machine learning models is promising but still under exploration. In this aspect, the hybrid model (Category 4) and other variants in other categories deserve more exploration although more design efforts are needed. Challenges and Opportunities {#SEC-CHL} ============================ Until now, few machine learning approaches have been realized in real-world power system operation. The key limitations of these approaches, therefore, should be carefully analyzed and handled. With this purpose, this section points out three major challenges and introduces some latest developments accordingly. Data Bottleneck --------------- Collecting clean and reliable data is essential to every machine learning application, and the data requirement for the most recent deep learning models is even higher. Two special features for the power system data are worthy of attention: First, very few real-world data sets are public available (due to privacy concerns or confidentiality requirements [@RV1]), and simulation data are widely applied as an alternative [@RV2]. Second, imbalanced data sets are very common, and in many cases, those rare parts are extremely important (e.g., unstable system conditions [@RV10]). Data issues may have adverse impacts on all the proposed categories, and thus become the major bottleneck for real-world applications. It will apparently be more risky for those that are in need of larger data volume. In addition to enlarging the data sets by policy or mechanism efforts, there are also some emerging technologies that could help tackle these data issues—data augmentation and few-shot learning. Data augmentation is a strategy to significantly increase data volume by a series of transform operations. Reference [@RN102] collected a list of useful resources, including classical techniques, papers and Github repositories. Let us take the time series data augmentation as an example. The mainstream techniques include simple operations (warping, jittering and perturbing) and advanced operations (embedding space and generative approaches). Few-shot learning intends to train machine learning models with limited amount of data. The basic idea is to use prior knowledge to alleviate the unreliable performance of empirical risk minimizer. In this aspect, the authors of [@RN103] reviewed some model-based methods (constrain the model complexity) and algorithm-based methods (constrain the search strategy for optimal parameters). Robustness and Prediction Errors -------------------------------- Power system optimization imposes high requirements on accuracy as well as robustness. To design a learning-assisted optimization system, it is crucial to take care of the specific vulnerabilities that deteriorates the robust performances. Occasionally, some small input changes may lead to a significant accuracy drop of the machine learning models. There are two perspectives to analyze and understand the robustness issue: - The output changes of machine learning models with a fluctuation in model inputs. - The solution changes of optimization models with a fluctuation in machine learning model outputs. For Category 3, these two perspectives merge into one, while for others, the second perspective is important but it often lacks sufficient attention. We will next introduce some latest research explorations in the above perspectives. Within the first perspective, adversarial examples are fairly helpful to examine the robustness of machine learning approaches. Reference [@RN104] studied the worst-case adversarial perturbations and found the robustness might be badly harmed. A recent work [@RN101] argued that, in fact, these adversarial examples were learned features rather than bugs. The authors further analyzed and found that there existed both robust and non-robust features, and the latter ones were the main cause for some specific vulnerabilities. The second perspective is highly related to the optimization model characteristics. We can categorize the model parameters and optimization options into two parts: error-tolerant parameters or error-sensitive parameters. Fig. \[fig-pred-error\] gives an illustrative example to show the difference. Our main focus is on the overall optimization performance when machine learning models make positive and negative prediction errors. It is shown that an error-tolerant parameter can robustly guarantee a shorter running time for optimization. This reflects that when designing the models for Category 1, 2 and 4, an error-tolerant coordination parameter could benefit the robustness of the whole system. ![Illustration of error-tolerant parameters and error-sensitive parameters. Error-tolerant parameters, represented by the blue curve, can ensure a shorter running time for optimization than error-sensitive parameters.[]{data-label="fig-pred-error"}](pred_error2.pdf){width="0.94\linewidth"} Interpretability ---------------- Interpretability is the degree to which people can understand the decisions made by machine learning approaches. Many advanced machine learning models, e.g., neural networks, are widely regarded as “black box’’ models [@RV1]. Some ensemble learning methods further combine several models to achieve higher prediction accuracy at the expense of interpretability. As a consequence, low interpretability also results in low acceptance in the power industry. There are two perspectives of the interpretability: - The explanation of those parameters learned by machine learning approaches. - The explanation of why machine learning model outputs can boost the optimization performance. Similar as the previous subsection, for Category 3, these two perspectives merge into one, while for Category 1, the second perspective is intuitive (closer to ground truth is better). For the remaining two categories, both the two perspectives are important factors that should be carefully paid attention to. The first perspective concerns about a conventional issue that have been discussed for years in machine learning community. Reference [@RN100] introduced most of the important progresses in this domain which are also shown in Fig. \[fig-interp\]. Basically, there are two options to achieve interpretability: applying an interpretable model or making further processing on a black box model. Model-agnostic interpretation methods, with many important advances in recent years, are able to extract more human-friendly features and visualization results. The second perspective, specific for learning-assisted optimization, can be probably handled with help of optimization theory. Often, a set of optimal configurations are very beneficial to boost the optimization performance, and machine learning approaches can approximate those configurations calculated from optimization theory. A typical example is the near-optimal step size selection in [@RN62]. ![Flowchart of achieving human-friendly interpretability. Two options—applying an interpretable model or making further processing on a black box model—are illustrated to translate the original data to some easy-to-understand explanations.[]{data-label="fig-interp"}](interpretability2.pdf){width="1\linewidth"} Conclusion {#SEC-CONCL} ========== This paper conducts a comprehensive review of learning-assisted power system optimization. A novel and well-designed taxonomy is proposed in this paper to better categorize the existing articles by their methodological features. The latest research progresses and key technologies are thoroughly summarized and discussed, together with the further comments on the key challenges in real-world applications. We strongly realize that the deep integration of machine learning and power system optimization is a future trend. This review is expected to offer some useful information as well as deep thoughts in this domain. [^1]: G. Ruan, H. Zhong, G. Zhang, Y. He and X. Wang are with the State Key Lab of Power Systems, Department of Electrical Engineering, Tsinghua University, Beijing 100084, China. T. Pu is with the China Electric Power Research Institute, Beijing 100192, China. [^2]: Corresponding author: H. Zhong ([email protected]).

ArXiv

--- author: - | Angli Liu\ \ Jiangtao Li\ \ Guobin Shen\ \ - | Chao Sun\ \ Liqun Li\ \ Feng Zhao\ \ bibliography: - 'ourbib.bib' date: 25 November 2014 title: ': Enabling Low-power Duplex Visible Light Communication' ---

ArXiv

--- abstract: 'Hints for sizable $\sin^2 \theta_{13}$ have been reported in earlier global neutrino oscillation data analyses as well as will be reported in this work, and quite recently by the Double Chooz experiment. However, as we enter the era of precision neutrino oscillation experiments, terms linear in $\sin\theta_{13}$ will no longer be negligible, and its sign would affect the extraction of other oscillation parameters. The sign of $\sin\theta_{13}$ also plays a crucial role in the determination of the CP-violating phase. In this work we show that by adopting an alternative parametrization for the Pontecorvo-Maki-Nakagawa-Sakata (PMNS) mixing matrix, one already has a chance to infer the sign of each mixing angle in the conventional parametrization using existing global neutrino data. A weak preference for negative $\sin \theta_{13}$ is found. In particular, the solar data suggest that $\sin\theta_{13} > 0$ while all other data the opposite. This leads to the speculation on whether the Mikheyev-Smirnov-Wolfenstein (MSW) effect is responsible. In this work we found that in the new mixing matrix parametrization, the $68\%$ CL constraints on the three mixing angles are comparable to those estimated in the conventional parametrization adopted in the literature. Owing to the strong correlations among the three mixing angles in the new parametrization, the advantages of doing the global neutrino oscillation analysis using data from past, current, and near future neutrino oscillation experiments shall become manifest.' author: - Melin Huang - 'S. D. Reitzner' - 'Wei-Chun Tsai' - Huitzu Tu title: | [ **Global Neutrino Data Analysis and the Quest to Pin Down\ $\sin\theta_{13}$ in Different Mixing Matrix Parametrizations** ]{} --- Introduction {#sec:intro} ============ The experiments involving solar, reactor, atmospheric and accelerator neutrinos have established a picture of neutrino oscillations caused by non-zero neutrino masses and mixing among different neutrino flavors (see e.g. Ref. [@Nakamura:2010zzi]). The phenomenology of neutrino oscillations arising from the mismatch between the weak and the mass eigenstates can be described by the Pontecorvo-Maki-Nakagawa-Sakata (PMNS) mixing matrix [@Maki:1962mu; @Pontecorvo:1967fh]. This matrix can be parametrized in various ways as seen in the literature (e.g., [@Schechter:1980gr; @Fritzsch:1997st; @Choubey:2000bf; @Giunti_Kim; @Zheng:2010kp; @Huang:2011by]). All experimental data except those from the LSND [@Aguilar:2001ty] can be well described assuming three active neutrinos. In the case of Dirac neutrinos, the $3 \times 3$ unitary mixing matrix is characterized by three Euler angles and one physical phase, and can be expressed as a product of three rotation matrices. The physical phase can be responsible for CP violation in the neutrino sector. As mentioned in Ref. [@Fritzsch:1997st], the CP-violating phase can be associated with the sine or cosine of any mixing angle or with the identity entry in any of the three rotation matrices. In this work, we will follow the standard Cabibbo-Kobayashi-Maskawa (CKM) matrix [@Chau:1984fp] for the assignment of the CP-violating phase. Define $$R_{23} = \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & c_{23} & s_{23} \\ 0 & -s_{23} & c_{23} \end{array} \right), \hskip0.5cm R_{13} = \left( \begin{array}{ccc} c_{13} & 0 & s_{13} \\ 0 & 1 & 0 \\ -s_{13} & 0 & c_{13} \end{array} \right), \vspace{-0.3cm}$$ $$R_{12} = \left( \begin{array}{ccc} c_{12} & s_{12} & 0 \\ -s_{12} & c_{12} & 0 \\ 0 & 0 & 1 \end{array} \right), \vspace{-0.3cm} \label{eq:Gfit_01}$$ and $$W_{23} = \left( \begin{array}{ccc} 1 & 0 & 0 \\ 0 & c_{23} & s_{23} e^{-i{\delta_{cp}}} \\ 0 & -s_{23} e^{i{\delta_{cp}}} & c_{23} \end{array} \right), \nonumber \\$$ $$W_{13} = \left( \begin{array}{ccc} c_{13} & 0 & s_{13} e^{-i{\delta_{cp}}} \\ 0 & 1 & 0 \\ -s_{13} e^{i{\delta_{cp}}} & 0 & c_{13} \end{array} \right), \nonumber \\$$ $$W_{12} = \left( \begin{array}{ccc} c_{12} & s_{12} e^{-i{\delta_{cp}}} & 0 \\ -s_{12} e^{i{\delta_{cp}}} & c_{12} & 0 \\ 0 & 0 & 1 \end{array} \right), \label{eq:Gfit_02}$$ where $\theta_{ij}$ and ${\delta_{cp}}$ are the mixing angles and the CP-violating phase, respectively, $c_{ij} \equiv \cos\theta_{ij}$ and $s_{ij} \equiv \sin\theta_{ij}$. As the combination $R_{23}W_{13}R_{12}$ is the standard choice for describing the quark mixing, it has been adopted to be the conventional parametrization for the mixing matrix for Dirac neutrinos. This choice of the parametrization was actually made prior to the era when neutrino oscillation data became available. Later, the solar neutrino experiments [@Aharmim:2009gd; @Aharmim:2011vm; @Hosaka:2005um; @Cravens:2008zn], reactor [@Gando:2010aa; @Apollonio:2002gd], long-baseline (LBL) accelerator [@Ahn:2006zza; @Adamson:2011ig; @MINOS_nue_app_2011], as well as the atmospheric neutrino experiments [@Wendell:2010md; @Aharmim:2009zm; @Hosaka:2006zd] all choose this parametrization to present their results. Using existing global neutrino data, most phenomenology works [@Fogli:2011qn; @Schwetz:2011qt; @GonzalezGarcia:2010er; @Roa:2009wp; @Balantekin:2008zm; @Ge:2008sj; @Goswami:2004cn; @Choubey:2003uw] also employ this parametrization to determine neutrino oscillation parameters. In the conventional parametrization, the three mixing angles happen to nearly decouple for solar, atmospheric/accelerator, and reactor neutrino oscillation experiments due to the different neutrino energies and traveling distances involved. The $\theta_{12}$ and $\theta_{23}$ angles are well determined by solar and atmospheric experiments, respectively. Current-generation short-baseline reactor experiments such as the Daya Bay [@Guo:2007ug], Double Chooz [@Ardellier:2006mn], RENO [@Joo:2007zzb] and the Angra [@Anjos:2005pg] experiments are exploited to pin down the yet unknown $\theta_{13}$ value. Non-zero or sizable $\theta_{13}$ values are predicted by many neutrino mass models (see e.g. Ref. [@Chen:2008eq] for a nice compilation of existing model predictions), and supported by global neutrino data analyses (see e.g. Ref [@Fogli:2011qn; @Schwetz:2011qt; @GonzalezGarcia:2010er; @Ge:2008sj] and this work). Very recently, the Double Chooz experiment [@double_chooz_2011] has reported their preliminary results of $\sin^2 2\theta_{13} = 0.085 \pm 0.051$ (68$\%$ CL). When this is confirmed in the future by other experiments with even better sensitivities, it will be good news for near-future experiments such as T2K [@T2K_expt], NO$\nu$A [@Ayres:2004js], T2HK [@Itow:2001ee], T2KK [@Hagiwara:2006vn]. Their goal of measuring the CP-violating effect will be more reachable. This is not the end of the story for the neutrino oscillation community. Another issue is the determination of the sign of $\sin\theta_{13}$. As we enter the era of precision neutrino oscillation experiments, terms linear in $\sin\theta_{13}$ may no longer be negligible in fitting the mixing angles. From the Jarlskog invariant quantity [@Jarlskog:1985ht; @Wu:1985ea] of CP violation, one sees that the sign of $\sin\theta_{13}$ would also have an impact on the ${\delta_{cp}}$ determination. However, in the conventional parametrization, the reactor and solar experiments are only sensitive to $\sin^2 \theta_{13}$. For LBL accelerator or atmospheric experiments, the observable has terms linear in $\sin\theta_{13}$, but only sub-dominant. Despite the unfavorable situation one faces when working in the conventional parametrization, a first attempt to determine the sign of $\sin\theta_{13}$ has been made in Ref. [@Roa:2009wp] using the LBL accelerator, CHOOZ, and atmospheric neutrino data. In fact the conventional parametrization is not the only way to establish the mixing matrix. As proposed by several authors [@Schechter:1980gr; @Fritzsch:1997st; @Zheng:2010kp; @Giunti_Kim; @Choubey:2000bf; @Huang:2011by], the corresponding mixing parameters may be more accessible, without sacrificing accuracy, in other parametrizations. In this work we follow the approach of Ref. [@Huang:2011by] to explore this possibility. Besides the conventional parametrization, we perform a global neutrino oscillation analysis adopting an alternative parametrization, $R_{13} W_{12} R_{23}$, in which the observables have leading terms linear in any of the three mixing angles. Equipped with the analysis results obtained in these two parametrizations, we will be eligible to address a couple of issues. Can one determine the sign of $\sin \theta_{13}$? Are there other parametrizations which can provide comparable sensitivities in extracting neutrino oscillation parameters as the conventional one? How are the three mixing angles correlated with each other therein? Do the matter (MSW) effects [@Wolfenstein_msw; @Mik_Smir_msw] affect the sign of $\sin \theta_{13}$? This paper is organized as follows. In Section \[sec:anal\] we briefly describe our data fitting procedure in the conventional parametrization (to be denoted by ${\mathcal{A}}$). In Section \[sec:anal\_results\] we present and discuss our results obtained from a similar analysis done in an alternative parametrization (to be denoted by ${\mathcal{D}}$). Section \[sec:summary\] gives our summary and outlook. Individual analysis approaches for each neutrino oscillation experiment are detailed in Appendix \[sec:apdx\_anal\_solar\] through \[sec:apdx\_anal\_atmos\]. Analysis {#sec:anal} ======== Table I in Ref. [@Huang:2011by] indicates that the predicted neutrino mixing angles in the parametrizations ${R_{23}}{W_{13}}{R_{12}}$, ${R_{23}}{W_{12}}{R_{13}}$, and ${R_{13}}{W_{23}}{R_{12}}$ come up with similar values, while those in ${R_{13}}{W_{12}}{R_{23}}$, ${R_{12}}{W_{23}}{R_{13}}$, and ${R_{12}}{W_{13}}{R_{23}}$ end up with similar values. We thus preform a global fit of the neutrino mixing parameters using parameterizations $R_{23}W_{13}R_{12}$ and $R_{13}W_{12}R_{23}$, taking advantage of the fact that these two parameterizations will have dissimilar outcomes. Following Ref. [@Huang:2011by], we denote them as the ${\mathcal{A}}$ and ${\mathcal{D}}$ ’representations’ respectively, where representation ${\mathcal{A}}$ corresponds to the conventional mixing matrix parametrization. In what follows, notations (${\Delta^{\mathcal{A}} m^{2}_{ij}}$, ${\theta^{\mathcal{A}} }_{ij}$) and (${\Delta^{\mathcal{D}} m^{2}_{ij}}$, ${\theta^{\mathcal{D}} }_{ij}$) stand for the oscillation parameters directly determined using representations ${\mathcal{A}}$ and ${\mathcal{D}}$, respectively. On the other hand, the notation ${\theta^{\mathcal{DA}} }_{ij}$ symbolizes the three mixing angles that are initially extracted from representation ${\mathcal{D}}$ and then translated to representation ${\mathcal{A}}$ by the transformation method described in Ref. [@Huang:2011by]. The transformation from representation ${\mathcal{D}}$ to ${\mathcal{A}}$ is outlined in Appendix \[apdx:Gfit\_NumParamSol\_DA\]. The global neutrino data used in this work include [*(i)*]{} solar data from rates measured in chlorine [@Cleveland:1998nv] and gallium [@Abdurashitov:2009tn] experiments, the rate of $^{7}$Be solar neutrinos measured in the Borexino [@Arpesella:2008mt] experiment, Super-Kamiokande (SK) phase I & II day/night spectra [@Hosaka:2005um; @Cravens:2008zn], and SNO phases I & II $\nu_e$ survival probability [@Aharmim:2009gd]; [*(ii)*]{} reactor data from KamLAND [@Gando:2010aa] and CHOOZ [@Apollonio:2002gd]; [*(iii)*]{} LBL accelerator data from K2K [@Ahn:2006zza] and MINOS $\nu_{\mu}$ disappearance channel [@Adamson:2011ig] and $\nu_e$ appearance channel [@MINOS_nue_app_2011]; and [*(iv)*]{} atmospheric data from SK phase I [@Hosaka:2006zd] and SNO [@Aharmim:2009zm]. In this work, the re-evaluated ${\bar{\nu}_e}$ flux from nuclear power plants [@Mueller:2011nm] is not taken into consideration for the reactor data. In addition, the large number of bins in the atmospheric data of SK phases II and III, as well as the lack of information, prevents us from reproducing their results. Therefore we do not include the SK-II and SK-III atmospheric data in our work. Since our purpose is to investigate the neutrino mixing phenomenology in different parametrizations using the same data sets and analysis conditions, the absence of these factors should not have any impact on the conclusions of this work. Described below is our analysis of the global neutrino data. Appendix \[sec:apdx\_anal\_solar\] outlines the analyses for each solar experiment employed in this work which includes chlorine, gallium, Borexino, SK and SNO. We use the Bahcall solar neutrino spectra [@Solar_Nu_Eng_Bahcall] except for the $^{8}$B neutrino spectrum, which is from Ref. [@Winter:2004kf]. Neutrino survival probabilities in the Sun are estimated using the BS05(OP) model [@Bahcall:2004pz] for the neutrino production rates at different solar radii. We do not use the more recent BPS09(GS) or BPS09(AGSS09) models [@Serenelli:2009yc] because of the conservative model uncertainties. For estimating neutrino survival probabilities inside the Earth, the Earth density profile of PEM-C [@PEM-C] (rather than PREM [@PREM]) is adopted in this work. It assumes the continental crust for the outer most layer of the Earth where the solar neutrinos enter the detectors. In our analysis, the flux of $^{8}$B solar neutrinos is a nuisance parameter while the fluxes of other solar neutrinos are taken from the BS05 model [@Bahcall:2004pz]. Neutrino oscillation probabilities are calculated using the adiabatic approximation [@Ioannisian_2004]. It has been verified to yield equivalent results as the numerical calculation does for several sets of neutrino oscillation parameters in the LMA region. Our analyses of oscillation parameter fitting using the KamLAND and the CHOOZ ${\bar{\nu}_e}$ oscillation events are delineated in Appendix \[sec:apdx\_anal\_reactor\]. Owing to the short distances between the source and the detector, matter effects are generally insignificant and oscillation probabilities in vacuum suffice. This is also the case for accelerator neutrino analyses. Appendix \[sec:apdx\_anal\_accel\] describes the analyses of oscillation parameter fitting using the K2K and the MINOS ${\nu_{\mu}}$ disappearance channel as well as the MINOS ${\nu_{e}}$ appearance channel. Atmospheric neutrino data from the SK [@Hosaka:2006zd] and the SNO [@Aharmim:2009zm] experiments are included in our analysis as described in Appendix \[sec:apdx\_anal\_atmos\]. We first employ the NUANCE package [@Casper:2002sd] to simulate atmospheric neutrino events assuming no oscillation effects. Neutrino oscillations in the atmosphere and inside the Earth are then incorporated by the “weighting factors”, Eq. (\[eq:weighting\_factors\]), with the matter effects taken into account by following the prescription of Ref. [@Barger:1980tf]. We apply similar criteria and cuts on the kinematics of the simulated events so that we achieve the same selection efficiency as Ref. [@Hosaka:2006zd] does for SK atmospheric data and as Ref. [@Aharmim:2009zm] does for SNO atmospheric data. -------------------------- ----------------------------------------- --------------------------------------- ---------------------------------------------- ---------------------------------------------- ---------------------------------------------- Data Sample $|{{\Delta^{\mathcal{A}} m^{2}_{21}}}|$ $|{\Delta^{\mathcal{A}} m^{2}_{32}}|$ ${\theta^{\mathcal{A}} }_{12}$ ${\theta^{\mathcal{A}} }_{23}$ $|{\theta^{\mathcal{A}} }_{13}|$ \[$10^{-5}$ eV$^2$\] \[$10^{-3}$ eV$^2$\] \[${\tan^{2}{\theta^{\mathcal{A}} }_{12}}$\] \[${\sin^{2}{\theta^{\mathcal{A}} }_{23}}$\] \[${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$\] Solar-only $6.55^{+3.21}_{-2.57}$ $33.587^{+2.156}_{-1.603}$ $5.444^{+8.735}_{-5.444}$ \[$0.441^{+0.077}_{-0.051}$\] \[$0.009^{+0.051}_{-0.009}$\] KamLAND-only $7.57^{+0.27}_{-0.22}$ $34.659^{+5.459}_{-5.010}$ $8.723^{+7.495}_{-8.723}$ \[$0.478^{+0.232}_{-0.154}$\] \[$0.023^{+0.055}_{-0.023}$\] Solar+KamLAND $7.57^{+0.26}_{-0.22}$ $34.001^{+0.881}_{-0.900}$ $9.458^{+2.651}_{-3.718}$ \[$0.455^{+0.031}_{-0.030}$\] \[$0.027^{+0.020}_{-0.017}$\] Accel+Atmos $2.30^{+0.21 }_{-0.11 }$ $43.739^{+4.391}_{-4.391}$ $5.132^{+3.591}_{-5.132}$ +CHOOZ \[$0.478^{+0.058}_{-0.058}$\] \[$0.008^{+0.015}_{-0.008}$\] Global $7.492^{+1.231}_{-1.753}$ $(\Delta \chi^2 = 1.0)$ \[$0.017^{+0.006}_{-0.007}$\] Global $7.492^{+1.787}_{-3.049}$ $(\Delta \chi^2 = 2.30)$ \[$0.017^{+0.009}_{-0.011}$\] -------------------------- ----------------------------------------- --------------------------------------- ---------------------------------------------- ---------------------------------------------- ---------------------------------------------- We first perform a global analysis using representation ${\mathcal{A}}$ in order to compare with existing results from the literature. As the current global neutrino data is insensitive to ${\delta_{cp}}$, we have selected ${\delta_{cp}}=0$. According to the study presented in Ref. [@Huang:2011by], this choice also turns out to be adequate for representation ${\mathcal{D}}$ since a rephasing still gives ${\delta_{cp}}=0$ therein. Table \[table:best\_fits\_A\] summarizes our best-fit results for the oscillation parameters from various three-flavor analyses in this case. Assuming CPT invariance, the results from the KamLAND reactor experiment and all solar experiments are combined in the $\chi^2$ function: $\chi^2_{\rm sol + KL} \equiv \chi^2_{\rm solar} + \chi^2_{\rm KL}$. We minimize $\chi^2_{\rm sol + KL}$ with respect to the oscillation parameters ${\Delta^{\mathcal{A}} m^{2}_{21}}$, ${\tan^{2}{\theta^{\mathcal{A}} }_{12}}$, and ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$, as well as the $^{8}$B solar neutrino flux in a three-flavor hypothesis. The other two parameters are fixed at ${\Delta^{\mathcal{A}} m^{2}_{32}}= 2.4 \times10^{-3}$ eV$^2$ and ${\sin^{2}{\theta^{\mathcal{A}} }_{23}}= 0.5$. The best-fit values are found to agree with the results presented in Refs. [@Aharmim:2009gd; @Gando:2010aa]. Figure \[fig:SolKL\_caseA\] shows the allowed regions in the planes of (${\tan^{2}{\theta^{\mathcal{A}} }_{12}}$, ${\Delta^{\mathcal{A}} m^{2}_{21}}$), (${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$, ${\Delta^{\mathcal{A}} m^{2}_{21}}$), and (${\tan^{2}{\theta^{\mathcal{A}} }_{12}}$, ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$). The $\chi^2$ functions for K2K, MINOS ${\nu_{\mu}}$ disappearance, MINOS ${\nu_{e}}$ appearance, CHOOZ, as well as SK and SNO atmospheric neutrino data are also summed together: $\chi^2_{\rm acc + CHOOZ + atm} \equiv \chi^2_{\rm acc} + \chi^2_{\rm CHOOZ} + \chi^2_{\rm atm}$. By minimizing this $\chi^2$ with respect to the oscillation parameters ${\Delta^{\mathcal{A}} m^{2}_{32}}$, ${\sin^{2}{\theta^{\mathcal{A}} }_{23}}$, and ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$ while fixing ${\Delta^{\mathcal{A}} m^{2}_{21}}= 7.67 \times10^{-5}$ eV$^2$ and ${\tan^{2}{\theta^{\mathcal{A}} }_{12}}= 0.427$, and assuming a normal mass hierarchy, the best-fit values are found to be consistent with the most recent results in Ref. [@Schwetz:2011qt; @Fogli:2011qn]. Figure \[fig:numu\_caseA\] shows the likelihood contours in the planes of (${\sin^{2}{\theta^{\mathcal{A}} }_{23}}$, ${\Delta^{\mathcal{A}} m^{2}_{32}}$), (${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$, ${\Delta^{\mathcal{A}} m^{2}_{32}}$), and (${\sin^{2}{\theta^{\mathcal{A}} }_{23}}$, ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$). Finally, for individual data sample mentioned above, i.e. the solar-only, KamLAND, solar$+$KamLAND, and accelerator$+$CHOOZ$+$atmospheric, the differences ($\Delta \chi^2$) between $\chi^2$ and the minimum $\chi^2_{\rm min}$ as a function of ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$ are obtained by marginalizing the other two oscillation parameters. Our results are shown in Fig. \[fig:Dchi2\_caseA\] for $\Delta \chi^2_{\rm solar}$, $\Delta \chi^2_{\rm KL}$, $\Delta \chi^2_{\rm sol + KL}$ and $\Delta \chi^2_{\rm acc + CHOOZ + atm}$. The global result is then $$\Delta \chi^2_{\rm global} ({\sin^{2}{\theta^{\mathcal{A}} }_{13}}) \equiv \Delta \chi^2_{\rm sol + KL} + \Delta \chi^2_{\rm acc + CHOOZ + atm}\, .$$ Our result for ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$ is consistent with those from other global neutrino data analyses (e.g., [@Aharmim:2009gd; @Gando:2010aa; @Fogli:2011qn; @Schwetz:2011qt; @GonzalezGarcia:2010er]). We also find a weak hint for non-zero ${\sin^{2}{\theta^{\mathcal{A}} }_{13}}$ at 95$\%$ CL, but it is not easy to determine the sign of $\sin {\theta^{\mathcal{A}} }_{13}$. Results and Discussions {#sec:anal_results} ======================= Changing to representation ${\mathcal{D}}$ and fixing ${\delta_{cp}}=0$, we perform a similar $\chi^2$ analysis as that in previous section using exactly the same data sets. After determining the oscillation parameters ${\Delta^{\mathcal{D}} m^{2}_{ij}}$ and ${\theta^{\mathcal{D}} }_{ij}$, with the transformation strategy described in Appendix \[apdx:Gfit\_NumParamSol\_DA\], we also perform a translation of the best-fit mixing angles back to representation ${\mathcal{A}}$, denoted by ${\theta^{\mathcal{DA}} }_{ij}$. Both results are given in Table \[table:best\_fits\_DA\] for a comparison with the corresponding ${\Delta^{\mathcal{A}} m^{2}_{ij}}$ and ${\theta^{\mathcal{A}} }_{ij}$ retrieved directly in representation ${\mathcal{A}}$. -------------------------- ----------------------------------------- --------------------------------------- ----------------------------------- --------------------------------- ----------------------------------- ----------------------------------- --------------------------------- ----------------------------------- Data Sample $|{{\Delta^{\mathcal{D}} m^{2}_{21}}}|$ $|{\Delta^{\mathcal{D}} m^{2}_{32}}|$ ${\theta^{\mathcal{D}} }_{12}$ ${\theta^{\mathcal{D}} }_{23}$ ${\theta^{\mathcal{D}} }_{13}$ ${\theta^{\mathcal{D}} }_{12}$ ${\theta^{\mathcal{D}} }_{23}$ ${\theta^{\mathcal{D}} }_{13}$ \[$10^{-5}$ eV$^2$\] \[$10^{-3}$ eV$^2$\] \[${\theta^{\mathcal{DA}} }_{12}$ ${\theta^{\mathcal{DA}} }_{23}$ ${\theta^{\mathcal{DA}} }_{13}$\] \[${\theta^{\mathcal{DA}} }_{12}$ ${\theta^{\mathcal{DA}} }_{23}$ ${\theta^{\mathcal{DA}} }_{13}$\] Solar-only $6.00^{+3.20}_{-2.60}$ 25.0 57.0 -24.0 \[33.662 49.827 5.870\] KamLAND-only $7.60^{+0.23}_{-0.25}$ 36.0 50.0 -46.0 \[55.362 38.814 -8.604\] Solar+KamLAND $7.60^{+0.23}_{-0.25}$ 11.0 57.0 -34.0 \[34.300 56.689 -9.898\] Accel+Atmos $2.40^{+0.18}_{-0.12}$ 27.0 50.0 -38.0 +CHOOZ \[44.978 43.445 -6.990\] Global 7.6 2.4 $19.0^{+1.8}_{-3.2}$ $46.0^{+4.6}_{-4.0}$ $-29.0^{+1.5}_{-2.5}$ $28.5^{+2.0}_{-2.5}$ $51.5^{+3.5}_{-4.5}$ $-20.0^{+2.5}_{-2.0}$ $(\Delta \chi^2 = 2.30)$ \[33.461 43.326 $-7.582$\] \[33.509 43.980 7.932\] \[$\pm$0.891 $\pm$4.391 $\pm$2.418\] \[$\pm$1.760 $\pm$4.118 $\pm$2.502\] -------------------------- ----------------------------------------- --------------------------------------- ----------------------------------- --------------------------------- ----------------------------------- ----------------------------------- --------------------------------- ----------------------------------- We found that in representation ${\mathcal{D}}$, individual data sets, i.e. the global solar data, , , and the combined data, cannot constrain the three mixing angles well. However, when combined altogether, they infer bounds at 68$\%$ confidence level (CL) on the three mixing angles ${\theta^{\mathcal{D}} }_{ij}$ which are comparable to those on ${\theta^{\mathcal{A}} }_{ij}$ (cf. Table \[table:best\_fits\_DA\]), except for ${\theta^{\mathcal{D}} }_{12}$. We therefore give only the bounds obtained from the combined global neutrino data. Using the global solar data, we find that the best-fit values of ${\theta^{\mathcal{DA}} }_{12}$ and ${\theta^{\mathcal{DA}} }_{13}$ are consistent with those of ${\theta^{\mathcal{A}} }_{12}$ and ${\theta^{\mathcal{A}} }_{13}$, respectively. In representation ${\mathcal{A}}$, the solar data are not sensitive to ${\theta^{\mathcal{A}} }_{23}$, thus one should not naively compare the translated best-fit result of ${\theta^{\mathcal{DA}} }_{23} \sim 50^{\circ}$ to the input value of ${\theta^{\mathcal{A}} }_{23} = 45^{\circ}$. The KamLAND data are insensitive to ${\theta^{\mathcal{A}} }_{23}$ too and have less constraining power on ${\theta^{\mathcal{A}} }_{12}$. These two facts explain why the translated values of ${\theta^{\mathcal{DA}} }_{12}$ and ${\theta^{\mathcal{DA}} }_{23}$ are very different from the expected values, ${\theta^{\mathcal{A}} }_{12}$ and ${\theta^{\mathcal{A}} }_{23}$. Surprisingly, the translated best-fit ${\theta^{\mathcal{DA}} }_{13}$ turns out to be almost the same as ${\theta^{\mathcal{A}} }_{13}$ in magnitude, but with an opposite sign. Like the KamLAND-only results, the combined solar and KamLAND results produce best-fit values where ${\theta^{\mathcal{DA}} }_{12} \sim {\theta^{\mathcal{A}} }_{12}$ and $|{\theta^{\mathcal{DA}} }_{13}| \sim |{\theta^{\mathcal{A}} }_{13}|$, and ${\theta^{\mathcal{DA}} }_{13} < 0$. Similarly, combining the LBL accelerator, CHOOZ, and atmospheric neutrino data, the translated best-fit ${\theta^{\mathcal{DA}} }_{23}$ and $|{\theta^{\mathcal{DA}} }_{13}|$ are consistent with the best-fit ${\theta^{\mathcal{A}} }_{23}$ and $|{\theta^{\mathcal{A}} }_{13}|$, respectively. The combined data are not sensitive to ${\theta^{\mathcal{A}} }_{12}$, so a comparison of the best-fit value of ${\theta^{\mathcal{DA}} }_{12} \sim 45^{\circ}$ with the input value of ${\theta^{\mathcal{A}} }_{12} = 33.2^\circ$ is not necessary. Interestingly for this case, it was found that the best fit ${\theta^{\mathcal{DA}} }_{13} < 0$ as well. This finding has been discussed in Ref. [@Roa:2009wp] using a similar data sample. In spite of the poor constraints on ${\theta^{\mathcal{D}} }_{ij}$, the parameter ${\Delta^{\mathcal{D}} m^{2}_{21}}$ determined by analyzing the solar, KamLAND and solar$+$KamLAND data, as well as ${\Delta^{\mathcal{D}} m^{2}_{32}}$ from the combined analysis of the LBL accelerator, CHOOZ and the atmospheric data agree well with those directly extracted in representation ${\mathcal{A}}$. Equipped with the above individual $\chi^2$ calculation, we perform a global analysis straightforwardly via $$\Delta \chi^2_{\rm global} ({\theta^{\mathcal{D}} }_{12}, {\theta^{\mathcal{D}} }_{23}, {\theta^{\mathcal{D}} }_{13}) \equiv \Delta \chi^2_{\rm sol + KL} + \Delta \chi^2_{\rm acc + CHOOZ + atm}\, .$$ Here we fix ${\Delta^{\mathcal{D}} m^{2}_{21}}= 7.6 \times 10^{-5}$ eV$^{2}$ and ${\Delta^{\mathcal{D}} m^{2}_{32}}= 2.4 \times 10^{-3}$ eV$^{2}$, the best-fit values determined from the and the analysis, respectively. Our results at 68$\%$ CL are given in Table \[table:best\_fits\_DA\], and Fig. \[fig:Global\_CaseD\_deg\] shows the projected likelihood contour plots in the planes of $({\theta^{\mathcal{D}} }_{12}, {\theta^{\mathcal{D}} }_{13})$, $({\theta^{\mathcal{D}} }_{23}, {\theta^{\mathcal{D}} }_{13})$, and $({\theta^{\mathcal{D}} }_{12}, {\theta^{\mathcal{D}} }_{23})$. As already advertised in the beginning of this section, we find that the 68$\%$ CL constraints on ${\theta^{\mathcal{D}} }_{ij}$ are comparable to those on ${\theta^{\mathcal{A}} }_{ij}$, with the exception of ${\theta^{\mathcal{D}} }_{12}$. Sign of $\sin {\theta^{\mathcal{A}} }_{13}$ {#sign-of-sin-thetamathcala-_13 .unnumbered} -------------------------------------------- As we enter the era of precision neutrino oscillation experiments, terms linear in $\sin \theta_{13}$ can no longer be neglected. As pointed out in Ref. [@Roa:2009wp], it will be necessary to perform a parameter fitting in the $\sin {\theta^{\mathcal{A}} }_{13}<0$ regime as well. The sign of $\sin \theta_{13}$ also plays the decisive role to the ${\delta_{cp}}$ determination. This can be seen in the Jarlskog invariant quantity of CP violation [@Jarlskog:1985ht; @Wu:1985ea]. For any representation ${\mathcal{H}}$, it is defined as $${\mathcal{J}} = \sin 2{\theta^{\mathcal{H}} }_{12} \sin 2{\theta^{\mathcal{H}} }_{23} \sin 2{\theta^{\mathcal{H}} }_{13} \cos{\theta^{\mathcal{H}} }_{ij} \sin{\delta_{cp}^{\mathcal{H}} }\, , \label{eq:jarlskog_inv}$$ where ${\theta^{\mathcal{H}} }_{ij}$ is the mixing angle seated in the middle of the three rotation matrices in the mixing matrix. For representation ${\mathcal{H}}= {\mathcal{A}}$ or ${\mathcal{D}}$, both $\sin 2{\theta^{\mathcal{H}} }_{12}$ and $\sin 2{\theta^{\mathcal{H}} }_{23}$ are positive. However, for the existing global neutrino data, the leading terms of the observables in representation ${\mathcal{A}}$ are linear in $\sin^2 \theta_{13}$ while the terms linear in $\sin \theta_{13}$ are in the sub-leading terms. This makes difficult to determine the sign of $\sin \theta_{13}$ in representation ${\mathcal{A}}$. An interesting feature is seen when analyzing neutrino data in representation ${\mathcal{D}}$. Two local minima exist in the global $\chi^2$, one corresponding to ${\theta^{\mathcal{DA}} }_{13} < 0$ (the global minimum) and the other one ${\theta^{\mathcal{DA}} }_{13} > 0$ (the second minimum). Apart from this, both minima have very similar values of ${\theta^{\mathcal{DA}} }_{12}$, ${\theta^{\mathcal{DA}} }_{23}$, and $|{\theta^{\mathcal{DA}} }_{13}|$, which also agree well with the best-fit ${\theta^{\mathcal{A}} }_{12}$, ${\theta^{\mathcal{A}} }_{23}$, and $|{\theta^{\mathcal{A}} }_{13}|$ (cf. Table \[table:best\_fits\_A\]), respectively. This suggests that by doing global neutrino data analysis in representation ${\mathcal{D}}$, one has a chance to identify the sign of $\sin {\theta^{\mathcal{DA}} }_{13}$. In the current case, the $\chi^2$ values for the global minimum and the second minimum differ by $\sim$2. While this preference for negative $\sin \theta_{13}$ is weak, the inclusion of data from the current and the near-future neutrino experiments may help strengthen the evidence. Compared among the various data samples used, the global solar data point to ${\theta^{\mathcal{DA}} }_{13} > 0$ while all the other data ${\theta^{\mathcal{DA}} }_{13} < 0$. We do not have the explanation for this difference yet. One may investigate whether the matter (MSW) effects [@Wolfenstein_msw; @Mik_Smir_msw] influence the sign of ${\theta^{\mathcal{DA}} }_{13}$ during the propagation of solar neutrinos through the Sun and the Earth. Atmospheric neutrinos propagating through the Earth are also subject to the MSW effects (see Appendix \[sec:apdx\_anal\_atmos\]), but to a much less degree. This issue will be further studied in a separate work. Error Correlations {#error-correlations .unnumbered} ------------------ Another issue to address is how any two mixing angles correlate with each other. In principle, the 68$\%$ CL constraints on ${\theta^{\mathcal{DA}} }_{ij}$ should be expected to be the same as those on ${\theta^{\mathcal{A}} }_{ij}$ since they are extracted from the same data sample. By virtue of this fact, one can estimate the correlation coefficients between any two mixing angles in each representation. For two different representations, say ${\mathcal{G}}$ and ${\mathcal{H}}$, any mixing angle in ${\mathcal{G}}$, ${\theta^{\mathcal{G}} }_{ab}$, can be expressed as a function of the three mixing angles, ${\theta^{\mathcal{H}} }_{ij}$, in ${\mathcal{H}}$ as: $${\theta^{\mathcal{G}} }_{ab} = F({\theta^{\mathcal{H}} }_{12}, {\theta^{\mathcal{H}} }_{23}, {\theta^{\mathcal{H}} }_{13})\, .$$ By applying the error propagation $$\begin{aligned} & & \hskip-1.0cm (\Delta {\theta^{\mathcal{G}} }_{ab} )^2 = \sum_{i \neq j} ( \frac{\partial F}{\partial {\theta^{\mathcal{H}} }_{ij}} )^2 (\Delta {\theta^{\mathcal{H}} }_{ij} )^2 \nonumber \\ & & \hskip-0.6cm + 2 \sum_{i \neq j, j \neq k} ( \frac{\partial F}{\partial {\theta^{\mathcal{H}} }_{ij}} ) ( \frac{\partial F}{\partial {\theta^{\mathcal{H}} }_{jk}} ) (\Delta {\theta^{\mathcal{H}} }_{ij} ) (\Delta {\theta^{\mathcal{H}} }_{jk} ) {\rho^{\mathcal{H}} }({\theta^{\mathcal{H}} }_{ij}, {\theta^{\mathcal{H}} }_{jk})\, , \label{eq:AD_corr_coef}\end{aligned}$$ the correlation coefficients $\rho$ can be analytically solved using Eq. (\[eq:AD\_corr\_coef\]). With the best-fit results from the global minimum in ${\mathcal{D}}$ and the corresponding translated results in ${\mathcal{A}}$, along with the average of upper and lower bounds of uncertainties from ${\theta^{\mathcal{D}} }_{ij}$ and ${\theta^{\mathcal{A}} }_{ij}$, one can estimate the correlation coefficients $\rho$ of any two mixing angles in representations ${\mathcal{A}}$ and ${\mathcal{D}}$, respectively. Two situations are compared, [*(i)*]{} the 68$\%$ CL constraints taken at $\Delta \chi^2 = 2.3$ for all ${\theta^{\mathcal{A}} }_{ij}$ (or ${\theta^{\mathcal{DA}} }_{ij}$) and ${\theta^{\mathcal{D}} }_{ij}$; and [*(ii)*]{} the 68$\%$ CL constraints taken at $\Delta \chi^2 = 1.0$ for ${\theta^{\mathcal{A}} }_{13}$ while keeping the rest at $\Delta \chi^2 = 2.3$. The results of the estimated correlation coefficients are presented in Table \[table:corr\_coef\_AD\]. [cccc]{} Rep. & $\rho(\theta_{12}, \theta_{23})$ & $\rho(\theta_{23}, \theta_{13})$ & $\rho(\theta_{12}, \theta_{13})$\ \ \ ${\mathcal{A}}$ & $ 0.288$ & $ 0.086$ & $0.290$\ ${\mathcal{D}}$ & $-0.854$ & $-0.114$ & $0.623$\ \ \ ${\mathcal{A}}$ & $1.861$ & $-0.554$ & $3.000$\ ${\mathcal{D}}$ & $-1.082$ & $-0.409$ & $0.632$\ It is interesting to note that the correlations between any two mixing angles in representations ${\mathcal{A}}$ and ${\mathcal{D}}$ are bigger if case [*(ii)*]{} is applied. The outcome of the correlations for case [*(ii)*]{} seems to conflict with existing neutrino data which indicate the three mixing angles in representation ${\mathcal{A}}$ are nearly decoupled. However, the results of the correlations in case [*(i)*]{} are more consistent with the nearly-decoupled feature among the three mixing angles in representation ${\mathcal{A}}$. This suggests that the 68$\%$ CL constraints on ${\theta^{\mathcal{A}} }_{13}$ seem to require $\Delta \chi^2 = 2.3$ rather than $\Delta \chi^2 = 1.0$ since ${\theta^{\mathcal{A}} }_{13}$ are also correlated with the other mixing angles (cf. Table \[table:corr\_coef\_AD\]). In addition, the correlations among the three mixing angles in representation ${\mathcal{A}}$ are not completely zero though small, indicating that the three mixing angles in representation ${\mathcal{A}}$ are not really decoupled. Nevertheless, the correlations among the three mixing angles in representation ${\mathcal{A}}$ are found to be smaller than those in ${\mathcal{D}}$, as expected. With the estimated ${\rho^{\mathcal{D}} }$, the 68$\%$ CL constraints on ${\theta^{\mathcal{DA}} }_{ij}$ for the second minimum are presented in Table \[table:best\_fits\_DA\]. Summary and Outlook {#sec:summary} =================== We have performed global neutrino oscillation data analyses in two representations for the neutrino mixing matrix. Individual data samples we used include those from solar, KamLAND, long-baseline accelerator, CHOOZ and atmospheric experiments. We found that individually they do not constrain the three mixing angles in representation ${\mathcal{D}}$ as well as those in representation ${\mathcal{A}}$. However, the 68$\%$ CL constraints on ${\theta^{\mathcal{D}} }_{ij}$ retrieved from combined global neutrino data in representation ${\mathcal{D}}$ are almost as good as those on ${\theta^{\mathcal{A}} }_{ij}$ determined in representation ${\mathcal{A}}$, with the exception of ${\theta^{\mathcal{D}} }_{12}$ which is a little bit worse than ${\theta^{\mathcal{A}} }_{12}$ by $\sim$1.5$^{\circ}$. Nevertheless, the best-fit ${\theta^{\mathcal{D}} }_{ij}$ results, as expected, turn out to be significantly large. This results provide a higher sensitivity to the CP-violating phase determination as can be seen from the Jarlskog invariant quantity. In addition, the translated best-fit result of each ${\theta^{\mathcal{DA}} }_{ij}$ from representation ${\mathcal{D}}$ to ${\mathcal{A}}$ are found to be in very good agreement with the corresponding angle, ${\theta^{\mathcal{A}} }_{ij}$, directly retrieved from representation ${\mathcal{A}}$. In other words, ${\theta^{\mathcal{DA}} }_{12}$, ${\theta^{\mathcal{DA}} }_{23}$, and $|{\theta^{\mathcal{DA}} }_{13}|$ are respectively consistent with ${\theta^{\mathcal{A}} }_{12}$ determined from the combined result of solar$+$KamLAND, ${\theta^{\mathcal{A}} }_{23}$ obtained from the combined result of the [accelerator$+$CHOOZ$+$atmospheric]{} data sets, and $|{\theta^{\mathcal{A}} }_{13}|$ extracted from the combined result of global neutrino data in representation ${\mathcal{A}}$. As we enter the era of precision neutrino oscillation experiments, observables with terms linear in $\sin\theta_{13}$ may be no longer negligible in fitting the mixing angles. The sign of $\sin\theta_{13}$ plays a key role in the determination of the sign of ${\delta_{cp}}$, and both $\sin\theta_{13}$ and ${\delta_{cp}}$ are important in the establishment of the mixing matrix. In this work we have shown that one has a chance to identify the sign of ${\theta^{\mathcal{DA}} }_{13}$ through an oscillation parameter fitting performed in representation ${\mathcal{D}}$. We found two local minima when analyzing global neutrino data in representation ${\mathcal{D}}$, one corresponding to ${\theta^{\mathcal{DA}} }_{13} < 0$ (the global minimum) and the other one ${\theta^{\mathcal{DA}} }_{13} > 0$ (the second minimum). A weak preference for negative $\sin \theta_{13}$ is found. It is interesting to note that the global solar data point to ${\theta^{\mathcal{DA}} }_{13} > 0$ while all the other data ${\theta^{\mathcal{DA}} }_{13} < 0$. We do not have interpretation to this difference yet. One may investigate whether the matter (MSW) effects have impacts on the sign of ${\theta^{\mathcal{DA}} }_{13}$ during the propagation of solar neutrinos through the Sun and the Earth. This issue will be further studied in a separate work. In addition, we also provide the information for the correlation of any two mixing angles. It is found that the correlations in representation ${\mathcal{A}}$ are less than those in ${\mathcal{D}}$, but the three mixing angles are not completely decoupled in representation ${\mathcal{A}}$. In conclusion, owing to the strong correlations among the three mixing angles in the new parametrization, the advantages of doing the global neutrino oscillation analysis using data from past, current, and near future neutrino oscillation experiments shall become manifest. Acknowledgements ================ First of all, we are indebted to Jen-Chieh Peng for inspiring this idea and his valuable discussion and suggestions. MH and SDR are grateful to Bruce T. Cleveland, Bernhard G. Nickel, Jimmy Law, and Ian T. Lawson for their useful discussion and suggestions, especially to Bernhard G. Nickel and Jimmy Law for their generosity in providing us with the adiabatic codes. We thank Hai-Yang Cheng, Guey-Lin Lin, and Jen-Chieh Peng for careful reading the manuscript and valuable comments. MH and WCT are thankful to Pisin Chen for his support and useful comments. SDR appreciates Pisin Chen for his hospitality and financial support during her stay at NTU, Taiwan. Last but not least, WCT and HT thank Yung-Shun Yeh for his help with installing and running the NUANCE package. This research was supported by Taiwan National Science Council under Project No. NSC 98-2811-M-002-501 and NSC 99-2811-M-002-064, Canadian Natural Sciences and Engineering Research Council, and Fermi Research Alliance, LLC under the U.S. Department of Energy contract No. DE-AC02-07CH11359. This work was also made possible by the facilities of the Shared Hierarchical Academic Research Computing Network (SHARCNET) and Canada and by the FermiGrid facilities for the availability of their resources. Solar Neutrino Sector {#sec:apdx_anal_solar} ===================== Cl and Ga Rate Neutrino Data ---------------------------- The rates for the Chlorine and Gallium experiments are expressed in terms of SNUs (1 Solar Neutrino Unit = one interaction per 1036 target atoms per second). The predicted rate for the Chlorine/Gallium experiment is given by $$R_{Cl/Ga} = \int^{\infty}_{E_{th}} dE_{\nu} \phi_{\nu_e}(E_{\nu}) P_{ee}(E_{\nu}) \sigma_{Cl/Ga}(E_{\nu}) , \label{eq:solar_ClGa_1}$$ where $E_{\nu}$ and $E_{th}$ are respectively the neutrino energy and the threshold energy for neutrinos captured by chlorine or gallium, $\phi_{\nu_e}(E_{\nu})$ is the flux of electron neutrinos arriving at the detector, including the flux from all solar neutrino reactions, $P_{ee}(E_{\nu})$ is the survival probability of $\nu_e \rightarrow \nu_e$, $\sigma(E_{\nu})$ is the cross section of electron neutrino with the target (either chlorine or gallium). The Cl rate of $2.56 \pm 0.l6 {\rm(statistical)} \pm 0.16 {\rm(systematic)}$ SNU from Ref. [@Cleveland:1998nv] and the Ga rate of $66.1 \pm 3.1$ SNU from SAGE, Gallex, and GNO [@Abdurashitov:2009tn] are used in this analysis. Borexino Rate Neutrino Data --------------------------- The predicted rate calculation of $^{7}$Be solar neutrinos measured in Borexino experiment is calculated as $$R_{Bx} = R^{0}_{Bx} \left\{ S_{e}(E_{\nu}, E_e) p_e + S_{x}(E_{\nu}, E_e) (1 - p_e) \right\} , \label{eq:solar_Bx_1}$$ where $R^{0}_{Bx}$ is the expected rate for non-oscillated solar $\nu_e$ that is $74 \pm 4$ counts/(day $\times$ 100 ton), $S_{e,x}(E_{\nu},E_e)$ describe the probability that the elastic scattering of a $\nu_{e,x}$ of energy $E_{\nu}$ with electrons produces a recoil electron of energy $E_e$, and $p_e$ is the survival probability of $\nu_e \rightarrow \nu_e$. The measured rate of $49 \pm 3_{stat} \pm 4_{syst}$ counts/(day $\cdot$ 100 ton) from Borexino experiment [@Arpesella:2008mt] is adopted in this analysis. Super-Kamiokande Solar Neutrino Data ------------------------------------ The measured day/night spectra of Super-Kamiokande (SK) phases I & II [@Hosaka:2005um; @Cravens:2008zn] are employed in this analysis. To fit for the oscillation parameters, we follow the methods described in Ref. [@Hosaka:2005um] to build up the predicted rate calculation for SK solar day/night spectra, which is given by $$\begin{aligned} R_{SK} &=& \int^{E_1}_{E_0} dE \int_{E_{\nu}} dE_{\nu} I(E_{\nu}) \int_{E_e} dE_{e} R(E_e, E) \times \nonumber \\ & & \left\{ S_{e}(E_{\nu}, E_e) p_e + S_{x}(E_{\nu}, E_e) (1 - p_e) \right\} , \label{eq:solar_SK_1}\end{aligned}$$ where $I(E_{\nu})$ is the spectrum of $^{8}$B or [*hep*]{} neutrino, $R(E_e, E)$ is the detector response function representing the probability that a recoil electron of energy $E_e$ is reconstructed with energy $E$, and $S_{e,x}(E_{\nu}, E_e)$ and $p_e$ have the same meanings as given in previous paragraph. Both rates due to $^{8}$B and [*hep*]{} interactions are taken into consideration in this work. SNO Solar Neutrino Data ----------------------- For SNO solar data, the fraction of the extracted $CC$ (charged current) spectra as a function of one unoscillated SSM [@Bahcall:2004pz] (Fig. 28 in Ref. [@Aharmim:2009gd]) is used for the mixing parameter fitting. The theoretical calculation of the $CC$ flux fraction for a set of oscillation parameter in order for comparison with SNO solar data is described as follows. The $CC$ flux fraction in the electron kinetic energy between $T_1$ and $T_2$, denoted as $f(T_1, T_2)$, is calculated as the number of events observed in the data set contributed to $CC$ interactions by the signal extraction with electron kinetic energies between $T_1$ and $T_2$ divided by the number of all $CC$ events that would be observed above the threshold kinetic energy, $T_{th}$. It is formulated as [@Aharmim:2005gt] $$\begin{aligned} & & \hskip-0.45cm f(T_1, T_2) \equiv \nonumber \\ & & \hskip-0.45cm \frac{\int^{\infty}_0 \int^{\infty}_0 \int^{T_{2}}_{T_{1}} \phi(E_{\nu}) P_{ee}(E_{\nu}) \frac{d \sigma}{dT_e}(E_{\nu}, T_e) R(T_e, T^{'}_{e}) dE_{\nu} dT_e dT^{'}_{e} } {\int^{\infty}_0 \int^{\infty}_0 \int^{\infty}_{T_{th}} \phi(E_{\nu}) \frac{d \sigma}{dT_e}(E_{\nu}, T_e) R(T_e, T^{'}_{e}) dE_{\nu} dT_e dT^{'}_{e} } \nonumber \\ \label{eq:solar_SNO_1}\end{aligned}$$ where $P_{ee}$ is the survival probability of $\nu_e \rightarrow \nu_e$, $\phi(E_{\nu})$ includes both fluxes of $^{8}$B and [*hep*]{} solar neutrinos as a function of neutrino energy, $d\sigma / dT_e$ is the differential cross section for $CC$ interactions, and $R(T_e, T^{'}_{e})$ is the energy resolution function, describing the probability of seeing an apparent kinetic energy $T^{'}_{e}$ for a given true energy $T_e$, whose expression can be found in Ref. [@Aharmim:2009gd]. Global Solar Neutrino Data -------------------------- To fit for oscillation parameters using global solar data, the combined $\chi^2$ for the observables and predictions is given by $$\begin{aligned} & & \hskip-1.0cm \chi^2_{solar} = \nonumber \\ & & \hskip-0.8cm \sum^{N}_{i, j = 1} ( \mathcal{O}_{i} - \mathcal{O}^{exp}_{i} ) [ \sigma^2_{i j}(tot) ]^{-1} ( \mathcal{O}_{j} - \mathcal{O}^{exp}_{j} ) , \label{eq:apdx_solar_chi2}\end{aligned}$$ where $i$, $j$ are indices to sum over all the observables. The quantities $\mathcal{O}_{i}$ $$\mathcal{O}_{i} = \mathcal{O}_{i}(\Delta m^2, \tan^2 \theta, \ldots)$$ is the theoretical prediction for that observable, and $\mathcal{O}^{exp}_{j}$ represents a series of observables from a number of solar neutrino experiments that include rate measurements from Cl and Ga (e.g. Homestake, Gallex, GNO, and SAGE) experiments, rate of $^{7}$Be solar neutrino measurement from Borexino experiment, a number of spectral shape measurements from SK and SNO as mentioned above. The total error matrix, $\sigma^2(tot)$, is a sum of contributions from the rate and spectral measurements, which is given as $$\begin{aligned} & & \hskip-1.0cm \sigma^2(tot) = \nonumber \\ & & \hskip-0.8cm \sigma^2(exp) + \sigma^2_R + \sigma^2_R(Bx) + \sigma^2_{S}(SNO) + \sigma^2_{S}(SK) , \label{eq:solar_sigma_2}\end{aligned}$$ where $\sigma^2(exp)$ is a diagonal matrix containing the statistical and systematic errors from the rate measurements and the statistical errors from the spectral measurements; $\sigma^2_R$ is the rate error matrix, handling the correlations between the rate CL and Ga experiments using the procedure described in Ref. [@Fogli:1994nn]; $\sigma^2_R(Bx)$ is treated as independent from $\sigma^2_R$ in our analysis. Furthermore, the spectral correlation matrix $\sigma^2_{S}(SNO)$ for SNO measurements is assumed to be uncorrelated with $\sigma^2_{S}(SK)$ for SK measurements, which are uncorrelated with the CL, Ga, and Borexino measurements. One can refer to Refs. [@Fogli:1994nn; @Fogli:1999zg; @Garzelli:2000tn; @Garzelli:2001zu] for detailed discussion on the covariance error matrix, $\sigma^2(tot)^{-1}$. By minimizing $\chi^2_{solar}$ with three-flavor neutrino oscillation scheme while fixing ${\Delta m^{2}_{32}}= 2.4 \times10^{-3}$ eV$^2$ and ${\sin^{2}\theta_{23}}= 0.5$, the values of ${\Delta m^{2}_{21}}$, ${\tan^{2}\theta_{12}}$, and ${\sin^{2}\theta_{13}}$ were found to reproduce the results of global solar data as reported in Ref. [@Aharmim:2009gd] when the conventional mixing matrix parametrization is used. Reactor Neutrino Sector {#sec:apdx_anal_reactor} ======================= KamLAND Reactor ${\bar{\nu}_e}$ Data ------------------------------------ For KamLAND reactor data, the prompt energy spectrum of $\bar{\nu}_e$ events of energy (as shown in Fig. (1) of Ref. [@Gando:2010aa]) is used. We follow the strategy described in Ref. [@Fogli:2005qa] for this analysis. Since the distance from the reactor source to the KamLAND detector is $\sim$180 Km, the treatment of vacuum oscillation is considered in this analysis. With a little modification, the number of expected events per unit of the prompt position energy is given by $$\begin{aligned} & & \hskip-1.0cm N(T^{'}_{e}) = \xi \Delta t \cdot \epsilon(T^{'}_{e}) \cdot \int dE_{\nu} \frac{d \phi}{d E_{\nu}} \nonumber \\ & & \times \int dT_e \frac{ d\sigma(E_{\nu}, T_e) }{ dT_e } R(T_e, T^{'}_{e}) \; . \label{eq:reactor_KL_1}\end{aligned}$$ The values $\xi = 5.98 \times 10^{32}$ protons and $\Delta t = 2135$ days are the total number of target protons and livetime, respectively. $\epsilon(T^{'}_{e})$ is the detection efficiency as a function of measured prompt energy, $T^{'}_{e}$ . $d \phi / d E_{\nu}$ is the time-averaged differential neutrino flux at the KamLAND detector, where the fission flux as a function of distance can be found at the website [@KL_2nd_fission_flux] and the relative fission yields, $^{235}$U:$^{238}$U:$^{239}$Pu:$^{241}$Pu, is taken from [@Gando:2010aa], and flux from Korean reactors is not taken into account since only $\sim$3% of the total flux contributes to KamLAND signal. In the presence of oscillation, each $j$th reactor term in $d \phi / d E_{\nu}$ must be multiplied by the corresponding neutrino survival probability $P_{ee}(E_{\nu}, L_j)$ for neutrinos of energy $E_{\nu}$ and taveling distance of $L_j$ between each reactor source and the KamLAND. $R(T_e, T^{'}_{e})$ is the energy resolution function with Gaussian width equal to 6.4%$\sqrt{T_e/MeV}$, which has the same meaning as described in the solar sector. $ d\sigma(E_{\nu}, T_e) / dT_e $ is the inverse beta decay cross section [@Fogli:2005qa; @Vogel:1999zy]. To fit for the oscillation parameters, the ${\chi^{2}}$ function is of the Gaussian form [@Fogli:2005qa] that includes both the total number of events and the spectral shape: $$\begin{aligned} \chi^2_{KL} &=& \left( \frac{N^{theo}_{tot} - N^{obs}_{tot}}{\sigma^{rate}_{tot}} \right)^2 \nonumber \\ & & + \sum_i \left( \frac{N^{theo}_i - N^{obs}_i}{\sigma^{rate}_i} \right)^2 + \left( \frac{\alpha}{\sigma^{\alpha}} \right)^2 , \label{eqn:chi_KL}\end{aligned}$$ where $N^{obs}_{tot} = 1780$ is the total number of observed events after substracting all background events, $N^{theo}_{tot}$ is the predicted total events; $N^{theo}_i$ and $N^{obs}_i$ are the numbers of observed events and predicted events respectively in $i$th energy bin, the total error, $\sigma^{rate}_{tot}$, is the quadrature sum of the statistical and systematic uncertainties, and $\sigma^{rate}_i$ is the error for $i$th bin by adding statistical and systematic uncertainties in quadrature. Minimizing the $\chi^2_{KL}$ while fixing ${\Delta m^{2}_{32}}= 2.4\times10^{-3}$ eV$^2$ and ${\sin^{2}\theta_{23}}= 0.5$, the values for ${\Delta m^{2}_{21}}$, ${\tan^{2}\theta_{12}}$, and ${\sin^{2}\theta_{13}}$ were found to reproduce the KamLAND results in Ref. [@Gando:2010aa] if the conventional mixing matrix parametrization is applied. CHOOZ Reactor ${\bar{\nu}_e}$ Data ---------------------------------- The CHOOZ results [@Apollonio:2002gd] reported the observed positron energy from the neutrino interactions. For this experiment, the distance between the detector and the neutrino source is relatively small (1 km) and thus the neutrino propagation through vacuum can be used to calculate the anticipated positron spectrum. Using the procedures reported in Ref. [@Apollonio:2002gd], the expected positron yield for the $k$-th reactor and the $j$-th energy spectrum bin is parametrized as $$\begin{aligned} & & \hskip-0.8cm \bar{X}(E_j, L_k, \theta, \delta m^2) = \tilde{X}(E_j) \cdot \bar{P}(E_j, L_k, \theta, \delta m^2_{32}), \nonumber \\ & & \hskip2.3cm (j = 1, \ldots , 7, \hskip0.2cm k = 1, 2) , \label{eq:reactor_Chooz_1}\end{aligned}$$ where $\tilde{X}(E_j)$ is the distance-independent positron yield for no presence of neutrino oscillation, $L_k$ is the reactor-detector distance, and $\bar{P}(E_j, L_k, \theta, \delta m^2_{32})$ is the oscillation probability averaged over the energy bin and the core sizes of the detector and the reactor. For the experimental data, the results reported in Table 8 in Ref. [@Apollonio:2002gd] were used. These results consist of seven positron energy bins for each of the two reactors for a total of 14 bins. The covariant matrix $V^{-1}_{ij}$ defined in equation 54 in Ref. [@Apollonio:2002gd] is applied to the $\chi^2$ function. This matrix accounts for any correlations between the energy bins of the two reactors. We then minimize the $\chi^2$ function with respect to the neutrino oscillation parametrization and we can reproduce the CHOOZ results as presented in Ref. [@Apollonio:2002gd] when the conventional mixing matrix parametization is applied. Long-Baseline Accelerator Neutrino Sector {#sec:apdx_anal_accel} ========================================= MINOS {#apdx_accel_MINOS} ----- The analysis for MINOS used results from the ${\nu_{\mu}}$ disappearance channel [@Adamson:2011ig] and the ${\nu_{e}}$ appearance channel [@MINOS_nue_app_2011]. Results from both channels consisted of the reconstructed neutrino energy as seen in the far detector. The distance of 735 km between the near and far detectors was considered to be too small for matter effects to have any significant contribution to the results. Vacuum oscillation is thus applied in this analysis. The predicted number of neutrinos at the far detector is calculated as: $$N^{osc} = f( P_{\mu\mu} + \epsilon_{\tau} P_{\mu\tau} ) N^{no\ osc} + g N^{NC} \label{eqn:chi_minos_k2k_1}$$ where $P_{\mu\mu}$ is the survival probability for ${\nu_{\mu}}$, $P_{\mu\tau}$ is the probability for ${\nu_{\tau}}$ appearance, $\epsilon_{\tau}$ is the ${\nu_{\tau}}$ detection efficiency, $N^{NC}$ is the $NC$ (neutral current) spectrum, $f$ is the normalization factor, and $g$ is the $NC$ scaling factor. Like the ${\nu_{\mu}}$ data, the ${\nu_{e}}$ data consists of events binned by reconstructed energy over a range of 1 to 8 GeV. The predicted neutrino spectrum of ${\nu_{e}}$ appearance channel is calculated using: $$N^{osc} = f^{'} P_{\mu e} N^{no\ osc} + g^{'} N^{NC} \label{eqn:chi_minos_k2k_2}$$ where $P_{\mu e}$ is the probability for ${\nu_{e}}$ appearance, and $f^{'}$ $(g^{'})$ has the same meaning as $f$ $(g)$ but with a different value. The value for $N^{no\ osc}$ is taken from [@Adamson:2011ig]. The ${\chi^{2}}$ function is of the Poisson form and includes both the spectral shape and the total number of neutrino events, which is given by: $$\begin{aligned} \chi^2 &=& 2[ N^{osc}_{tot} - N^{obs}_{tot} ( 1+\ln( N^{osc}_{tot}/N^{obs}_{tot} )) ] \nonumber \\ & & + \sum_{i} 2[ N^{osc}_{i} - N^{obs}_{i} ( 1+\ln( N^{osc}_{i}/D_i )) ] \nonumber \\ & & + \sum_{j} \left( \frac{\xi_j}{\sigma_j} \right)^2 , \label{eqn:chi_minos_k2k}\end{aligned}$$ where $N^{osc}_{tot}$ and $N^{obs}_{tot}$ are the predicted and detected total number of neutrinos respectively; $N^{osc}_{i}$ and $N^{obs}_{i}$ are the predicted and detected number of neutrinos in the $i^{th}$ energy bin. The systematics for the energy scale, normalization factor and scale factor for the $NC$ events are represented by $\xi_j$ with uncertainty of $\sigma_j$. Minimizing the ${\chi^{2}}$ for the ${\nu_{\mu}}$ disappearance channel while fixing ${\Delta m^{2}_{21}}= 7.67\times10^{-5}$ eV$^2$, ${\tan^{2}\theta_{12}}= 0.427$, ${\sin^{2}\theta_{13}}= 0.0$ and assuming a normal mass hierarchy, the values for ${\Delta m^{2}_{32}}$ and ${\sin^{2}\theta_{23}}$ where found to reproduce the MINOS results in Ref. [@Adamson:2011ig] when the mixing matrix parametrization ${\mathcal{A}}$ is applied. K2K {#apdx_accel_K2K} --- The results from K2K [@Ahn:2006zza] used in this analysis were the reconstructed ${\nu_{\mu}}$ energy spectrum for one-ring $\mu$-like sample. Like MINOS, the K2K’s 250-km oscillation length is considered to be too short to manifest matter enhanced oscillation effects. Vacuum oscillation can thus be applied to this analysis. The predicted no-oscillation neutrino spectrum represents the true neutrino energy spectrum at the near detector. To extract the predicted far detector oscillation spectrum from the no-oscillation spectrum, the neutrino interaction cross sections and detection efficiencies must be applied. Furthermore, both $CC$ and $NC$ events are present in the neutrino spectrum and hence are accounted for separately. The expected number of neutrino events for oscillation neutrinos is given by $$N^{osc}_j = N^{no\ osc}_j \left[ \epsilon_{cc} P_{cc}\sigma_{cc} + \epsilon_{nc}\sigma_{nc} \right]$$ where $\sigma_{cc}$ and $\sigma_{nc}$ are the cross sections for the $CC$ and $NC$ interactions respectively, $\epsilon_{cc}$ and $\epsilon_{nc}$ are the detection efficiencies for the $CC$ and $NC$ interactions, and $P_{cc}$ is the survival probability for $\nu_{\mu}$. The energy response function is applied to obtain the probability of seeing the measured energy for a given true energy, same meaning as employed in the solar analyses. The form of the energy response function is employed from Ref. [@Fogli:2003th]: $$N^{obs} = \sum_{i} N^{osc}_{i} * e^{ -\frac{1}{2} ( (E_{obs} - E_{i} + 0.05)/\sigma )^2 }$$ where $E_{obs}$ and $E_{i}$ are the measured and true neutrino energies respectively, and $\sigma$ is defined as [@Ahn:2006zza]: $$\sigma = 0.2 E_{i} ( 1-e^{ (0.2-E_{i}) / 0.8} )$$ The K2K analysis uses the same ${\chi^{2}}$ function as that used in the MINOS analysis. By minimizing the ${\chi^{2}}$ for the ${\nu_{\mu}}$ while fixing ${\Delta m^{2}_{21}}= 7.67 \times10^{-5}$ eV$^2$, ${\tan^{2}\theta_{12}}= 0.427$, ${\sin^{2}\theta_{13}}= 0.0$, and assuming a normal mass hierarchy, the values for ${\Delta m^{2}_{32}}$ and ${\sin^{2}\theta_{23}}$ were found to reproduce the K2K results in Ref. [@Ahn:2006zza] when the conventional mixing parametrization is applied. Atmospheric Neutrino Sector {#sec:apdx_anal_atmos} =========================== Super-Kamiokande Atmospheric Data --------------------------------- The Super-Kamiokande (SK) detector is located deep under the peak of Mt Ikenoyama, where the 1,200-m rock overburden can reduce the flux of cosmic rays reaching the detector down to about 3 Hz (see e.g. Ref. [@Wendell:2008zz]). Atmospheric neutrinos penetrating the Earth interact with the nucleus or nucleons in the SK water tank or in the surrounding rock, giving rise to partially contained (PC), fully contained (FC), or upward-going muon (UP$\mu$) events. Two and three-flavour neutrino oscillation analyses [@Wendell:2010md; @Hosaka:2006zd; @Ashie:2005ik] have been performed using the accumulated data from SK-I, II and III phases. We try to reproduce the results of the 3-flavour analysis published in Ref. [@Hosaka:2006zd] to include in our global analysis. Therein, the 1489 live-days of PC and FC, and the 1646 days UP$\mu$ data collected during the SK-I period (1996-2001) are used. Based on the types of the out-going leptons, their energy deposited in the SK detector, and their zenith angle ($-1 < \cos \theta_{\rm zenith} < 1$ for PC and FC; $-1 < \cos \theta_{\rm zenith} < 0$ for UP$\mu$), the selected events are divided in 370 bins. In our work, however, lack of information forced us to follow the approach of Ref. [@Fogli:2003th], where only 55 energy and zenith angle bins of SK atmospheric data are used. Here we briefly describe our analysis. First, the NUANCE package [@Casper:2002sd] is used to simulate atmospheric neutrino events assuming no oscillation effects. For achieving good statistics we have run the simulations for 200-year operation time. Neutrino oscillation effects in the atmosphere and inside the Earth are then incorporated by the “weighting” factors [@Wendell:2008zz] as: $$\begin{aligned} & & \hskip-1.0cm w_{e} = P(\nu_{e} \rightarrow \nu_{e}) + \frac{\phi_{\mu}}{\phi_{e}}P(\nu_{\mu} \rightarrow \nu_{e})\, , \hspace{0.4cm} {\rm for} \hspace{0.1cm} \nu_{e}\, , \nonumber \\ & & \hskip-1.0cm w_{\mu} = P(\nu_{\mu} \rightarrow \nu_{\mu}) + \frac{\phi_{e}}{\phi_{\mu}}P(\nu_{e} \rightarrow \nu_{\mu})\, , \hspace{0.4cm} {\rm for} \hspace{0.1cm} \nu_{\mu}\, . \label{eq:weighting_factors}\end{aligned}$$ For the (energy and angular-dependent) incident atmospheric neutrino fluxes $\phi_{e, \mu}$, we also adopt the Honda three-dimensional calculation [@Honda:2004yz]. We take into account the effects when neutrinos oscillating in matter following the prescription of Ref. [@Barger:1980tf], and approximate the Earth as a four-layer division as Ref. [@Hosaka:2006zd] did. Each layer has a constant density (inner core: $R \leq 1221$ km, $\rho=13.0$ g/cm$^{3}$; outer core: $1221 < R \leq 3480$ km, $\rho=11.3$ g/cm$^{3}$; mantle: $3480 < R \leq 5701$ km, $\rho=5.0$ g/cm$^{3}$; crust: $5701 < R \leq 6371$ km, $\rho=3.3$ g/cm$^{3}$). We apply similar criteria and cuts on the kinematics of the simulated events so that we achieve the same selection efficiency as Ref. [@Hosaka:2006zd] does. Furthermore, one can include the systematics by using the “pull technique” [@Fogli:2003th; @Fogli:2002pt]. Due to correlated systematic uncertainties, the event rate in the $n$-th bin predicted by the MC simulation, $R^{MC}_n$, is shifted by an amount as $$R^{MC}_n \rightarrow \widetilde{R_{n}^{MC}} \equiv R_{n}^{MC} \left( 1 + \sum_{k=1}^{11} S_{n}^{k}\xi_{k} \right)\, .$$ Here $S_{n}^{k}$ is the 1$\sigma$ error associated to the $k$-th source of systematics, and $\xi_{k}$’s are a set of univariate Gaussian random variables. Ref. [@Fogli:2003th] gives a summary and detailed discussion of the 11 systematics sources, which we also use in our work. The $\chi^2$ function thus becomes $$\begin{aligned} & & \hskip-1.0cm \chi_{\rm SK}^{2} = {\min}_{\{\xi_k\}} \left[ \sum_{n=1}^{55}\left( \frac{\widetilde{R_{n}^{MC}}-R_{n}^{ex}} {\sigma_{n}^{stat}} \right)^{2} \right. \nonumber \\ & & \hskip1.1cm \left. + \sum_{k,h=1}^{11}\xi_{k}[\rho^{-1}]_{kh}\xi_{h} \right]\, , \label{eq:ChiSq_formula} \end{aligned}$$ for the observed event rate $R_{n}^{ex}$ in each bin. Here $\sigma_{n}^{stat}$ is the statistical error of the $n$-th bin, and $\rho^{-1}$ the inverse of the correlation matrix, the value of which can also be found in Ref. [@Fogli:2003th]. We minimize the $\chi^{2}_{\rm SK}$ value with respect to the oscillation parameters by solving $\frac{\partial \chi^{2}}{\partial \xi_{i}}=0$, which is equivalent to solving a set of linear equations: $$\begin{aligned} & & \hskip-0.5cm \sum_{k=1}^{11} \left[ \sum_{n=1}^{55}\left(\frac{R_{n}^{MC}} {\sigma_{n}^{stat}}\right)^{2}\times S_{n}^{i}S_{n}^{k}+ \rho^{-1}_{ik}\right]\xi_{k} \nonumber \\ & & \hskip0.5cm = \sum_{n=1}^{55}\left[ \frac{R_{n}^{ex}R_{n}^{MC}-(R_{n}^{MC})^{2}} {(\sigma_{n}^{stat})^{2}} \right]S_{n}^{i}\, . \label{eq:LiEq_to_minize_Xi}\end{aligned}$$ Under normal hierarchy and ${\mathcal{A}}$ parametrization, our best-fit values for ($\Delta m^{2}_{32}$, $\sin^{2}\theta_{23}$, $\sin^{2}\theta_{13}$) are ($2.4\times10^{-3}\mbox{ eV}^{2}$, $0.45$, $0.0$), and the $90\%$ confidence levels ($\Delta \chi^{2}=4.61$) are $1.5\times10^{-3}\mbox{ eV}^{2}<\Delta m^{2}_{32}<3.3\times10^{-3} \mbox{ eV}^{2}$, $0.35<\sin^{2}\theta_{23}<0.59$, and $\sin^{2}\theta_{13}<0.17$, which is consistent with those published in Ref. [@Hosaka:2006zd]. SNO Atmospheric Data -------------------- Thanks to its deep location and flat overburden, the SNO detector can observe atmospheric neutrinos over a wide range of zenith angle, via their charged-current interactions in the surrounding rock. Angular distribution of through-going muons having $-1 \leq \cos \theta_{\rm zenith} < 0.4$ can be used to infer neutrino oscillation parameters and incident neutrino flux, while data above this cutoff provide access to the study of the cosmic-ray muon flux. Below we describe briefly our approach to including the latest SNO atmospheric neutrino data from Ref. [@Aharmim:2009zm]. Details of the analyses can be found in T. Sonley’s PhD thesis [@Sonley:2009]. We first use the NUANCE package [@Casper:2002sd] to simulate atmospheric neutrino induced through-going muons and muons generated in SNO detector’s D$_2$O and H$_2$O regions for 100 year operation time, assuming no oscillation. Oscillation effects are then added with the help of the “weighting” factor. For the incident atmospheric neutrino fluxes, we adopt the Bartol three-dimensional calculation [@Barr:2004br]. Due to lack of a detector simulation package such as the SNOMAN, we apply simple cuts on the kinematics of simulated events, separately for different event types ($\nu_\mu$-induced through-going muons, $\nu_\mu$ water interactions, and $\nu_\mu$ and $\nu_e$ internal interactions). We require (i) the impacat parameter $b \leq 830$ cm; (ii) the muon energy when entering the detector $E_{\mu} \gtrsim$ 1 GeV, and adjust the “trigger efficiency” so that the yearly event rate match that given in Ref. [@Aharmim:2009zm]. The trigger efficiency we defined summarises all other instrumental cuts which we are not able to apply. In addition, the systematics are taken into account using the generalised “pull technique”. Following Ref. [@Aharmim:2009zm; @Sonley:2009], the likelihood function used in our analysis is $$\begin{aligned} & & \hskip-0.8cm L_{\rm total} \equiv - \ln {\cal L} = \nonumber \\ & & \hskip-0.1cm 2 \sum_{{\rm bins}\, i} \left[N^{\rm data}_i\, \ln \frac{N^{\rm data}_i}{N^{\rm MC}_ {0\, , i}}\, + \left(N^{\rm MC}_{0\, ,i} - N^{\rm data}_i \right) \right]\, \nonumber \\ & & \hskip-0.1cm - \vec{\alpha}_{\rm min}\, {\cal S}^2\, \vec{\alpha}_{\rm min}\, , \label{eq:apdx_sno_atmos_1}\end{aligned}$$ where $N^{\rm data}_i$ is the measured event number in zenith angle bin $i$, and $N^{\rm MC}_{0\, , i}$ denotes the expected event number in bin $i$ with the systematics $\vec{\alpha}$ equal to zero. We include 8 systematics error sources, and calculate their values which minimise the likelihood function $$\vec{\alpha}_{\rm min} = \sum_i\, \vec{\beta}^{\cal T}_i\, \left(N^{\rm data}_i - N^{\rm MC}_{0\, ,i} \right)\, {\cal S}^{-2}\, . \label{eq:apdx_sno_atmos_2}$$ Here the coefficients $\vec{\beta}_i$ are available in Ref. [@Sonley:2009], and the new error matrix is $${\cal S}^2 = \sigma^{-2} + \sum_i\, N^{\rm data}_i\, \vec{\beta}_i\, \times \vec{\beta}^{\cal T}_i\, \label{eq:apdx_sno_atmos_3}$$ with $\sigma^{-2}$ the diagonal error matrix whose entries represent the size of the systematic errors. We first compare the results of our two-flavour analysis with those in Ref. [@Aharmim:2009zm; @Sonley:2009]. Normalisation of the Bartol three-dimensional atmospheric neutrino flux $\Phi_0$ is determined simultaneously with the neutrino oscillation parameters. Our best fit values for $({\Delta m^{2}_{32}}\, , \sin^2 2\theta_{23}\, , \Phi_0)$ are $(2.0 \times 10^{-3}\, , 1.0\, , 1.30)$ in the normal hierarchy scheme. These are close to those obtained in Ref. [@Sonley:2009], while still within $1\sigma$ of those from Ref. [@Aharmim:2009zm]. To save CPU time, we then do the three-flavour analysis by keeping $\Phi_0$ fixed at 1.30. Matter effects induced when neutrinos pass through different Earth layers are considered following the prescription of Ref. [@Barger:1980tf]. As expected, SNO atmospheric neutrino data are not sensitvie to $\theta_{13}$. Since, unlike the Super-Kamiokande experiment, SNO only observes muons plus a few possible $\nu_e$-induced internal events. $({\theta^{\mathcal{A}} }_{ij}, {\delta_{cp}^{\mathcal{A}} })$ Solutions in Terms of $({\theta^{\mathcal{D}} }_{ij}, {\delta_{cp}^{\mathcal{D}} })$ {#apdx:Gfit_NumParamSol_DA} ==================================================================================================================================================== Following the procedure presented in Ref. [@Huang:2011by], the transformation of $({\theta^{\mathcal{D}} }_{ij}, {\delta_{cp}^{\mathcal{D}} })$ from representation ${\mathcal{D}}$ to ${\mathcal{A}}$ is again briefly summarized here by only listing the solutions of the nine parameters. To solve for $({\theta^{\mathcal{A}} }_{ij}, {\delta_{cp}^{\mathcal{A}} })$ in representation ${\mathcal{A}}$ when the $({\theta^{\mathcal{D}} }_{ij}, {\delta_{cp}^{\mathcal{D}} })$ are known in representation ${\mathcal{D}}$, begin with $$U = RWR({\theta^{\mathcal{A}} }_{ij}, {\delta_{cp}^{\mathcal{A}} }) = D^L \cdot RWR({\theta^{\mathcal{D}} }_{ij}, {\delta_{cp}^{\mathcal{D}} }) \cdot D^R \; ,$$ \[eq:Append\_D\_DA\_1\] where $$\begin{aligned} D^L({\Phi^{\mathcal{A}} }_{Li}) &=& diag \left( e^{i{\Phi^{\mathcal{A}} }_{L1}}, e^{i{\Phi^{\mathcal{A}} }_{L2}}, e^{i{\Phi^{\mathcal{A}} }_{L3}} \right) \; , \nonumber \\ D^R({\Phi^{\mathcal{A}} }_{Ri}) &=& diag \left( e^{i{\Phi^{\mathcal{A}} }_{R1}}, e^{i{\Phi^{\mathcal{A}} }_{R2}}, 1 \right) \; . \label{eq:Append_D_DA_2}\end{aligned}$$ Through the nine real parts and the nine imaginary parts of Equation (\[eq:Append\_D\_DA\_1\]), the solutions to the nine parameters are listed as follows: $$\begin{aligned} {\Phi^{\mathcal{A}} }_{L2} &=& 0 , \nonumber \\ {\Phi^{\mathcal{A}} }_{L1} + {\Phi^{\mathcal{A}} }_{R1} &=& 0 . \label{eq:Append_D_DA_41}\end{aligned}$$ $$\sin^2{{\Phi^{\mathcal{A}} }_{L3}} = \frac{ (a~\sin{\delta_{cp}^{\mathcal{D}} })^2} { (a~\sin{\delta_{cp}^{\mathcal{D}} })^2 + (a~\cos{\delta_{cp}^{\mathcal{D}} }- b)^2} \; , \label{eq:Append_D_DA_42}$$ where $a = {s^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {s^{\mathcal{D}} }_{13}$ and $b = {c^{\mathcal{D}} }_{23} {c^{\mathcal{D}} }_{13}$ . $$\sin^2{{\Phi^{\mathcal{A}} }_{15}} = \frac{ (a^{'} \sin{\delta_{cp}^{\mathcal{D}} })^2 } { (a^{'} \sin{\delta_{cp}^{\mathcal{D}} })^2 + (a^{'} \cos{\delta_{cp}^{\mathcal{D}} }- b^{'})^2 } \; , \label{eq:Append_D_DA_43}$$ where $a^{'} = {c^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {c^{\mathcal{D}} }_{13}$, $b^{'} = {s^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{13}$, and ${\Phi^{\mathcal{A}} }_{15} \equiv {\Phi^{\mathcal{A}} }_{L1} + {\Phi^{\mathcal{A}} }_{R2}$. With these phases, the three mixing angles can be extracted in representation ${\mathcal{A}}$: $$\begin{aligned} & & \hskip-1.2cm \tan {\theta^{\mathcal{A}} }_{23} = \frac{ {s^{\mathcal{D}} }_{23} {c^{\mathcal{D}} }_{12} \cos{\Phi^{\mathcal{A}} }_{L2} } { {c^{\mathcal{D}} }_{23} {c^{\mathcal{D}} }_{13} \cos{\Phi^{\mathcal{A}} }_{L3} -{s^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {s^{\mathcal{D}} }_{13} \cos({\Phi^{\mathcal{A}} }_{L3} - {\delta_{cp}^{\mathcal{D}} }) } \label{eq:Append_D_DA_44}\end{aligned}$$ $$\begin{aligned} & & \hskip-1.2cm \cos {\theta^{\mathcal{A}} }_{12} = \frac{ {c^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {c^{\mathcal{D}} }_{13} \cos({\Phi^{\mathcal{A}} }_{15} - {\delta_{cp}^{\mathcal{D}} }) -{s^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{13} \cos{\Phi^{\mathcal{A}} }_{15} } { {c^{\mathcal{D}} }_{12} {c^{\mathcal{D}} }_{13} \cos({\Phi^{\mathcal{A}} }_{L1} + {\Phi^{\mathcal{A}} }_{R1}) } \label{eq:Append_D_DA_45}\end{aligned}$$ $$\begin{aligned} & & \hskip-1.2cm \cos {\theta^{\mathcal{A}} }_{13} = \frac{ {c^{\mathcal{D}} }_{12} {c^{\mathcal{D}} }_{13} \cos({\Phi^{\mathcal{A}} }_{L1} + {\Phi^{\mathcal{A}} }_{R1}) } { {c^{\mathcal{A}} }_{12} } \label{eq:Append_D_DA_46}\end{aligned}$$ The remaining parameters thus can be determined as follows. $$\begin{aligned} & & \hskip-1.2cm \sin^2{{\Phi^{\mathcal{A}} }_{R1}} \nonumber \\ & & \hskip-1.2cm = \frac{ (\sin{\Phi^{\mathcal{A}} }_{L3} - d \sin{\delta_{cp}^{\mathcal{D}} })^2 } { (\sin{\Phi^{\mathcal{A}} }_{L3} - d \sin{\delta_{cp}^{\mathcal{D}} })^2 + (d \cos{\delta_{cp}^{\mathcal{D}} }- \cos{\Phi^{\mathcal{A}} }_{L3})^2 } , \label{eq:Append_D_DA_47}\end{aligned}$$ where $$d = \frac{ {c^{\mathcal{A}} }_{23} {s^{\mathcal{D}} }_{12} }{ {s^{\mathcal{A}} }_{23} {c^{\mathcal{D}} }_{12} {s^{\mathcal{D}} }_{13} } .$$ Therefore, ${\Phi^{\mathcal{A}} }_{L1} = -{\Phi^{\mathcal{A}} }_{R1}$ and ${\Phi^{\mathcal{A}} }_{R2} = {\Phi^{\mathcal{A}} }_{15} - {\Phi^{\mathcal{A}} }_{L1}$ can be determined. Finally, the CP-violating phase in representation ${\mathcal{A}}$, ${\delta_{cp}^{\mathcal{A}} }$, can be resolved using those conditions associated with $\sin {\delta_{cp}^{\mathcal{A}} }$: $$\begin{aligned} & & \hskip-1.0cm \sin {\delta_{cp}^{\mathcal{A}} }\nonumber \\ & & \hskip-1.0cm = \frac{ {s^{\mathcal{D}} }_{12} \sin({\Phi^{\mathcal{A}} }_{L2} + {\Phi^{\mathcal{A}} }_{R1} + {\delta_{cp}^{\mathcal{D}} }) } { {s^{\mathcal{A}} }_{23} {c^{\mathcal{A}} }_{12} {s^{\mathcal{A}} }_{13} } \nonumber \\ & & \hskip-1.0cm = \frac{ {c^{\mathcal{D}} }_{12} {s^{\mathcal{D}} }_{13} \sin({\Phi^{\mathcal{A}} }_{L3} + {\Phi^{\mathcal{A}} }_{R1}) } { {c^{\mathcal{A}} }_{23} {c^{\mathcal{A}} }_{12} {s^{\mathcal{A}} }_{13} } \nonumber \\ & & \hskip-1.0cm = \frac{-{c^{\mathcal{D}} }_{23} {c^{\mathcal{D}} }_{12} \sin({\Phi^{\mathcal{A}} }_{L2} + {\Phi^{\mathcal{A}} }_{R2}) } { {s^{\mathcal{A}} }_{23} {s^{\mathcal{A}} }_{12} {s^{\mathcal{A}} }_{13} } \nonumber \\ & & \hskip-1.0cm = \frac{ {c^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {s^{\mathcal{D}} }_{13} \sin(\eta - {\delta_{cp}^{\mathcal{D}} }) +{s^{\mathcal{D}} }_{23} {c^{\mathcal{D}} }_{13} \sin\eta } { {c^{\mathcal{A}} }_{23} {s^{\mathcal{A}} }_{12} {s^{\mathcal{A}} }_{13} } \nonumber \\ & & \hskip-1.0cm = \frac{ {s^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{12} {c^{\mathcal{D}} }_{13} \sin({\Phi^{\mathcal{A}} }_{L1} - {\delta_{cp}^{\mathcal{D}} }) +{c^{\mathcal{D}} }_{23} {s^{\mathcal{D}} }_{13} \sin{\Phi^{\mathcal{A}} }_{L1} } { {s^{\mathcal{A}} }_{13} } \label{eq:Append_D_DA_5}\end{aligned}$$ where $\eta \equiv {\Phi^{\mathcal{A}} }_{L3} + {\Phi^{\mathcal{A}} }_{R2}$. Consistency in the value of $\sin {\delta_{cp}^{\mathcal{A}} }$ calculated from the five different expressions in Eq.(\[eq:Append\_D\_DA\_5\]) serves as a means to check whether the values of the nine parameters are correct. [99]{} K. Nakamura [*et al.*]{} \[ Particle Data Group Collaboration \], J. Phys. G [**G37**]{}, 075021 (2010). Z. Maki, M. Nakagawa and S. Sakata, Prog. Theor. Phys.  [**28**]{}, 870 (1962). B. Pontecorvo, Sov. Phys. JETP [**26**]{}, 984 (1968) \[Zh. Eksp. Teor. Fiz.  [**53**]{}, 1717 (1967)\]. J. Schechter and J. W. F. Valle, Phys. Rev.  D [**22**]{}, 2227 (1980). H. Fritzsch, Z. -z. Xing, Phys. Rev.  [**D57**]{}, 594-597 (1998) \[hep-ph/9708366\]. S. Choubey, S. Goswami, K. Kar, Astropart. Phys.  [**17**]{}, 51-73 (2002) \[arXiv:hep-ph/0004100 \[hep-ph\]\]. Carlo Giunti and Chung W. Kim, *Fundamentals of Neutrino Physics and Astrophysics*, published by Oxford University Press Inc. New York (2007). Y. j. Zheng, Phys. Rev.  D [**81**]{}, 073009 (2010) \[arXiv:1002.0919 \[hep-ph\]\]. M. Huang, D. Liu, J. C. Peng, S. D. Reitzner and W. C. Tsai, arXiv:1108.3906 \[hep-ph\]. A. Aguilar [*et al.*]{} \[LSND Collaboration\], Phys. Rev.  D [**64**]{}, 112007 (2001) \[arXiv:hep-ex/0104049\]. L. -L. Chau, W. -Y. Keung, Phys. Rev. Lett.  [**53**]{}, 1802 (1984). B. Aharmim [*et al.*]{} \[SNO Collaboration\], Phys. Rev.  C [**81**]{}, 055504 (2010) \[arXiv:0910.2984 \[nucl-ex\]\]. [*et al.*]{} \[SNO Collaboration\], \[arXiv:1109.0763 \[nucl-ex\]\]. J. Hosaka [*et al.*]{} \[Super-Kamkiokande Collaboration\], Phys. Rev.  D [**73**]{}, 112001 (2006) \[arXiv:hep-ex/0508053\]. J. P. Cravens [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Rev.  D [**78**]{}, 032002 (2008) \[arXiv:0803.4312 \[hep-ex\]\]. A. Gando [*et al.*]{} \[The KamLAND Collaboration\], Phys. Rev.  D [**83**]{}, 052002 (2011) \[arXiv:1009.4771 \[hep-ex\]\]. M. Apollonio [*et al.*]{} \[CHOOZ Collaboration\], Eur. Phys. J.  C [**27**]{}, 331 (2003) \[arXiv:hep-ex/0301017\]. M. H. Ahn [*et al.*]{} \[K2K Collaboration\], Phys. Rev.  D [**74**]{}, 072003 (2006) \[arXiv:hep-ex/0606032\]. P. Adamson [*et al.*]{} \[The MINOS Collaboration\], Phys. Rev. Lett.  [**106**]{}, 181801 (2011) \[arXiv:1103.0340 \[hep-ex\]\]. L. Whitehead \[for MINOS Collaboration\], “Recent results from MINOS”, Joint Experimental-Theoretical Seminar (24 June 2011, Fermilab, USA); websites: theory.fnal.gov/jetp, http://www-numi.fnal.gov/pr\_plots/ R. Wendell [*et al.*]{} \[Kamiokande Collaboration\], Phys. Rev.  [**D81**]{}, 092004 (2010), \[arXiv:1002.3471 \[hep-ex\]\]. B. Aharmim [*et al.*]{} \[SNO Collaboration\], Phys. Rev.  D [**80**]{}, 012001 (2009), \[arXiv:0902.2776 \[hep-ex\]\]. J. Hosaka [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Rev.  [**D74** ]{}, 032002 (2006), \[hep-ex/0604011\]. G. L. Fogli, E. Lisi, A. Marrone, A. Palazzo, A. M. Rotunno, \[arXiv:1106.6028 \[hep-ph\]\]. T. Schwetz, M. Tortola, J. W. F. Valle, New J. Phys.  [**13**]{}, 063004 (2011), \[arXiv:1103.0734 \[hep-ph\]\]. M. C. Gonzalez-Garcia, M. Maltoni, J. Salvado, JHEP [**1004**]{}, 056 (2010), \[arXiv:1001.4524 \[hep-ph\]\]. J. E. Roa, D. C. Latimer, D. J. Ernst, Phys. Rev.  [**C81**]{}, 015501 (2010), \[arXiv:0904.3930 \[nucl-th\]\]. A. B. Balantekin, D. Yilmaz, J. Phys. G [**G35**]{}, 075007 (2008), \[arXiv:0804.3345 \[hep-ph\]\]. H. L. Ge, C. Giunti, Q. Y. Liu, Phys. Rev.  [**D80**]{}, 053009 (2009), \[arXiv:0810.5443 \[hep-ph\]\]. S. Goswami and A. Y. Smirnov, Phys. Rev.  D [**72**]{}, 053011 (2005), \[arXiv:hep-ph/0411359\]. S. Choubey, J. Phys. G [**G29**]{}, 1833-1837 (2003). X. Guo [*et al.*]{} \[Daya Bay Collaboration\], \[hep-ex/0701029\]. F. Ardellier [*et al.*]{} \[Double Chooz Collaboration\], \[hep-ex/0606025\]. K. K. Joo \[ RENO Collaboration \], Nucl. Phys. Proc. Suppl.  [**168**]{}, 125-127 (2007). J. C. Anjos [*et al.*]{} \[Angra Collaboration\], Nucl. Phys. Proc. Suppl.  [**155**]{}, 231 (2006), \[arXiv:hep-ex/0511059\]. M. -C. Chen, K. T. Mahanthappa, Nucl. Phys. Proc. Suppl.  [**188**]{}, 315-320 (2009). \[arXiv:0812.4981 \[hep-ph\]\]. http://doublechooz.in2p3.fr/Status\_and\_News/status\_and\_news.php http://j-parc.jp/NuclPart/pac\_0606/pdf/p11-Nishikawa.pdf D. S. Ayres [*et al.*]{} \[NOvA Collaboration\], \[arXiv:hep-ex/0503053\]. Y. Itow [*et al.*]{} \[The T2K Collaboration\], \[arXiv:hep-ex/0106019\]. K. Hagiwara, N. Okamura and K. i. Senda, Phys. Rev.  D [**76**]{}, 093002 (2007), \[arXiv:hep-ph/0607255\]. C. Jarlskog, Phys. Rev. Lett.  [**55**]{}, 1039 (1985). D. -d. Wu, Phys. Rev.  [**D33**]{}, 860 (1986). L. Wolfenstein, Phys. Rev. D **17**, 2369 (1978). S. P. Mikheyev and A.Yu. Smirnov, Sov. J. Nucl. Phys. **42**, 913 (1985). B. T. Cleveland [*et al.*]{}, Astrophys. J.  [**496**]{}, 505 (1998). J. N. Abdurashitov [*et al.*]{} \[SAGE Collaboration\], Phys. Rev.  C [**80**]{}, 015807 (2009), \[arXiv:0901.2200 \[nucl-ex\]\]. C. Arpesella [*et al.*]{} \[The Borexino Collaboration\], Phys. Rev. Lett. [**101**]{}, 091302 (2008), \[arXiv:0805.3843 \[astro-ph\]\]. T. .A. Mueller, D. Lhuillier, M. Fallot, A. Letourneau, S. Cormon, M. Fechner, L. Giot, T. Lasserre [*et al.*]{}, Phys. Rev.  [**C83**]{}, 054615 (2011), \[arXiv:1101.2663 \[hep-ex\]\]. http://www.sns.ias.edu/ jnb/SNdata/sndata.html W. T. Winter, S. J. Freedman, K. E. Rehm and J. P. Schiffer, Phys. Rev.  C [**73**]{}, 025503 (2006), \[arXiv:nucl-ex/0406019\]. J. N. Bahcall, A. M. Serenelli and S. Basu, Astrophys. J.  [**621**]{}, L85 (2005), \[arXiv:astro-ph/0412440\]. A. Serenelli, S. Basu, J. W. Ferguson and M. Asplund, Astrophys. J.  [**705**]{}, L123 (2009), \[arXiv:0909.2668 \[astro-ph.SR\]\]. A. Dziewonski, Physics of the Earth and Planetary Interiors [**10**]{}, 12 (1975). A. M. Dziewonski and D. L. Anderson, Physics of the Earth and Planetary Interiors [**25**]{}, 297 (1981). A. N. Ioannisiana and A. Yu. Smirnov, Phys. Rev. Lett. [**93**]{}, 241801 (2004). D. Casper, Nucl. Phys. Proc. Suppl.  [**112**]{}, 161 (2002), \[arXiv:hep-ph/0208030\]. V. D. Barger, K. Whisnant, S. Pakvasa and R. J. N. Phillips, Phys. Rev.  D [**22**]{}, 2718 (1980). B. Aharmim [*et al.*]{} \[SNO Collaboration\], Phys. Rev.  C [**72**]{}, 055502 (2005), \[arXiv:nucl-ex/0502021\]. G. L. Fogli and E. Lisi, Astropart. Phys.  [**3**]{}, 185 (1995). G. L. Fogli, E. Lisi, D. Montanino and A. Palazzo, Phys. Rev.  D [**62**]{}, 013002 (2000), \[arXiv:hep-ph/9912231\]. M. V. Garzelli and C. Giunti, Phys. Lett.  B [**488**]{}, 339 (2000), \[arXiv:hep-ph/0006026\]. M. V. Garzelli and C. Giunti, JHEP [**0112**]{}, 017 (2001), \[arXiv:hep-ph/0108191\]. G. L. Fogli, E. Lisi, A. Palazzo and A. M. Rotunno, Phys. Lett.  B [**623**]{}, 80 (2005), \[arXiv:hep-ph/0505081\]. http://www.awa.tohoku.ac.jp/KamLAND/datarelease/fission\_flux\_distance.dat P. Vogel and J. F. Beacom, Phys. Rev.  D [**60**]{}, 053003 (1999), \[arXiv:hep-ph/9903554\]. G. L. Fogli, E. Lisi, A. Marrone and D. Montanino, Phys. Rev.  D [**67**]{}, 093006 (2003), \[arXiv:hep-ph/0303064\]. R. A. Wendell, Ph.D. Thesis, University of North Carolina (2008). Y. Ashie [*et al.*]{} \[Super-Kamiokande Collaboration\], Phys. Rev.  D [**71**]{} (2005) 112005 \[arXiv:hep-ex/0501064\]. M. Honda, T. Kajita, K. Kasahara and S. Midorikawa, Phys. Rev.  D [**70**]{} (2004) 043008 \[arXiv:astro-ph/0404457\]. G. L. Fogli, E. Lisi, A. Marrone, D. Montanino and A. Palazzo, Phys. Rev.  D [**66**]{} (2002) 053010 \[arXiv:hep-ph/0206162\]. T. J. Sonley, PhD thesis (2009) G. D. Barr, T. K. Gaisser, P. Lipari, S. Robbins and T. Stanev, Phys. Rev.  D [**70**]{} (2004) 023006 \[arXiv:astro-ph/0403630\].

ArXiv

--- abstract: 'Sparse subspace clustering methods, such as Sparse Subspace Clustering (SSC) [@ElhamifarV13] and $\ell^{1}$-graph [@YanW09; @ChengYYFH10], are effective in partitioning the data that lie in a union of subspaces. Most of those methods use $\ell^{1}$-norm or $\ell^{2}$-norm with thresholding to impose the sparsity of the constructed sparse similarity graph, and certain assumptions, e.g. independence or disjointness, on the subspaces are required to obtain the subspace-sparse representation, which is the key to their success. Such assumptions are not guaranteed to hold in practice and they limit the application of sparse subspace clustering on subspaces with general location. In this paper, we propose a new sparse subspace clustering method named $\ell^{0}$-graph. In contrast to the required assumptions on subspaces for most existing sparse subspace clustering methods, it is proved that subspace-sparse representation can be obtained by $\ell^{0}$-graph for arbitrary distinct underlying subspaces almost surely under the mild i.i.d. assumption on the data generation. We develop a proximal method to obtain the sub-optimal solution to the optimization problem of $\ell^{0}$-graph with proved guarantee of convergence. Moreover, we propose a regularized $\ell^{0}$-graph that encourages nearby data to have similar neighbors so that the similarity graph is more aligned within each cluster and the graph connectivity issue is alleviated. Extensive experimental results on various data sets demonstrate the superiority of $\ell^{0}$-graph compared to other competing clustering methods, as well as the effectiveness of regularized $\ell^{0}$-graph.' bibliography: - 'egbib.bib' title: 'Learning with $\ell^{0}$-Graph: $\ell^{0}$-Induced Sparse Subspace Clustering' --- Introduction ============ Clustering is a common unsupervised data analysis method which partitions data into a set of self-similar clusters. High dimensionality of data often imposes difficulty on clustering. For example, model-based clustering methods, such as Gaussian Mixture Model (GMM) that models the data by a mixture of parametric distributions, suffer from the curse of dimensionality when fitting a statistical model to the data [@Fraley02]. Based on the observation that high dimensional data often lie in a set of low-dimensional subspaces in many practical scenarios, subspace clustering algorithms [@Vidal11] aim to partition the data such that data belonging to the same subspace are identified as one cluster. Among various subspace clustering algorithms, the ones that employ sparsity prior, such as Sparse Subspace Clustering (SSC) [@ElhamifarV13] and $\ell^{1}$-graph [@YanW09; @ChengYYFH10], have been proven to be effective in separating the data in accordance with the subspaces that the data lie in under certain assumptions. Sparse subspace clustering methods construct sparse similarity graph by sparse representation of the data, where the vertices represent the data, and an edge is between two vertices whenever one participates the spare representation of the other. Thanks to the subspace-sparse representation, the nonzero elements in the sparse representation of each datum in a subspace correspond to the data points in the same subspace, so that vertices corresponding to different subspaces are disconnected in the sparse similarity graph, leading to their compelling performance with spectral clustering [@Ng01] applied on such graph. [@ElhamifarV13] proves that when the subspaces are independent or disjoint, then subspace-sparse representations can be obtained by solving the canonical sparse coding problem using data as the dictionary under certain conditions on the rank, or singular value of the data matrix and the principle angle between the subspaces respectively. Under the independence assumption on the subspaces, low rank representation [@LiuLY10; @Liu12] is also proposed to recover the subspace structures. Relaxing the assumptions on the subspaces to allowing overlapping subspaces, the Greedy Subspace Clustering [@ParkCS14] and the Low-Rank Sparse Subspace Clustering [@Wang13] achieve subspace-sparse representation with high probability. However, their results rely on the semi-random model which assumes the data in each subspace are generated i.i.d. uniformly on the unit sphere in that subspace as well as certain additional conditions on the size and dimensionality of the data. In addition, the geometric analysis in [@Soltanolkotabi2012] also adopts the semi-random model and it handles overlapping subspaces. To avoid the non-convex optimization problem incurred by $\ell^{0}$-norm, most of the sparse subspace clustering or sparse graph based clustering methods use $\ell^{1}$-norm [@YanW09; @ChengYYFH10; @ElhamifarV11; @ElhamifarV13; @YYZRl1graphBMVC2014] or $\ell^{2}$-norm with thresholding [@Peng2015robust] to impose the sparsity on the constructed similarity graph. In addition, $\ell^{1}$-norm has been widely used as a convex relaxation of $\ell^{0}$-norm for efficient sparse coding algorithms [@jenatton2010proximal; @Mairal2010; @MairalBPSZ08]. On the other hand, sparse representation methods such as [@Mancera2006] that directly optimize objective function involving $\ell^{0}$-norm demonstrate compelling performance compared to its $\ell^{1}$-norm counterpart. It remains an interesting question whether sparse subspace clustering equipped with $\ell^{0}$-norm, which is the origination of the sparsity that counts the number of nonzero elemens, has advantage in obtaining the subspace-sparse representation. In this paper, we propose $\ell^{0}$-graph which employs $\ell^{0}$-norm to enforce the sparsity of the similarity graph. This paper offers three contributions: - **Theoretical Results on $\ell^{0}$-Induced Almost Surely Subspace-Sparse Representation** We present the theory of the $\ell^{0}$-induced sparse subspace clustering by $\ell^{0}$-graph, which shows that $\ell^{0}$-graph renders subspace-sparse representation almost surely under minimum assumptions on the underlying subspaces the data lie in, i.e. subspaces are distinct. To the best of our knowledge, this is the mildest assumption on the subspaces compared to most existing sparse subspace clustering methods. Furthermore, our theory assumes that the data in each subspace are generated i.i.d. from arbitrary continuous distribution supported on that subspace, which is milder than the assumption of semi-random model in [@ParkCS14] and [@Wang13] that assume the data are i.i.d. uniformly distributed on the unit sphere in each subspace. - **Efficient Optimization** The optimization problem of $\ell^{0}$-graph is NP-hard and it is impractical to pursue the global optimal solution. Instead, we develop an efficient proximal method to obtain a sub-optimal solution with convergence guarantee. - **Regularized $\ell^{0}$-Graph** In order to obtain a sparse similarity graph where neighboring data have similar neighbors so as to encourage the graph connectivity within each cluster, we propose Regularized $\ell^{0}$-graph that incorporates an regularization term into the objective of $\ell^{0}$-graph. Moreover, we have implemented both $\ell^{0}$-graph and regularized $\ell^{0}$-graph in CUDA C programming language for significant speedup by parallel computing. Note that SSC-OMP [@Dyer13a] adopts Orthogonal Matching Pursuit (OMP) [@Tropp04] to choose neighbors for each datum in the sparse similarity graph, which can be interpreted as approximately solving a $\ell^{0}$ problem. However, SSC-OMP does not present the theoretical properties of the $\ell^{0}$-induced sparse subspace clustering, and the experimental results show the significant performance advantage of $\ell^{0}$-graph over the OMP-graph. OMP-graph solves the $\ell^{0}$ problem of $\ell^{0}$-graph by OMP, so that it is equivalent to SSC-OMP for clustering. Although our optimization algorithm only obtains a sub-optimal solution to the objective of $\ell^{0}$-graph, we give theory about $\ell^{0}$-induced subspace structures and extensive experimental results show the effectiveness of our model. The remaining parts of the paper are organized as follows. The representative subspace subspace clustering methods, SSC and $\ell^{1}$-graph, are introduced in the next subsection, and then the detailed formulation of $\ell^{0}$-graph and regularized $\ell^{0}$-graph is illustrated. We then show the clustering performance of the proposed models, and conclude the paper. We use bold letters for matrices and vectors, and regular lower letter for scalars throughout this paper. The bold letter with superscript indicates the corresponding column of a matrix, and the bold letter with subscript indicates the corresponding element of a matrix or vector. $\|\cdot\|_F$ and $\|\cdot\|_p$ denote the Frobenius norm and the $\ell^{p}$-norm, and ${\rm diag}(\cdot)$ indicates the diagonal elements of a matrix. Sparse Subspace Clustering and $\ell^{1}$-Graph ----------------------------------------------- Sparse coding methods represent an input signal by a linear combination of only a few atoms of a dictionary, and the sparse coefficients are named sparse code. Sparse coding has been broadly applied in machine learning and signal processing, and sparse code is extensively used as a discriminative and robust feature representation [@YangYGH09; @ChengSRL2013; @ZhangGLXA13; @YYZRl1graphBMVC2014] SSC [@ElhamifarV13] and $\ell^{1}$-graph [@YanW09; @ChengYYFH10] employ sparse representation of the data to construct the sparse similarity graph. With the data ${\bm X}=[ {{\bx_1},\ldots ,{\bx_n}} ] \in {\R^{d \times n}}$ where $n$ is the size of the data and $d$ is the dimensionality, SSC and $\ell^{1}$-graph solves the following sparse coding problem: $$\begin{aligned} \label{eq:ssc-l1} \mathop {\min }\limits_{{\balpha}} {\| {{\balpha}} \|_1}\quad s.t.\;{\bmX} = {{\bmX}}{\balpha},\,\, {\rm diag}(\balpha) = \bzero\end{aligned}$$ Both SSC and $\ell^{1}$-graph construct a sparse similarity graph $G = ( {{\bm X},{\mathbf W}} )$ where the data ${\bmX}$ are represented as vertices, $\bW$ is the graph weight matrix of size $n \times n$ and $\bW_{ij}$ indicates the similarity between $\bx_i$ and $\bx_j$, $\bW$ is set by the sparse codes $\balpha$ as below: $$\begin{aligned} \label{eq:W} {\bW_{ij}}=({|{\balpha_{ij}}|+|{\balpha_{ji}}|})/{2} \quad 1 \le i,j \le n\end{aligned}$$ Furthermore, suppose the underlying subspaces that the data lie in are independent or disjoint, SSC [@ElhamifarV13] proves that the optimal solution to (\[eq:ssc-l1\]) is the subspace-sparse representation under several additional conditions. *The sparse representation $\balpha$ is called subspace-sparse representation if the nonzero elements of $\balpha^i$, namely the sparse representation of the datum $\bx_i$, correspond to the data points in the same subspace as $\bx_i$*. Therefore, vertices corresponding to different subspaces are disconnected in the sparse similarity graph. With the subsequent spectral clustering [@Ng01] applied on such sparse similarity graph, compelling clustering performance is achieved. Allowing some tolerance for inexact representation, the literature often turns to solve the following problem for SSC and $\ell^{1}$-graph: $$\begin{aligned} \mathop {\min }\limits_{{\balpha}} {\| {{\balpha}} \|_1}\quad s.t.\;\|{\bmX} - {{\bmX}}{\balpha}\|_F \le \delta,\,\, {\rm diag}(\balpha) = \bzero\end{aligned}$$ which is equivalent to the following problem $$\begin{aligned} \label{eq:ssc-l1-lasso} \mathop {\min }\limits_{{\balpha}} {\|\bmX - \bmX \balpha\|_F^2 + {\lambda_{\ell^{1}}}\|{\balpha}\|_1} \quad s.t. \,\, {\rm diag}(\balpha) = \bzero\end{aligned}$$ where ${\lambda_{\ell^{1}}}>0$ is a weighting parameter for the $\ell^{1}$ term. Assumption on Subspaces Explanation ------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------- -- -- -- -- -- -- -- -- -- $S_1$:Independent Subspaces ([@LiuLY10; @Liu12]) ${\rm Dim} [\cS_1 \otimes \cS_2 \ldots \cS_K] = \sum\limits_k {\rm Dim}[\cS_k]$ $S_2$:Disjoint Subspaces ([@ElhamifarV13]) $\cS_k \cap \cS_{k'} = \bzero$ for $k \neq k'$ $S_3$:Overlapping Subspaces ([@ParkCS14; @Wang13; @Soltanolkotabi2012]) ${\rm Dim} [\cS_k \cap \cS_{k'}] < \min\{{\rm Dim}[\cS_k],{\rm Dim}[\cS_{k'}]\}$ for $k \neq k'$ $S_4$:Distinct Subspaces ($\ell^{0}$-Graph) $\cS_k \neq \cS_{k'}$ for $k \neq k'$ Assumption on Random Data Generation Explanation $D_1$:Semi-Random Model ([@ParkCS14; @Wang13; @Soltanolkotabi2012]) The data in each subspace are generated i.i.d. uniformly on the unit sphere in that subspace. $D_2$:IID ($\ell^{0}$-Graph) The data in each subspace are generated i.i.d. from arbitrary continuous distribution supported on that subspace. \[table:assumptions\] $\ell^{0}$-Induced Sparse Subspace Clustering ============================================= In this paper, we investigate $\ell^{0}$-induced sparse subspace clustering method, which solves the following $\ell^{0}$ problem: $$\begin{aligned} \label{eq:ssc-l0} \mathop {\min }\limits_{{\balpha}} {\| {{\balpha}} \|_0}\quad s.t.\;{\bmX} = {{\bmX}}{\balpha},\,\, {\rm diag}(\balpha) = \bzero\end{aligned}$$ We then give the theorem about $\ell^{0}$-induced almost surely subspace-sparse representation, and the proof is presented in the supplementary document for this paper. \[theorem::l0-ssc\] (*$\ell^{0}$-Induced Almost Surely Subspace-Sparse Representation*) Suppose the data ${\bm X}=[ {{\bx_1},\ldots ,{\bx_n}} ] \in {\R^{d \times n}}$ lie in a union of $K$ distinct subspaces $\{\cS_k\}_{k=1}^K$ of dimensions $\{d_k\}_{k=1}^K$, i.e. $\cS_k \neq \cS_{k'}$ for $k \neq k'$. Let $\bmX^{(k)} \in \R^{d \times n_k}$ denotes the data that belong to subspace $\cS_k$, and $\sum\limits_{k=1}^K n_k = n$. When $n_k \ge d_k+1$, if the data belonging to each subspace are generated i.i.d. from some unknown distribution supported on that subspace, then with probability $1$, the optimal solution to (\[eq:ssc-l0\]), denoted by $\balpha^*$, is a subspace-sparse representation, i.e. nonzero elements in ${\balpha^*}^i$ corresponds to the data that lie in the same subspace as $\bx_i$. Based on the above theorem, we propose $\ell^{0}$-graph that solves (\[eq:ssc-l0\]) and uses the sparse representation to build the sparse similarity graph for clustering. According to Theorem \[theorem::l0-ssc\], $\ell^{0}$-induced sparse subspace clustering method (\[eq:ssc-l0\]) obtains the subspace-sparse representation almost surely under minimum assumption on the subspaces, i.e. it only requires that the subspaces be distinct. To the best of our knowledge, this is the mildest assumption on the subspaces for most existing sparse subspace clustering methods. Moreover, the only assumption on the data generation is that the data in each subspace are i.i.d. random samples from arbitrary continuous distributions supported on that subspace. In the light of assumed data distribution, such assumption on the data generation is much milder than the assumption of the semi-random model in ([@ParkCS14; @Wang13; @Soltanolkotabi2012]) (note that the data can always be normalized to have unit norm and reside on the unit sphere). Table \[table:assumptions\] summarizes different assumptions on the subspaces and random data generation for different sparse subspace clustering methods. It can be seen that $\ell^{0}$-graph has mildest assumption on both subspaces and the random data generation. Optimization of $\ell^{0}$-Graph ================================ We introduce the optimization algorithm for $\ell^{0}$-graph in this section. Similar to the case of SSC and $\ell^{1}$-graph, by allowing tolerance for inexact representation, we turn to optimize the following $\ell^{0}$ problem $$\begin{aligned} \label{eq:l0graph} \mathop {\min }\limits_{{\balpha}} L(\balpha) = {\|\bmX - \bmX \balpha\|_F^2 + {\lambda}\|{\balpha}\|_0} \quad s.t. \,\, {\rm diag} = \bzero\end{aligned}$$ Problem (\[eq:l0graph\]) is NP-hard, and it is impractical to seek for its global optimal solution. The literature extensively resorts to approximate algorithms, such as Orthogonal Matching Pursuit [@Tropp04], or that uses surrogate functions [@Hyder09], for $\ell^{0}$ problems. Inspired by recent advances in solving non-convex optimization problems by proximal linearized method [@BoltePAL2014] and the application of this method to $\ell^{0}$-norm based dictionary learning [@BaoJQS14], we propose an iterative proximal method to optimize (\[eq:l0graph\]) and obtain a sub-optimal solution with proved convergence guarantee. In the following text, the superscript with bracket indicates the iteration number of the proposed proximal method. In $t$-th iteration of our proximal method for $t \ge 1$, gradient descent is performed on the squared loss term of (\[eq:l0graph\]), i.e. $Q(\balpha) = \|\bmX - \bmX \balpha\|_F^2$, to obtain $$\begin{aligned} \label{eq:l0graph-proximal-step1} \tilde {\balpha}^{(t)} = {\balpha}^{(t-1)} - \frac{2}{{\tau}s} ({\bmX^\top}{\bmX}{\balpha^{(t-1)}}-{\bmX^\top}{\bmX})\end{aligned}$$ where $\tau$ is any constant that is greater than $1$, and $s$ is the Lipschitz constant for the gradient of function $Q(\cdot)$, namely $$\begin{aligned} \label{eq:lipschitz-L} \|\nabla Q(\bY) - \nabla Q(\bZ)\|_F \le s \|\bY-\bZ\|_F, \,\, \forall \, \bY,\bZ \in \R^{n \times n}\end{aligned}$$ Then ${\balpha}^{(t)}$ is the solution to the following $\ell^{0}$ regularized problem: $$\begin{aligned} \label{eq:l0graph-subprob} &{\balpha}^{(t)} = \argmin \limits_{\bv \in \R^{n \times n}} {\frac{{\tau} s}{2}\|\bv - {\tilde {\balpha}^{(t)}}\|_F^2 + {\lambda}\|\bv\|_0} \\ &\quad s.t. \,\, {\rm diag} (\bv) = \bzero \nonumber\end{aligned}$$ It can be verified that (\[eq:l0graph-subprob\]) has closed-form solution, i.e. $$\begin{aligned} \label{eq:l0graph-proximal-step2} &{\balpha}_{ij}^{(t)} = \left\{ \begin{array} {r@{\quad:\quad}l} 0 & {|{\tilde {\balpha}_{ij}^{(t)}}| < \sqrt{\frac{2\lambda}{{\tau}s}} \,\, {\rm or } \,\, i = j } \\ {\tilde {\balpha}_{ij}^{(t)}} & {\rm otherwise} \end{array} \right.\end{aligned}$$ for $1 \le i,j \le n$. The iterations start from $t=1$ and continue until the sequence $\{L({\balpha}^{(t)})\}$ converges or maximum iteration number is achieved. We initialize $\balpha$ as ${\balpha}^{(0)} = \balpha_{\ell^{1}}$ and $\balpha_{\ell^{1}}$ is the sparse codes generated by SSC or $\ell^{1}$-graph via solving (\[eq:ssc-l1-lasso\]) with some proper weighting parameter $\lambda_{\ell^{1}}$. In all the experimental results of this paper, we empirically set $\lambda_{\ell^{1}} = 0.1$ when initializing $\ell^{0}$-graph. The data clustering algorithm by $\ell^{0}$-graph is described in Algorithm \[alg:l0graph\]. Also, the following theorem shows that each iteration of the proposed proximal method decreases the value of the objective function $L(\cdot)$ in (\[eq:l0graph\]), therefore, our proximal method always converges. \[theorem::sufficient-decrease\] Let $s = 2 \sigma_{\max}({\bmX^\top}{\bmX})$ where $\sigma_{\max}(\cdot)$ indicates the largest eigenvalue of a matrix, then the sequence $\{L(\balpha^{(t)})\}$ generated by the proximal method with (\[eq:l0graph-proximal-step1\]) and (\[eq:l0graph-proximal-step2\]) decreases, and the following inequality holds for $t \ge 1$: $$\begin{aligned} \label{eq:l0graph-proximal-sufficient-decrease} &L(\balpha^{(t)}) \le L(\balpha^{(t-1)}) - \frac{(\tau-1)s}{2} \|{\balpha}^{(t)} - {\balpha}^{(t-1)}\|_F^2\end{aligned}$$ And it follows that the sequence $\{L(\balpha^{(t)})\}$ converges. Furthermore, we show that if the sequence $\{\balpha^{(t)}\}$ generated by the proposed proximal method is bounded, then it is a Cauchy sequence and it converges to a critical point of the objective function $L$ in (\[eq:l0graph\]). \[theorem::converge-to-critical-point\] Suppose that the sequence $\{\balpha^{(t)}\}$ generated by the proximal method with (\[eq:l0graph-proximal-step1\]) and (\[eq:l0graph-proximal-step2\]) is bounded, then 1) $\sum\limits_{t=1}^{\infty} \|\balpha^{(t)} - \balpha^{(t-1)}\|_F < \infty$ 2) $\{\balpha^{(t)}\}$ converges to a critical point [^1] of the function $L(\cdot)$ in (\[eq:l0graph\]). [@BoltePAL2014] shows that the $\ell^{0}$-norm function $\|\cdot\|_0$ is a semi-algebraic function. The conclusions of this theorem directly follows from Theorem 1 in [@BoltePAL2014]. The detailed proofs of Theorem \[theorem::sufficient-decrease\] and Theorem \[theorem::converge-to-critical-point\] are included in the supplementary document.   \ The data set ${\bmX}=\{\bx_i\}_{i=1}^{n}$, the number of clusters $c$, the parameter $\lambda$ for $\ell^{0}$-graph, $\lambda_{\ell^{1}}$ for the initialization of the the $\ell^{0}$-graph, maximum iteration number $M$, stopping threshold $\varepsilon$\ $t=1$, initialize the coefficient matrix as ${\balpha}^{(0)} = \balpha_{\ell^{1}}$, $s = 2 \sigma_{\max}({\bmX^\top}{\bmX})$. The cluster label of $\bx_i$ is set as the cluster label of the $i$-th row of $\bv$, $1 \le i \le n$. Regularized $\ell^{0}$-Graph ============================ While the subspace-sparse representation separates the data belonging to different subspaces in the constructed sparse similarity graph, it is not guaranteed that the data points in the same subspace form a connected component. This is the well known graph connectivity issue in the sparse subspace clustering literature [@ElhamifarV13; @Nasihatkon11] which is the only gap that prevents a sparse similarity graph with subspace-sparse representation from forming the perfect clustering result, i.e. the data belonging to each subspace form a single connected component in the sparse similarity graph. SSC [@ElhamifarV13] suggests alleviating the graph connectivity issue by promoting common neighbors across the data in each subspace. In this section we propose Regularized $\ell^{0}$-Graph by adding a regularization term to (\[eq:l0graph\]) which employs $\ell^{0}$-distance between the sparse representation of the data so as to impose the sparsity of the representation and encourage common neighbors for nearby data simultaneously. Regularized $\ell^{0}$-graph solves the following problem $$\begin{aligned} \label{eq:rl0graph} &\mathop {\min }\limits_{\balpha} \|{\bmX} - {\bmX}\balpha\|_F^2 + \gamma R_{\bS}(\balpha)\end{aligned}$$ where $R_{\bS}(\balpha) = \sum\limits_{i,j=1}^n {\bS_{ij} \|\balpha^i - \balpha^j\|_0 }$ is the regularization term, $\bS$ is the adjacency matrix of the KNN graph and $\bS_{ij} = 1$ if and only if $\bx_i$ is among the $K$ nearest neighbors of $\bx_j$ in the sense of Euclidean distance. It should be emphasized that such KNN graph is a widely used strategy to identify nearby data for graph regularization in sparse coding [@Zheng11; @YYZRl1graphBMVC2014]. $\gamma>0$ is the weighting parameter for the regularization term. Since the co-located elements of two sparse codes $\balpha^i$ and $\balpha^j$ are not exactly the same in most cases, their $\ell^{0}$-distance $\|\balpha^i - \balpha^j\|_0$ is almost always the sum of their difference in support and the number of their co-located nonzero elements, and the support of a vector is defined to be the indices of its nonzero elements. Therefore, the regularization term $R_{\bS}(\balpha)$ encourages both sparsity and common neighbors across nearby data. We use coordinate descent to optimize (\[eq:rl0graph\]) with respect to $\balpha^i$ in each step of the coordinate descent, with all the other sparse codes $\{\balpha^j\}_{j \neq i}$ fixed. The optimization problem for $\balpha^i$ in each step is presented below: $$\begin{aligned} \label{eq:rl0graph-cdi} &\mathop {\min }\limits_{\balpha^i} F(\balpha^i) = \|{\bx_i} - {\bmX}\balpha^{i}\|_2^2 + \gamma R_{\tilde \bS}(\balpha^i)\end{aligned}$$ where $R_{\tilde \bS}(\balpha^i) = \sum\limits_{j=1}^n {\tilde \bS_{ij} \|\balpha^i - \balpha^j\|_0 }$, where $\tilde \bS = \bS + \bS^\top$. (\[eq:rl0graph-cdi\]) can also be optimized by the proximal method in a similar manner to $\ell^{0}$-graph. In $t$-th ($t \ge 1$) iteration of our proximal method for the problem (\[eq:rl0graph-cdi\]), gradient descent on the squared loss term of the objective function of (\[eq:rl0graph-cdi\]) is performed by (\[eq:rl0-graph-proximal-step1\]): $$\begin{aligned} \label{eq:rl0-graph-proximal-step1} \tilde {\balpha^i}^{(t)} = {\balpha^i}^{(t-1)} - \frac{2}{{\tau}s} ({\bmX^\top}{\bmX}{{\balpha^i}^{(t-1)}}-{\bmX^\top}{\bx_i})\end{aligned}$$ where $\tau$ and $s$ are the same as that in (\[eq:l0graph-proximal-step1\]). Then ${\balpha^i}^{(t)}$ is obtained as the solution to the following $\ell^{0}$ regularized problem: $$\begin{aligned} \label{eq:rl0graph-cdi-subprob} &{\balpha^i}^{(t)} = \nonumber \\ &{\argmin}_{{\bv \in \R^{n}, \bv_{i} = 0}} {\frac{{\tau} s}{2}\|\bv - \tilde {\balpha^i}^{(t)}\|_2^2 + \gamma R_{\tilde \bS}(\bv) }\end{aligned}$$ Proposition \[proposition::sol-to-subprob\] below shows the closed form solution to the subproblem (\[eq:rl0graph-cdi-subprob\]): \[proposition::sol-to-subprob\] Define $F_k(v) = \frac{{\tau} s}{2}\|v - \tilde {\balpha}_{ki}^{(t)}\|_2^2 + \gamma R_{\tilde \bS}(v)$ for $v \in \R$ and $R_{\tilde \bS}(v) \triangleq \sum\limits_{j=1}^n {\tilde \bS_{ij} \|v - \balpha_{kj}\|_0 }$. Let $\bv^{*}$ be the optimal solution to (\[eq:rl0graph-cdi-subprob\]), then the $k$-th element of $\bv^{*}$ is $$\begin{aligned} \label{eq:rl0graph-proximal-step2} &\bv_k^{*} = \left\{ \begin{array} {r@{\quad:\quad}l} \argmin_{v \in \{\tilde {\balpha}_{ki}^{(t)}\} \cup \{\balpha_{kj}\}_{\{j: \tilde \bS_{ij} \neq 0\}} } F_k(v) & { k \neq i} \\ 0 & k = i \end{array} \right.\end{aligned}$$ Proposition \[proposition::sol-to-subprob\] suggests an efficient way of obtaining the solution to (\[eq:rl0graph-cdi-subprob\]). According to (\[eq:rl0graph-proximal-step2\]), ${\balpha^i}^{(t)} = \bv^{*}$ can be obtained by searching over a candidate set of size $K+1$, where $K$ is the number of nearest neighbors to construct the KNN graph $\bS$ for regularized $\ell^{0}$-graph. Similar to Theorem \[theorem::sufficient-decrease\], the sequence $\{F({\balpha^i}^{(t)})\}_t$ is decreasing. The iterative proximal method starts from $t=1$ and continue until the sequence $\{F({\balpha^i}^{(t)})\}_t$ converges or maximum iteration number is achieved. When the proximal method converges or terminates for each $\balpha^i$, the step of coordinate descent for $\balpha^i$ is finished and the optimization algorithm proceeds to optimize other sparse codes. Each iteration of coordinate descent solves (\[eq:rl0graph-cdi\]) for $i=1 \ldots n$ sequentially, and it terminates when maximum iteration number is reached or converges under some stopping threshold on the change of the objective function (\[eq:rl0graph\]). Data Set Measure KM SC $\ell^{1}$-Graph SMCE OMP-Graph $\ell^{0}$-Graph ---------- --------- -------- -------- ------------------ -------- ----------- ------------------ AC 0.7097 0.7350 0.5128 0.6809 0.6353 **0.7692** NMI 0.1287 0.2155 0.1165 0.0871 0.0299 **0.2609** AC 0.5621 0.4922 0.4948 0.5784 0.5754 **0.6590** NMI 0.5113 0.4755 0.5210 0.6332 0.5463 **0.6709** \[table:uci-mnist\] [|c|c|c|c|c|c|c|c|]{} COIL-20 ------------- \# Clusters &Measure & KM & SC &$\ell^{1}$-Graph &SMCE &OMP-Graph &$\ell^{0}$-Graph\ &AC &0.6632 &0.6701 &1.0000 &0.7639 &0.9271 &**1.0000**\ &NMI &0.5106 &0.5455 &1.0000 &0.6741 &0.8397 &**1.0000**\ &AC &0.5130 &0.4462 &0.7986 &0.5365 &0.6753 &**0.9705**\ &NMI &0.5354 &0.4947 &0.8950 &0.6786 &0.7656 &**0.9638**\ &AC &0.5885 &0.4965 &0.7697 &0.6806 &0.5475 &**0.8310**\ &NMI &0.6707 &0.6096 &0.8960 &0.8066 &0.6316 &**0.9149**\ &AC &0.6579 &0.4271 &0.8273 &0.7622 &0.3481 &**0.9002**\ &NMI &0.7555 &0.6031 &0.9301 &0.8730 &0.4520 &**0.9552**\ &AC &0.6554 &0.4278 &0.7854 &0.7549 &0.3389 &**0.8472**\ &NMI &0.7630 &0.6217 &0.9148 &0.8754 &0.4853 &**0.9428**\ \[table:coil20\] [|c|c|c|c|c|c|c|c|]{} COIL-100 ------------- \# Clusters &Measure & KM & SC &$\ell^{1}$-Graph &SMCE &OMP-Graph &$\ell^{0}$-Graph\ &AC &0.5850 &0.4514 &0.5757 &0.6208 &0.4243 &**0.9264**\ &NMI &0.7456 &0.6700 &0.7980 &0.7993 &0.5258 &**0.9681**\ &AC &0.5791 &0.4139 &0.5934 &0.6038 &0.2340 &**0.8472**\ &NMI &0.7691 &0.6681 &0.7962 &0.7918 &0.4378 &**0.9471**\ &AC &0.5371 &0.3389 &0.5657 &0.5887 &0.1905 &**0.8326**\ &NMI &0.7622 &0.6343 &0.8162 &0.7973 &0.3690 &**0.9352**\ &AC &0.5048 &0.3115 &0.5271 &0.5835 &0.2247 &**0.7899**\ &NMI &0.7474 &0.6088 &0.8006 &0.8006 &0.4173 &**0.9218**\ &AC &0.4996 &0.2835 &0.5275 &0.5639 &0.1667 &**0.7683**\ &NMI &0.7539 &0.5923 &0.8041 &0.8064 &0.3757 &**0.9182**\ \[table:coil100\] [|c|c|c|c|c|c|c|c|]{} Yale-B ------------- \# Clusters &Measure & KM & SC &$\ell^{1}$-Graph &SMCE &OMP-Graph &$\ell^{0}$-Graph\ &AC &0.1782 &0.1922 &0.7580 &0.3672 &0.7375 &**0.8406**\ &NMI &0.0897 &0.1310 &0.7380 &0.3266 &0.7468 &**0.7695**\ &AC &0.1554 &0.1706 &0.7620 &0.3761 &0.7532 &**0.7987**\ &NMI &0.1083 &0.1390 &0.7590 &0.3593 &0.7943 &**0.8183**\ &AC &0.1200 &0.1466 &0.7930 &0.3526 &0.7813 &**0.8273**\ &NMI &0.0872 &0.1183 &0.7860 &0.3771 &0.8172 &**0.8429**\ &AC &0.1096 &0.1209 &0.8210 &0.3470 &0.7156 &**0.8633**\ &NMI &0.1159 &0.1338 &0.8030 &0.3927 &0.7260 &**0.8762**\ &AC &0.0954 &0.1077 &0.7850 &0.3293 &0.6529 &**0.8480**\ &NMI &0.1258 &0.1485 &0.7760 &0.3812 &0.7024 &**0.8612**\ \[table:yaleb\] {width="46.00000%"} {width="46.00000%"} {width="46.00000%"} {width="46.00000%"} Experimental Results ==================== The superior clustering performance of $\ell^{0}$-graph is demonstrated in this section with extensive experimental results, and we also show the effectiveness of regularized $\ell^{0}$-graph. We compare our $\ell^{0}$-graph to K-means (KM), Spectral Clustering (SC), $\ell^{1}$-graph, Sparse Manifold Clustering and Embedding (SMCE) [@ElhamifarV11]. Moreover, we derive the OMP-graph, which builds the sparse graph in the same way as $\ell^{0}$-graph except that it solves the following optimization problem by Orthogonal Matching Pursuit (OMP) to obtain the sparse code: $$\begin{aligned} \label{eq:ompgraph} \mathop {\min }\limits_{{\balpha^i}} \|\bx_i - \bmX \balpha^{i}\|_F^2 \quad s.t. \,\, \|{\balpha^{i}}\|_0 \le T, {\balpha}_i^i = 0, \,\, i=1,\ldots,n\end{aligned}$$ $\ell^{0}$-graph is also compared to OMP-graph to show the advantage of the proposed proximal method in the previous sections. By adjusting the parameters, $\ell^{1}$-graph and SSC solve the same problem and generate equivalent results, so we report their performance under the same name “$\ell^{1}$-graph”. Evaluation Metric ----------------- Two measures are used to evaluate the performance of the clustering methods, i.e. the accuracy and the Normalized Mutual Information(NMI) [@Zheng04]. Let the predicted label of the datum $\bx_i$ be $\hat y_i$ which is produced by the clustering method, and $y_i$ is its ground truth label. The accuracy is defined as $$\begin{aligned} \label{eq:accuracy} &{Accuracy} = \frac{\1_{{\Omega}(\hat y_i) \ne y_i}}{n}\end{aligned}$$ where $\1$ is the indicator function, and $\Omega$ is the best permutation mapping function by the Kuhn-Munkres algorithm [@plummer1986]. The more predicted labels match the ground truth ones, the more accuracy value is obtained. Let $\hat X$ be the index set obtained from the predicted labels $\{\hat y_i\}_{i=1}^n$ and $X$ be the index set from the ground truth labels $\{y_i\}_{i=1}^n$. The mutual information between ${\hat X}$ and $X$ is $$\begin{aligned} \label{eq:MI} MI( {\hat X,X}) = \sum\limits_{\hat x \in \hat X,x \in X} {p( {\hat x,x} ){{\log }_2}( {\frac{{p( {\hat x,x} )}}{{p( {\hat x} )p( x )}}} )}\end{aligned}$$ where $p(\hat x)$ and $p(x)$ are the margined distribution of $\hat X$ and $X$ respectively, induced from the joint distribution $p(\hat x, x)$ over $\hat X$ and $X$. Let $H( {\hat X} )$ and $H( X )$ be the entropy of $\hat X$ and $X$, then the normalized mutual information (NMI) is defined as below: $$\begin{aligned} \label{eq:NMI} &NMI( {\hat X,X} ) = \frac{{MI( {\hat X,X} )}}{{\max \{ {H( {\hat X} ),H( X )}\}}}\end{aligned}$$ It can be verified that the normalized mutual information takes values in $[0,1]$. The accuracy and the normalized mutual information have been widely used for evaluating the performance of the clustering methods [@Zheng11; @ChengYYFH10; @Zheng04]. Clustering on UCI Data Set and MNIST Handwritten Digits Database ---------------------------------------------------------------- In this subsection, we conduct experiments on the Ionosphere data from UCI machine learning repository [@Asuncion07] and the MNIST database of handwritten digits. The information of these two data sets are in Table \[table:datasets\]. MNIST handwritten digits database has a total number of $70000$ samples for digits from $0$ to $9$. The digits are normalized and centered in a fixed-size image. For MNIST data set, we randomly select $500$ samples for each digit to obtain a subset of MNIST data consisting of $5000$ samples. The random sampling is performed for $10$ times and the average clustering performance is recorded. The clustering results on the two data sets are shown in Table \[table:uci-mnist\]. Heart Ionosphere MNIST ----------------- ------- ------------ ------- -- -- -- -- -- -- -- \# of instances 270 351 70000 Dimension 13 34 1024 \# of classes 2 2 10 : Two UCI data sets and MNIST Handwritten Digits Database in the experiments \[table:datasets\] Clustering On COIL-20 and COIL-100 Database ------------------------------------------- COIL-20 Database has $1440$ images of $20$ objects in which the background has been removed, and the size of each image is $32 \times 32$, so the dimension of this data is $1024$. COIL-100 Database contains $100$ objects with $72$ images of size $32 \times 32$ for each object. The images of each object were taken $5$ degrees apart when the object was rotated on a turntable. The clustering results on these two data sets are shown in Table \[table:coil20\] and Table \[table:coil100\] respectively. We observe that $\ell^{0}$-graph performs consistently better than all other competing methods. On COIL-100 Database, SMCE renders slightly better results than $\ell^{1}$-graph on the entire data due to its capability of modeling non-linear manifolds. Clustering On Extended Yale Face Database B ------------------------------------------- The Extended Yale Face Database B contains face images for $38$ subjects with $64$ frontal face images taken under different illuminations for each subject. The clustering results are shown in Table \[table:yaleb\]. We can see that $\ell^{0}$-graph achieves significantly better clustering result than $\ell^{1}$-graph, which is the second best method on this data. Improved $\ell^{0}$-Graph with Regularization --------------------------------------------- In this subsection, we investigate the performance of regularized $\ell^{0}$-graph. We empirically set $\bS$ to be the the adjacency matrix of $5$-NN graph and $\gamma = 0.1$ as the default parameter setting for regularized $\ell^{0}$-graph in (\[eq:rl0graph\]). We conduct comparison experiments on the UCI Heart data whose information is in Table \[table:uci-mnist\], the Extended Yale Face Database B and UMIST Face Database. The UMIST Face Database consists of $575$ images of size $112 \times 92$ for $20$ people. Each person is shown in a range of poses from profile to frontal views. The clustering results are shown in Table \[table:rl0graph\]. The better results of regularized $\ell^{0}$-graph are due to the fact that it promotes common neighbors for nearby data so as to produce a more aligned similarity graph and alleviate the graph connectivity issue. Data Set Measure $\ell^{0}$-Graph R$\ell^{0}$-Graph ---------- --------- ------------------ ------------------- -- -- -- -- AC 0.5111 **0.6444** NMI 0.0064 **0.0590** AC 0.8480 **0.8521** NMI 0.8612 **0.8634** AC 0.6730 **0.7078** NMI 0.7924 **0.8153** : Clustering Performance of Regularized $\ell^{0}$-Graph \[table:rl0graph\] Parameter Setting ----------------- We use the sparse codes generated by $\ell^{1}$-graph with the weighting parameter $\lambda_{\ell^{1}} = 0.1$ in (\[eq:ssc-l1-lasso\]), which is the default value suggested in [@ElhamifarV13], to initialize $\ell^{0}$-graph, and set $\lambda=0.5$ for $\ell^{0}$-graph empirically throughout all the experiments in this section. We observe that the average number of non-zero elements of the sparse code for each data point is around $3$ for most data sets. The maximum iteration number $M = 100$ and the stopping threshold $\varepsilon = 10^{-6}$. For OMP-graph, we tune the parameter $T$ in (\[eq:ompgraph\]) to control the sparsity of the generated sparse codes such that the aforementioned average number of non-zero elements of the sparse code matches that of $\ell^{0}$-graph. For $\ell^{1}$-graph, the weighting parameter for the $\ell^{1}$-norm is chosen from $[0.1,1]$ for the best performance. We investigate how the clustering performance on the Extended Yale Face Database B and COIL-20 Database changes by varying the weighting parameter $\lambda$ for $\ell^{0}$-graph, and illustrate the result in Figure \[fig:yaleb-lambda\] and Figure \[fig:coil20-lambda\] respectively. We observe that the performance of $\ell^{0}$-graph is much better than other algorithms over a relatively large range of $\lambda$, revealing the robustness of our algorithm with respect to the weighting parameter $\lambda$. Efficient Parallel Computing by CUDA Implementation --------------------------------------------------- We have implemented $\ell^{0}$-graph, regularized $\ell^{0}$-graph in CUDA C programming language on NVIDIA K$40$. Both the MATLAB and CUDA implementation will be available for downloading. We compare the running time of $\ell^{0}$-graph in MATLAB implementation and CUDA C implementation on the Extended Yale Face Database B data, on a workstation with $2$ Intel Xeon X$5650$ $2.67$ GHz CPU, $48$ GB memory and one NVIDIA K$40$ graphics card. MATLAB implementation takes $48.51$ seconds while the CUDA implementation only takes $1.68$ seconds, with a speedup of $28.87$ times. Due to the limited space, we have put additional experimental results in the supplementary document for this paper, such as the application of $\ell^{0}$-graph on semi-supervised learning, and the parameter sensitivity for regularized $\ell^{0}$-graph. Conclusion ========== We propose a novel $\ell^{0}$-graph for data clustering in this paper. In contrast to the existing sparse subspace clustering method such as Sparse Subspace Clustering and $\ell^{1}$-graph, $\ell^{0}$-graph features $\ell^{0}$-induced almost surely subspace-sparse representation under milder assumptions on the subspaces and random data generation. The objective function of $\ell^{0}$-graph is optimized using a proposed proximal method. Convergence of this proximal method is proved, and extensive experimental results on various real data sets demonstrate the effectiveness and superiority of $\ell^{0}$-graph over other competing methods. To improve the graph connectivity, we propose regularized $\ell^{0}$-graph whose effectiveness is also demonstrated on real data sets. [^1]: $x$ is a critical point of function $f$ if $0 \in \partial f (x)$, where $\partial f(x)$ is the limiting-subdifferential of $f$ at $x$. Please refer to more detailed definition in [@BoltePAL2014].

ArXiv

--- abstract: 'Intermediate-mass black holes (IMBHs) of mass ${\Mbh \approx 10^{2} \textendash 10^{5}}$ solar masses, $M_{\odot}$, are the long-sought missing link[@mil+04b] between stellar black holes, born of supernovae[@abb+16a], and massive black holes[@gra16], tied to galaxy evolution by the empirical $\Msig$ correlation[@fer+00; @geb+00]. We show that low-mass black hole seeds that accrete stars from locally dense environments in galaxies following a universal $\Msig$ relation[@mcc+13; @bos16] grow over the age of the Universe to be above ${\Mzo\approx3\times10^{5}\Mo}$ ($5\%$ lower limit), independent of the unknown seed masses and formation processes. The mass $\Mzo$ depends weakly on the uncertain formation redshift, and sets a universal minimal mass scale for present-day black holes. This can explain why no IMBHs have yet been found[@gra16], and it implies that present-day galaxies with ${\ss<\Szo\approx40\,\mathrm{km\,s}^{-1}}$ lack a central black hole, or formed it only recently. A dearth of IMBHs at low redshifts has observable implications for tidal disruptions[@ree88] and gravitational wave mergers[@ama+13].' author: - 'Tal Alexander[^1] and Ben Bar-Or[^2]' bibliography: - 'nat.bib' title: 'A universal minimal mass scale for present-day central black holes' --- ß[\_]{} The early stages of massive black hole growth are poorly understood[@vol12]. High-luminosity active galactic nuclei at very high redshift[@mor+11] $z$ further imply rapid growth soon after the Big Bang. Suggested formation mechanisms typically rely on the extreme conditions found in the early Universe (very low metallicity, very high gas or star density). It is therefore plausible that these black hole seeds were formed in dense environments, at least a Hubble time ago (${z\!>\!1.8}$ for a time of ${t_{H}\!=\!10}$ Gyr)[@ben+14]. The relation ${M_{\bullet}=M_{s}{(\sigma_{\star}/\sigma_{s})}^{\beta}}$ between black hole mass, $\Mbh$, and stellar velocity dispersion, $\ss$, that is observed in the local Universe over more than about three decades in massive black hole mass, correlates $\Mbh$ and $\ss$ on scales that are well outside the massive black hole’s radius of dynamical influence[@gra16], ${\rh\approx G\Mbh/\ss^{2}}$. Recent analyses of large heterogeneous galaxy samples find that a universal $\Msig$ relation holds for all galaxy types[@mcc+13; @bos16], although the scope of this relation and its evolution with redshift remain controversial[@kor+13]. Here we adopt the empirical fit[@bos16] [$\log_{10}(\!\Mbh/\Mo\!)\!=\!8.32\!\pm\!0.04\!\pm\!\delta_{\epsilon}\!\!+\!\!(5.35\!\pm\!0.23)\!\log_{10}(\!\ss\!/200\,\mathrm{km\,s}^{-1}\!)$]{}, where ${\delta_{\epsilon}=0.49\pm0.03}$ is the root mean square of the intrinsic scatter. We assume that this universal $\Msig$ holds at all redshifts[@she+15], and that the black hole seeds grow in a locally (within a few $\rh$) dense stellar environment. By fixing $\rh$, the $\Msig$ relation then imposes tight connections between the black hole and the dynamical properties of its stellar surroundings[@ale11], and specifically the rate at which it consumes stars (see ). A central black hole grows by (1) the accretion of stars, compact remnants and dark matter particles that are deflected toward it on nearly radial orbits, and either fall whole through the event horizon or are tidally disrupted outside it, and then accreted; (2) viscosity-driven accretion of interstellar gas; and (3) mergers with other black holes. Of these growth channels, only the accretion of stars must follow from the existence of a central black hole in a stellar system. Moreover, the tidal disruption event (TDE) rate in steady-state can be estimated from first principles, for given boundary conditions at $\rh$ (ref. [@bar+16]). It has been noted that typical steady-state TDE rates, $\Gamma$, around ${10^{-4}\,\mathrm{yr}^{-1}}$ (Fig. \[f:Rprc\]), imply by simple dimensional analysis that massive black holes (MBHs) with low mass, ${\lesssim10^{7}M_{\odot}}$, may acquire a substantial fraction of their mass from TDEs over the Hubble time $t_{\mathrm{H}}$, or equivalently, that linear growth by TDEs has a typical mass scale[@mag+99; @mer13; @sto+16b], ${\Mbh^{\mathrm{TDE}}\sim\Ms\Gamma t_{\mathrm{H}}\sim10^{6}\Mbh}$ (however, the growth equation is generally nonlinear, and therefore $\Mbh^{\mathrm{TDE}}$ can significantly mis-estimate $\Mbh(t_{H})$; see ). Previous studies have usually focused on the rates and prospects of TDE detection, and not on black hole growth. Although it was recently argued that $\Mbh^{\mathrm{TDE}}$ arises as a minimal black hole mass in a specific formation scenario[@sto+16b], the commonly held assumption remains that IMBHs with ${\Mbh\ll\Mbh^{\mathrm{TDE}}}$ do exist, and that this must constrain formation scenarios, or set an upper bound on the efficiency of TDE accretion, rather than a lower bound on IMBH masses[@mer13]. Here, we argue that IMBHs are transient objects, which no longer exist in the present-day Universe. We derive a universal lower bound on the present-day mass scale of central black holes, $\Mzo$, that follows directly from the universal $\Msig$ relation, and is independent of the unknown seed masses and their formation processes. We use the $\Msig$ relation to set the boundary conditions, and show that the nonlinear growth equation for black holes can be bounded from below by a simple inequality that includes only growth by TDEs. We translate the intrinsic scatter in the $\Msig$ relation to a probability distribution for the lower bounds $\Mzo$ and $\Szo$, and show that $\Mzo$ lies just below the lightest MBHs yet discovered[@gra16], ${\Mzo\lesssim\min(\Mbh^{\mathrm{obs}})\sim10^{6}\Mo}$. Stars around a central black hole are constantly scattered in angular momentum to nearly radial orbits below a critical (“loss-cone”) value, ${j_{lc}=\sqrt{1-e^{2}}}$ ($e$ is the orbital eccentricity), which approach the black hole closer than the tidal disruption radius, ${r_{t}\simeq{(\Mbh/\Ms)}^{1/3}\Rs}$ ($\Ms$ and $\Rs$ are the stellar mass and radius), where they are destroyed. Main sequence stars are disrupted outside an IMBH’s event horizon, and a fraction $f_{a}$ of about $1/4$ to $1/2$ of their mass is ultimately accreted by the black hole[@aya+00]. The TDE rate depends on the number of stars near the black hole and on the competition between the two-body relaxation time $T_{R}$ (equation ) and the orbital time in supplying and draining loss-cone orbits. The integrated contribution in steady-state from all radii is a function of $\Mbh$ and of the boundary conditions at $\rh$, fixed by $\ss$. The TDE rate is well-approximated by a power-law $\Gamma\simeq\Gamma_{\star}{(\Mbh/\Ms)}^{b}$, whose index $b$ is a function of the $\Msig$ index $\beta$, and changes across a critical mass scale $M_{c}\sim10^{6}\Mo$ (Fig. \[f:Rprc\]; see ). ![\[f:Rprc\]**Plunge rates as function of the black hole mass** $\protect\Mbh$**.** Rates (solid lines, see ) are plotted for ${\Ms=1\,\Mo}$, ${\Rs=1\,\Ro}$, ${f_{a}=3/8}$, ${\alpha=7/4}$, and ${\beta=4,5}$, normalized to the empirical $\Msig$ mass parameter ${M_{s}=2.1\times10^{8}\Mo}$ [@bos16]. The corresponding velocity dispersion $\sigma_{\star}$ is displayed for a ${\beta=5}$ $\Msig$ relation. The rates are well-approximated by power-laws (dashed lines, equation ). Radiation back-reaction above the Eddington limit $L_{E}$ by the mean accretion luminosity ${L_{TD}=\eta f_{a} \Ms c^{2}\Gamma}$ (${\eta=0.1}$ assumed for the radiative efficiency) is only relevant for the lowest black hole masses, where it may slow down the initial black hole growth.](Figure1a){width="1\columnwidth"} The index ${b\ll1}$ for the empirical range ${4\lesssim\beta<6}$ (refs [@gra16; @kor+13]). Let us assume that a black hole seed forms with an initial mass $M_{i}$ large enough to dominate its radius of influence, in a central stellar system that is massive enough to allow it to grow: that is, ${M_{\mathrm{sys}}\!\gg\! \rh^{3}\Ms\ns(\rh)\!>\!\Mbh\!\gg\!\Ms}$ at all times ($\ns$ is the stellar density). Consider first the case where the black hole grows only by accreting stars. The black hole growth equation is $$\dot{\Mbh}=f_{a}\Ms\Gamma_{\star}{(\Mbh/\Ms)}^{b}\equiv\dot{M}_{\bullet}^{\star}\,,\,\,\,\,\,\,\Mbh(0)=M_{i}\,.\label{e:dMbhdt}$$ The solution for ${b<1}$ in the ${t\gg t_{\infty}\!=\!{(M_{i}/\Ms)}^{1-b}\!/\!(|1-b|f_{a}\Gs\!)}$ limit (equation ), ${\Mbh(t)\simeq{[(1-b)f_{a}\Gs t]}^{1/(1-b)}\Ms\!\equiv\!M_\bullet^\star(t)}$, is independent of $M_{i}$. Because ${t_{\infty}\ll t_{H}}$ for ${M_{i}\lesssim 10^{5}\Mo}$, all seeds reach the same mass scale after $\mathcal{O}(t_{H})$ (Fig. \[f:Mz3zi\]). ![ \[f:Mz3zi\] **Cosmological growth of the minimal black hole mass $\Mz$ as function of redshift.** The evolution of the minimal black hole mass (equation ) is plotted for several values of the black hole seed mass $M_{i}$ at different formation redshifts $z_{i}$ ( times[@ben+14] $t_{i}$), assuming the empirical $\Msig$ relation[@bos16] without scatter, an ${\alpha=7/4}$ cusp of Solar type stars and an accreted mass fraction ${f_{a}=3/8}$. The convergence to evolution that is independent of initial mass is rapid in redshift. The earlier the formation, the higher is $\Mzo$. The range ${z_{i}=0.1-10}$ translates to a lower mass limit on present-day MBHs ${\Mzo\sim10^{(5.6\pm0.5)\Mo}}$ (gray band).](Figure2a){width="1\columnwidth"} Consider next a realistic black hole that grows also by gas accretion and/or mergers, ${\dot{M}_{\bullet}=\dot{M}_{\bullet}^{\star}+\dot{M}_{\bullet}^{+}}$, where ${\dot{M}_{\bullet}^{+}\ge0}$ is the accretion rate by the non-stellar channels. The full growth equation is $$\dot{M}_{\bullet}=\dot{M}_{\bullet}^{\star}+\dot{M}_{\bullet}^{+}\ge f_{a}\Ms\Gamma_{\star}{(\Mbh/\Ms)}^{b}\,,\,\,\,\,\,\Mbh(0)=M_{i}\,.\label{e:mutot}$$ The solution $M_{\bullet}^{\star}(t)$ of the stars-only growth (equation ) then provides a lower limit ** on ** the actual mass $\Mbh(t)$ of the growing black hole. Note that $M_{\bullet}^{\star}$ is not necessarily a lower limit on the actual stellar mass contribution $M^{\star}$ to $\Mbh$: ${M_{\bullet}^{\star}\le M^{\star}\le\Mbh}$ for ${b\ge0}$, but ${M^{\star}<M_{\bullet}^{\star}\le\Mbh}$ for ${b<0}$. The universal minimal mass scale of a central black hole at $t_{H}$, and the corresponding minimal velocity dispersion scale are then $$\label{e:Mh} \small{\!\!\Mzo\!=\!M_{\bullet}^{\star}(t_{H}\!)\!=\!{[(1\!-\!b)f_{a}\Gs t_{H}\!]}^{1\!/\!(1\!-\!b)}\!\Ms,\,\,\Szo\!=\!{(\Mzo/M_{s}\!)}^{1\!/\!\beta}\!\sigma_{s}},$$ with the index $b$ for the ${\Mbh<M_{c}}$ branch of equation . This implies that galaxies with ${\ss<\Szo}$ do not have a central black hole, or have formed it only recently, for otherwise the coevolution of the black hole and nucleus over $t_{H}$ would have driven $\ss$ to a much larger present-day value. Assuming the universal $\Msig$ relation and Solar type stars, the range in $\Mzo$ values is due to the variety and uncertainty in the properties of galactic nuclei and of tidal disruption, and to the intrinsic scatter in the relation. $\Mzo$ increases with the index $\alpha$ of the stellar cusp (${\ns\propto r^{-\alpha}}$) inside $\rh$, and with the accreted mass fraction $f_{a}$, from ${10^{5}\Mo}$ for the shallowest possible cusp of unbound stars (${\alpha=1/2}$, ${f_{a}=1/4}$) to ${10^{6}\Mo}$ for a steep isothermal cusp (${\alpha=2}$, ${f_{a}=1/2}$). ${\Mzo\simeq10^{6}\Mo}$ for the parameters adopted here: a dynamically relaxed cusp[@bar+13] where ${\alpha=7/4}$, and an accreted mass fraction ${f_{a}=3/8}$. The scatter around the $\Msig$ relation likely reflects intrinsic physical differences between galaxies beyond the measurement errors on the $\Msig$ parameters[@bos16]. This induces roughly Gaussian probability distributions for $\Mzo$ and $\Szo$: ${\Mzo=(1.1\pm0.8)\times10^{6}\Mo}$ and ${\Szo=79\pm35\,\mathrm{km\,s}^{-1}}$ $(1\sigma)$. The lowest-$\ss$ galaxies known to harbor active galactic nuclei[@xia+11] (and hence black holes, with estimated masses ${10^{5}\lesssim\Mbh\lesssim10^{6}\Mo}$), have ${\ss\sim30-40\,\mathrm{km\,s}^{-1}}$. This corresponds to the 5% lower limits ${\Szo\lesssim40\,\mathrm{km\,s}^{-1}}$ and ${\Mzo\lesssim3\times10^{5}\Mo}$, which we adopt here as representative lower limits. Lighter black holes are much rarer yet: e.g. ${\Mbh\le10^{4}\Mo}$ is below the $0.0001\%$ limit. The agreement ${\Mzo\lesssim\min(\Mbh^{\mathrm{obs}})\sim10^{6}\Mo}$ follows directly from basic local physics (tidal disruption and loss-cone dynamics) and empirical global properties of the Universe (its age and a universal $\Msig$ relation). Our derivation of $\Mzo$ rests on four assumptions. (1) There is effective accretion of tidally disrupted stars ($f_{a}$ is a few ${\!\times\,0.1}$)[@aya+00]. (2) Most black hole seeds were formed early, at a time ${t_{i}\sim{\mathcal{O}}(t_{H})}$. (3) Black hole growth is not typically mass or density limited; that is, the growing black hole is embedded in a stellar system with ${M_{\mathrm{sys}}\gg\rh^{3}\Ms\ns>\Mbh}$ for a substantial fraction of $t_{H}$. (4) The boundary conditions at $\rh$ are set by the universal $\Msig$ relation at all times. An early start for black hole seeds, and the requirement that a system that can form and retain a seed black hole should be dense and massive enough, are both physically plausible and possibly even essential[@vol12]. Such a system can be approximated as embedded in an isothermal density distribution, and is dynamically relaxed (see ). Furthermore, the accretion rate of stars in a system with $\Ns$ stars inside $\rh$, ${\mathrm{d}\Mbh/\mathrm{dt}\simeq f_{a}\Ms\Ns/(\log(1/j_{lc})T_{R})}$ (equation ), is slow enough to allow it to remain near equilibrium as it grows, as the timescale for growth by order of the stellar mass, ${(\mathrm{d}\Mbh/\mathrm{dt})/(\Ms N_{\star})}$, is longer by a factor ${\log(1/j_{lc})/f_{a}\gg1}$ than the timescale to return to steady-state, ${T_{SS}\simeq T_{R}/4}$ (ref. [@bar+13]). The least secure assumption is that a universal $\Msig$ relation holds near its present-day value as the black hole grows. However, this is broadly consistent with observations of active galactic nuclei[@she+15; @sal+13] up to ${z\sim1}$, and with simulations of large scale structure evolution[@sij+15; @tay+16] up to ${z\sim4}$. Mergers between two cental black holes increase the black hole mass, but also affect the dynamics around it. Mergers initially enhance the TDE rate[@che+11], but later they may strongly suppress it by ejecting the cusp. However, steady-state is quickly re-established around IMBHs, ${T_{SS}(\Mbh\!\lesssim\!10^{5}\Mo)<\mathcal{O}(10^{8}\,\mathrm{yr})}$ (ref. [@bar+13]). Additional growth channels thus only increase present-day black hole masses, and reinforce the conclusion that central black holes with ${\Mbh<\Mzo}$ are rare. Figure \[f:Mz3zi\] shows the evolution of the lower mass limit $\Mz$, which increases rapidly with decreasing redshift to its present-day value $\Mzo$. Present-day IMBHs may exist in recently formed systems, in mass-limited ones (for example globular clusters with ${M<10^{6}\Mo}$), in sub-galactic systems where the $\Msig$ relation need not apply (such as globular clusters or super star clusters), or in very low density galaxies (such as cored dwarfs[@wal+09]). However, it is unlikely that such systems can form a black hole seed to start with[@mil+04b; @vol12], and therefore the black hole occupation fraction there is probably low. Candidate IMBHs have been reported in globular clusters and dwarf galaxies, including recently[@bal+15; @kiz+17], but the evidence remains inconclusive. Early TDE-driven growth and the suppression of the cosmic black hole mass function below $\Mzo$ have implications for black hole seed evolution, for the cosmic rates and properties of TDEs[@ree88], and for gravitational waves (GWs) from IMBH-IMBH mergers and intermediate-mass-ratio inspirals into IMBHs. We conclude by listing these briefly. A high rate of TDEs can allow black hole seeds to continue growing despite the ejection of the ambient gas by supernovae feedback[@dub+15]. The lack of IMBHs at low redshifts means that electromagnetic searches will have to reach very deep to detect TDEs from IMBHs (jetted TDEs may provide an opportunity[@fia+16]). The prospects of detecting exotic processes related to IMBHs, such as tidal detonations of white dwarfs in the steep tidal field of a low-mass black hole[@ros+09], will be low. The mean observed TDE rate per galaxy, ${\Gamma\sim10^{-5}\,\mathrm{yr^{-1}\,gal^{-1}}}$, is much lower than predicted rates[@sto+16]. A dearth of black holes below $\Mzo$ may partially resolve the rate discrepancy. IMBHs produce GWs by intermediate-mass-ratio inspirals and by IMBH mergers. Detection of intermediate-mass-ratio inspirals by planned space-borne GW observatories[@ama+13; @yag13] is limited to redshifts below a few ${\!\times\,0.1}$, and is therefore unlikely. However, IMBH mergers can be detected to very high redshifts. A GW search for IMBH mergers and intermediate-mass-ratio inspirals can reveal the formation history of black holes. We predict that black hole seeds are driven early on to higher mass by the accretion of stars, and therefore IMBHs are rare in the present-day Universe, but will be found near their high formation redshifts. This supplement summarizes results from loss-cone theory used to derive the equation for black hole growth by stellar disruptions, and discusses the properties of its solutions. We first present, without derivation, a recipe for the approximate power-law growth rate equation (equation ), which has the advantage of leading to simple analytic results. We then comment on the general properties of its solutions, to clarify under what circumstances, and to what extent can simple dimensional analysis be used to estimate the minimal mass limit $\Mzo$. We then describe how the intrinsic scatter in the $\Msig$ relation is propagated through the growth equation to obtain the probability distributions for the lower limits $\Mzo$ and $\Szo$. Finally, we present for completeness and reproducibility an outline of the derivation of the full growth rate equation (used to verify our approximations, see Figure \[f:Rprc\]) and of its power-law approximation. Approximate power-law black hole growth rate equation. ------------------------------------------------------ We focus here on a steady-state stellar system around a black hole[@bah+76], which has a density cusp ${\ns\propto r^{-\alpha}}$ with ${\alpha=7/4}$ inside the radius of influence ${\rh=GM_{\bullet}/\sigma_{\star}^{2}}$. We further assume that the cusp is embedded in an external isothermal stellar distribution, ${\rho(r)=\sigma_{\star}^{2}/(2\pi Gr^{2})}$, so that the stellar mass enclosed inside $\rh$ is twice the black hole mass[@mer13]. Under the assumption of a universal $\Msig$ relation ${\Mbh={M_{s}(\ss/\sigma_{s})}^{\beta}}$, the dynamics leading to tidal disruption are characterized by a critical mass scale ${M_{c}\sim10^{6}\Mo}$. Tidal disruptions are dominated by stars originating from ${\sim \rh(\Mbh)}$ for ${\Mbh\ge M_{c}}$, and by stars originating from an inner critical radius ${r_{c}(\Mbh)<\rh(\Mbh)}$ for ${\Mbh<M_{c}}$ (see below for more details)[@lig+77]. The TDE rate is well-approximated by a broken power-law (Figure \[f:Rprc\]) $$\Gamma\simeq\Gamma_{\star}{(\Mbh/\Ms)}^{b}\,,\label{e:TDEratePL}$$ whose index $b$ changes across $M_{c}$, which is given by (see equations (\[e:Rp\]\[e:Gammas\]) for the general case), $$M_{c}\simeq M_{\star}{\left(\frac{16}{5s^{2}}\right)}^{3\beta/(6+\beta)}\,,\label{e:TDEQc}$$ in terms of the dimensionless velocity dispersion scale ${s={(M_{s}/M_{\star})}^{-1/\beta}\sigma_{s}/v_{\star}}$, where ${v_{\star}^{2}=G\Ms/\Rs}$ and $M_{\star}$ and $R_{\star}$ are the mass and radius of a typical star in the system, and where we approximated the logarithmic term (equation ) appearing in the general expressions by a typical value $\Lambda_{lc}=2$. The index $b$ is (see equation  for the general case), $$\begin{aligned} b & =\begin{cases} (105-23\beta)/27\beta & M_{\bullet}\le M_{c}\\ (3-\beta)/\beta & M_{\bullet}>M_{c} \end{cases}\,.\label{e:TDEb74}\end{aligned}$$ Note that ${b\ll1}$ for the empirically determined range of the $\Msig$ relation index[@gra16; @kor+13], ${4\lesssim\beta<6}$. Defining ${t_{\star}=\sqrt{R_{\star}^{3}/G\Ms}}$, the rate factor is $$\Gamma_{\star}\simeq\frac{5/4}{t_{\star}}s^{40/9}\begin{cases} 1 & M_{\bullet}\le M_{c}\\ {(M_{c}/M_{s})}^{4(6+\beta)/27\beta} & M_{\bullet}>M_{c} \end{cases}\,.\label{eq:TDEratesPL}$$ To summarize, the approximate power-law TDE rate for a black hole with mass $M_{\bullet}$ is calculated as follows. (1) Calculate the critical mass $M_{c}$ (equation ). (2) Calculate the power-law index $b$ (equation ) according to the low or high mass branch, depending on $M_{\bullet}$, and similarly calculate the rate factor $\Gamma_{\star}$ (equation ). (3) Use equation  to obtain the TDE rate from $\Gamma_{\star}$ and $b$. Properties of the black hole growth solutions. ---------------------------------------------- The general solution of the growth equation (equation ) with the initial condition ${M_{\bullet}(t=0) = M_{i}}$ is $$M_{\bullet}(t)=\begin{cases} M_{\star}{[{(M_{i}/M_{\star})}^{1-b}+(1-b)t/t_{a}]}^{1/(1-b)} & b\ne1\\ M_{i}e^{t/t_{a}} & b=1 \end{cases}\,,\label{e:PLsol}$$ where ${t_{a}={(f_{a}\Gs)}^{-1}}$ is the accretion timescale. The growth solution has three branches. The solution for ${b=1}$ diverges exponentially to infinity in infinite time. When ${b>1}$, $\Mbh$ diverges on a finite timescale $$t_{\infty}=t_{a}{(M_{i}/M_{\star})}^{1-b}/|1-b|\,,\label{e:tinf}$$ and is supra-exponential. The ${b<1}$ branch is sub-exponential and diverges slowly as a power-law. Exponential growth describes, for example, radiation pressure-regulated accretion of gas at the Eddington limit. Supra-exponential growth describes Hoyle-Lyttleton wind accretion[@hoy+39], spherical Bondi accretion[@bon52], or their generalization of accretion on an accelerating black hole[@ale+14b]. Sub-exponential growth, which is the relevant case for tidal accretion with the universal $\Msig$ relation, at ${t\gg t_{\infty}}$ asymptotically approaches a power law that is independent of seed mass $M_{i}$ $$M_{\bullet}(t)/M_{\star} \simeq {[(1-b)t/t_{a}]}^{1/(1-b)}\,.\label{e:asymptote}$$ There are two mass scales in the growth equation: the initial mass $M_{i}$, and the natural mass scale[@mur+91; @fre+02] $M^{\mathrm{TDE}}$, obtained by solving ${M^{\mathrm{TDE}}/M_{\star}=f_{a}\Gamma t_{H}=f_{a}\Gamma_{\star}t_{H}{(M^{\mathrm{TDE}}/M_{\star})}^{b}}$ with ${t_{H}=10^{10}\,\mathrm{yr}}$. This mass scale was used in past studies[@mag+99; @mer13; @sto+16] to estimate $M_{\bullet}(t_{H})$. It is instructive to analyze the role of $M_{\bullet}^{\mathrm{TDE}}$ in the growth solutions, and identify when it can provide a relevant estimate for the black hole mass. The exponential and supra-exponential solutions (${b\ge1}$) are functions of $M_{i}$ on all timescales, and $M_{\bullet}^{\mathrm{TDE}}$ plays there a role related to the exponential or divergence timescales. Because these solutions diverge, the black hole mass at any finite time is generally unrelated to either $M_{i}$ or $M_{\bullet}^\mathrm{TDE}$. The asymptotic sub-exponential solution (${b<1}$) can be written as ${M_{\bullet}(t_{H})=M_{\bullet}^\mathrm{TDE}{(1-b)}^{1/(1-b)}}$. In this case, $M_{\bullet}^{\mathrm{TDE}}$ provides a reasonable approximation for $M_{\bullet}$ as long as ${|b|\ll1}$. This is the indeed case for the empirical universal $\Msig$ relation, where ${b(\alpha=7/4,\beta=5.40)\simeq-0.125}$. However, other combinations of cusp and $\Msig$ indices can lead to arbitrarily large disparities: for example ${b(\alpha=3/2,\beta=3)\simeq0.6}$, results in ${M_{\bullet}\simeq0.1M_{\bullet}^{\mathrm{TDE}}}$. It should be emphasized that the solution branch that describes the black hole growth is not determined solely by the assumed growth channel  tidal disruptions in this case  but also by the choice of boundary conditions, which here are determined by an empirical relation. Other possible values of cusp and $\Msig$ indices would imply very different relations between $M_{\bullet}$ and $M_{\bullet}^{\mathrm{TDE}}$. For example, the transition to the exponential and supra-exponential solutions (${b\ge1}$) occurs for ${\beta\le2.1}$ (for ${\alpha=7/4}$) or for ${\beta\le3}$ (for ${\alpha=1/2}$). Therefore, it is not generally true that $M_{\bullet}^{\mathrm{TDE}}$ estimates the black hole mass. Its relevance depends on the specific solution and on the adopted boundary conditions, and cannot be assumed a priori. Intrinsic $\Msig$ scatter and distribution of lower limits. ----------------------------------------------------------- The observed intrinsic scatter in the $\Msig$ relation at ${z\simeq0}$, with r.m.s $\delta_{\epsilon}$, can be interpreted as reflecting a variance in the initial conditions of individual galaxies at their formation, a Hubble time $t_{H}$ ago, or a variance that developed gradually over their individual evolutionary histories and reached the observed rms value at $t_{H}$. We assume that the estimation errors in the parameters $\alpha$ and $\beta$ of the $\Msig$ relation, [${\log(\!\Mbh\!/\!\Ms\!)\! =\! (\bar{\alpha}\!\pm\!\delta_{\alpha}\!)\! +\! (\bar{\beta}\!\pm\!\delta_{\beta}\!)\!\log(\sigma\!/\!\sigma_{s}\!)\!\pm\!\delta_{\epsilon}}$]{}, can be approximated by a correlated bi-Gaussian distribution, ${(\alpha,\beta)\sim G_{2}(\bar{\alpha},\delta_{a},\bar{\beta},\delta_{\beta},\rho_{\alpha\beta})}$, where for an arbitrarily chosen low reference velocity dispersion ${\sigma_{s}\ll\sigma}$, the correlation coefficient ${\rho_{\alpha\beta}\to-1}$ (ref. [@tre+02]), and that the intrinsic scatter $\epsilon$ is drawn from a Gaussian distribution, ${\epsilon\sim G(0,\delta_{\epsilon})}$. We approximate the evolution of the scatter by assuming $n_{t}$ discrete time steps of duration ${\Delta t=t_{H}/n_{t}}$, where the accumulated change in $\alpha$ due to scatter, $\Delta\epsilon$, is modified in a random walk fashion by ${\Delta\epsilon\to\Delta\epsilon+\epsilon/\sqrt{n_{t}}}$. We then evolve the black hole mass over time $\Delta t$ by the growth equation (equation ), and repeat until ${t=t_{H}}$. The joint and marginal probability distributions for $\Mzo$ and $\Szo$ at $t_{H}$ are obtained by Monte Carlo simulations over randomly drawn values of $\alpha$ and $\beta$. The limit ${n_{t}=1}$ corresponds to scatter that is determined by the galaxy’s initial conditions, whereas ${n_{t}\gg1}$ corresponds to scatter that is determined by the galaxy’s evolution. We find that the probability distributions for $\Mzo$ and $\Szo$ do not depend strongly on the choice of $n_{t}$, and that they converge rapidly for ${n_{t}>3}$ to an asymptotic form. The values quoted in this study, $5\%$ lower limits of ${\Mzo=2.8\times10^{5}\Mo}$ and ${\Szo=38\,\mathrm{km\,s}^{-1}}$, correspond to the asymptotic evolutionary scatter case ($n_{t}=5$), whereas the initial scatter case ($n_{t}=1)$ differs only slightly, with $5\%$ lower limits of $\Mzo=1.9\times10^{5}\Mo$ and $\Szo=36\,\mathrm{km\,s}^{-1}$. Full black hole growth rate equation. ------------------------------------- The tidal disruption (plunge) rate can be approximated by the flux of stars into the black hole from from the boundary between the inner region, where stars slowly diffuse into the loss-cone (the empty loss-cone) and the outer region, where stellar scattering is strong enough that the loss-cone is effectively full (the full loss-cone)[@lig+77]. The boundary is at a critical radius, $a_{c}$, that satisfies $$q=\frac{{[J_{c}(a_{c})/J_{lc}]}^{2}P(a_{c})}{\log(J_{c}(a_{c})/J_{lc})T_{R}(a_{c})}=1\,,\label{e:qfactor}$$ where ${P=2\pi\sqrt{a^{3}/GM_{\bullet}}}$ is the orbital period, ${J_{c}=\sqrt{GM_{\bullet}a}}$ is the circular angular momentum at $a$, $J_{lc}$ is angular momentum of the loss-cone (${J_{lc}\simeq\sqrt{2G\Mbh Q^{1/3}\Rs}}$ for tidal disruption, so ${J_{c}/J_{lc}\simeq\sqrt{(a/\Rs)/2Q^{1/3}}}$, where ${Q=M_{\bullet}/M_{\star}}$). $T_{R}$ is the 2-body (non-resonant) relaxation time[@bar+13], $$T_{R}(a)=\frac{5}{8}\frac{Q^{2}P(a)}{\Ns(a)\log(Q)}\,,\label{e:TR}$$ where ${\Ns(a)=\mu_{h}Q{(a/\rh)}^{3-\alpha}}$ is the number of stars enclosed in $r$, and ${\rh=\eta_{h}GM_{\bullet}/\ss^{2}}$ is the radius of influence. The numeric prefactors are conventionally assumed to be ${\mu_{h}=2}$ and ${\eta_{h}=1}$. We further assume that ${\alpha<9/4}$. The exact solution for the critical radius can be written by the implicit equation $$\begin{aligned} a_{c}/\rh = {} & {(A_{c}\sigma_{\star}^{2}/v_{\star}^{2})}^{1/(4-\alpha)}Q^{1/(12-3\alpha)} \nonumber \\ = {} & {(A_{c}s^{2})}^{1/(4-\alpha)}Q^{(\beta+6)/\beta(12-3\alpha)}\,,\label{e:qexact}\end{aligned}$$ where ${A_{c}=5/(4\mu_{h}\Lambda_{lc}\eta_{h})}$ and $$\label{e:Llc} \Lambda_{lc}(Q,a)=\log Q/\log(J_{c}(a)/J_{lc})\,,$$ where the last equality in equation  assumes the $\Msig$ relation in terms of the dimensionless velocity dispersion scale ${s={(M_{s}/M_{\star})}^{-1/\beta}\sigma_{s}/v_{\star}}$, where ${v_{\star}=\sqrt{GM_{\star}/R_{\star}}}$. When ${a_{c}>\rh}$, the rate is estimated[@sye+99] at $\rh$. The transition occurs above a critical black hole mass such that ${a_{c}(M_{c})=\rh(M_{c})}$, $$M_{c} = M_{\star}{(A_{c}\sigma_{\star}^{2}/v_{\star}^{2})}^{-3}=\Ms{(A_{c}s^{2})}^{-3\beta/(6+\beta)}\,,\label{eq:Mc}$$ which is independent of $\alpha$ and almost independent of $\beta$. The plunge rate can then be conservatively approximated by the empty loss-cone rate at ${a_{e}=\min(a_{c},\rh)}$, $$\Gamma\simeq\frac{\Ns(a_{e})}{\log(J_{c}(a_{e})/J_{lc})T_{R}(a_{e})}=\frac{8}{5}\frac{\Lambda_{lc}\mu_{h}^{2}}{P(\rh)}{\left(\frac{a_{e}}{\rh}\right)}^{(9-4\alpha)/2}\,.\label{e:Rpexact}$$ The actual rate, including the contribution from the full loss-cone regime, can be up to twice as high as this as this[@sye+99]. Using ${\rh=\eta_{h}GM_\bullet/\sigma_{\star}^{2}}$, the plunge rate can be represented as $$\small{ \Gamma^{\Lambda}\!=\!\frac{1}{\pi t_{\star}}\!\begin{cases} \gamma_{c}^{\Lambda}Q^{(2\alpha-15)/6(4-\alpha)}{\left(\!\frac{\ss}{v_{\star}}\!\right)}^{7(3-\alpha)/(4-\alpha)} & \!\!M_{\bullet}\!\le\! M_{c}\\ \gamma_{h}^\Lambda Q^{-1}{\left(\!\frac{\ss}{v_{\star}}\!\right)}^3 & \!\!M_{\bullet}\!>\!M_{c} \end{cases}},$$ where ${t_\star=\sqrt{R_{\star}^{3}/GM_{\star}}}$, $$\gamma_{c}^\Lambda = {\left(\frac{4\Lambda_{lc}}{5}\right)}^{(2\alpha-1)/(8-2\alpha)} \! {\left(\frac{\mu_{h}}{\eta_{h}^{(3-\alpha)}}\right)}^{7/(8-2\alpha)}\,,$$ and $$\gamma_{h}^{\Lambda}=\frac{4\mu_{h}^{2}\Lambda_{lc}}{5\eta_{h}^{3/2}}\,.$$ When $\ss$ is given by the $\Msig$ relation, the rate can be expressed as $$\Gamma=\Gamma_{\star}^{\Lambda}Q^{b}\,,\label{e:Rp}$$ where $$\label{e:Gammas} \small{ \Gamma_{\star}^{\Lambda} \!\!=\!\! \frac{\gamma_{c}^{\Lambda}}{\pi t_{\star}} s^{7(3-\alpha)/(4-\alpha)}\!\! \begin{cases} 1 & \!\!\!M_{\bullet}\!\le\! M_{c}\\ \!\!{\left(\!\frac{M_{c}}{M_{\star}}\!\right)}^{\!(6+\beta)(9-4\alpha)/6(4-\alpha)\beta} & \!\!\!M_{\bullet}\!>\!M_{c} \end{cases}}.$$ The notation $\Gamma_{\star}^{\Lambda}$ denotes the weak functional dependence on $Q$ via the logarithmic term ${\Lambda_{lc}\simeq2}$. The index is $$\small{ b\!=\!\begin{cases} 7(3-\alpha)/\beta(4-\alpha)-(15/2-\alpha)/3(4-\alpha) & a_{c}\le \rh\\ (3-\beta)/\beta & a_{c}>\rh \end{cases}}. \label{e:bRp}$$ A simpler power-law approximation can be obtained by choosing this typical value for the logarithmic term ${\Lambda_{lc}\simeq2}$. Then, ${\Gamma\simeq\Gamma_{\star}Q^{b}}$ (see above), where the normalization ${\Gamma_{\star}=\Gamma_{\star}^\Lambda(\Lambda_{lc}=2)}$ is not a function of $Q$. Figure \[f:Rprc\] shows the TDE rates for ${\mu_{h}=2}$, ${\eta_{h}=1}$ in the two dynamical regimes ${a_{c} < \rh}$ and ${a_{c} > \rh}$ (${Q < Q_{c}}$ and ${Q > Q_{c}}$). At the lower mass end, the mean mass accretion rate may rise above the Eddington rate. For example, for ${\beta=5}$, ${L_{TD}/L_{E}\lesssim6}$ at ${\Mbh=10^{3}\Mo}$, but falls below ${L_{TD}/L_{E}=1}$ for ${\Mbh>5\times 10^{3}\Mo}$. Depending on the exact value of the logarithmic slope of the $\Msig$ relation, $\beta$, the TDE rate on the low-mass branch, can either rise or fall with $\Mbh$. The numeric results that support the plots within this paper and other findings of this study are available from the corresponding author upon reasonable request. We are grateful for helpful discussions with Y. Alexander, J. Gair, A. Gal-Yam, J. Green, J. Guillochon, M. MacLeod, N. Neumayer, T. Piran, E. Rossi, A. Sesana, J. Silk, N. Stone and B. Trakhtenbrot. TA acknowledges support by the I-CORE Program of the Planning and Budgeting Committee and The Israel Science Foundation (grant No 1829/12). BB acknowledges support by NASA (grant NNX14AM24G) and by the NSF (grant AST-1406166). TA and BB developed the ideas presented in this paper together and collaborated in its writing. Correspondence and requests for materials should be addressed to Tal Alexander ([email protected]) or Ben Bar-Or ([email protected]). The authors declare that they have no competing financial interests. [^1]: Department of Particle Physics & Astrophysics, Weizmann Institute of Science, Rehovot 76100, Israel. [^2]: Institute for Advanced Study, Einstein Drive, Princeton, NJ 08540, USA

ArXiv

--- abstract: 'We show that the relational theory of intersection types known as BCD has the finite model property; that is, BCD is complete for its finite models. Our proof uses rewriting techniques which have as an immediate by-product the polynomial time decidability of the preorder $\subseteq $ (although this also follows from the so called beta soundness of BCD).' author: - Rick Statman title: A Finite Model Property for Intersection Types --- Barendregt, Coppo, and Dezani ============================= BCD is the relational theory of intersection types presented by Henk Barendregt, Mario Coppo, and Mariangiola Dezani in [@2]. Here we consider the theory, without top element, as about a preorder $\subseteq$, $a \subseteq a$ $a \subseteq b\ \& \ \subseteq c {\Rightarrow}a \subseteq c$ $a \wedge b \subseteq a$ $a \wedge b \subseteq b$ $c \subseteq a\ \&\ c \subseteq b {\Rightarrow}c \subseteq a \wedge b,$ and a contravariant-covariant operation ${\rightarrow}$, $c \subseteq a \ \&\ b \subseteq d {\Rightarrow}a {\rightarrow}b \subseteq c {\rightarrow}d$ satisfying the weak distributive law $(c{\rightarrow}a) \wedge (c{\rightarrow}b)\ \subseteq c {\rightarrow}(a \wedge b).$ Of course it is well known that if the points of such a preorder are partitioned by the congruence $\sim$ defined by $ a \sim b {\Leftrightarrow}a \subseteq b \ \& \ b \subseteq a $ we obtain a semilattice with $\wedge$, that is, $\begin{array}{rcl} a \wedge (b \wedge c) & \sim & (a \wedge b) \wedge c \\ a \wedge b & \sim & b \wedge a \\ a & \sim & a \wedge a \end{array}$ where the quotient partial order can be recovered $a \subseteq b\ {\Leftrightarrow}\ a \sim a \wedge b.$ In addition, the quotient satisfies the distributive law $c {\rightarrow}(a \wedge b) \sim (c {\rightarrow}a) \wedge (c {\rightarrow}b)$ and an absorption law $a {\rightarrow}b \sim (a {\rightarrow}b) \wedge ((a \wedge c) {\rightarrow}b).$ Now if a semilattice is given and $a \subseteq b$ is defined by $$a \subseteq b {\Leftrightarrow}a = a \wedge b$$ then $\subseteq$ is a preorder with a meet operation. In addition, if the distributive law and the absorption law are satisfied then the ${\rightarrow}$ operation enjoys the contravariant-covariant property. There is also a derived absorption law $c {\rightarrow}(a \wedge b) = (c{\rightarrow}a) \wedge (c {\rightarrow}(a\wedge b))$ which proves useful. In this way we have an equational presentation of BCD. Expressions and their rewriting =============================== We define the notion of an expression as follows. @,$p,q,r,\ldots$ are atomic expressions. If $A$ and $B$ are expressions then so are $(A {\rightarrow}B)$ and $(A \wedge B)$. Even though we write infix notation we say that these expressions begin with ${\rightarrow}$ and $\wedge$ respectively. The notions of positive, negative, and strictly positive are defined recursively by $A$ is positive and strictly positive in $A$. If $C$ is positive in $B$ then $C$ is positive in $A{\rightarrow}B$ and negative in $B {\rightarrow}A$. If $C$ is strictly positive in $B$ then $C$ is strictly positive in $A{\rightarrow}B$. If $C$ is positive in $A$ or $B$ then $C$ is positive in $A \wedge B$. If $C$ is strictly positive in $A$ or $B$ then $C$ is strictly positive in $A \wedge B$. If $C$ is negative in $B$ then $C$ is negative in $A {\rightarrow}B$ and positive in $B {\rightarrow}A$. If $C$ is negative in $A$ or $B$ then $C$ is negative in $A \wedge B$. A single occurrence of $B$ as a subexpression of $A$ will be indicated $A[B]$. An expression can be thought of as a rooted oriented binary tree with atoms at its leaves and either ${\rightarrow}$ or $\wedge$ at each internal vertex. For each subexpression $B$ of $A$ there is a unique path from the root of $A$ to the root of $B$. The ebb of $B$ in $A$ is the number of ${\rightarrow}$ verticies on the path from the root of $A$ to the root of $B$; so $C {\rightarrow}D$ has ebb $=1 $ in $C {\rightarrow}D$. The ${\rightarrow}$ depth of $A$ is the maximum ebb of a subexpression of $A$. With an equational presentation we can associate a set of rewrite rules. The one step rewrite of an expression $A$ by the rule $R$ to the expressiion $B$ is denoted $A$ $R$ $B$. This is the replacement of exactly one occurrence of the left hand side of the rule as a subexpression of $A$, the redex, by the right hand side. Sets of rules can be combined by the regular operations $+$ (union) and $*$ (reflexive-transitive closure). Now fix $n$ to be a natural number or infinity $= o$(mega). We define rewrites --------- ------------------------------- ------- ----------------------------------------------------------- (asso.) $A \wedge (B \wedge C)$ asso. $(A \wedge B) \wedge C$ (asso.) $(A \wedge B) \wedge C$ asso. $A \wedge (B \wedge C)$ (comm.) $A \wedge B$ comm. $B \wedge A$ (idem.) $A$ idem. $A \wedge A$ (absp.) $A {\rightarrow}B$ absp. $(A {\rightarrow}B) \wedge ((A \wedge C) {\rightarrow}B)$ (dist.) $A {\rightarrow}(B \wedge C)$ dist. $(A {\rightarrow}B) \wedge (A {\rightarrow}C$ (dept.) $A[B]$ dept. $A [@]$ if $B$ lies at ebb $> n$ in $A[B]$ --------- ------------------------------- ------- ----------------------------------------------------------- and we set semi. $=$ asso. $+$ comm., and slat. $=$ semi. $+$ idem. Let redn. $=$ slat. $+$ absp. $+$ dist. $+$ dept. . Of course, when $n =$ infinity dept. is trivial and redo. generates the congruence on expressions induced by BCD. Given a reduction $A(1)$ redn. $\ldots$ redn. $A(k)$ an occurrence of ${\rightarrow}$ in $A(k)$ has a unique ancestor in each $A(i)$ except when $A(i)$ absp. $A(i+1)$ and the $A(i+1)$ ancestor of ${\rightarrow}$ lies in $C$. Similarly for atoms. Properties of the rewriting system ================================== 1. idem. can be restricted to atoms. $A \wedge B$ idem. $(A \wedge A) \wedge B$ idem. $(A\wedge A) \wedge (B \wedge B)$ asso.\* $ A \wedge ((A \wedge B) \wedge B)$ comm. $A \wedge ((B\wedge A) \wedge B)$ asso.\* $(A \wedge B) \wedge (A \wedge B).\ A {\rightarrow}B$ idem. $A {\rightarrow}(B \wedge B)$ dist. $(A {\rightarrow}B) \wedge (A {\rightarrow}B)$. End of proof. 2. comm. can be restricted to atoms and expressions beginning with ${\rightarrow}$. All permutations can be done by adjacent transpositions. End of proof. 3. dept. can be restricted to intersections of atoms and ${\rightarrow}$ of @’s. If $B$ lies at ebb $>n$ in $A$ then any longest ${\rightarrow}$ path in $B$ ends in an intersection of atoms. Indeed, since it is longest, it is either $C$ or $D$ in a subexpression $C {\rightarrow}D$ of $B$, where the other of $C$ and $D$ is similar. If such an intersection is non-trivial or $p$ it can be replaced by @. Similarly for the other. Otherwise, we have a subterm @ ${\rightarrow}$ @ of $B$ which dept. @. End of proof. From here on we assume that the restrictions in (1), (2), and (3) are obeyed in all reductions. 1. Every dist. reduction terminates. The ebb of $\wedge$’s decreases. End of proof. 2. Every dept. reduction terminates. Either length decreases or atoms change to @. End of proof. 3. idem. expedition. If $A$ slat.$^*\ B$ then there exists $C$ such that $$A \ {\rm idem.}^*\ C\ {\rm semi.}^*\ B$$ Each idem. redex has a unique ancestor in $A$ to which idem. can be applied. End of proof. 4. dept. postponement. If $A$ dept.$^*\ B\ R^*\ C$, where $R \in\ ${slat., dist., absp.} then there exists $D$ s.t. $$A\ R^*\ D\ {\rm dept.}^*\ C.$$ A dept. redex is either an intersection of atoms or @ ${\rightarrow}$ @. It has either one or two descendants in the result of any $R$ reduction these are also dept. redexes. End of proof. 5. dist. has the weak diamond property. 6. dept. has the weak diamond property. 7. Parallel moves lemma. This lemma has the form: if $A\ R^*\ B$ and $A\ S^*\ C$ then, for some $D,\ B\ S^*\ D$ and $C$ redn.$^*D$ for various $R,S \in $ {slat., absp., dist., dept.} so there are 16 possible cases. We denote these cases $R/S$. There are several exceptional cases; these are 10.9, 10.13, and 10.14. These cases must be accounted for separately so they fit together in a strip lemma argument for the Church-Rosser theorem. We begin with the special case $R$/idem. which is trivial. (10.1-4) slat./S -------------- ------------------------------------------------------------------- slat./slat.; by idem. expedition. slat./dist.; If $A$ idem.$^*\ B$ and $A$ dist.$^*\ C$ then there exist $D,E$ such that $B$ dist.$^*\ D,\ C$ idem.$^*\ E$ and $D$ semi.$^*\ E$. If $A$ semi.$^*\ B$ and $A$ dist.$^*\ C$ then there exists $D,E$ such that $C$ dist.$^*\ D,\ B$ dist.$^*\ E$ and $D$ semi.$^*\ E.$ slat./absp.; The strong diamond property holds for sets of non-overlapping redexes. slat./dept.; by idem. expedition. -------------- ------------------------------------------------------------------- (10.5-7) remaining R/R -------------- ---------------------------------------------------------------- dist./dist.; \(4) and (8) give us the strong diamond property. dept./dept.; \(5) and (9) give us the strong diamond property. absp./absp.; If $A$ absp.\* $B$ and $A$ absp.\* $C$ then there exists $D,E$ such that $B$ absp.\* $D$, $C$ absp.\* $E$ and $E$ semi.\* $D$ -------------- ---------------------------------------------------------------- (10.8-9) remaining R/dist. -------------- --------------------------------------------------------------------- absp./dist.; If $A$ absp.$^*\ B$ and $A$ dist.$^*\ C$ then there exists $D,E$ such that $B$ dist.$^*\ D$, $C$ absp.$^*\ E$ and $E$ semi.$^*\ D$ dept./dist.; If $A$ dept.$^*\ B$ and $A$ dist.$^*\ C$ then there exists $D,E$ such that $A$ idem.$^*\ D$ dept.$^*\ E$ and $B$ dept.$^*\ E$. This is an exceptional case. -------------- --------------------------------------------------------------------- (10.10-11) remaining R/dept. -------------- -------------------------------------------------------------------- dist./dept.; This is the same as 10.9 but here it is not exceptional. absp./dept.; If $A$ absp.\* $B$ and $A$ dept.\* $C$ then there exist $D,E$ such that $B$ dept.\* $E$, $C$ absp.\* $D$, and $D$ idem.\* $E$. -------------- -------------------------------------------------------------------- (10.12-13) remaining R/absp. -------------- ------------------------------------------------------------------------------------------------------------------------- dist./absp.; This is the same as 10.8 and is not exceptional since semi. is bidirectional. dept./absp.; There is one special case which is exceptional. @ ${\rightarrow}$ @ dept. @ and @ ${\rightarrow}$ @ absp. (@ ${\rightarrow}$) $\wedge$ ((@ $\wedge C$) ${\rightarrow}$ @) so @ idem. @ $\wedge$ @ and (@ ${\rightarrow}$) $\wedge$ ((@ $\wedge C$) ${\rightarrow}$ @) dept.$^*$ @ $\wedge$ @. So, in general $A$ dept.$^*\ B$ and $A$ absp.$^*\ C$ then there exist $D$ that $B$ (absp. $+$ idem.)$^*\ D,\ C$ dept.$^*\ D$. -------------- ------------------------------------------------------------------------------------------------------------------------- (10.14-16) remaining R/slat. -------------- ---------------------------------------------------------------------- dist./slat.; First consider the case $A$ dist. $B$ and $A$ semi. $C$. Then there exists $D$, $E$ such that $B$ dist.\* $D$, $C$ dist.\* $E$ and $D$ semi.\* $E$. Now use idem. expedition. This is an exceptional case. absp./slat.; As in 10.3. dept./slat.; As in 10.4. -------------- ---------------------------------------------------------------------- End of proof. We may divide redn. reductions into alternating segments slat.\*, absp.\*, dist.\*, and dept.\*. Such a reduction has the pointedness property if --------------- ------------------------------------------------------- (pointedness) Every dist.\* segment ends in a dist. normal form and every dept.\* segment ends in a dept. normal form --------------- ------------------------------------------------------- A reduction is said to be focused if --------- ------------------------------------------------------------ (focus) The reduction has the pointedness property and every segment either ends with a dist. and dept. normal form or is followed by a dist. segment and then a dept. segment or vice versa. --------- ------------------------------------------------------------ 8. Focus lemma. If $A$ redn.\* $B$ then there is a dist., dept. normal form $C$ of $B$ and a focused reduction from $A$ to $C$. Given a reduction from $A$ to $B$ repeatedly apply parallel moves R/dept. to the segments of the reduction where the dept.\* is to normal form. Now repeatedly apply parallel moves R/dist. to the segments of the reduction where the dist.\* is to normal form. In the exceptional case 10.9 we have an expression $X$ reduced on the one hand to dept. normal form $Y$ and on the other hand reduced to dist. normal form $Z$. Thus the dist. normal form $W$ of $Y$ has $W$ idem.\* $Z$. Now we continue the process with $W$. In the end all the extra idem.\*’s are pushed to the end. Note that this does not change the status of the final expression although an extra reduction to dist. normal form could be added anyway. End of proof. 9. Strip lemma. If $A$ redn.$^*\ B$ by a focused reduction and $A\ R^*\ C$ for $R \in$ {slat., dist., dept.} then there exists $D$ such that $C$ redn.$^*\ D$ and $B\ S^*\ D$ for $S \in$ {slat., dist., dept.}. In addition, if $A$ redn.$^*\ B$ by a focused reduction and $A$ idem.$^*$ absp.$^*\ C$ then there exists $D$ such that $C$ redn.$^*\ D$ and $B$ idem.$^*$ absp.\* $D$. We have divided redn.$^*$ into alternating segments slat.$^*$, absp.$^*$, dist.$^*$, and dept.$^*$. The proof is by induction on the number of such segments. Clearly it suffices to assume that if $R$ is dist. then $C$ is in dist. normal form and similarly for dept. The basis case is just the parallel moves lemma together with the observation that 1. If the case is 10.9 then the $R$ changes to slat. 2. If the case is 10.13 then the hypothesis considers the exception. 3. If the case is 10.14 then $B = D$ since $B$ is dist. normal and the case is not really exceptional. For the induction step we suppose that $A\ S^*\ B^{\prime}$ redn. $B$. We can apply the basis step to $A\ S^*\ B^{\prime}$, and we can apply the induction hypothesis to the reduction $B^{\prime}$ redn.$^*\ B$. Since the original reduction was pointed these compose to give the result. 10. Church-Rosser property. Let conv. be the congruence generated by redn. We need to show that if $A$ conv. $B$ then there exists $C$ such that $A$ redn.$^*\ C$ and $B$ redn.$^*\ C$. The proof is by induction on the length of a conversion from $A$ to $B$. With the strip and focus lemmas completing the proof is routine. End of proof. The models $F(n)$ ================= 1. Conservation lemma. If the ${\rightarrow}$ depth of $A$ is $< n+1$, $A$ redn.\* $B$, and $B$ is in dept. normal form then $A$ redo.\* $B$. By dept. postponement we may assume that the reduction $A$ to $B$ has all dept. reductions at the end and we have a $C'''$ such that $A$ redo.\* $C'''$ dept.\* $B$. Now the dept. redex of $C'''$ contracted next lies in a subexpression $C' {\rightarrow}C''$ of ebb at least $n+1$ in $C'''$. Since $A$ has ${\rightarrow}$ depth $< n+1$ the subexpression $C' {\rightarrow}C''$ has a unique ancestor which is a subexpression of the $C$ occurring on the right hand side of the absp. reduction rule applied to some redex in the reduction of $A$ to $C'''$. Now every descendant of this ancestor has ebb at least $n+1$ so the choice of $C$ can be modified to the result of replacing the ancestor subterm by @, without changing the dept. normal form. End of proof. Now the conv. congruence has an equational presentation and thus a free model $F(n)$ consisting of congruence classes of expressions. We adopt the customary notation $F(n)\ \models A = B$ to signify that $A$ and $B$ belong to the same congruence class of $F(n)$. The stack of 2’s function $s(n,m)$ is defined by $\begin{array}{rcl} s(0,m) & = & m \\ s(n+1,m) & = & 2^{s(n,m)} \end{array} $ 2. Finiteness lemma. If there are $m$ atoms and $n$ is finite then $F(n)$ has at most $s(n+1,m+n)$ elements. It suffices to over estimate the number of dept. normal forms. End of proof. 3. Completeness of the $F(n)$ If the ${\rightarrow}$ depths of $A$ and $B$ are both $< n+1$ then $F(n)\ \models A = B$ implies $A$ and $B$ are congruent in BCD. Suppose that $F(n)\ \models A = B$. Then $A$ conv. $B$ so by the Church-Rosser theorem there exists $C$ such that $A$ redn.$^*\ C$ and $B$ redn.$^*\ C$. By (5) we can assume $C$ is dept. normal. Thus, by the conservation lemma we have both $A$ redo.$^*\ C$ and $B$ redo.$^*\ C$ so $A$ and $B$ are congruent in BCD. End of proof. Polynomial time decidability of $\subseteq$ =========================================== First we remark that the beta soundness lemma ([@1]) is a simple consequence of the Church-Rosser theorem. 1. Weak standardization of redo. If $A$ redo.\* $B$ then there is a reduction from $A$ to $B$ where no strictly positive redex is contracted after one which is not strictly positive. Is straightforward. End of proof. If $A$ has no strictly positive dist. redex then $A$ is an intersection, under some association, of expressions of the form $A(1) {\rightarrow}(\ldots (A(t) {\rightarrow}p) \ldots )$ where $p$ is an atom (here we do not distinguish @). We call these expressions the factors of $A$. We define the set of factors of an expression $E$ more generally by recursion -------------------------------- --- ------------------------------------------------------- factors$(p)$ = { p} factors$(E'\wedge E'')$ = factors $(E')$  $\cup$ factors$(E'')$ factors$(E' {\rightarrow}E'')$ = $\{E' {\rightarrow}E'''\ |\ E'''$ : factors$(E'') \}$ -------------------------------- --- ------------------------------------------------------- 2. Complete invariants lemma. If there is a strictly positive reduction from $E'$ to $E''$ then each factor of $E'$ is a factor of $E''$ and for each factor $$E''(1) {\rightarrow}( \ldots (E''(t) {\rightarrow}p )\ldots )$$ of $E''$ there exists a factor $$E'(1) {\rightarrow}(\ldots (E'(t) {\rightarrow}p) \ldots )$$ of $E'$ and expressions $D(1) , \ldots , D(t)$ such that $E''(i)$ slat.\* $E'(i) \wedge D(i)$ for $i = 1, \ldots , t$. [*Remark.*]{} If we allow each $D(i)$ to be empty then slat. can be replaced by assoc. By inspection of the rewrite rules. End of proof. 3. Beta soundness ([@1], [@4] Lemma 2) If both $A$ and $B$ have no strictly positive dist. redexes and $A$ conv. $B$ then for each factor $A(1) {\rightarrow}(\ldots(A(t) {\rightarrow}p) \ldots )$ of $A$ there exists a factor $B(1) {\rightarrow}( \ldots (B(t) {\rightarrow}p) \ldots )$ of $B$ and expressions $C(1), \ldots , C(t)$ such that $A(i)$ conv. $B(i) \wedge C(i)$ for $i=1, \ldots , t$. By the Church-Rosser theorem there exists $C$ such that both $A$ and $B$ redo.\* $C$. By (4) we may assume that $C$ has no strictly positive dist. redex and by weak standardization there exist $A',B'$ such that $A$ redo.\* $A'$ by only strictly positive reductions, $B$ redo.\* $B'$ by only strictly positive reductions, $A'$ redo.\* $C$ with no strictly positive reductions, and $B'$ redo.\* $C$ with no strictly positive reductions. In particular, $A'$ and $B'$ have no strictly positive dist. redexes so their factors are actually subexpressions. By the complete invariants lemma with $E' = A$ and $E'' = A'$ for each factor $A(1) {\rightarrow}(\ldots (A(t) {\rightarrow}p) \ldots )$ of $A$ there exists a factor $B'(1) {\rightarrow}(\ldots (B'(t) {\rightarrow}p) \ldots )$ of $B'$ such that $A(i)$ conv. $B'(i)$ for $i = 1, \ldots , t$. Again by the complete invariants lemma with $E'' = B'$ and $E' = B$ each factor $B'(1) {\rightarrow}(\ldots (B'(t) {\rightarrow}p) \ldots ) $ of $B'$ there exists a factor $B(1) {\rightarrow}(\ldots (B(t) {\rightarrow}p) \ldots )$ of $B$ and expressions $C(1) , \ldots , C(t)$ such that $B(i)$ conv. $B'(i) \wedge C(i)$. End of proof. We conclude that $B \subseteq A {\Leftrightarrow}$ for each factor $A(1){\rightarrow}(\ldots (A(t) {\rightarrow}p) \ldots )$ of $A$ there exists a factor $B(1) {\rightarrow}(\ldots(B(t) {\rightarrow}p)\ldots )$ of $B$ such that $A(i) \subseteq B(i)$ for $i=1,\ldots , t$. Now we present a polynomial time algorithm for determining whether $A$ conv. $B$. A different algorithm is proposed in [@5]. Clearly, it suffices, given an expression $A$, to determine in polynomial time whether any two subexpressions are interconvertible. We suppose that the binary tree $A$ has $n$ nodes and these are numbered by depth first search so subexpressions of $A$ have lower numbers than their subexpressions. We construct an $n \times n$ Boolean matrix whose $(i,j)$ entry is $1$ if the $i$th node of $A\ \subseteq $ the $j$th node of $A$ and is $0$ otherwise. We shall fill in the $n \times n$ entries in time polynomial in $n$. We suppose that we wish to fill in the entry $(i,j)$ and that the entries filled in for all pairs $(k,l)$ with $k+l > i+j$. Let $B$ be the $i$th subexpression of $A$ and $C$ the $j$th. The factors of $B$ are in 1-1 correspondence with its strictly positive atoms; similarly for $C$. For each pair of factors $B(1) {\rightarrow}( \ldots (B(t) {\rightarrow}p)\ldots )$, $C(1) {\rightarrow}( \ldots (C(t) {\rightarrow}p) \ldots )$, we consider the entries for pairs of nodes corresponding to the pairs of expressions $C(k),B(k)$, for $k = 1, \ldots ,t$, already in the matrix. This takes time $O(n^3)$. If for each factor of $B$ the procedure succeeds for some factor of $C$ we enter a $1$ in $(i,j)$; otherwise, we enter a $0$. The entire algorithm runs in time $O(n^5)$. It is correct by beta soundness. [99]{} Allesi, F. and Lusin, S ., Simple easy terms, [*ITRS ’02*]{}, [**70**]{}, ENTCS (2002), [[doi: ](http://dx.doi.org/\detokenize{10.1016/S1571-0661(04)80487-0})]{}. Barendregt, H., Coppo, M., and Dezani-Ciancaglini, M., A filter lambda model and the completeness of type assignment [*J.S.L. 48*]{}, [**4**]{} (1983), pp 938-940, [[doi: ](http://dx.doi.org/\detokenize{10.2307/2273659})]{}. Barendregt, H., Dekkers, W., and Statman, R., “Lambda Calculus with Types”, Cambridge University Press, (2013), [[doi: ](http://dx.doi.org/\detokenize{10.1017/CBO9781139032636})]{}. D[ü]{}dder, B., Martens, M., Rehof, J., Urzyczyn, P., Bounded combinatory logic, CSL 2013, [[doi: ](http://dx.doi.org/\detokenize{10.4230/LIPIcs.CSL.2012.243})]{}. Rehof, J., Urczyczyn, P., Finite combinatory logic with intersection types, TLCA 2011, [[doi: ](http://dx.doi.org/\detokenize{10.1007/978-3-642-21691-6_15})]{}. Rehof, J., Urzyczyn, P., The complexity of inhabitation with explicit intersection, LNCS 7230, [[doi: ](http://dx.doi.org/\detokenize{10.1007/978-3-642-29485-3_16})]{}.

ArXiv

--- abstract: 'We introduce a new construction for the balancing of non-binary sequences that make use of Gray codes for prefix coding. Our construction provides full encoding and decoding of sequences, including the prefix. This construction is based on a generalization of Knuth’s parallel balancing approach, which can handle very long information sequences. However, the overall sequence—composed of the information sequence, together with the prefix—must be balanced. This is reminiscent of Knuth’s serial algorithm. The encoding of our construction does not make use of lookup tables, while the decoding process is simple and can be done in parallel.' author: - 'Elie N. Mambou and Theo G. Swart, [^1][^2] [^3]' title: 'A Construction for Balancing Non-Binary Sequences Based on Gray Code Prefixes' --- Balanced sequence, DC-free codes, Gray code prefix. Introduction {#sec:1} ============ use of balanced codes is crucial for some information transmission systems. Errors can occur in the process of storing data onto optical devices due to the low frequency of operation between structures of the servo and the data written on the disc. This can be avoided by using encoded balanced codes, as no low frequencies are observed. In such systems, balanced codes are also useful for tracking the data on the disc. Balanced codes are also used for countering cut-off at low frequencies in digital transmission through capacitive coupling or transformers. This cut-off is caused by multiple same-charge bits, and results in a DC level that charges the capacitor in the AC coupler [@immink2004]. In general, the suppression of low-frequency spectrum can be done with balanced codes. A large body of work on balanced codes is derived from the simple algorithm for balancing sequences proposed by Knuth [@knuth1986]. According to Knuth’s parallel algorithm, a binary sequence, ${\mbox{\boldmath $x$}}$, of even length $k$, can always be balanced by complementing its first or last $i$ bits, where $0 \leq i \leq k$. The index $i$ is then encoded as a balanced prefix that is appended to the data. The decoder can easily recover $i$ from the prefix, and then again complementing the first or last $i$ bits to obtain the original information. For Knuth’s serial (or sequential) algorithm, the prefix is used to provide information regarding the information sequence’s initial weight. Bits are sequentially complemented from one side of the overall sequence, until the information sequence and prefix together are balanced. Since the original weight is indicated by the prefix, the decoder simply has to sequentially complement the bits until this weight is attained. Al-Bassam [@bassam1990phd] presented a generalization of Knuth’s algorithm for binary codes, non-binary codes and semi-balanced codes (the latter occur where the number of 0’s and 1’s differs by at most a certain value in each sequence of the code). The balancing of binary codes with low DC level is based on DC-free coset codes. For the design of non-binary balanced codes, symbols in the information sequence are $q$-ary complemented from one side, but because this process does not guarantee balancing, an extra redundant symbol is added to enforce the balancing (similar to our approach later on). Information regarding how many symbols to complement is sent by using a balanced prefix. Capocelli *et al*. [@capocelli1991] proposed using two functions that must satisfy certain properties to encode any $q$-ary sequence into balanced sequences. The first function is similar to Knuth’s serial scheme: it outputs a prefix sequence depending on the original sequence’s weight. Additionally, all the $q$-ary sequences are partitioned into disjointed chains, where each chain’s sequences have unique weights. The second function is then used to select an alternate sequence in the chain containing the original information sequence, such that the chosen prefix and the alternate sequence together are balanced. Tallini and Vaccaro [@tallini1999] presented another construction for balanced $q$-ary sequences that makes use of balancing and compression. Sequences that are close to being balanced are encoded with a generalization of Knuth’s serial scheme. Based on the weight of the information sequence, a prefix is chosen. Symbols are then “complemented in stages”, one at a time, until the weight that balances the sequence and prefix together is attained. Other sequences are compressed with a uniquely decodable variable length code and balanced using the saved space. Swart and Weber [@swart2009] extended Knuth’s parallel balancing scheme to $q$-ary sequences with parallel decoding. However, this technique does not provide a prefix code implementation, with the assumption that small lookup tables can be used for this. Our approach aims to implement these prefixes via Gray codes. Swart and Weber’s scheme will be expanded on in Section \[sec:2.2\], as it also forms the basis of our proposed algorithm. Swart and Immink [@swart2013] described a prefixless algorithm for balancing of $q$-ary sequences. By using the scheme from [@swart2009] and applying precoding to a very specific error correction code, it was shown that balancing can be achieved without the need for a prefix. Pelusi *et al*. [@pelusi2015] presented a refined implementation of Knuth’s algorithm for parallel decoding of $q$-ary balanced codes, similar to [@swart2009]. This method significantly improved [@capocelli1991] and [@tallini1999] in terms of complexity. The rest of this paper is structured as follows. In Section \[sec:2\], we present the background for our work, which includes Swart and Weber’s balancing scheme for $q$-ary sequences [@swart2009] and non-binary Gray code theory [@guan1998]. In Section \[sec:3\], a construction is presented for sequences where $k=q^t$. Section \[sec:4\] extends on our proposed construction to sequences with $k\neq q^t$. Finally, Section \[sec:5\] deals with the redundancy and complexity of our construction compared to prior art constructions, our conclusions are presented in Section \[sec:6\]. Preliminaries {#sec:2} ============= Let ${\mbox{\boldmath $x$}}= (x_1 x_2 \dots x_k)$ be a $q$-ary information sequence of length $k$, where $x_i \in \{0,1,\dots q-1\}$ is from a non-binary alphabet. A prefix of length $r$ is appended to ${\mbox{\boldmath $x$}}$. The prefix and information together are denoted by ${\mbox{\boldmath $c$}}= (c_1 c_2 \dots c_n)$ of length $n=k+r$, where $c_i \in \{0,1,\dots q-1\}$. Let $w({\mbox{\boldmath $c$}})$ refer to the weight of ${\mbox{\boldmath $c$}}$, that is the algebraic sum of symbols in ${\mbox{\boldmath $c$}}$. The sequence ${\mbox{\boldmath $c$}}$ is said to be balanced if $$w({\mbox{\boldmath $c$}}) = \sum_{i=1}^n c_i = \frac{n(q-1)}{2}.$$ Let $\beta_{n,q}$ represent this value obtained at the balancing state. For the rest of the paper, the parameters $k$, $n$, $q$ and $r$ are chosen in such a way that the balancing value, $\beta_{n,q}=n(q-1)/2$, leads to a positive integer. Balancing of $q$-ary Sequences {#sec:2.2} ------------------------------ Any information sequence, ${\mbox{\boldmath $x$}}$ of length $k$ and alphabet size $q$, can always be balanced by adding (modulo $q$) to that sequence one sequence from a set of balancing sequences [@swart2009]. The balancing sequence, ${\mbox{\boldmath $b$}}_{s,p} = (b_1 b_2 \dots b_k)$ is derived as $$b_i = \begin{cases} s, & i > p, \\ s+1 \pmod q, & i \leq p, \end{cases}$$ where $s$ and $p$ are positive integers with $0 \leq s \leq q-1$ and $0 \leq p \leq k-1$. Let $z$ be the iterator through all possible balancing sequences, such that $z = sn + p$ and $0 \leq z \leq kq-1$. Let ${\mbox{\boldmath $y$}}$ refer to the resulting sequence when adding (modulo $q$) the balancing sequence to the information sequence, ${\mbox{\boldmath $y$}}= {\mbox{\boldmath $x$}}\oplus_q {\mbox{\boldmath $b$}}_{s,p}$, where $\oplus_q$ denotes modulo $q$ addition. The cardinality of balancing sequences equals $kq$ and amongst them, at least one leads to a balanced output ${\mbox{\boldmath $y$}}$. Since $s$ and $p$ can easily be determined for the $z$-th balancing sequence using $z = sn + p$, we will use the simplified notation ${\mbox{\boldmath $b$}}_z$ to denote ${\mbox{\boldmath $b$}}_{s,p}$.\ \[ex1\] Let us consider the balancing of the 3-ary sequence 2101, of length 4. The encoding process is illustrated below, with weights in bold indicating that the sequences are balanced. $$\begin{array}{c@{\quad\quad}c@{\;}c@{\;}c@{\;}c@{\;}c@{\quad\quad}c} z & {\mbox{\boldmath $x$}}& \oplus_q & {\mbox{\boldmath $b$}}_z & = & {\mbox{\boldmath $y$}}& w({\mbox{\boldmath $y$}}) \\ \hline 0 & (2101) & \oplus_3 & (0000) & = & (2101) & \mathbf{4} \\ 1 & (2101) & \oplus_3 & (1000) & = & (0101) & 2 \\ 2 & (2101) & \oplus_3 & (1100) & = & (0201) & 3 \\ 3 & (2101) & \oplus_3 & (1110) & = & (0211) & \mathbf{4} \\ 4 & (2101) & \oplus_3 & (1111) & = & (0212) & 5 \\ 5 & (2101) & \oplus_3 & (2111) & = & (1212) & 6 \\ 6 & (2101) & \oplus_3 & (2211) & = & (1012) & \mathbf{4} \\ 7 & (2101) & \oplus_3 & (2221) & = & (1022) & 5 \\ 8 & (2101) & \oplus_3 & (2222) & = & (1020) & 3 \\ 9 & (2101) & \oplus_3 & (0222) & = & (2020) & \mathbf{4} \\ 10 & (2101) & \oplus_3 & (0022) & = & (2120) & 5 \\ 11 & (2101) & \oplus_3 & (0002) & = & (2100) & 3 \\ \end{array}$$ For this example, there are four occurrences of balanced sequences. A $(\gamma, \tau)$-random walk refers to a path with random increases of $\gamma$ and decreases of $\tau$. In our case, a random walk graph is the plot of the function of $w({\mbox{\boldmath $y$}})$ versus $z$. In general, the random walk graph of $w({\mbox{\boldmath $y$}})$ always forms a $(1, q-1)$-random walk [@swart2009]. Fig. \[fig:ex1\] presents the $(1,2)$-random walk for Example \[ex1\]. The dashed line indicates the balancing value $\beta_{4,3} = 4$. ![Random walk graph of $w({\mbox{\boldmath $y$}})$ for Example \[ex1\][]{data-label="fig:ex1"}](example1) This method, as presented in [@swart2009], assumed that the $z$ indices can be sent using balanced prefixes, but the actual encoding of these was not taken into account. For instance, in Example \[ex1\] indices $z=0$, $3$, $6$ and $9$ must be encoded into balanced prefixes, in order to send overall balanced sequences. Non-binary Gray Codes {#sec:2.3} --------------------- Binary Gray codes were first proposed by Gray [@gray1953] for solving problems in pulse code communication, and have been extended to various other applications. The assumption throughout this paper is that a Gray code is mapped from a set of possible sequences appearing in the normal lexicographical order. This ordering results in the main property of binary Gray codes: two adjacent codewords differ in only one bit. The $(r',q)$-Gray code is a set of $q$-ary sequences of length $r'$ such that any two adjacent codewords differ in only one symbol position. This set is not unique, as any permutation of a symbol column within the code could also generate a new $(r',q)$-Gray code. In this work, a unique set of $(r',q)$-Gray codes is considered, as presented by Guan [@guan1998]. This set possesses an additional property: the difference between any two consecutive sequences’ weights is $\pm1$. This same set of Gray codes was already determined in [@er1984] through a recursive method. Let ${\mbox{\boldmath $d$}}= (d_1 d_2 \ldots d_{r'})$ be any sequence within the set of all $q$-ary sequences of length $r'$, listed in the normal lexicographic order. These sequences are mapped to $(r', q)$-Gray code sequences, ${\mbox{\boldmath $g$}}= (g_1 g_2 \ldots g_{r'})$, such that any two consecutive sequences are different in only one symbol position. Table \[gray\] shows a $(3,3)$-Gray code, where ${\mbox{\boldmath $d$}}$ is the 3-ary representation of the index $z \in \{0,1,\ldots,26\}$ and ${\mbox{\boldmath $g$}}$ is the corresponding Gray code sequence. We see that for ${\mbox{\boldmath $g$}}$, the adjacent sequences’ weights differ by $+1$ or $-1$. $z$ ${\mbox{\boldmath $d$}}$ ${\mbox{\boldmath $g$}}$ $z$ ${\mbox{\boldmath $d$}}$ ${\mbox{\boldmath $g$}}$ $z$ ${\mbox{\boldmath $d$}}$ ${\mbox{\boldmath $g$}}$ ----- -------------------------- -------------------------- ----- -------------------------- -------------------------- ----- -------------------------- -------------------------- 0 $(000)$ $(000)$ 9 $(100)$ $(122)$ 18 $(200)$ $(200)$ 1 $(001)$ $(001)$ 10 $(101)$ $(121)$ 19 $(201)$ $(201)$ 2 $(002)$ $(002)$ 11 $(102)$ $(120)$ 20 $(202)$ $(202)$ 3 $(010)$ $(012)$ 12 $(110)$ $(110)$ 21 $(210)$ $(212)$ 4 $(011)$ $(011)$ 13 $(111)$ $(111)$ 22 $(211)$ $(211)$ 5 $(012)$ $(010)$ 14 $(112)$ $(112)$ 23 $(212)$ $(210)$ 6 $(020)$ $(020)$ 15 $(120)$ $(102)$ 24 $(220)$ $(220)$ 7 $(021)$ $(021)$ 16 $(121)$ $(101)$ 25 $(221)$ $(221)$ 8 $(022)$ $(022)$ 17 $(122)$ $(100)$ 26 $(222)$ $(222)$ : Example of $(3,3)$-Gray code[]{data-label="gray"} We will make use of the following encoding and decoding algorithms from [@guan1998]. ### Encoding algorithm for $(r',q)$-Gray code Let ${\mbox{\boldmath $d$}}= (d_1 d_2 \ldots d_{r'})$ and ${\mbox{\boldmath $g$}}= (g_1 g_2 \ldots g_{r'})$ denote respectively a $q$-ary sequence of length $r'$ and its corresponding Gray code sequence. Let $S_i$ be the sum of the first $i-1$ symbols of ${\mbox{\boldmath $g$}}$, with $2\leq i \leq r'$ and $g_1=d_1$. Then $$S_i=\sum_{j=1}^{i-1}g_j, \quad\text{and}\quad g_i = \begin{cases} d_i, & \text{if } S_i \text{ is even}, \\ q-1-d_i, & \text{if } S_i \text{ is odd}. \end{cases}$$ The parity of $S_i$ determines ${\mbox{\boldmath $g$}}$’s symbols from ${\mbox{\boldmath $d$}}$. If $S_i$ is even then the symbol stays the same, otherwise the $q$-ary complement of the symbol is taken. ### Decoding algorithm for $(r',q)$-Gray code Let ${\mbox{\boldmath $g$}}$, ${\mbox{\boldmath $d$}}$ and $S_i$ be defined as before, with $2\leq i \leq r'$ and $d_1=g_1$. Then $$S_i=\sum_{j=1}^{i-1}g_j, \quad\text{and}\quad d_i = \begin{cases} g_i, & \text{if } S_i \text{ is even}, \\ q-1-g_i, & \text{if } S_i \text{ is odd}. \end{cases}$$ Construction for $k=q^t$ {#sec:3} ======================== For the sake of simplicity, we will briefly explain the construction for information lengths limited to $k=q^t$, with $t$ being a positive integer. More details can be found in our conference paper [@mambou2016]. In the next section we will show how this restriction can be avoided. The main component of this technique is to encode the balancing indices, $z$, into Gray code prefixes that can easily be encoded and decoded. The prefix together with the information sequence must be balanced. The condition, $k=q^t$, is enforced so that the cardinality of the $(r',q)$-Gray code is equal to that of the balancing sequences, making $r' = \log_q(kq) = \log_q(q^{t+1}) = t+1$. ### Encoding Let ${\mbox{\boldmath $c$}}' = ({\mbox{\boldmath $g$}}|{\mbox{\boldmath $y$}}) = (g_1 g_2 \ldots g_{r'} y_1 y_2 \ldots y_k)$ be the concatenation of the Gray code prefix with ${\mbox{\boldmath $y$}}$, with $|$ representing the concatenation. As stated earlier, for the sequences ${\mbox{\boldmath $y$}}$ we obtain a $(1,q-1)$-random walk, and for the Gray codes ${\mbox{\boldmath $g$}}$ we have a $(1,1)$-random walk. Therefore, when we concatenate the two sequences together, the random walk graph of ${\mbox{\boldmath $c$}}'$ forms a $(\{0;2\}, \{q-2;q\})$-random walk, i.e. increases of 0 or 2 and decreases of $q-2$ or $q$. This concatenation of a Gray code prefix, ${\mbox{\boldmath $g$}}$, with an output sequence, ${\mbox{\boldmath $y$}}$, does not guarantee the balancing of the overall sequence, since the increases of 2 in the random walk graph do not guarantee that it will pass through a specific point. An extra symbol $u$ is added to ensure overall balancing, with $u = \beta_{n,q} - w({\mbox{\boldmath $c$}}')$ if $0 \leq u \leq q-1$, otherwise $u=0$, thus forcing the random graph to a specific point. The overall sequence is the concatenation of $u$, ${\mbox{\boldmath $g$}}$ and ${\mbox{\boldmath $y$}}$, i.e. ${\mbox{\boldmath $c$}}= (u|{\mbox{\boldmath $g$}}|{\mbox{\boldmath $y$}}) = (u g_1 g_2 \ldots g_{r'} y_1 y_2 \ldots y_k)$. The length of ${\mbox{\boldmath $c$}}$ is $n=k+r'+1$. In summary, the balancing of any $q$-ary sequence of length $k$, where $k=q^t$, can be achieved by adding (modulo $q$) an appropriate balancing sequence, ${\mbox{\boldmath $b$}}_z$, and prefixing a redundant symbol $u$ with a Gray code sequence, ${\mbox{\boldmath $g$}}$. The construction relies on finding a Gray code prefix to describe $z$, and at the same time be balanced together with ${\mbox{\boldmath $y$}}$.\ \[ex2\] Let us consider the encoding of the ternary sequence, 201 of length 3. Since $t=1$, the length of Gray code prefixes will be $r'=2$. The overall length is $n=6$ and the balancing value is $\beta_{6,3}=6$. The encoding process below is followed. $$\begin{array}{c@{\quad\quad}c@{\;}c@{\;}c@{\;}c@{\;}c@{\quad\quad}c@{\quad\quad}c} z & {\mbox{\boldmath $x$}}& \oplus_q & {\mbox{\boldmath $b$}}_z & = & {\mbox{\boldmath $y$}}& {\mbox{\boldmath $c$}}& w({\mbox{\boldmath $c$}}) \\ \hline 0 & (201) & \oplus_3 & (000) & = & (201) & (\underline{\mathbf{0}00} 201) & 3 \\ 1 & (201) & \oplus_3 & (100) & = & (001) & (\underline{\mathbf{0}01} 001) & 2 \\ 2 & (201) & \oplus_3 & (110) & = & (011) & (\underline{\mathbf{2}02} 011) & \mathbf{6} \\ 3 & (201) & \oplus_3 & (111) & = & (012) & (\underline{\mathbf{0}12} 012) & \mathbf{6} \\ 4 & (201) & \oplus_3 & (211) & = & (112) & (\underline{\mathbf{0}11} 112) & \mathbf{6} \\ 5 & (201) & \oplus_3 & (221) & = & (122) & (\underline{\mathbf{0}10} 122) & \mathbf{6} \\ 6 & (201) & \oplus_3 & (222) & = & (120) & (\underline{\mathbf{1}20} 120) & \mathbf{6} \\ 7 & (201) & \oplus_3 & (022) & = & (220) & (\underline{\mathbf{0}21} 220) & 7 \\ 8 & (201) & \oplus_3 & (002) & = & (220) & (\underline{\mathbf{1}21} 200) & \mathbf{6} \\ \end{array}$$ The underlined symbols represent the appended prefix, the bold underlined symbol is $u$, which is chosen such that $\beta_{6,3}$ is obtained whenever possible, and the bold weights indicate that balancing was achieved. Fig. \[fig:ex2\] presents the random walk graph for the weight of the overall sequence, ${\mbox{\boldmath $c$}}$, with the shaded area indicating the possible weights as a result of the flexibility in choosing $u$. ![Random walk graph of $w({\mbox{\boldmath $c$}})$ for Example \[ex2\][]{data-label="fig:ex2"}](example2) ### Decoding The decoding consists of recovering the index $z$ from the Gray code prefix, ${\mbox{\boldmath $g$}}$, and finding $s$ and $p$ to reconstruct ${\mbox{\boldmath $b$}}_z$. The original sequence is then obtained as ${\mbox{\boldmath $x$}}= {\mbox{\boldmath $y$}}\ominus_q {\mbox{\boldmath $b$}}_z$, where $\ominus_q$ represents modulo $q$ subtraction. As an example, Table \[tab:a1\] shows the decoding of every Gray code sequence into balancing sequences using the $(2,3)$-Gray code set. Gray code (${\mbox{\boldmath $g$}}$) Sequence (${\mbox{\boldmath $d$}}$) $z$ $s,p$ ${\mbox{\boldmath $b$}}_z$ -------------------------------------- ------------------------------------- ----- ------- ---------------------------- $(00)$ $(00)$ 0 $0,0$ $(000)$ $(01)$ $(01)$ 1 $0,1$ $(100)$ $(02)$ $(02)$ 2 $0,2$ $(110)$ $(12)$ $(10)$ 3 $1,0$ $(111)$ $(11)$ $(11)$ 4 $1,1$ $(211)$ $(10)$ $(12)$ 5 $1,2$ $(221)$ $(20)$ $(20)$ 6 $2,0$ $(222)$ $(21)$ $(21)$ 7 $2,1$ $(022)$ $(22)$ $(22)$ 8 $2,2$ $(002)$ : Decoding of $(2,3)$-Gray codes for $3$-ary sequences of length 2[]{data-label="tab:a1"} \ \[ex3\] Consider the received ternary sequence ${\mbox{\boldmath $c$}}= (012012)$ of length $n=6$ (one of the balanced sequences from Example \[ex2\]). The $(2,3)$-Gray code prefixes were used in encoding the original sequence. The first symbol in ${\mbox{\boldmath $c$}}$, $u=0$ is dropped, then the Gray code prefix is ${\mbox{\boldmath $g$}}=(12)$. This Gray code corresponds to ${\mbox{\boldmath $d$}}=(10)$ as presented in Table \[tab:a1\]. This implies that $z=3$, leading to $s=1$, $p=0$ and therefore ${\mbox{\boldmath $b$}}_3 = (111)$. The original sequence is recovered as $${\mbox{\boldmath $x$}}= {\mbox{\boldmath $y$}}\ominus_q {\mbox{\boldmath $b$}}_z = (012) \ominus_3 (111) = (201).$$ Thus, the information sequence from Example \[ex2\] is recovered. Construction for $k \neq q^t$ {#sec:4} ============================= We will now generalize the technique described in the previous section to sequences of any length, i.e. $k \neq q^t$. The idea is to use a subset of the $(r',q)$-Gray code with an appropriate length to encode the $z$ indices that represent the $kq$ balancing sequences. Therefore, the cardinality of $(r',q)$-Gray code prefixes must be greater than that of the balancing sequences, i.e. $q^{r'} > kq$ or $r' > \log_q k + 1$. However, the challenge is to find the appropriate subset of $(r',q)$-Gray code prefixes that can uniquely match the $kq$ balancing sequences, and still guarantee balancing when combined with $u$ and ${\mbox{\boldmath $y$}}$. $(r',q)$-Gray code prefixes for $q$ odd {#sec:subset-q-odd} --------------------------------------- When examining the random walk graph for Gray codes with $q$ odd, one notices that the random walk forms an odd function around a specific point. Fig. \[fig:graycode1\] presents the $(4,3)$-Gray code random walk graph, with $G$ being the intersection point between the horizontal line, $w({\mbox{\boldmath $g$}})=4$, and the vertical line, $z = 40$. The graph forms an odd function around this point $G$. In general, for $(r',q)$-Gray codes where $q$ is odd, the random walk of the Gray codes gives an odd function centered around $w({\mbox{\boldmath $g$}})=\beta_{r',q}$ and $z = \lfloor \frac{q^{r'}}{2} \rfloor$, where $\lfloor \cdot \rfloor$ represents the floor function. ![$(4,3)$-Gray code random walk graph[]{data-label="fig:graycode1"}](graycode1){width="1\linewidth"} \ \[lem:1\] The random walk graph of $(r',q)$-Gray codes where $q$ is odd forms an odd function around the point $G$. It was proved in [@er1984] that any $(r',q)$-Gray code, where $q$ is odd, is reflected. That is, the random walk graph of the $(r',q)$-Gray code forms an odd function centered around the point $G$. This implies that any subset of an $(r',q)$-Gray code around the center of its random walk graph, where the information sequence is such that $kq$ is odd (i.e. $k$ is odd), always has an average weight equal to $\beta_{r',q}$. As we need a unique subset of Gray code sequences for any case, we choose $kq$ elements from the “middle” values of $z \in [0,q^{r'}-1]$ and call it the $z$-centered subset. The index for this subset is denoted by $z'$, with $z' \in [z_1, z_2]$. When $kq$ is even (i.e. $k$ is even), it is not guaranteed that the subset of $(r',q)$-Gray codes’ average weight around the center equals exactly $\beta_{r',q}$. However, it will be very close to it, with a rounded value that is equal to $\beta_{r',q}$. We formalize these observations in the subsequent lemma. Let $\mathcal{G}$ denote the subset of $kq$ Gray code sequences that are used to encode the index $z'$, let $\overline{w}(\cdot)$ denote the average weight of a set of sequences and let $\lVert\cdot\rVert$ denote rounding to the nearest integer.\ \[lem:2\] For an $(r',q)$-Gray code subset, $\mathcal{G}$, where $q$ is odd and the $z'$-th codewords are chosen with $z' \in [z_1, z_2]$, the following holds: - if $k$ is odd with $z_1 = \lfloor \frac{q^{r'}}{2} \rfloor - \lfloor \frac{kq}{2} \rfloor$ and $z_2 = \lfloor \frac{q^{r'}}{2} \rfloor + \lfloor \frac{kq}{2} \rfloor$, then $\overline{w}(\mathcal{G}) = \beta_{r',q}$, - if $k$ is even with $z_1 = \lfloor \frac{q^{r'}}{2} \rfloor - \frac{kq}{2}$ and $z_2 = \lfloor \frac{q^{r'}}{2} \rfloor + \frac{kq}{2}-1$, then $\lVert\overline{w}(\mathcal{G})\rVert = \beta_{r',q}$. To simplify notation in this proof, we simply use $\beta$ to represent $\beta_{r',q}$ throughout. If $k$ is odd, it follows directly from Lemma \[lem:1\] that choosing $kq$ sequences (where $kq$ is odd) from $z=\lfloor \frac{q^{r'}}{2} \rfloor - \lfloor \frac{kq}{2} \rfloor$ to $z=\lfloor \frac{q^{r'}}{2} \rfloor + \lfloor \frac{kq}{2} \rfloor$, centered around $z=\lfloor \frac{q^{r'}}{2} \rfloor$, will result in $\overline{w}(\mathcal{G}) = \beta$, since the random walk forms an odd function around this point. In cases where $k$ is even, if $z_2$ was chosen as $\lfloor \frac{q^{r'}}{2} \rfloor + \frac{kq}{2}$, we would have exactly $\overline{w}(\mathcal{G}) = \beta$ (using the same reasoning as for the case where $k$ is odd), as we use $\frac{kq}{2}$ elements to the left of $\lfloor \frac{q^{r'}}{2} \rfloor$ and $\frac{kq}{2}$ elements to the right of it. However, this would mean that $kq+1$ elements are being used. Thus, $z_2 = \lfloor \frac{q^{r'}}{2} \rfloor + \frac{kq}{2}-1$ must be used. Let $\alpha$ be the weight of the $(z_2+1)$-th Gray code, then $$\overline{w}(\mathcal{G}) = \frac{(kq+1)\beta - \alpha}{kq} = \beta + \frac{\beta-\alpha}{kq}.$$ The lowest possible value of $\alpha$ is $\alpha_{\min} = 0$, and its highest possible value is $\alpha_{\max} = k(q-1)$. Thus, $$\beta + \frac{\beta-\alpha_{\max}}{kq} \leq \overline{w}(\mathcal{G}) \leq \beta + \frac{\beta-\alpha_{\min}}{kq}$$ and with some manipulations it can be shown that $$\beta \left(1 - \frac{\frac{2k}{r'}-1}{kq}\right) \leq \overline{w}(\mathcal{G}) \leq \beta\left(1 + \frac{1}{kq}\right).$$ Finally, where $q$ is odd, we have $q \geq 3$, and rounding to the nearest integer results in $\lVert\overline{w}(\mathcal{G})\rVert = \beta$. $(r',q)$-Gray code prefixes for $q$ even {#sec:subset-q-even} ---------------------------------------- For the encoding of sequences that make use of $(r',q)$-Gray code prefixes where $q$ is even, a different approach is followed. The subset of Gray code prefixes is obtained by placing a sliding window of length $kq$ over the random walk graph of the $(r',q)$-Gray code sequences, and shifting it until we obtain a subset with an average weight value of $\beta_{r',q}$. Fig. \[fig:graycode2\] shows the $(6,2)$-Gray code random walk graph. However, this process does not always guarantee a subset of Gray code prefixes with an average weight value of exactly $\beta_{r',q}$. Since we have flexibility in choosing $u$, we can choose the average weight for the subset to be close to $\beta_{r',q}$, and adjust $u$ as necessary to obtain exact balancing. ![$(6,2)$-Gray code random walk graph[]{data-label="fig:graycode2"}](graycode2){width="1\linewidth"} \ \[lem:3\] An $(r', q)$-Gray code subset, $\mathcal{G}$, where $q$ is even, can be chosen such that $\lVert\overline{w}(\mathcal{G})\rVert = \beta_{r',q}$. A similar reasoning as in the proof of Lemma \[lem:2\], where a symbol with weight $\alpha$ is repeatedly removed from the set, can be used to find $\lVert\overline{w}(\mathcal{G})\rVert$. Encoding -------- Having presented all the required components, we now propose our encoding algorithm. The length of the required Gray code prefix is $$\label{eq1} r' = \lceil \log_qk \rceil + 1,$$ where $\lceil \cdot \rceil$ represents the ceiling function. The cardinality of $(r',q)$-Gray codes equals $q^{r'}$. This implies that $q^{r'-1} < kq < q^{r'}$. The encoding will make use of a subset of $kq$ Gray code sequences from the $q^{r'}$ available ones.\ Any $q$-ary sequence can be balanced by adding (modulo $q$) an appropriate balancing sequence, ${\mbox{\boldmath $b$}}_z$, and prefixing a redundant symbol, $u$, with a Gray code sequence, ${\mbox{\boldmath $g$}}$, taken from the subset of $(r',q)$-Gray code prefixes. Let $\mathcal{U}$ denote the set of possible symbols for $u$, i.e. $\mathcal{U} = \{0,1,\ldots,q-1\}$, let $\mathcal{G}$ denote the subset of Gray code sequences, and let $\mathcal{Y}$ denote the set of $kq$ output sequences after the $kq$ balancing sequences are added to the information sequence. It is easy to see that $$\overline{w}(\mathcal{U}) = \frac{(q-1)}{2}.$$ From Lemmas \[lem:2\] and \[lem:3\], the subset of $(r',q)$-Gray code prefixes that corresponds to the $kq$ balancing sequences is chosen such that $$\lVert\overline{w}(\mathcal{G})\rVert = \frac{r'(q-1)}{2} = \beta_{r',q}.$$ It was proved in [@swart2009] that the average weight of the $kq$ sequences, ${\mbox{\boldmath $y$}}= {\mbox{\boldmath $x$}}\oplus_q {\mbox{\boldmath $b$}}_z$, is such that $$\overline{w}(\mathcal{Y}) = \frac{k(q-1)}{2} = \beta_{k,q}.$$ By considering ${\mbox{\boldmath $c$}}= (u|{\mbox{\boldmath $g$}}|{\mbox{\boldmath $y$}})$, with length $n = k+r'+1$, as the overall sequence to be transmitted, it follows that: $$\begin{aligned} \overline{w}(\mathcal{U}) + \lVert\overline{w}(\mathcal{G})\rVert + \overline{w}(\mathcal{Y}) &= \frac{(q-1)}{2} + \frac{r'(q-1)}{2} + \frac{k(q-1)}{2} \\ &= \frac{(k+r'+1)(q-1)}{2} \\ &= \beta_{n,q}.\end{aligned}$$ This implies that there is at least one ${\mbox{\boldmath $c$}}$ for which $w({\mbox{\boldmath $c$}}) \leq \beta_{n,q}$ and at least one other ${\mbox{\boldmath $c$}}$ for which $w({\mbox{\boldmath $c$}}) \geq \beta_{n,q}$. Taking the random walk’s increases into account, as well as the flexibility in choosing $u$, we can conclude that there is at least one ${\mbox{\boldmath $c$}}$ such that $w({\mbox{\boldmath $c$}}) = \beta_{n,q}$. The encoding algorithm consists of the following steps: 1. Obtain the correct Gray code length $r'$ by using . Then find the corresponding subset of $(r',q)$-Gray code prefixes, $z' \in [z_1, z_2]$, using the methods discussed in Section \[sec:subset-q-odd\] where $q$ is odd and in Section \[sec:subset-q-even\] where $q$ is even. 2. Incrementing through $z$, determine the balancing sequences, ${\mbox{\boldmath $b$}}_z$, and add them to the information sequence ${\mbox{\boldmath $x$}}$ to obtain outputs ${\mbox{\boldmath $y$}}$. 3. For each increment of $z$, append every ${\mbox{\boldmath $y$}}$ with the corresponding Gray code prefix ${\mbox{\boldmath $g$}}$ following the lexicographic order, with ${\mbox{\boldmath $g$}}$ obtained from the $q$-ary representations of the $z'$ indices. 4. Finally, set $u = \beta_{n,q}-w({\mbox{\boldmath $y$}})-w({\mbox{\boldmath $g$}})$ if $u \in \{0, 1, \ldots, q-1\}$, otherwise set $u=0$. We illustrate the encoding algorithm with the following two examples, one for an odd value of $q$ and the other for an even value of $q$.\ \[ex4\] Consider encoding the ternary sequence, $(21120)$, of length 5. Since $r' = \lceil \log_3 5 \rceil + 1 = 3$, we require $(3,3)$-Gray code prefixes to encode the $z'$ indices. The overall sequence length is $n = k+r'+1 = 9$, and the balancing value is $\beta_{9,3} = 9$. The cardinality of the $(3,3)$-Gray code is 27 and the required $z$-centered subset of prefixes containing $kq=15$ elements is such that $z' \in [5,19]$. The following process shows the possible sequences obtained. Again the underlined symbols represent the appended prefix, the bold underlined symbol is $u$, and the bold weights indicate balancing. $$\begin{array}{c@{\;}ccc@{\;}c} z & z' & {\mbox{\boldmath $x$}}\oplus_q {\mbox{\boldmath $b$}}_z ={\mbox{\boldmath $y$}}& {\mbox{\boldmath $c$}}&w({\mbox{\boldmath $c$}})\\ \hline 0 & 5 & (21120) \oplus_3 (00000) = (21120) & (\underline{\mathbf{2}010}21120) & \mathbf{9} \\ 1 & 6 & (21120) \oplus_3 (10000) = (01120) & (\underline{\mathbf{0}020}01120) & 6 \\ 2 & 7 & (21120) \oplus_3 (11000) = (02120) & (\underline{\mathbf{1}021}02120) & \mathbf{9} \\ 3 & 8 & (21120) \oplus_3 (11100) = (02220) & (\underline{\mathbf{0}022}02220) & 10 \\ 4 & 9 & (21120) \oplus_3 (11110) = (02200) & (\underline{\mathbf{0}122}02200) & \mathbf{9} \\ 5 & 10 & (21120) \oplus_3 (11111) = (02201) & (\underline{\mathbf{0}121}02201) & \mathbf{9} \\ 6 & 11 & (21120) \oplus_3 (21111) = (12201) & (\underline{\mathbf{0}120}12201) & \mathbf{9} \\ 7 & 12 & (21120) \oplus_3 (22111) = (10201) & (\underline{\mathbf{0}110}10201) & 6 \\ 8 & 13 & (21120) \oplus_3 (22211) = (10001) & (\underline{\mathbf{0}111}10001) & 5 \\ 9 & 14 & (21120) \oplus_3 (22221) = (10011) & (\underline{\mathbf{2}112}10011) & \mathbf{9} \\ 10 & 15 & (21120) \oplus_3 (22222) = (10012) & (\underline{\mathbf{2}102}10012) & \mathbf{9} \\ 11 & 16 & (21120) \oplus_3 (02222) = (20012) & (\underline{\mathbf{2}101}20012) & \mathbf{9} \\ 12 & 17 & (21120) \oplus_3 (00222) = (21012) & (\underline{\mathbf{2}100}21012) & \mathbf{9} \\ 13 & 18 & (21120) \oplus_3 (00022) = (21112) & (\underline{\mathbf{0}200}21112) & \mathbf{9} \\ 14 & 19 & (21120) \oplus_3 (00002) = (21122) & (\underline{\mathbf{0}201}21122) & 11 \\ \end{array}$$ \[ex5\] Consider encoding the 4-ary sequence, (312), of length 3. As before, $r' = \lceil \log_4 3 \rceil + 1 = 2$, requiring $(2,4)$-Gray code prefixes to be used. The overall sequence length is $n = 6$, and the balancing value is $\beta_{6,4} = 9$. The cardinality of the $(2,4)$-Gray code equals 16. The $z'$-subset is found by employing a sliding window of length $kq = 12$ over the random walk graph of the $(2,4)$-Gray code prefixes, shown in Fig. \[fig:graycode3\]. A suitable subset is found where $z_1 = 1$ and $z_2 = 12$, with an average weight value of 3, which equals $\beta_{2,4} = 3$. ![$(2,4)$-Gray code random walk graph with chosen subset[]{data-label="fig:graycode3"}](graycode3){width="1\linewidth"} The encoding process for the 4-ary sequence is shown next. $$\begin{array}{c@{\quad}c@{\quad\quad}c@{\;}c@{\;}c@{\;}c@{\;}c@{\quad\quad}c@{\quad\quad}c} z & z' & {\mbox{\boldmath $x$}}& \oplus_q & {\mbox{\boldmath $b$}}_z & = & {\mbox{\boldmath $y$}}& {\mbox{\boldmath $c$}}& w({\mbox{\boldmath $c$}}) \\ \hline 0 & 1 & (312) & \oplus_4 & (000) & = & (312) & (\underline{\mathbf{2}01}312) & \mathbf{9} \\ 1 & 2 & (312) & \oplus_4 & (100) & = & (012) & (\underline{\mathbf{0}02}012) & 5 \\ 2 & 3 & (312) & \oplus_4 & (110) & = & (022) & (\underline{\mathbf{2}03}022) & \mathbf{9} \\ 3 & 4 & (312) & \oplus_4 & (111) & = & (023) & (\underline{\mathbf{0}13}023) & \mathbf{9} \\ 4 & 5 & (312) & \oplus_4 & (211) & = & (123) & (\underline{\mathbf{0}12}123) & \mathbf{9} \\ 5 & 6 & (312) & \oplus_4 & (221) & = & (133) & (\underline{\mathbf{0}11}133) & \mathbf{9} \\ 6 & 7 & (312) & \oplus_4 & (222) & = & (130) & (\underline{\mathbf{0}10}130) & 5 \\ 7 & 8 & (312) & \oplus_4 & (322) & = & (230) & (\underline{\mathbf{2}20}230) & \mathbf{9} \\ 8 & 9 & (312) & \oplus_4 & (332) & = & (200) & (\underline{\mathbf{0}21}200) & 5 \\ 9 & 10 & (312) & \oplus_4 & (333) & = & (201) & (\underline{\mathbf{2}22}201) & \mathbf{9} \\ 10 & 11 & (312) & \oplus_4 & (033) & = & (301) & (\underline{\mathbf{0}23}301) & \mathbf{9} \\ 11 & 12 & (312) & \oplus_4 & (003) & = & (311) & (\underline{\mathbf{0}33}311) & 11 \\ \end{array}$$ Decoding -------- Fig. \[fig:dec2\] presents the decoding process of our proposed scheme, for any $q$-ary information sequence. The decoding algorithm consists of the following steps: 1. The redundant symbol $u$ is dropped, then the following $r'$ symbols are extracted as the Gray code prefix, ${\mbox{\boldmath $g$}}$, converted to ${\mbox{\boldmath $d$}}$ and used to find $z'$. 2. From $z'$, the corresponding $z$ index is computed as $z=z'-z_1$. 3. $z$ is used to find the parameters $s$ and $p$, then ${\mbox{\boldmath $b$}}_z$ is derived. 4. Finally, the original sequence is recovered through ${\mbox{\boldmath $x$}}= {\mbox{\boldmath $y$}}\ominus_q {\mbox{\boldmath $b$}}_z$. ![Decoding process for any $q$-ary information sequence[]{data-label="fig:dec2"}](decoding) \[ex6\] Consider the decoding of the sequence, $(\underline{\mathbf{2}100}121200)$ (the underlined symbols are the prefix and the bold underlined symbol is $u$), where $n=10$ and $q=3$, that was encoded using $(3,3)$-Gray code prefixes. The first symbol $u=2$ is dropped, then the Gray code prefix is extracted as $(100)$, which corresponds to $z'=17$, and the $z'$-subset of $(3,3)$-Gray code prefixes is $z' \in [4,21]$, thus $z=13$. This can be seen from Table \[tab:a2\], where the decoding of all $(3,3)$-Gray codes is shown. This implies that $s=2$ and $p=1$, resulting in ${\mbox{\boldmath $b$}}_{13}=(022222)$. Finally, the original information sequence is extracted as ${\mbox{\boldmath $x$}}= {\mbox{\boldmath $y$}}\ominus_q {\mbox{\boldmath $b$}}_z = (121200) \ominus_3 (022222) = (102011)$. Gray code (${\mbox{\boldmath $g$}}$) Sequence (${\mbox{\boldmath $d$}}$) $z'$ $z$ $s,p$ ${\mbox{\boldmath $b$}}_z$ -------------------------------------- ------------------------------------- ------ ----- ------- ---------------------------- $(000)$ $(000)$ 0 — — — $(001)$ $(001)$ 1 — — — $(002)$ $(002)$ 2 — — — $(012)$ $(010)$ 3 — — — $(011)$ $(011)$ 4 0 $0,0$ $(000000)$ $(010)$ $(012)$ 5 1 $0,1$ $(100000)$ $(020)$ $(020)$ 6 2 $0,2$ $(110000)$ $(021)$ $(021)$ 7 3 $0,3$ $(111000)$ $(022)$ $(022)$ 8 4 $0,4$ $(111100)$ $(122)$ $(100)$ 9 5 $0,5$ $(111110)$ $(121)$ $(101)$ 10 6 $1,0$ $(111111)$ $(120)$ $(102)$ 11 7 $1,1$ $(211111)$ $(110)$ $(110)$ 12 8 $1,2$ $(221111)$ $(111)$ $(111)$ 13 9 $1,3$ $(222111)$ $(112)$ $(112)$ 14 10 $1,4$ $(222211)$ $(102)$ $(120)$ 15 11 $1,5$ $(222221)$ $(101)$ $(121)$ 16 12 $2,0$ $(222222)$ $(100)$ $(122)$ 17 13 $2,1$ $(022222)$ $(200)$ $(200)$ 18 14 $2,2$ $(002222)$ $(201)$ $(201)$ 19 15 $2,3$ $(000222)$ $(202)$ $(202)$ 20 16 $2,4$ $(000022)$ $(212)$ $(210)$ 21 17 $2,5$ $(000002)$ $(211)$ $(211)$ 22 — — — $(210)$ $(212)$ 23 — — — $(220)$ $(220)$ 24 — — — $(221)$ $(221)$ 25 — — — $(222)$ $(222)$ 26 — — — : Decoding of $(3,3)$-Gray codes for ternary sequences with $k=6$ and $z' \in [4,21]$[]{data-label="tab:a2"} Redundancy and Complexity {#sec:5} ========================= In this section we compare the redundancy and complexity of our proposed scheme with some existing ones. Redundancy {#sec:5.1} ---------- Let $\mathcal{F}_q^k$ denote the cardinality of the full set of balanced $q$-ary sequences of length $k$. According to [@star1975], $$\mathcal{F}_q^k = q^{k} \sqrt{\frac{6}{\pi r(q^2-1)}} \left(1+\mathcal{O}\left(\frac{1}{k}\right)\right).$$ The information sequence length, $k$, in terms of the redundancy, $r$, for the construction in [@swart2009] is $$\label{eq_c1} k \leq \frac{\mathcal{F}_q^r}{q} \approx q^{r-1} \sqrt{\frac{6}{\pi r(q^2-1)}}.$$ In [@capocelli1991], two schemes are presented for $k$ information symbols, where one satisfies the bound $$\label{eq_c3} k \leq \frac{q^r-1}{q-1},$$ and the other one satisfies $$\label{eq_c2} k \leq 2\frac{q^r-1}{q-1}-r.$$ The construction in [@tallini1999] presents the information sequence length in terms of the redundancy as $$k \leq \frac{1}{1-2\gamma}\frac{q^r-1}{q-1}-a_1(q, \gamma)r-a_2(q, \gamma),$$ with $\gamma \in [0,\frac{1}{2})$, where $a_1$ and $a_2$ are scalars depending on $q$ and $\gamma$. If the compression aspect is ignored, the information sequence length is the same as in . The prefixless scheme presented in [@swart2013] has information sequence length that satisfies $$\label{eq_c4} k \leq q^{r-1} - r.$$ Two constructions with parallel decoding are presented in [@pelusi2015]. The first construction, where the prefixes are also balanced as in [@swart2009], has its information length as a function of $r$ as $$\label{eq_c5} k \leq \frac{\mathcal{F}_q^r - \{ q \bmod 2 + [(q-1)k] \bmod 2\}}{q-1}.$$ The second construction, where the prefixes need not be balanced, is a refinement of the first and has an information length the same as . As presented in Section \[sec:4\], the redundancy of our new construction is given by $r = \lceil \log_q k \rceil + 2$. Therefore, the information sequence length in terms of redundancy is $$k = q^{r-2}.$$ Fig. \[fig:redundancies\] presents a comparison of the information length, $k$, versus the redundancy, $r$, for various constructions as discussed above. For all $q$, our construction is only comparable to the information lengths from and , although it does slightly improve on both. However, the trade-off is that as the redundancy becomes greater, the complexity of our scheme tends to remain constant, as we see in the next section. ![Comparison of information sequence length vs. redundancy for various schemes[]{data-label="fig:redundancies"}](redundancies){width="1\linewidth"} Complexity {#sec:5.2} ---------- We estimate the complexity of our proposed scheme and compare it to that of existing algorithms. The techniques in [@capocelli1991] and [@tallini1999] both require $\mathcal{O}(qk\log_qk)$ digit operations for the encoding and decoding. The method from [@swart2009] takes $\mathcal{O}(qk\log_qk)$ digit operations for the encoding and $\mathcal{O}(1)$ digit operations for the decoding. A refined design of the parallel decoding method is presented in [@pelusi2015], where the complexity equals $\mathcal{O}(k\sqrt{\log_qk})$ in the encoding case and $\mathcal{O}(1)$ digit operations in the decoding process. The following pseudo code presents the steps of our encoding method: Input: Information sequence, x of length k. Output: Encoded sequence, y of length n=k+r. for i=0:kq; for j=z1:z2; y(i) = [u | g(j) | x + b(s,p)(i)]; If (w(y(i))==beta) // Testing for balanced sequence. exit(); //Terminate the program. end; end; In the above code, $i$ is the iterator through the $kq$ output sequences and also through the balancing sequences, while $j$ is the iterator through the subset of Gray code sequences, ranging from $z_1 = \lfloor \frac{q^{r'}}{2} \rfloor - \lfloor \frac{kq}{2} \rfloor$ to $z_2= z_1 + kq-1$. The symbol ‘$\mid$’ denotes the concatenation. Our encoding scheme is based on the construction in [@swart2009] that has an encoding complexity of $\mathcal{O}(qk\log_qk)$, and it takes $\mathcal{O}(\log_qk)$ to encode Gray code prefixes as presented in [@guan1998]. Therefore the encoding of our algorithm requires $\mathcal{O}(qk\log_qk)$ digit operations. The decoding process consists of very simple steps: the recovery of the index $z'$ from the Gray code requires $\mathcal{O}(\log_q k)$ digit operations [@guan1998]. After obtaining the index $z'$ from the Gray code prefix, the balancing sequence ${\mbox{\boldmath $b$}}_z$ is found and then the original information sequence is recovered through the operation, ${\mbox{\boldmath $y$}}= {\mbox{\boldmath $x$}}\ominus_q {\mbox{\boldmath $b$}}_z$, which can be performed in parallel, resulting in a complexity of $\mathcal{O}(1)$. Therefore the overall complexity for the decoding is $\mathcal{O}(\log_q k)$ digit operations. Table \[tab:cmpl\] summarizes the complexities for various constructions, where the orders of digit operations it takes to complete the encoding/decoding are compared. Algorithm Encoding order Decoding order ------------------ -------------------------------- -------------------------- [@tallini1999] $\mathcal{O}(qk\log_qk)$ $\mathcal{O}(qk\log_qk)$ [@capocelli1991] $\mathcal{O}(qk\log_qk)$ $\mathcal{O}(qk\log_qk)$ [@swart2009] $\mathcal{O}(qk\log_qk)$ $\mathcal{O}(1)$ [@pelusi2015] $\mathcal{O}(k\sqrt{\log_qk})$ $\mathcal{O}(1)$ Our scheme $\mathcal{O}(qk\log_qk)$ $\mathcal{O}(\log_qk)$ : Complexities of various schemes (orders are in digit operations)[]{data-label="tab:cmpl"} Conclusion {#sec:6} ========== An efficient construction has been proposed for balancing non-binary information sequences. By making use of Gray codes for the prefix, no lookup tables are used, only linear operations are needed for the balancing and the Gray code implementation. The encoding scheme has a complexity of $\mathcal{O}(qk\log_qk)$ digit operations. For the decoding process, once the Gray code prefix is decoded using $\mathcal{O}(\log_qk)$ digit operations, the balancing sequence is determined and the rest of the decoding process is performed in parallel. This makes the decoding fast and efficient. Possible future research directions include finding a mathematical procedure to determine the subset of Gray code sequences for $q$ even, given that it was found manually, by using a sliding window over the random walk graph. Practically, the redundant symbol $u$ only needs to take on values of zero (when the random walk falls on the balancing value) or one (when the random walk falls just below the balancing value). Thus, unnecessary redundancy is contained in $u$, especially for large values of $q$. However, the flexibility over $u$ increases the occurrences of balanced sequences. These additional balanced outputs could potentially be used to send auxiliary data that could reduce the redundancy. This property was proved for the binary case [@weber2010]. Additionally, given that the random walk graph passes through other weights in the region of the balancing value, the scheme can be extended to the construction of constant weight sequences with arbitrary weights. [11]{} K. A. S. Immink, *Codes for Mass Data Storage Systems*, 2nd ed., Shannon Foundation Publishers, Eindhoven, The Netherlands, 2004. D. E. Knuth, “Efficient balanced codes,” *IEEE Transactions on Information Theory*, vol. 32, no. 1, pp. 51–53, Jan. 1986. S. Al-Bassam, “Balanced codes,” Ph.D. dissertation, Oregon State University, USA, Jan. 1990. R. M. Capocelli, L. Gargano and U. Vaccaro, “Efficient $q$-ary immutable codes,” *Discrete Applied Mathematics*, vol. 33, no. 1–3, pp. 25–41, Nov. 1991. T. G. Swart and J. H. Weber, “Efficient balancing of $q$-ary sequences with parallel decoding,” in *Proceedings of the IEEE International Symposium on Information Theory*, Seoul, Korea, 28 Jun.–3 Jul. 2009, pp. 1564–1568. T. G. Swart and K. A. S. Immink, “Prefixless $q$-ary balanced codes with ECC,” in *Proceedings of the IEEE Information Theory Workshop*, Seville, Spain, Sep. 9–13, 2013. D. Pelusi, S. Elmougy, L. G. Tallini and B. Bose, “$m$-ary balanced codes with parallel decoding,” *IEEE Transactions on Information Theory*, vol. 61, no. 6, pp. 3251–3264, Jun. 2015. L. G. Tallini and U. Vaccaro, “Efficient $m$-ary immutable codes,” *Discrete Applied Mathematics*, vol. 92, no. 1, pp. 17–56, Mar. 1999. F. Gray, “Pulse code communication,” *U. S. Patent 2632058*, Mar. 1953. D.-J. Guan, “Generalized Gray codes with applications,” in *Proceedings of National Science Council, Republic of China, Part A*, vol. 22, no. 6, Apr. 1998, pp. 841–848. M. C. Er, “On generating the N-ary reflected Gray codes,” *IEEE Transactions on Computers*, vol. 33, no. 8, pp. 739–741, Aug. 1984. J. H. Weber and K. A. S. Immink, “Knuth’s balancing of codewords revisited,” *IEEE Transactions on Information Theory*, vol. 56, no. 4, pp. 1673–1679, Apr. 2010. E. N. Mambou, and T. G. Swart, “Encoding and decoding of balanced $q$-ary sequences using a Gray code prefix,” in *Proceedings of the IEEE International Symposium on Information Theory*, Barcelona, Spain, Jul. 10–15, 2016, pp. 380–384. K. A. S. Immink and J. H. Weber, “Very efficient balanced codes,” *IEEE Transactions on Information Theory*, vol. 28, no. 2, pp. 188–192, Feb. 2010. L. G. Tallini and B. Bose, “Balanced codes with parallel encoding and decoding,” *IEEE Transactions on Computers*, vol. 48, no. 8, pp. 794–814, Aug. 1999. B. Bose, “On unordered codes,” *Proceedings of the International Symposium on Fault-Tolerant Computing*, Pittsburgh, PA, 1987, pp. 102–107. Z. Star, “An asymptotic formula in the theory of compositions,” *Aequationes Mathematicae*, vol. 13, no. 1, pp. 279–284, Feb. 1975. [^1]: This paper was presented in part at the IEEE International Symposium on Information Theory, Barcelona, Spain, July, 2016. [^2]: The authors are with the Department of Electrical and Electronic Engineering Science, University of Johannesburg, P. O. Box 524, Auckland Park, 2006, South Africa (e-mails: {emambou, tgswart}@uj.ac.za). [^3]: This work is based on research supported in part by the National Research Foundation of South Africa (UID 77596).

ArXiv

--- address: | Eötvös Loránd University, Department of Atomic Physics,\ Pázmány Péter sétány, 1111 Budapest, Hungary author: - 'Gábor I. Veres on behalf of the CMS and ATLAS Collaborations' title: 'Heavy ion physics at CMS and ATLAS: hard probes' --- Transport properties, parton energy loss ======================================== In heavy ion physics it is common to compare A+A and p+p interactions in order to isolate physical phenomena unique to large colliding systems. For the interpretation of such comparisons, it is necessary to quantify the modification of parton distribution functions in nuclei, including gluon saturation at low $x$. Nuclei can be probed with p+Pb collisions at the LHC, and recent results in this area include high-energy photons and dijets with a large rapidity separation. The ATLAS collaboration [@atlaspaper] has recently measured the nuclear modification factors, $R_{\rm pA}$, of isolated photons in p+Pb collisions, and concluded that the data disfavor a large amount of (initial state) energy loss, and impose constraints on the nuclear PDFs [@atlasphotons]. Dijets measured in p+Pb collisions with a large rapidity separation can probe partons at low x (between $10^{-4} - 10^{-5}$). No broadening was observed in the azimuthal angle correlations for such dijets. Jet pairs, where both jets had a high rapidity in the proton-going direction (i.e. sampling low-x partons in the Pb nucleus) were found to be suppressed with respect to p+p collisions [@atlasdijets]. Departing from the baseline of nPDFs, one can study final state suppression using the nuclear modification factors in Pb+Pb and Xe+Xe collisions, as measured by the CMS collaboration [@cmspaper] for charged hadrons [@xexe]. On the left panel of Fig. \[gammajet\] the charged-particle $R_{\rm AA}$ is shown for Xe+Xe collisions at $\sqrt{s_{\rm NN}}=5.44$ TeV for the 5% most central collisions, together with earlier data on $R_{\rm AA}$ in Pb+Pb collisions at 5.02 TeV. The data may indicate a slight difference in suppression at high $p_T$. Comparing $R_{\rm AA}$ values at the same number of participating nucleons, there is a hint of a greater suppression in Xe+Xe collisions, probably due to a geometrical effect. ![ Left: The charged-particle $R_{\rm AA}$ for Xe+Xe collisions at $\sqrt{s_{\rm NN}}=5.44$ TeV for the 5% most central collisions$^5$, together with an earlier measurement of $R_{\rm AA}$ in Pb+Pb collisions at 5.02 TeV. Right: photon-jet $p_T$-balance distributions$^7$ in Pb+Pb events (red circles) in different centrality bins compared to that in p+p events (blue squares) for $p_T^\gamma$ = 79.6-100 GeV, where $x_J\gamma = p_T^{\rm jet}/ p_T^\gamma$. []{data-label="gammajet"}](CMS-HIN-18-004_Figure_004-a "fig:"){width="0.37\linewidth"} ![ Left: The charged-particle $R_{\rm AA}$ for Xe+Xe collisions at $\sqrt{s_{\rm NN}}=5.44$ TeV for the 5% most central collisions$^5$, together with an earlier measurement of $R_{\rm AA}$ in Pb+Pb collisions at 5.02 TeV. Right: photon-jet $p_T$-balance distributions$^7$ in Pb+Pb events (red circles) in different centrality bins compared to that in p+p events (blue squares) for $p_T^\gamma$ = 79.6-100 GeV, where $x_J\gamma = p_T^{\rm jet}/ p_T^\gamma$. []{data-label="gammajet"}](fig_04b "fig:"){width="0.62\linewidth"} The final state suppression of jets was also measured by the ATLAS experiment with an unprecedented precision recently, showing that the nuclear modification factor increases from low to high $p_T$ and from central to peripheral Pb+Pb collisions [@atlasjetraa]. Since photons are not affected by final state interactions, the energy loss of jets can be more precisely characterized by selecting photon-jet events. The recent data published by the ATLAS collaboration [@gammajetatlas] is corrected for accidental pairings and unfolded for energy resolution. The result can be seen on the right panel of Fig. \[gammajet\], in terms of the distribution of the photon-jet $p_T$-balance distributions in different centrality bins, where $x_J\gamma = p_T^{\rm jet}/ p_T^\gamma$. One can conclude that while many of the jets lose a significant amount of energy in the most central Pb+Pb collisions, there still remain some relatively symmetric photon-jet pairs, producing a peak-like structure close to unity in $x_J\gamma$. Medium temperature, quarkonium states, heavy flavor =================================================== Quarkonium production is a sensitive gauge of the temperature in the colored medium created in heavy ion collisions. These heavy mesons have a modest binding energy and a large radius, and the Debye-screening in the quark-gluon matter may cause their dissociation. The weakly bound states (like $\Upsilon(2S)$ and $\Upsilon(3S)$) are expected to suffer a stronger suppression, in comparison to p+p collisions, than more tightly bound ones, like $\Upsilon(1S)$. The dissociation temperatures are predicted to be at $2T_c$, $1.2T_c$ and $T_c$ for these three mesons, where $T_c$ is the critical temperature. Indeed, this successive suppression of the $\Upsilon$ states, measured in their dimuon decay channel was observed by the CMS collaboration using a high-statistics data set of Pb+Pb collisions [@ycms]. The invariant dimuon mass spectrum can be seen on the left panel of Fig. \[quarkonia\]. The result of the fit to the data including the three $\Upsilon$ states and the non-resonant background is shown as a solid blue line. The dashed red line represents the result of the same fit, but with the $\Upsilon$ yield for each state respectively divided by their measured $R_{\rm AA}$ value (i.e. their measured suppression with respect to p+p collisions recorded at the same center-of-mass energy). The suppression is also found to be gradually strengthening with increasing collision centrality. It was also shown by the recent analysis of the CMS collaboration that the excited prompt $\Psi(2S)$ is more suppressed in Pb+Pb collisions than the $J/\Psi$ ground state [@cmsjpsi]. The $J/\Psi$ mesons also constitute an important tool to characterize the b-quark energy loss, since b-decays to $J/\Psi$ can be measured separately from prompt $J/\Psi$ production making use of the long lifetime of the b quark. The ATLAS collaboration has measured the nuclear modification factors of prompt and non-prompt $J/\Psi$ particles at high $p_T$, and found a strong suppression in both cases [@jpsiatlas], as can be seen on the right panel of Fig. \[quarkonia\]. Prompt $J/\Psi$ mesons are suppressed to a similar extent as inclusive charged particles, while the non-prompt states experience less suppression in the $p_T<20$ GeV range, owing to the more modest energy loss of b-quarks compared to light quarks. A further confirmation of this phenomena is that the suppression of non-prompt $\Psi(2S)$ and non-prompt $J/\Psi$ states, both originating from b-decays, were found to be equal. A similar conclusion can be drawn from a recent analysis of muons originating from heavy quark decays by the ATLAS experiment [@atlasmuons], and from the CMS measurement of the non-prompt $D^0\rightarrow K^-\pi^+$ mesons (coming from b-hadron decays), which exhibit significantly less suppression at low $p_T$ compared to charged hadrons [@cmsd0]. It is also interesting to measure the $B_s^0$ state in Pb+Pb collisions to test if beauty and strange quarks can coalesce in the environment abundant in $s$ quarks, possibly leading to an increase of the $B_s^0/B^+$ ratio. The CMS collaboration has published the first result on that recently in the $B_s^0\rightarrow \mu^+\mu^- K^+ K^-$ decay channel, with a possible indication of such an increase [@bs0]. ![ Quarkonium results in Pb+Pb collisions at $\sqrt{s_{\rm NN}}=5.02$ TeV. Left: invariant mass distribution of muon pairs$^8$ for the kinematic range $p_T^{\mu^+\mu^-}<30$ GeV and $|y^{\mu^+\mu^-}|<2.4$. The result of the fit to the data is shown as a solid blue line. The dashed red line is also the result of the same fit but with the $\Upsilon$ yield for each state divided by the measured $R_{\rm AA}$. Right: Comparison of prompt and non-prompt $J/\Psi$ $R_{\rm AA}$ with the $R_{\rm AA}$ of charged particles$^{10}$.[]{data-label="quarkonia"}](CMS-HIN-16-023_Figure_001-b "fig:"){width="0.49\linewidth"} ![ Quarkonium results in Pb+Pb collisions at $\sqrt{s_{\rm NN}}=5.02$ TeV. Left: invariant mass distribution of muon pairs$^8$ for the kinematic range $p_T^{\mu^+\mu^-}<30$ GeV and $|y^{\mu^+\mu^-}|<2.4$. The result of the fit to the data is shown as a solid blue line. The dashed red line is also the result of the same fit but with the $\Upsilon$ yield for each state divided by the measured $R_{\rm AA}$. Right: Comparison of prompt and non-prompt $J/\Psi$ $R_{\rm AA}$ with the $R_{\rm AA}$ of charged particles$^{10}$.[]{data-label="quarkonia"}](fig_06a "fig:"){width="0.49\linewidth"} Jet substructure ================ After considering the spectacular jet quenching (parton energy loss) results obtained from heavy ion data, the next immediate question to ask is about the possible changes of the jet structure with respect to p+p collisions. The CMS experiment has measured the transverse shape (energy density) of jets tagged by isolated photons, as a function of the distance $r$ from the jet axis for various centrality categories [@jetshapecms], as shown on the left panel of Fig. \[jetshape\]. A jet broadening can be observed for central Pb+Pb collisions in these photon-jet events, while no depletion is visible in the $0.1<r<0.2$ region, as opposed to inclusive jets. The longitudinal structure of jets, the fragmentation functions, were published by the ATLAS experiment [@atlaslongit], and show that there is an enhancement of particles with a small or very large fraction of the jet momentum, and a suppression of particles with an intermediate momentum fraction. The jet fragmentation functions were also measured in photon-tagged jets [@jetfragatlas], and a ratio with respect to those in p+p collisions are shown on the right panel of Fig. \[jetshape\]. The jets with a photon partner, dominated by quark jets (blue data points), and inclusive jets (red points) are modified in a different way in central Pb+Pb events, although the interpretation of the data is complicated due to the different selection biases. ![ Left panel, upper plots: the differential jet shape, $\rho(r)$, for jets associated with an isolated photon for (from left to right) 50-100%, 30-50%, 10-30%, 0-10% Pb+Pb (full circles), and p+p (open circles) collisions and from PYTHIA simulation (histogram). Left panel, lower plots: the ratios of the Pb+Pb and p+p distributions$^{14}$. Right: ratio of the fragmentation function in jets azimuthally balanced by a high-$p_T$ photon: 30-80% Pb+Pb collisions to p+p collisions (left panel) and 0-30% Pb+Pb collisions to p+p collisions (right panel). Results are shown as a function of charged-particle transverse momentum $p_T$, for $\gamma$-tagged jets (this measurement, full markers) and for inclusive jets in 2.76 TeV Pb+Pb collisions (open markers)$^{16}$.[]{data-label="jetshape"}](CMS-HIN-18-006_Figure_001){width="0.47\linewidth"} Finally, the CMS experiment has employed the jet grooming technique to remove large-angle, soft radiation, and extract the hard subjets. Since the opening angle between those subjets is sensitive to medium induced modifications, the distribution of the jet mass of these groomed jets is an important observable. These results indicate that available model calculations overestimate the yield of jets with a large groomed mass relative to the jet $p_T$[@jetgroom]. In summary, the heavy ion research program at the LHC provides a precise and detailed set of experimental data, challenging many of the phenomenological models concerning jet substructure modifications, and supporting various expectations about nuclear PDFs, energy loss of light and heavy quarks and quarkonium dissociation. [**Acknowledgments.**]{} The author is grateful for the support of the MTA Momentum program LP 2015-7/2015 and the support by the grants NKFIA K\_18 128713, NKFIA K\_17 124845 and NKFIA FK\_17 123842. References {#references .unnumbered} ========== [99]{} ATLAS Collaboration [*JINST*]{} [**3**]{} S08003 (2008). CMS Collaboration [*JINST*]{} [**3**]{} S08004 (2008). ATLAS Collaboration, submitted to [*Phys. Lett.*]{} B, [arXiv:1903.02209]{} (2019). ATLAS Collaboration, submitted to [*Phys. Rev.*]{} C, [arXiv:1901.10440]{} (2019). CMS Collaboration [*JHEP*]{} [**10**]{} 138 (2018). ATLAS Collaboration [*Phys. Lett.*]{} B [ **790**]{} 108 (2019). ATLAS Collaboration [*Phys. Lett.*]{} B [ **789**]{} 167 (2019). CMS Collaboration [*Phys. Lett.*]{} B [**790**]{} 270 (2019). CMS Collaboration [*Eur. Phys. J.*]{} C [**78**]{} 509 (2018). ATLAS Collaboration [*Eur. Phys. J.*]{} C [**78**]{} 762 (2018). ATLAS Collaboration [*Phys. Rev.*]{} C [**98**]{} 044905 (2018). CMS Collaboration, submitted to [*Phys. Rev. Lett.*]{}, [arXiv:1810.11102]{} (2018). CMS Collaboration, submitted to [*Phys. Lett.*]{} B, [arXiv:1810.03022]{} (2018). CMS Collaboration [*Phys. Rev. Lett.*]{} [**122**]{} 152001 (2019). ATLAS Collaboration [*Phys. Rev.*]{} C [**98**]{} 024908 (2018). ATLAS Collaboration, submitted to [*Phys. Rev. Lett.*]{}, [arXiv:1902.10007]{} (2019). CMS Collaboration [*JHEP*]{} [**10**]{} 161 (2018).

ArXiv

**Localized electron state in a T-shaped confinement potential** L. A. Openov *Moscow State Engineering Physics Institute (Technical University),* *Kashirskoe sh. 31, Moscow 115409, Russia* **Abstract** > We consider a simple model of an electron moving in a T-shaped confinement potential. This model allows for an analytical solution that explicitly demonstrates the existence of laterally bound electron states in quantum wires obtained by the cleaved edge overgrowth technique. In recent years, semiconductor nanostructures have attracted much attention due to both a great variety of their possible applications in electronic devices and a new interesting physics emerged to describe their peculiar characteristics [@meso]. Among other research directions, there are intensive studies of the so called quantum wires [@wires], in which electrons are confined in two spatial dimensions within a nanometer-size region, while being allowed to move freely in the third direction (an axis of the wire). High-quality quantum wires can be fabricated by the molecular-beam epitaxy using the cleaved edge overgrowth (CEO) technique [@Pfeiffer], as proposed by Esaki [*et al.*]{} [@Esaki]. Such structures are formed at the T-shaped intersection of two semiconductor quantum wells, see fig. 1. The electronic band structure of realistic T-shaped AlGaAs/GaAs quantum wires has been studied numerically using different methods [@theory]. Note, however, that the form of the T-shaped confinement potential, fig. 1, does not allow for analytical solutions of the one-electron continuum Schrödinger equation even in the limiting case of infinite energy barriers and isotropic electron effective mass. Meanwhile, the existence of electron states localized in the region of intersection of two quantum wells within the plane perpendicular to the axis of CEO quantum wire is not obvious [*a priori*]{}. In this paper, we propose a simple model of the two-dimensional T-shaped structure and obtain an analytical solution for both the lowest energy eigenvalue and corresponding localized eigenfunction of the one-electron Schrödinger equation. Let us consider two intersecting atomic chains, one of which (along [*x*]{}-axis) is infinite, while another (along [*y*]{}-axis) is semi-infinite, see fig. 2. Such a system provides a model for the description of electron motion in the [*xy*]{}-plane perpendicular to the axis of a CEO quantum wire (fig. 1) since an electron is confined to the chains and can not escape into the interjacent region. This model corresponds to the limiting case of a CEO quantum wire formed by two quantum wells with one-atomic-layer width and an infinite value of the energy barrier $V=E_c^A-E_c^B$ preventing an electron from going to the conduction band of the semiconductor A, see fig. 1. In a tight-binding approach, one-electron states of the system under consideration are described by the following Hamiltonian: $$\hat{H}(t)= \epsilon_0\sum_{i}{\hat{a}^{+}_{i}\hat{a}^{}_{i}}- t\sum_{<i,j>}{(\hat{a}^{+}_{i}\hat{a}^{}_{j}+h.c)}~, \label{Ham}$$ where $\epsilon_0$ is an on-site potential, $t$ is a matrix element for electron hopping between nearest sites $i$ and $j$, $\hat{a}^{+}_{i}$ $(\hat{a}^{}_{i})$ is the operator of electron creation (annihilation) at the site $i$, and $<i,j>$ means the summation over nearest neighbors only. We do not specify the spin index explicitly since we consider the states of a single electron. Expanding the one-electron wave function $\Psi$ into atomic electron states $|i>$, $$\Psi=\sum_{i}A_i|i>~, \label{Psi}$$ we have from the Schrödinger equation $\hat{H}\Psi=E\Psi$ a set of algebraic equations for coefficients $A_i$ whose squared absolute values $|A_i|^2$ give the probabilities to find an electron at a particular site $i$: $$\tilde{E}A_i=\sum_{j=nn(i)} A_j~, \label{Ai}$$ where $$\tilde{E}=\frac{\epsilon_0-E}{t} \label{Etilde}$$ and $j=nn(i)$ means the summation over sites $j$ nearest to the site $i$. Let $i=n$ for atoms in the chain running along [*x*]{}-axis and $i=m$ for atoms in the chain along [*y*]{}-axis ($n = 0, \pm 1, \pm 2, \pm 3,... ; m = 0, 1, 2, 3, ...$), see fig. 2. For the atom belonging to both chains (at the intersection point) one has $n=m=0$. We consider three different regions: (1) $n \ge 0$; (2) $n \le 0$; (3) $m \ge 0$. In those regions, we have from eq. (\[Ai\]): $$\begin{aligned} \tilde{E}A_n=A_{n-1}+A_{n+1}~,~n\ge 1~, \nonumber \\ \tilde{E}A_n=A_{n-1}+A_{n+1}~,~n\le -1~, \nonumber \\ \tilde{E}A_m=A_{m-1}+A_{m+1}~, m\ge 1~. \label{Anm}\end{aligned}$$ General solutions of eqs. (\[Anm\]) are: $$\begin{aligned} A_n = A_0 \exp(-\alpha n)~,~n\ge 0~, \nonumber \\ A_n = B_0 \exp(\beta n)~,~n\le 0~, \nonumber \\ A_m = C_0 \exp(-\gamma n)~,~m\ge 0~, \label{Anm1}\end{aligned}$$ where $$\tilde{E}=2\cosh(\alpha)=2\cosh(\beta)=2\cosh(\gamma)~. \label{Etilde1}$$ Since, due to normalization condition $\sum_{i}|A_i|^2=1$, the coefficients $A_n$ and $A_m$ should be finite at $n\rightarrow \pm \infty$ and $m\rightarrow \infty$ respectively, one has $Re(\alpha) \ge 0,~Re(\beta) \ge 0,~Re(\gamma) \ge 0$. Hence, we see from eq. (\[Etilde1\]) that $\alpha=\beta=\gamma$, while from eqs. (\[Anm1\]) we have $A_0=B_0=C_0$ since $A_{n=0}=A_{m=0}$ at the intersection point. Next, from eq. (\[Ai\]) for the coefficient $A_0$ we obtain $$\tilde{E}A_0=A_{n=1}+A_{n=-1}+A_{m=1}~. \label{A0}$$ Taking eqs. (\[Anm1\]) into account, we have $$\tilde{E}=3\exp(-\alpha)~. \label{Etilde2}$$ The lowest energy solution of eqs. (\[Etilde1\]) and (\[Etilde2\]) is $\tilde{E}=3/\sqrt{2}$ and $\exp(-\alpha)=1/\sqrt{2}$, and so from eq. (\[Etilde\]) we have the minimum one-electron energy of the system under consideration: $$E_{min}=\epsilon_0-\frac{3}{\sqrt{2}}t~, \label{E}$$ while the corresponding normalized wave function is $$\Psi=\sum_{n=-\infty}^{\infty}\frac{1}{2^{|n|/2+1}}|n>+ \sum_{m=1}^{\infty}\frac{1}{2^{m/2+1}}|m>~. \label{Psi1}$$ One can see from eq. (\[Psi1\]) that the wave function is localized exponentially in the vicinity of the intersection point $n=0, m=0$. The probability to find an electron at the site $n=m=0$ equals to 1/4, while the probabilities to find an electron at the nearest sites $n=-1, n=1, m=1$ are 1/8 each. The localization length $\xi$ defined by $A_n=\exp(-a|n|/\xi)$, where $a$ is the interatomic spacing, equals to $\xi=a/\alpha=2a/\ln(2)\approx 3a$. We note that the value of $E_{min}=\epsilon_0-3t/\sqrt{2}$ is by $\approx 0.12t$ lower than the edge $E{_0}=\epsilon_0-2t$ of the band of delocalized electron states in the infinite one-dimensional chain. An estimate of the hopping matrix element $t$ for a specific case of GaAs gives in a tight-binding approach $t=\hbar^2/2m^*a^2\approx$ 1.8 eV, where $m^*=0.067m_0$, $a=$ 0.565 nm, $m_0$ is the mass of a free electron. Then for the confinement energy one has $E_c=E{_0}-E_{min}\approx$ 200 meV. The physical reason for appearance of the localized state in the T-shaped geometry seems to be related with the effect of Anderson localization [@Anderson] since the semi-infinite chain along $y$-axis plays a role of defect for an electron moving in the infinite chain along $x$-axis. We have also studied the model under consideration by the exact numerical diagonalization of the Hamiltonian matrix. We made use of the complex conjugate method which allows one to find eigenvalues and eigenvectors of large sparse matrixes with any prescribed accuracy [@Openov]. For our purposes, we have restricted ourselves to several low-lying levels. We have considered the systems composed of $3N+1$ sites ($2N+1$ sites in the chain along $x$-axis and $N+1$ sites in the chain along $y$-axis, one site being common for both chains), with $N$ up to $10^3$. For the ground state, the calculated values of $E_{min}$ and $\alpha$ coincide with analytical results. The excited states have the energies $E_i\ge E{_0}$, where the value of $E{_0}$ tends to $\epsilon_0-2t$ as $N$ increases, e.g., $E{_0}=\epsilon_0-1.99903t$ for $N=$ 100 and $E{_0}=\epsilon_0-1.99976t$ for $N=$ 200. The mean level spacing of the excited states is of the order of $t/N$ and goes to zero as $N$ increases. An inspection of wave functions of low-lying excited states has shown that all of them are delocalized over the whole system, having the form of sines or cosines. Finally, along with the system shown in fig. 2, we have examined a more general case of quasi-one-dimensional chains of finite width, so that each chain had $N_0\ge 2$ sites in width ($N_0=$ 1 in a particular case studied analytically above). We have considered the systems composed of $3N\cdot N_0+N_0^2$ sites ($(2N+N_0)\cdot N_0$ sites in the chain along $x$-axis and $(N+N_0)\cdot N_0$ sites in the chain along $y$-axis, $N_0^2$ sites being common for both chains), with $N$ up to $10^4$ and $N_0$ up to 40. We have found that the ground state energy $E_{min}$ decreases with increasing $N_0$, see fig. 3, and remains lower than the edge of the band of delocalized states, $E{_0}$. Fig. 4 shows the confinement energy $E_c=E{_0}-E_{min}$ as a function of the chain width $N_0$. One can see that $E_c$ decreases with $N_0$ and tends to zero as $N_0$ goes to infinity. Taking the values of $t$ and $a$ for GaAs, see above, one has, e.g., $E_c\approx 32$ meV and 9 meV for the chains of 5 nm and 10 nm width respectively. These values of $E_c$ are about twice as large as those obtained for the conduction band of symmetrical T-shape wire structures with GaAs wells and Al$_{0.3}$Ga$_{0.7}$As barriers within a much more sophisticated model [@theory]. An apparent reason for this quantitative discrepancy is the infinite value of the conduction-band offset in our simple model. So, our approach provides an upper estimate for the electron confinement energy in the T-shaped quantum wires. From figs. 3 and 4 we find that at large $N_0$ both $E_{min}$ and $E{_0}$ approach -4$t$ which is just the value of $E{_0}=E_{min}$ for the infinite two-dimensional system in a tight-binding model. Again, the ground state wave function is always localized exponentially in the vicinity of the intersection region, while the wave functions of excited states are delocalized. The values of the localization length $\xi$ at different $N_0$ where found through the analysis of the asymptotic behavior of the ground state wave function at large distance from the intersection region. It was found that the decrease in $E_c$ with $N_0$ is accompanied by the corresponding increase in $\xi$ according to the general relation $E_c=\hbar^2/2m^*\xi^2=t(a/\xi)^2$. The dependence of $\xi$ on $N_0$ is shown in fig. 5. In conclusion, we have presented an analytically solvable model of the T-shaped two-dimensional confinement potential. Although being rather simple, this textbook model explicitly demonstrates the existence of localized electron states in the T-shaped geometry. It can be used for qualitative estimates of confinement energies and the extent of spatial localization of one-electron wave function in CEO quantum wires. The model can be easily generalized to account for electron hopping to atoms other than nearest neighbors only and/or for the finite value of the confinement potential. In such a case, however, it will be difficult to obtain an analytical solution of one-electron Schrödinger equation. This work was supported in part by the Russian Federal Program “Integration”, project No A0133. [5]{} =-1mm SCHULZ H. J., in [*Mesoscopic Quantum Physics*]{}, edited by E. AKKERMANS, G. MONTAMBEAUX, J.-L. PICHARD and J. ZINN-JUSTIN (Elsevier, Amsterdam) 1995. BEENAKKER C. W. J. and VON HOUTEN H., in [*Solid State Physics, Semiconductor Heterostructures and Nanostructures*]{}, edited by H. EHRENREICH and D. TURNBULL (Academic Press, New York) 1991. PFEIFFER L., WEST K. W., STÖRMER H. L., EISENSTEIN J. P., BALDWIN K. W., GERSHONI D. and SPECTOR J., Appl. Phys. Lett., [**56**]{} (1990) 1697. CHANG Y.-C., CHANG L. L. and ESAKI L., Appl. Phys. Lett., [**47**]{} (1985) 1324. For example, LANGBEIN W., GISLASON H. and HVAM J. M., Phys. Rev. B, [**54**]{} (1996) 14595; GOLDONI G., ROSSI F., MOLINARI E. and FASOLINO A., Phys. Rev. B, [**55**]{} (1997) 7110. ANDERSON P. W., Phys. Rev., [**109**]{} (1958) 1492. OPENOV L. A., Phys. Low-Dim. Struct., [**10/11**]{} (1995) 365. **FIGURE CAPTIONS** Fig. 1. Schematic view of a T-shaped semiconductor quantum wire formed by two intersecting quantum wells. The height of the energy barrier for electrons in the semiconductor B equals to the conduction-band offset $\Delta E_c=E_c^A-E_c^B$ between semiconductors A and B. Electrons are confined in the $xy$-plane in the vicinity of intersection of quantum wells and move freely along $z$-axis (an axis of the quantum wire) Fig. 2. Two intersecting atomic chains as a model of a T-shaped confinement potential for electrons in the plane perpendicular to the axis of CEO quantum wire. Numbers $n$ and $m$ numerate atoms in the chains along $x$- and $y$-axis respectively. Fig. 3. The energy $E_{min}$ of the localized state in units of the hopping matrix element $t$ versus the number of sites $N_0$ across the chain. Points are the results of analytical (for $N_0=$ 1) and numerical (for $N_0\ge 2$, $N=$ 800) calculations. The line is a guide to the eye. Fig. 4. The confinement energy $E_c=E{_0}-E_{min}$ in units of the hopping matrix element $t$ versus the number of sites $N_0$ across the chain. Points are the results of analytical (for $N_0=$ 1) and numerical (for $N_0\ge 2$, $N=$ 800) calculations. The line is a guide to the eye. Fig. 5. Localization length $\xi$ of the ground state wave function in units of the interatomic spacing $a$ versus the number of sites $N_0$ across the chain. Points are the results of analytical (for $N_0=$ 1) and numerical (for $N_0\ge 2$, $N=$ 800) calculations. The line is the curve $\xi=a(t/E_c)^{1/2}$, where $E_c=E{_0}-E_{min}$ is the confinement energy.

ArXiv

--- abstract: | Recent results on classical and quantum strings in a variety of black hole and cosmological spacetimes, in various dimensions, are presented. The curved backgrounds under consideration include the $2+1$ black hole anti de Sitter spacetime and its dual, the black string, the ordinary $D\geq 4$ black holes with or without a cosmological constant, the de Sitter and anti de Sitter spacetimes and static Robertson-Walker spacetimes. Exact solutions to the string equations of motion and constraints, representing circular strings, stationary open strings and dynamical straight strings, are obtained in these backgrounds and their physical properties (length, energy, pressure) are described. The existence of [*multi-string*]{} solutions, describing finitely or infinitely many strings, is shown to be a general feature of spacetimes with a positive or negative cosmological constant. Generic approximative solutions are obtained using the string perturbation series approach, and the question of the stability of the solutions is addressed. Furthermore, using a canonical quantization procedure, we find the string mass spectrum in de Sitter and anti de Sitter spacetimes. New features as compared to the string spectrum in flat Minkowski spacetime appear, for instance the [*fine-structure effect*]{} at all levels beyond the graviton in both de Sitter and anti de Sitter spacetimes, and the [*non-existence*]{} of a Hagedorn temperature in anti de Sitter spacetime. We discuss the physical implications of these results. Finally, we consider the effect of spatial curvature on the string dynamics in Robertson-Walker spacetimes. author: - | A.L. Larsen and N. Sánchez\ \ Observatoire de Paris, DEMIRM. Laboratoire Associé au CNRS\ UA 336, Observatoire de Paris et École Normale Supérieure.\ 61, Avenue de l’Observatoire, 75014 Paris, France. title: 'Strings and Multi-Strings in Black Hole and Cosmological Spacetimes[^1]' --- Introduction ============ The classical and quantum propagation of strings in curved spacetimes has attracted a great deal of interest in recent years. The main complication, as compared to the case of flat Minkowski spacetime, is related to the non-linearity of the equations of motion. It makes it possible to obtain the complete analytic solution only in a very few special cases like conical spacetime [@san3] and plane-wave/shock-wave backgrounds [@san4]. There are however also very general results concerning integrability and solvability for maximally symmetric spacetimes [@mic1; @four] and gauged WZW models [@bar; @six]. These are the exceptional cases; generally the string equations of motion in curved spacetimes are not integrable and even if they are, it is usually an extremely difficult task to actually separate the equations, integrate them and finally write down the complete solution in closed form. Fortunately, there are several different ways to “attack” a system of coupled non-linear partial differential equations. The systematic study of string dynamics in curved spacetimes and its associated physical phenomena was started in Refs.[@veg1; @san1]. Besides numerical methods, which will not be discussed here, approximative \[7-10\] and exact \[11-14\] methods for solving the string equations of motion and constraints in curved spacetimes, have been developed. Classical and quantum string dynamics have been investigated in black hole backgrounds \[15-18\], cosmological spacetimes \[7, 11-14, 17-21\], cosmic string spacetime [@san3], gravitational wave backgrounds [@san4], supergravity backgrounds (which are necessary for fermionic strings) [@san5], and near spacetime singularities [@san6]. Physical phenomena like the Hawking-Unruh effect in string theory [@san1; @san7], horizon string stretching [@san1; @san7], particle transmutation [@san2; @san8], string scattering [@san3; @san4; @san2], mass spectrum and critical dimension [@san3; @veg1; @san2; @all2], string instability \[7, 11-13, 16, 21\] and multi-string solutions \[11-13, 28\] have been found. In a generic $D$-dimensional curved spacetime with metric $g_{\mu\nu}$ and coordinates $x^\mu,\;(\mu=0,1,...,D-1),$ the string equations of motion and constraints are: $$\ddot{x}^\mu-x''^\mu+\Gamma^\mu_{\rho\sigma}(\dot{x}^\rho\dot{x}^\sigma- x'^\rho x'^\sigma)=0,$$ $$g_{\mu\nu}\dot{x}^\mu x'^\nu=g_{\mu\nu}(\dot{x}^\mu\dot{x}^\nu+x'^\mu x'^\nu) =0,$$ where dot and prime stand for derivative with respect to the world-sheet coordinates $\tau$ and $\sigma,$ respectively and $\Gamma^\mu_{\rho\sigma}$ are the Christoffel symbols with respect to the metric $g_{\mu\nu}.$ In the following we present recent results \[26-29\] on solutions to Eqs.(1.1)-(1.2) in a variety of curved spacetimes from cosmology, gravitation and string theory. We obtain explicit (exact and/or approximate) mathematical solutions, discuss the corresponding physical properties, we quantize the solutions in different ways (canonical quantization, semi-classical quantization) and find the physical content: the mass spectrum. The presentation is organized as follows: In Section 2, we discuss the classical string dynamics in the $2+1$ black hole anti de Sitter (BH-AdS) spacetime, recently found by Bañados et al [@ban1]. We compare with the string dynamics in ordinary cosmological and black hole spacetimes. This clarifies the geometry (as seen by a string) of the BH-AdS spacetime. In Section 3, generalizing results from Section 2, we derive the quantum string mass spectrum in ordinary $D$-dimensional anti de Sitter spacetime. We discuss, in particular, the sectors of low and very high mass states. New physical phenomena arise like the fine-structure effect at all levels beyond the graviton and the non-existence of a Hagedorn critical temperature. The results are compared with corresponding results obtained in Minkowski and de Sitter spacetimes. Sections 4 and 5 are devoted to the investigation of the more general underlying structure of solutions to Eqs.(1.1)-(1.2). In Section 4, we find new classes of exact string and multi-string solutions in cosmological and black hole spacetimes, while in Section 5, we consider the effects of a non-zero spatial curvature on the classical and quantum string dynamics in Friedmann-Robertson-Walker (FRW) universes. Classical String Dynamics in $2+1$ BH-AdS ========================================= Anti de Sitter (AdS) spacetime is often considered to be a spacetime of minor importance in a cosmological context. However, first of all, it [*is*]{} a FRW universe and as such should not be neglected, secondly; it serves as a simple and convenient spacetime for comparison and understanding of results obtained in (say) Minkowski or de Sitter spacetimes and third, it has a tendency to show up (in disguise) as a solution in various models of dilaton-gravity and string theory. An example of the latter is represented by the $2+1$ BH-AdS spacetime of Bañados et al [@ban1]. This spacetime background has arised much interest recently \[31-35\]. It describes a two-parameter family (mass $M$ and angular momentum $J$) of black holes in 2+1 dimensional general relativity with metric: $$ds^2=(M-\frac{r^2}{l^2})dt^2+(\frac{r^2}{l^2}-M+\frac{J^2}{4r^2})^{-1}dr^2 -Jdtd\phi+r^2 d\phi^2.$$ It has two horizons $r_\pm=\sqrt{\frac{Ml^2}{2}\pm\frac{l}{2}\sqrt{M^2 l^2-J^2}}$ and a static limit $r_{\mbox{erg}}=\sqrt{M}l,$ defining an ergosphere, as for ordinary Kerr black holes. The spacetime is not asymptotically flat; it approaches anti de Sitter spacetime asymptotically with cosmological constant $\Lambda=-1/l^2.$ The curvature is constant $R_{\mu\nu}=-(2/l^2)g_{\mu\nu}$ everywhere, except probably at $r=0,$ where it has at most a delta-function singularity. Notice the weak nature of the singularity at $r=0$ in 2+1 dimensions as compared with the power law divergence of curvature scalars in $D>3$ (We will not discuss here the geometry near $r=0.$ For a discusion, see Refs.[@ban2; @delta]). The spacetime, Eq.(2.1), is also a solution of the low energy effective action of string theory with zero dilaton field $\Phi=0,$ anti-symmetric tensor field $H_{\mu\nu\rho}=(2/l^2)\epsilon_{\mu\nu\rho}\;\; (\mbox{i.e.}\;B_{\phi t}=r^2/l^2)$ and $k=l^2$ [@hor1]. Moreover, it yields an exact solution of string theory in 2+1 dimensions, obtained by gauging the WZWN sigma model of the group $SL(2,R)\times R$ at level $k$ [@hor1; @kal] (for non-compact groups, $k$ does not need to be an integer, so the central charge $c=3k/(k-2)=26$ when k=52/23). This solution is the black string background [@hor2]: $$\begin{aligned} &d\tilde{s}^2=-(1-\frac{{\cal M}}{\tilde{r}})d\tilde{t}^2+(1-\frac{{\cal Q}^2} {{{\cal M}}\tilde{r}})d\tilde{x}^2+(1-\frac{{\cal M}}{\tilde{r}})^{-1} (1-\frac{{\cal Q}^2}{{{\cal M}}\tilde{r}})^{-1}\frac{l^2 d\tilde{r}^2} {4\tilde{r}^2},&\nonumber\end{aligned}$$ $$\begin{aligned} &\tilde{B}_{\tilde{x}\tilde{t}}=\frac{{\cal Q}}{r},\;\;\; \tilde{\Phi}=-\frac{1}{2} \log\tilde{r}l,&\end{aligned}$$ which is related by duality [@hor1; @bus] to the 2+1 BH-AdS spacetime, Eq.(2.1). It has two horizons $\tilde{r}_{\pm}=r_\pm,$ the same as the metric, Eq.(2.1), while the static limit is $\tilde{r}_{\mbox{erg}}=J/(2\sqrt{M}).$ We first investigate the string propagation in the $2+1$ BH-AdS background by considering the perturbation series around the exact center of mass of the string: $$x^\mu(\tau,\sigma)=q^\mu(\tau)+\eta^\mu(\tau,\sigma)+\xi^\mu(\tau,\sigma)+...$$ The original method of Refs.[@veg1; @san1] can be conveniently formulated in covariant form. It is useful to introduce $D-1$ normal vectors $n^\mu_R\;\;(R=1,..,D-1),$ (which can be chosen to be covariantly constant by gauge fixing), and consider comoving perturbations $\delta x_R,$ i.e. those seen by an observer travelling with the center of mass, thus $\eta^\mu= \delta x^Rn^\mu_R.$ After Fourier expansion, $\delta x^R(\tau,\sigma)= \sum_n C^R_n(\tau)e^{-in\sigma},$ the first order perturbations satisfy the matrix Schrödinger-type equation in $\tau$: $$\ddot{C}_{nR}+(n^2\delta_{RS}-R_{\mu\rho\sigma\nu}n^\mu_R n^\nu_S \dot{q}^\rho\dot{q}^\sigma)C^S_n=0.$$ Second order perturbations $\xi^\mu$ and constraints are similarly covariantly treated, $\xi^\mu$ also satisfying Schrödinger-type equations with source terms, see Ref.[@all1]. For our purposes here it is enough to consider the non-rotating $(J=0)$ 2+1 BH-AdS background and a radially infalling string. We have solved completely the c.m. motion $q^\mu(\tau)$ and the first and second order perturbations $\eta^\mu(\tau,\sigma)$ and $\xi^\mu(\tau,\sigma)$ in this background. Eqs.(2.4) become: $$\ddot{C}_{nR}+(n^2+\frac{m^2}{l^2})C_{nR}=0;\;\;\;R=\perp,\parallel$$ The first order perturbations are independent of the black hole mass, only the anti de Sitter (AdS) part emerges. All oscillation frequencies $\omega_n= \sqrt{n^2+m^2/l^2}$ are real; there are no unstable modes in this case. The perturbations: $$\delta x_R(\tau,\sigma)=\sum_n [ A_{nR}e^{-i(n\sigma+\omega_n\tau)}+\tilde{A}_{nR} e^{-i(n\sigma-\omega_n\tau)}]$$ are completely finite and regular. This is also true for the second order perturbations which are bounded everywhere, even for $r\rightarrow 0\;(\tau \rightarrow 0).$ We have also computed the conformal generators $L_n,$ (see Ref.[@all1]), and the string mass: $$m^2=2\sum_n (2n^2+\frac{m^2}{l^2})[A_{n\parallel} \tilde{A}_{-n\parallel}+ A_{n\perp}\tilde{A}_{-n\perp}].$$ The mass formula is modified (by the term $m^2/l^2$) with respect to the usual flat spacetime expression. This is due to the asymptotic (here AdS) behaviour of the spacetime. In ordinary $D\geq 4$ black hole spacetimes (without cosmological constant), which are asymptotically flat, the mass spectrum is the same as in flat Minkowski spacetime [@san2]. We compare with the string perturbations in the ordinary ($D\geq 4$) black hole anti de Sitter spacetime. In this case Eqs.(2.4) become: $$\ddot{C}_{nS\perp}+(n^2+m^2 H^2+\frac{Mm^2}{r^3}) C_{nS\perp}=0,\;\;\;S=1,2$$ $$\ddot{C}_{n\parallel}+(n^2+m^2 H^2-\frac{2Mm^2}{r^3}) C_{n\parallel}=0.$$ The transverse $\perp$-perturbations are oscillating with real frequencies and are bounded even for $r\rightarrow 0.$ For longitudinal $\parallel$-perturbations, however, imaginary frequencies arise and instabilities develop. The $(\mid n\mid=1)$-instability sets in at: $$r_{\mbox{inst.}}=(\frac{2Mm^2}{1+m^2H^2})^{1/3}$$ Lower modes become unstable even outside the horizon, while higher modes develop instabilities at smaller $r$ and eventually only for $r\approx 0.$ For $r\rightarrow 0$ (which implies $\tau\rightarrow\tau_0$) we find $r(\tau) \approx (3m\sqrt{M/2})^{2/3}(\tau_0-\tau)^{2/3}$ and: $$\ddot{C}_{nS\perp}+\frac{2}{9(\tau-\tau_0)^2}C_{nS\perp}=0,\;\;\;S=1,2$$ $$\ddot{C}_{n\parallel}-\frac{4}{9(\tau-\tau_0)^2}C_{n\parallel}=0.$$ For $\tau\rightarrow\tau_0$ the $\parallel$-perturbations blow up while the string ends trapped into the $r=0$ singularity. We see the important difference between the string evolution in the 2+1 BH-AdS background and the ordinary 3+1 (or higher dimensional) black hole anti de Sitter spacetime. We also compare with the string propagation in the 2+1 black string background, Eq.(2.2) (with $J=0$). In this case, Eqs.(2.4) become: $$\ddot{C}_{n\perp}+n^2 C_{n\perp}=0,$$ $$\ddot{C}_{n\parallel}+(n^2-\frac{2m^2M}{lr})C_{n\parallel}=0.$$ The $\perp$-modes are stable, while $C_{n\parallel}$ develop instabilities. For $r\rightarrow 0$ (which implies $\tau\rightarrow\tau_0$) we find $r(\tau) \approx\frac{m^2M}{l}(\tau_0-\tau)^2$ and: $$\ddot{C}_{n\parallel}-\frac{2}{(\tau_0-\tau)^2}C_{n\parallel}=0,$$ with similar conclusions as for the ordinary 3+1 (or higher dimensional) black hole anti de Sitter spacetime. In order to extract more information about the string evolution in these backgrounds, in particular exact properties, we consider the circular string ansatz: $$t=t(\tau),\;\;\;r=r(\tau),\;\;\;\phi=\sigma+f(\tau),$$ in the equatorial plane $(\theta=\pi/2)$ of the stationary axially symmetric backgrounds: $$ds^2=g_{tt}(r)dt^2+g_{rr}(r)dr^2+2g_{t\phi}(r)dtd\phi+g_{\phi\phi}(r)d\phi^2.$$ This includes all the cases of interest here: The 2+1 BH-AdS spacetime, the black string, as well as the equatorial plane of ordinary Einstein black holes. The string dynamics, determined by Eqs.(1.1)-(1.2), is then reduced to a system of second order ordinary differential equations and constraints, also described as a Hamiltonian system: $$\dot{r}^2+V(r)=0;\;\;\;\;\;V(r)=g^{rr}(g_{\phi\phi}+E^2g^{tt}),$$ $$\dot{t}=-Eg^{tt},\;\;\;\;\;\dot{f}=-Eg^{t\phi};\;\;\;\;E=-P_t=\mbox{const.},$$ which in all backgrounds considered here are solved in terms of either elementary or elliptic functions. The dynamics of the circular strings takes place at the $r$-axis in the $(r,V(r))$-diagram and from the properties of the potential $V(r)$ (minima, zeroes, asymptotic behaviour for large $r$ and the value $V(0)$), general knowledge about the string motion can be obtained. On the other hand, the line element of the circular string turns out to be: $$ds^2=g_{\phi\phi}(d\sigma^2-d\tau^2),\;\;\;\mbox{i.e.}\;\;S(\tau)=\sqrt {g_{\phi\phi}(r(\tau))},$$ $S(\tau)$ being the invariant string size. For all the static black hole AdS spacetimes ( 2+1 and higher dimensional): $\;S(\tau)=r(\tau).$ For the rotating 2+1 BH-AdS spacetime: $$V(r)=r^2(\frac{r^2}{l^2}-M)+\frac{J^2}{4}-E^2,$$ (see Fig.1). $V(r)$ has a global minimum $V_{\mbox{min}}<0$ between the two horizons $r_+,r_-$ (for $Ml^2\geq J^2,$ otherwise there are no horizons). The vanishing of $V(r)$ at $r=r_{01,2}$ (see Ref.[@all1]) determines three different types of solutions: (i) For $J^2>4E^2,$ there are two positive zeroes $r_{01}<r_{02},$ the string never comes outside the static limit, never falls into $r=0$ neither (there is a barrier between $r=r_{01}$ and $r=0$). The mathematical solution oscillates between $r_{01}$ and $r_{02}$ with $0<r_{01}<r_-<r_+<r_{02}<r_{\mbox{erg}}.$ It may be interpreted as a string travelling between the different universes described by the maximal analytic extension of the manifold. (ii) For $J^2<4E^2,$ there is only one positive zero $r_0$ outside the static limit and there is no barrier preventing the string from collapsing into $r=0$. The string starts at $\tau=0$ with maximal size $S^{(ii)}_{\mbox{max}}$ outside the static limit, it then contracts through the ergosphere and the two horizons and eventually collapses into a point $r=0.$ For $J\neq 0,$ it may be still possible to continue this solution into another universe like in the case (i). (iii) $J^2= 4E^2$ is the limiting case where the maximal string size equals the static limit: $S^{(iii)}_{\mbox{max}}=l\sqrt{M}.$ In this case $V(0)=0,$ thus the string contracts through the two horizons and eventually collapses into a point $r=0.$ The exact general solution in the three cases (i)-(iii) is given by: $$r(\tau)=\mid r_m-\frac{1}{c_1\wp(\tau-\tau_0)+c_2}\mid,$$ where: $$r_m=S_{\mbox{max}}=\sqrt{\frac{Ml^2}{2}}\sqrt{1+\sqrt{1-\frac{4V(0)}{Ml^2}}}, \;\;\;V(0)=\frac{J^2}{4}-E^2.$$ $c_1,c_2$ are constants in terms of $(l,M,r_m),$ given in Ref.[@all1], and $\wp$ is the Weierstrass elliptic $\wp$-function with invariants $(g_2,g_3),$ discriminant $\Delta$ and roots $(e_1,e_2,e_3),$ also given in Ref.[@all1]. The three cases (i)-(iii) correspond to the cases $\Delta>0,$ $\Delta<0$ and $\Delta=0,$ respectively. Notice that $S^{(ii)}_{\mbox{max}}>S^{(iii)}_{\mbox{max}}=l\sqrt{M}>S^{(i)}_{\mbox{max}}.$ In the case (i), $r(\tau)$ can be written in terms of the Jacobian elliptic function $\mbox{sn}[\tau^*,k],\;\tau^*=\sqrt{e_1-e_3}\;\tau,\;k=\sqrt{(e_2- e_3)/(e_1-e_3)}.$ It follows that the solution (i) oscillates between the two zeroes $r_{01}$ and $r_{02}$ of $V(r),$ with period $2\omega,$ where $\omega$ is the real semi-period of the Weierstrass function: $\omega=K(k)/\sqrt{e_1-e_3}\;,$ in terms of the complete elliptic integral of the first kind $K(k).$ We have: $$r(0)=r_m,\;\;\;r(\omega)=\sqrt{Ml^2-r_m^2},\;\;\;r(2\omega)=r_m,...$$ In the case (ii) ($\Delta<0$) two roots ($e_1,e_3$) become complex, the string collapses into a point $r=0$ and we have: $$r(0)=r_m,\;\;\;r(\frac{\omega_2}{2})=0,\;\;\;r(\omega_2)=r_m,...$$ where $\omega_2$ is the real semi-period of the Weierstrass function for this case. In the case (iii) ($\Delta=0$) the elliptic functions reduce to hyperbolic functions: $$r(\tau)=\frac{\sqrt{M}l}{\cosh(\sqrt{M}\tau)},$$ so that: $$r(-\infty)=0,\;\;\;r(0)=r_m=\sqrt{M}l,\;\;\;r(+\infty)=0.$$ Here, the string starts as a point, grows until $r=r_m$ (at $\tau=0$), and then it contracts until it collapses again ($r=0$) at $\tau=+\infty.$ In this case the string makes only one oscillation between $r=0$ and $r=r_m.$ Notice that for the static background ($J=0$), the only allowed motion is (ii), i.e. $r_m>r_{\mbox{hor}}=\sqrt{M}l$ (there is no ergosphere and only one horizon in this case), with: $$r_m=\sqrt{\frac{Ml^2}{2}}\sqrt{1+\sqrt{1+\frac{4E^2}{Ml^2}}}.$$ For $J=0,$ the string collapses into $r=0$ and stops there. The Penrose diagram of the 2+1 BH-AdS spacetime for $J=0$ is very similar to the Penrose diagram of the ordinary ($D\geq 4$) Schwarzschild spacetime, so the string motion outwards from $r=0$ is unphysical because of the causal structure. The coordinate time $t(\tau)$ can be expressed in terms of the incomplete elliptic integral of the third kind $\Pi,$ see Ref.[@all1]. The string has its maximal size $r_m$ at $\tau=0,$ passes the horizon at $\tau=\tau_{\mbox{hor}}$ (expressed in terms of an incomplete elliptic integral of the first kind) and falls into $r=0$ for $\tau= \omega_2/2,$ $\omega_2$ being the real semi-period of the Weierstrass function. That is, we have: $$\begin{aligned} &r(0)=r_m,\;\;\;r(\tau_{\mbox{hor}})=\sqrt{M}l,\;\;\;r(\omega_2/2) =0,&\nonumber\\ &t(0)=0,\;\;\;t(\tau_{\mbox{hor}})=\infty,&\end{aligned}$$ and $t(\omega_2/2)$ can be expressed in terms of the Jacobian zeta function $Z,$ see Ref.[@all1]. We also study the circular strings in the ordinary $D\geq 4$ spacetimes. In the generic 3+1 Kerr anti de Sitter (or Kerr de Sitter) spacetime, the potential $V(r)$ is however quite complicated, see Ref.[@all1]. The general circular string solution involves higher genus elliptic functions and it is not necessary to go into details here. We will compare with the non-rotating cases, only. It is instructive to recall [@vil4] the circular string in Minkowski spacetime (Min), for which $V(r)=r^2-E^2$ (Fig.2a), the string oscillates between its maximal size $r_m=E,$ and $r=0$ with the solution $r(\tau)=r_m\mid\cos\tau\mid.$ In the Schwarzschild black hole (S) $V(r)=r^2-2Mr-E^2$ (Fig.2b), the solution is remarkably simple: $r(\tau)=M+ \sqrt{M^2+E^2}\cos\tau.$ The mathematical solution oscillates between $r_m=M+\sqrt{M^2+E^2}$ and $M-\sqrt{M^2+E^2}<0,$ but because of the causal structure and the curvature singularity the motion can not be continued after the string has collapsed into $r=0.$ For anti de Sitter spacetime (AdS), we find: $V(r)=r^2(1+H^2r^2)-E^2$ (Fig.2c). The string oscillates between $r_m=\frac{1}{\sqrt{2}H}\sqrt{-1+\sqrt{1+4H^2E^2}}$ and $r=0$ with the solution: $$r(\tau)=r_m\mid\mbox{cn}[(1+4H^2E^2)^{1/4},k]\mid,$$ which is periodic with period $2\omega:$ $$\omega=\frac{K(k)}{(1+4H^2E^2)^{1/4}},\;\;\;\;k=\sqrt{\frac{\sqrt{1+4H^2E^2}-1} {2\sqrt{1+4H^2E^2}}}.$$ For Schwarzschild anti de Sitter spacetime (S-AdS), we find $V(r)= r^2(1+H^2r^2)-2Mr-E^2$ (Fig.2d) and: $$r(\tau)=r_m-\frac{1}{d_1\wp(\tau)+d_2},\;\;\;\;r(0)=r_m$$ $d_1,d_2$ are constants given in terms of $(M,H,r_m),$ see Ref.[@all1] ($r_m$ is the root of the equation $V(r)=0$, which has in this case exactly one positive solution). The invariants, the discriminant and the roots are also given explicitly in Ref.[@all1]. The string starts with $r=r_m$ at $\tau=0,$ it then contracts and eventually collapses into the $r=0$ singularity. The existence of elliptic function solutions for the string motion is characteristic of the presence of a cosmological constant. For $\Lambda=0=-3H^2$ the circular string motion reduces to simple trigonometric functions. From Fig.2 and our analysis we see that the circular string motion is qualitatively very similar in all these backgrounds (Min, S, AdS, S-AdS): the string has a maximal [*bounded*]{} size and then it contracts towards $r=0.$ There are however physical and quantitative differences: in Minkowski and pure anti de Sitter spacetimes, the string truly oscillates between $r_m$ and $r=0,$ while in the black hole cases (S, S-AdS), there is only one half oscillation, the string motion stops at $r=0.$ This also holds for the 2+1 BH-AdS spacetime with $J=0$ (Fig.1b). Notice also that in all these cases, $V(0)=-E^2<0$ and $V(r)\sim r^\alpha;\;(\alpha=2,4)$ for $r>>E.$ The similarity can be pushed one step further by considering small perturbations around the circular strings. We find [@all]: $$\ddot{C}_{n}+(n^2+\frac{r}{2}\frac{da(r)}{dr}+\frac{r^2}{2}\frac{d^2a(r)} {dr^2}-\frac{2E^2}{r^2})C_n=0,$$ determining the Fourier components of the comoving perturbations. For the spacetimes of interest here, $a(r)=1-2M/r+H^2r^2$ (Min, S, AdS, S-AdS), or $a(r)=r^2/l^2-1$ (2+1 BH-AdS), the comoving perturbations are regular except near $r=0,$ where we find (for all cases) $r(\tau)\approx-E(\tau-\tau_0)$ and: $$\ddot{C}_{n}-\frac{2}{(\tau-\tau_0)^2}C_n=0.$$ It follows that not only the unperturbed circular strings, but also the comoving perturbations around them behave in a similar way in all these non-rotating backgrounds (2+1 and higher dimensional). This should be contrasted with the string perturbations around the center of mass, which behave differently in these backgrounds. It must be noticed that for rotating ($J\neq 0$) spacetimes, the circular string behaviour is qualitatively different from the non-rotating ($J=0$) spacetimes. For large $J,$ both in the 2+1 BH-AdS as well as in the 3+1 ordinary Kerr-AdS spacetimes, non-collapsing circular string solutions exist. The potential $V(r)\rightarrow +\infty$ for $r\rightarrow 0$ and no collapse into $r=0$ is possible. The dynamics of circular strings in curved spacetimes is determined by the string tension, which tends to contract the string, and by the local gravity (which may be attractive or repulsive). In all the previous non-rotating backgrounds, the local gravity is attractive (i.e. $da(r)/dr>0$), and it acts together with the string tension in the sense of contraction. But in spacetimes with regions in which repulsion (i.e. $da(r)/dr<0$) dominates, the strings can expand with unbounded radius (unstable strings [@mic; @veg4]). It may also happen that the string tension and the local gravity be of the same order, i.e. the two opposite effects can balance, and the string is stationary. De Sitter spacetime provides an example in which all such type of solutions exist [@mic; @veg4]. In de Sitter spacetime, $V(r)$ is unbounded from below for $r\rightarrow\infty$ ($V(r)\sim -r^4$) and unbounded expanding circular strings are present. In addition, an interesting new feature appears in the presence of a positive cosmological constant: the existence of [*multi-string*]{} solutions \[9-11\]. The world-sheet time $\tau$ turns out to be a multi (finite or infinite) valued function of the physical time. That is, one single world-sheet where $-\infty\leq\tau\leq+\infty,$ can describe many (even infinitely many [@veg4]) different and independent strings (in flat spacetime, one single world-sheet describes only one string). In the S, AdS and S-AdS spacetimes, the multi-string feature for circular strings is absent. The main conclusions of this section are given in Tables 1, 2. Further details can be found in Ref.[@all1] Quantum String Spectrum in Ordinary AdS ======================================= In the first part of the previous section, the classical string dynamics was studied in the $2+1$ BH-AdS spacetime. In this section we go further in the investigation of the physical properties of strings in AdS spacetime, by performing string quantization. Since the $2+1$ BH-AdS spacetime is locally AdS, the results of Section 2 can be extended to classical strings in $D$-dimensional AdS spacetime as well. In AdS spacetime, the string motion is oscillatory in time and is [*stable*]{}; all fluctuations around the string center of mass are well behaved and bounded. Local gravity of AdS spacetime is always negative and string instabilities do not develop. The string perturbation series approach, considering fluctuations around the center of mass, is particularly appropriate in AdS spacetime, the natural dimensionless expansion parameter being $\lambda=\alpha'/l^2>0,$ where $\alpha'$ is the string tension and the ‘Hubble constant’ $H=1/l.$ The negative cosmological constant of AdS spacetime is related to the ‘Hubble constant’ $H$ by $\Lambda=-(D-1)(D-2)H^2/2.$ All the spatial ($\mu=1,...,D-1$) modes in $D$-dimensional AdS, oscillate with frequency $\omega_n=\sqrt{n^2+m^2\alpha'\lambda},$ which are real for all $n$ ($m$ being the string mass). In this section, we perform a canonical quantization procedure. From the conformal generators $L_n,\;\tilde{L}_n$ and the constraints $L_0=\tilde{L}_0=0,$ imposed at the quantum level, we obtain the mass formula (the quantum version of Eq.(2.7) written in units where $\alpha'$ appears explicitly): $$m^2\alpha'=(D-1)\sum_{n>0}\Omega_n(\lambda) +\sum_{n>0}\Omega_n(\lambda)\sum_{R=1}^{D-1}[(\alpha^R_n)^{\dag} \alpha^R_n+(\tilde{\alpha}^R_n)^{\dag}\tilde{\alpha}^R_n],$$ where: $$\Omega_n(\lambda)=\frac{2n^2+m^2\alpha'\lambda}{\sqrt{n^2+m^2\alpha'\lambda}},$$ and we have applied symmetric ordering of the operators. The operators $\alpha^R_n,\;\tilde{\alpha}^R_n$ satisfy: $$[\alpha^R_n,\;(\alpha^R_n)^{\dag}]=[\tilde{\alpha}^R_n,\;(\tilde{\alpha}^R_n)^ {\dag}]=1,\;\;\;\;\;\;\mbox{for all}\;\;n>0.$$ For $\lambda<<1,$ which is clearly fulfilled in most interesting cases, we have found the lower mass states $m^2\alpha'\lambda<<1$ and the quantum mass spectrum [@all2]. Physical states are characterized by the eigenvalue of the number operator: $$N=\frac{1}{2}\sum_{n>0} n\sum_{R=1}^{D-1}[(\alpha^R_n)^{\dag} \alpha^R_n+(\tilde{\alpha}^R_n)^{\dag}\tilde{\alpha}^R_n],$$ and the ground state is defined by: $$\alpha^R_n\mid 0>=\tilde{\alpha}^R_n\mid 0>=0,\;\;\;\;\;\; \mbox{for all}\;\;n>0.$$ We find that $m^2\alpha'=0$ is an [*exact*]{} solution of the mass formula in $D=25$ and that there is a graviton at $D=25,$ which indicates, as in de Sitter space [@veg1], that the critical dimension of AdS is 25 (although it should be stressed that the question whether these spacetimes are solutions to the full $\beta$-function equations remains open). As in Minkowski spacetime, the ground state is a tachyon. Remarkably enough, for $N\geq 2$ we find that a generic feature of all excited states beyond the graviton, is the presence of a [*fine structure*]{} effect: for a given eigenvalue $N\geq 2,$ the corresponding states have different masses. For the lower mass states the expectation value of the mass operator in the corresponding states (generically labelled $\mid {j}>$) turns out to have the form (see Ref.[@all2]): $$<{j}\mid m^2\alpha' \mid {j}>_{\mbox{AdS}}=a_{{j}} +b_{{j}}\lambda^2+c_{{j}}\lambda^3+{\cal O}(\lambda^4).$$ The collective index “j” generically labels the state $\mid {j}>$ and the coefficients $a_{{j}},\;b_{{j}},\;c_{{j}}$ are all well computed numbers, different for each state, their precise values are given in Ref.[@all2]. The corrections to the mass in Minkowski spacetime appear to order $\lambda^2.$ Therefore, the leading Regge trajectory for the lower mass states is: $$J=2+\frac{1}{2}m^2\alpha'+{\cal O}(\lambda^2).$$ In Minkowski spacetime the mass and number operator of the string are related by $m^2\alpha'=-4+4N.$ In AdS (as well as in de Sitter (dS)), there is no such simple relation between the mass and the number operators; the splitting of levels increases considerably for very large $N.$ The [*fine structure*]{} effect we find here is also present in de Sitter space. Up to order $\lambda^2,$ the lower mass states in dS and in AdS are the same, the differences appear to the order $\lambda^3.$ The lower mass states in de Sitter spacetime are given by Eq.(3.6) but with the $\lambda^3$-term getting an opposite sign ($-c_{{j}}\lambda^3$). For the very high mass spectrum, we find more drastic effects. States with very large eigenvalue $N,$ namely $N>>1/\lambda,$ have masses: $$<j\mid m^2\alpha' \mid j>_{\mbox{AdS}}\approx d_{{j}}\lambda N^2$$ and angular momentum: $$J^2\approx\frac{1}{\lambda}m^2\alpha',$$ where $d_{{j}}$ are well computed numbers different for each state, their precise values are given in Ref.[@all2]. Since $\lambda=\alpha'/l^2,$ we see from Eq.(3.8) that the masses of the high mass states are [*independent*]{} of $\alpha'.$ In Minkowski spacetime, very large $N$ states all have the same mass $m^2\alpha'\approx 4N,$ but here in AdS the masses of the high mass states with the same eigenvalue $N$ are [*different*]{} by factors $d_{{j}}.$ In addition, because of the fine structure effect, states with different $N$ can get mixed up. For high mass states, the level spacing [*grows*]{} with $N$ (instead of being constant as in Minkowski spacetime). As a consequence, the density of states $\rho(m)$ as a function of mass grows like $\mbox{Exp}[(m/\sqrt{\mid\Lambda\mid}\;)^{1/2}]$ (instead of $\mbox{Exp}[m\sqrt{\alpha'}]$ as in Minkowski spacetime), and [*independently*]{} of $\alpha'.$ The partition function for a gas of strings at a temperature $\beta^{-1}$ in AdS spacetime is well defined for all finite temperatures $\beta^{-1},$ discarding the existence of the Hagedorn temperature. For the low mass states ($m^2\alpha'\lambda<<1$) in anti de Sitter spacetime, our results can be written as [@all2]: $$<j\mid m^2(\alpha',l)\mid j>=\frac{4(N-1)}{\alpha'}+\frac{1}{l^2} \sum_{n=0}^{\infty}a_{jn}(\alpha'/l^2)^n,$$ where “$j$” is a collective index labelling the state $\mid j>.$ It is now important to notice that $a_{j0}=0$ for [*all*]{} the low mass states [@all2], i.e. there is no “constant” term on the right hand side of Eq.(3.10). A non-zero $a_{j0}$-term would give rise to a $\alpha'$-independent contribution to the string mass. Its absense, on the other hand, means that the first term on the right hand side of Eq.(3.10) is super-dominant (since, in all cases, $\alpha'/l^2=\lambda<<1$) and that the string scale is therefore set by $1/\alpha'.$ For the high mass states ($m^2\alpha'\lambda>>1$) we found instead [@all2]: $$<j\mid m^2(\alpha',l)\mid j>\approx\frac{d_j}{l^2}N^2,\;\;\;\;\;\;\mbox{for}\; N>>l^2/\alpha'$$ where the number $d_j$ depends on the state. The right hand side of Eq.(3.11) is exactly like a non-zero dominant $a_{j0}$-term in Eq.(3.10). For the high mass states the scale is therefore set by $1/l^2$ which is equal to the absolute value of the cosmological constant $\Lambda$ (up to a geometrical factor) and [*independent*]{} of $\alpha'.$ This suggests that for $\lambda<<1,$ the masses of [*all*]{} string states can be represented by a formula of the form (3.10). For the low mass states $a_{j0}=0,$ while for the high mass states $a_{j0}$ becomes a large positive number. In the black string background, Eq.(2.2), we have calculated explicitly the first and second order string fluctuations around the center of mass [@all2]. We then determined the world-sheet energy-momentum tensor and we derived the mass formula in the asymptotic region. The mass spectrum is equal to the mass spectrum in flat Minkowski spacetime. Therefore, for a gas of strings at temperature $\beta^{-1}$ in the asymptotic region of the black string background, the partition function goes like: $$Z(\beta)\sim\int^{\infty} dm\; e^{-m(\beta-\sqrt{\alpha'}\;)},$$ which is only defined for $\beta>\sqrt{\alpha'},$ i.e. there is a Hagedorn temperature: $$T_{\mbox{Hg}}=(\alpha')^{-1/2}.$$ In higher dimensional ($D\geq 4$) black hole spacetimes the next step now would be to set up a scattering formalism, where a string from an asymptotic in-state interacts with the gravitational field of the black hole and reappears in an asymptotic out-state [@san2]. However, this is not possible in the black string background. In the uncharged black string background under consideration here, [*all*]{} null and timelike geodesics incoming from spatial infinity pass through the horizon and fall into the physical singularity [@hor2]. No “angular momentum”, as in the case of scattering off the ordinary Schwarzschild black hole, can prevent a point particle from falling into the singularity. The string solutions considered here are based on perturbations around the string center of mass which follows, at least approximately, a point particle geodesic. Therefore, a string incoming from spatial infinity inevitably falls into the singularity in the black string background. New Classes of Exact String and Multi-String Solutions in Curved Spacetimes =========================================================================== In this section we return to the string equations of motion and constraints, Eqs.(1.1)-(1.2), to look for new classes of exact solutions. In most spacetimes, quite general families of exact solutions can be found by making an appropriate ansatz, which exploits the symmetries of the underlying curved spacetime. In axially symmetric spacetimes, a convenient ansatz corresponds to circular strings, as we saw in Section 2. Such an ansatz effectively decouples the dependence on the spatial world-sheet coordinate $\sigma,$ and the string equations of motion and constraints reduce to non-linear coupled ordinary differential equations, Eqs.(2.18)-(2.19). In this section we will make instead an ansatz which effectively decouples the dependence on the temporal world-sheet coordinate $\tau.$ This ansatz, which we call the “stationary string ansatz” is dual to the “circular string ansatz” in the sense that it corresponds to a formal interchange of the world-sheet coordinates $(\tau,\sigma),$ as well as of the azimuthal angle $\phi$ and the stationary time $t$ in the target space: $$\tau\;\leftrightarrow\;\sigma,\;\;\;\;\;\;\;\;t\;\leftrightarrow\;\phi.$$ The stationary string ansatz will describe stationary strings when $t$ (and $\tau$) are timelike, for instance in anti de Sitter spacetime (in static coordinates) and outside the horizon of a Schwarzschild black hole. On the other hand, if $t$ (and $\tau$) are spacelike, for instance inside the horizon of a Schwarzschild black hole or outside the horizon of de Sitter spacetime (in static coordinates), the stationary string ansatz will describe dynamical propagating strings. Considering for simplicity a static line element in the form: $$ds^2=-a(r)dt^2+\frac{dr^2}{a(r)}+r^2(d\theta^2+\sin^2\theta d\phi^2),$$ the stationary string ansatz reads explicitly: $$t=\tau,\;\;\;\;\;r=r(\sigma),\;\;\;\;\;\phi=\phi(\sigma),\;\;\;\;\;\theta= \pi/2.$$ The string equations of motion and constraints, Eqs.(1.1)-(1.2), reduce to two separated first order ordinary differential equations: $$\phi'=\frac{L}{r^2},\;\;\;\;\;\;r'^2+V(r)=0;\;\;\;\; V(r)=-a(r)[a(r)-\frac{L^2}{r^2}],$$ where $L$ is an integration constant. The qualitative features of the possible string configurations can be read off directly from the shape of the potential $V(r).$ Thereafter, the detailed analysis of the quantitative features can be performed by explicitly solving the (integrable) system of equations (4.4). The induced line element on the world-sheet is given by: $$ds^2=a(r)(-d\tau^2+d\sigma^2).$$ Thus, if $a(r)$ is negative, the world-sheet coordinate $\tau$ becomes spacelike while $\sigma$ becomes timelike and the stationary string ansatz (4.3) describes a dynamical string. If $a(r)$ is positive, the ansatz describes a stationary string. In this section we solve explicitly Eqs.(4.4) and we analyze in detail the solutions and their physical interpretation in Minkowski, de Sitter, anti de Sitter, Schwarzschild and $2+1$ black hole anti de Sitter spacetimes. In all these cases, the solutions are expressed in terms of elliptic (or elementary) functions. We furthermore analyze the physical properties, energy and pressure, of these solutions. The energy and pressure densities of the strings can be obtained from the $3+1$ dimensional spacetime energy-momentum tensor: $$\sqrt{-g}T^{\mu\nu}=\frac{1}{2\pi\alpha'}\int d\tau d\sigma (\dot{X}^\mu\dot{X}^\nu-X'^\mu X'^\nu)\delta^{(4)}(X-X(\tau,\sigma)).$$ After integration over a spatial volume that completely encloses the string [@energy], the energy-momentum tensor takes the form of a fluid: $$T^\mu\;_\nu=\mbox{diag.}(-E,P_1,P_2,P_3).$$ In Minkowski spacetime (Min), the potential defined in Eqs.(4.4) is given by: $$V_{\mbox{M}}(r)=\frac{L^2}{r^2}-1,$$ and the solution of Eqs.(4.4) describes one infinitely long straight string with “impact-parameter” $L.$ The equation of state, relating energy and pressure densities, takes the well-known form Ref.[@all3]: $$dE=-dP_2=\mbox{const.},\;\;P_1=P_3=0.$$ In anti de Sitter spacetime (AdS), the potential is given by: $$V_{\mbox{AdS}}(r)=-(1+H^2r^2)[1+H^2r^2-\frac{L^2}{r^2}].$$ The radial coordinate $r(\sigma)$ is periodic with finite period $T_\sigma,$ which can be expressed in terms of a complete elliptic integral, see Ref.[@all3]. For $\sigma\in[0,T_\sigma],$ the solution describes an infinitely long stationary string in the wedge $\phi\in\;]0,\Delta\phi[\;,$ where: $$\Delta\phi=2k\sqrt{\frac{1-2k^2}{1-k^2}}[\Pi(1-k^2,\;k)-K(k)]\;\in\;\;]0,\;\pi[$$ The elliptic modulus $k,$ which is a function of $HL$ [@all3], parametrizes the solutions, $k\in\;]0,1/\sqrt{2}[\;.$ The azimuthal angle is generally not a periodic function of $\sigma,$ thus when the spacelike world-sheet coordinate $\sigma$ runs through the range $]-\infty,+\infty[\;,$ the solution describes an [*infinite number*]{} of infinitely long stationary open strings. The general solution is therefore a [*multi-string*]{} solution. Until now multi-string solutions were only found in de Sitter spacetime \[11-13\]. Our results show that multi-string solutions are a general feature of spacetimes with a cosmological constant (positive or negative). The solution in anti de Sitter spacetime describes a [*finite*]{} number of strings if the following relation holds: $$N\Delta\phi=2\pi M.$$ Here $N$ and $M$ are integers, determining the number of strings and the winding in azimuthal angle, respectively, for the multi-string solution, see Fig.3. The equation of state for a full multi-string solution takes the form $(P_3=0)$: $$dP_1=dP_2=-\frac{1}{2}dE,\;\;\;\;\;\mbox{for}\;\;r\rightarrow\infty$$ corresponding to extremely unstable strings [@ven]. In de Sitter spacetime (dS), the potential is given by: $$V_{\mbox{dS}}(r)=-(1-H^2r^2)[1-H^2r^2-\frac{L^2}{r^2}].$$ In this case we have to distinguish between solutions inside the horizon (where $\tau$ is timelike) and solutions outside the horizon (where $\tau$ is spacelike). Inside the horizon, the generic solution describes one infinitely long open stationary string winding around $r=0.$ For special values of the constants of motion, corresponding to a relation, which formally takes the same form as Eq.(4.12), the solution describes a closed string of finite length $l=N\pi/H.$ The integer $N$ in this case determines the number of “leaves”, see Fig.4. The energy is positive and finite and grows with $N.$ The pressure turns out to vanish identically, thus the equation of state corresponds to cold matter. Outside the horizon, the world-sheet coordinate $\tau$ becomes spacelike while $\sigma$ becomes timelike, thus we define: $$\tilde{\tau}\equiv\sigma,\;\;\;\;\;\;\;\;\tilde{\sigma}\equiv\tau,$$ and the string solution is conveniently expressed in hyperboloid coordinates. The radial coordinate $r(\tilde{\tau})$ is periodic with a finite period $T_{\tilde{\tau}}.$ For $\tilde{\tau}\in[0,T_{\tilde{\tau}}],$ the solution describes a straight string incoming non-radially from spatial infinity, scattering at the horizon and escaping towards infinity again, Fig.5. The string length is zero at the horizon and grows indefinetely in the asymptotic regions. As in the case of anti de Sitter spacetime, the azimuthal angle is generally not a periodic function, thus when the timelike world-sheet coordinate $\tilde{\tau}$ runs through the range $]-\infty,+\infty[\;,$ the solution describes an [*infinite*]{} number of dynamical straight strings scattering at the horizon at different angles. The general solution is therefore a multi-string solution. In particular, a multi-string solution describing a [*finite*]{} number of strings is obtained if a relation of the form (4.12) is fulfilled. It turns out that the solution describes [*at least*]{} three strings. The energy and pressures of a full multi-string solution have also been computed. In the asymptotic region they fulfill an equation of state corresponding to extremely unstable strings, i.e. like Eq.(4.13). In the Schwarzschild black hole background (S), the potential is given by: $$V_{\mbox{S}}(r)=-(1-2m/r)[1-2m/r-\frac{L^2}{r^2}].$$ No multi-string solutions are found in this case. Outside the horizon the solution of Eqs.(4.4) describes one infinitely long stationary open string. This solution was already derived in Ref.[@fro1], and we shall not discuss it here. Inside the horizon, where $\tau$ becomes spacelike while $\sigma$ becomes timelike, we make the redefinitions (4.15) and the solution is conveniently expressed in terms of Kruskal coordinates, see Ref.[@all3]. The solution describes one straight string infalling [*non-radially*]{} towards the singularity, see Fig.6. At the horizon, the string length is zero and it grows [*indefinitely*]{} when the string approaches the spacetime singularity. Thereafter, the solution can not be continued, see Ref.[@all3]. In the $2+1$ black hole anti de Sitter spacetime (BH-AdS) [@ban1], the potential is given by: $$V_{\mbox{BH-AdS}}(r)=-(\frac{r^2}{l^2}-1)[\frac{r^2}{l^2}-1-\frac{L^2}{r^2}].$$ Outside the horizon, the solutions “interpolate” between the solutions found in anti de Sitter spacetime and outside the horizon of the Schwarzschild black hole. The solutions thus describe infinitely long stationary open strings. As in anti de Sitter spacetime, the general solution is a [*multi-string*]{} describing [*infinitely*]{} many strings. In particular, for certain values of the constants of motion, corresponding to the condition of the form (4.12), the solution describes a finite number of strings. In the simplest version of the $2+1$ BH-AdS background ($M=1,\;J=0$), it turns out that the solution describes [*at least*]{} seven strings, see Ref.[@all3]. Inside the horizon, we make the redefinitions (4.15) and the solution is conveniently expressed in terms of Kruskal-like coordinates, see Ref.[@all3]. The solution is similar to the solution found inside the horizon of the Schwarzschild black hole, but there is one important difference at $r=0.$ As in the Schwarzschild black hole background, the solution describes one straight string infalling non-radially towards $r=0,$ and beyond this point the solution can not be continued because of the global structure of the spacetime. At the horizon the string size is zero and during the fall towards $r=0,$ the string size [*grows*]{} but stays [*finite*]{}. This should be compared with the straight string inside the horizon of the Schwarzschild black hole, where the string size grows indefinetely. The physical reason for this difference is that the point $r=0$ is not a strong curvature singularity in the $2+1$ BH-AdS spacetime. We have considered also the Schwarzschild de Sitter and Schwarzschild anti de Sitter spacetimes. These spacetimes contain all the features of the spacetimes already discussed: singularities, horizons, positive or negative cosmological constants. All the various types of string solutions found in the other spacetimes (open, closed, straight, finitely and infinitely long, multi-strings) are therefore present in the different regions of the Schwarzschild-de Sitter and Schwarzschild-anti de Sitter spacetimes. The details are given in Ref.[@all3]. We close this section with a few remarks on the stability of the solutions. Generally, the question of stability must be addressed by considering small perturbations around the exact solutions. In Ref.[@fro], a covariant formalism describing physical perturbations propagating along an arbitrary string configuration embedded in an arbitrary curved spacetime, was developed. The resulting equations determining the evolution of the perturbations are however very complicated in the general case, although partial (analytical) results have been obtained in special cases for de Sitter [@all; @all6] and Schwarzschild black hole [@fro; @all6; @all7] spacetimes. The exact solutions found in this section fall essentially into two classes: dynamical and stationary. The dynamical string solutions outside the horizon of de Sitter (or S-dS) and inside the horizon of Schwarzschild (or S-dS, S-AdS) spacetimes, are already unstable at the zeroth order approximation (i.e. without including small perturbations), in the sense that their physical length grows indefinetely. For the stationary string solutions the situation is more delicate. The existence of the stationary configurations is based on an exact balance between the string tension and the local attractive or repulsive gravity. For that reason, it can be expected that the configurations are actually unstable for certain modes of perturbation, especially in strong curvature regions. The main conclusions of this section are given in Table 4. Further details can be found in Ref.[@all3]. Spatial Curvature Effects on the String Dynamics ================================================ The propagation of strings in Friedmann-Robertson-Walker (FRW) cosmologies has been investigated using both exact and approximative methods, see for example Refs.\[7, 11-14, 17-20\] (as well as numerical methods, which shall not be discussed here). Except for anti de Sitter spacetime, which has negative spatial curvature, the cosmologies that have been considered until now, have all been spatially flat. In this section we will consider the physical effects of a non-zero (positive or negative) curvature index on the classical and quantum strings. The non-vanishing components of the Riemann tensor for the generic D-dimensional FRW line element, in comoving coordinates: $$ds^2=-dt^2+a^2(t)\frac{d\vec{x} d\vec{x}}{(1+\frac{K}{4}\vec{x}\vec{x})^2},$$ are given by: $$R_{itit}=\frac{-aa_{tt}}{(1+\frac{K}{4}\vec{x}\vec{x})^2},\;\;\;\;\;\; R_{ijij}=\frac{a^2(K+a_{t}^2)}{(1+\frac{K}{4}\vec{x}\vec{x})^4};\;\;\;\;i\neq j$$ where $a=a(t)$ is the scale factor and $K$ is the curvature index. Clearly, a non-zero curvature index introduces a non-zero spacetime curvature; the exceptional case provided by $K=-a_{t}^2=\mbox{const.},$ corresponds to the Milne-Universe. From Eqs.(5.2), it is also seen that the curvature index has to compete with the first derivative of the scale factor. The effects of the curvature index are therefore most conveniently discussed in the family of FRW-universes with constant scale factor, the so-called static Robertson-Walker spacetimes. This is the point of view we take in the present section. We consider both the closed ($K>0$) and the hyperbolic ($K<0$) static Robertson-Walker spacetimes, and all our results are compared with the already known results in the flat ($K=0$) Minkowski spacetime. We determine the evolution of circular strings, derive the corresponding equations of state (using Eq.(4.6)), discuss the question of strings as self-consistent solutions to the Einstein equations [@self], and we perform a semi-classical quantization. We also find all the stationary string configurations in these spacetimes. The radius of a classical circular string in the spacetime (5.1), for $a=1,$ is determined by: $$\dot{r}^2+V(r)=0;\;\;\;\;\;\;\;\;V(r)=(1-Kr^2)(r^2-b\alpha'^2),$$ where $b$ is an integration constant ($b\alpha'^2=E,$ in the notation of Eqs.(2.18)-(2.19)). This equation is solved in terms of elliptic functions and all solutions describe oscillating strings (Fig.7 shows the potential $V(r)$ for $K>0,\;K=0,\;K<0$). For $K>0,$ when the spatial section is a hypersphere, the string either oscillates on one hemisphere or on the full hypersphere. The energy is always positive while the average pressure can be positive, negative or zero, depending on the precise value of an elliptic modulus; the equation of state is given explicitly in Table 5, in the different cases. Interestingly enough, we find that the circular strings provide a self-consistent solution to the Einstein equations with a selected value of the curvature index. Self-consistent solutions to the Einstein equations with string sources have been found previously in the form of power law expanding universes [@self]. We have semi-classically quantized the circular string solutions of Eq.(5.3). We used an approach developed in field theory by Dashen et. al. [@das], based on the stationary phase approximation of the partition function. The result of the stationary phase integration is expressed in terms of the function: $$W(m)\equiv S_{\mbox{cl}}(T(m))+m\;T(m),$$ where $S_{\mbox{cl}}$ is the action of the classical solution, $m$ is the mass and the period $T(m)$ is implicitly given by: $$\frac{dS_{\mbox{cl}}}{dT}=-m.$$ The bound state quantization condition then becomes [@das]: $$W(m)=2\pi n,\quad n \in N_{0}$$ for $n$ ‘large’. Parametric plots of $K\alpha'W$ as a function of $K\alpha'^2m^2$ in the different cases, are shown in Fig.8. The strings oscillating on one hemisphere give rise to a finite number $N_-$ of states with the following mass-formula Ref.[@all4]: $$m_-^2\alpha'\approx \pi\;n,\;\;\;\;\;\;\;\;N_-\approx\frac{4}{\pi K\alpha'}.$$ As in flat Minkowski spacetime, the scale of these string states is set by $\alpha'.$ The strings oscillating on the full hypersphere give rise to an infinity of more and more massive states with the asymptotic mass-formula Ref.[@all4]: $$m_+^2\approx K\;n^2.$$ The masses of these states are independent of $\alpha',$ the scale is set by the curvature index $K.$ Notice also that the level spacing grows with $n.$ A similar result was found recently for strings in anti de Sitter spacetime [@all5]. For $K<0,$ when the spatial section is a hyperboloid, both the energy and the average pressure of the oscillating strings are positive. The equation of state is given in Table 5. In this case, the strings can not provide a self-consistent solution to the Einstein equations. After semi-classical quantization, we find an infinity of more and more massive states. The mass-formula is given by (see Ref.[@all4]): $$\sqrt{-Km^2\alpha'^2}\;\log\sqrt{-Km^2\alpha'^2}\approx -\frac{\pi}{2}K\alpha'\;n$$ Notice that the level spacing grows faster than in Minkowski spacetime but slower than in the closed static Robertson-Walker spacetime. On the other hand, the stationary strings are determined by Eqs.(4.4): $$\phi'=\frac{L}{r^2},\;\;\;\;\;\;\;\; r'^2+V(r)=0;\;\;\;V(r)=(1-Kr^2)(\frac{L^2}{r^2}-1),$$ where $L$ is an integration constant (Fig.9 shows the potential $V(r)$ for $K>0,\;K=0,\;K<0$). For $K>0,$ all the stationary string solutions describe circular strings winding around the hypersphere. The equation of state is of the extremely unstable string type [@ven]. For $K<0,$ the stationary strings are represented by infinitely long open configurations with an angle between the two “arms” given by: $$\Delta\phi=\pi-2\arctan(\sqrt{-K}L).$$ The energy density is positive while the pressure densities are negative. No simple equation of state is found for these solutions. We also computed the first and second order fluctuations around a static string center of mass, using the string perturbation series approach [@veg1] and its covariant versions [@men; @all1]. Up to second order, the mass-formula for arbitrary values of the curvature index (positive or negative) is identical to the well-known flat spacetime mass-formula; all dependence on $K$ cancels out. Conclusion ========== We first studied the string propagation in the 2+1 BH-AdS background. We found the first and second order perturbations around the string center of mass as well as the mass formula, and compared with the ordinary black hole AdS spacetime and with the black string background as well. These results were then generalized to ordinary $D$-dimensional anti de Sitter spacetime. The classical string motion in anti de Sitter spacetime is stable in the sense that it is oscillatory in time with real frequencies and the string size and energy are bounded. Quantum mechanically, this reflects in the mass operator, which is well defined for any value of the wave number $n,$ and arbitrary high mass states (and therefore an infinite number of states) can be constructed. This is to be contrasted with de Sitter spacetime, where string instabilities develop, in the sense that the string size and energy become unbounded for large de Sitter radius. For low mass states (the stable regime), the mass operator in de Sitter spacetime is given by Eq.(3.1) but with $$\begin{aligned} \Omega_n(\lambda)_{\mbox{dS}}= \frac{2n^2-m^2\alpha'\lambda}{\sqrt{n^2-m^2\alpha'\lambda}}. \nonumber\end{aligned}$$ Real mass solutions can be defined only up to some [*maximal mass*]{} of the order $m^2\alpha'\approx 1/\lambda\;$ [@veg1]. For $\lambda<<1,$ real mass solutions can be defined only for $N\leq N_{\mbox{max}}\sim 0.15/\lambda$ (where $N$ is the eigenvalue of the number operator) and therefore there exists a [*finite*]{} number of states only. These features of strings in de Sitter spacetime have been recently confirmed within a different (semi-classical) quantization approach based on [*exact*]{} circular string solutions [@all5]. The presence of a cosmological constant $\Lambda$ (positive or negative) increases considerably the number of levels of different eigenvalue of the mass operator (there is a splitting of levels) with respect to flat spacetime. That is, a non-zero cosmological constant [*decreases*]{} (although does not remove) the degeneracy of the string mass states, introducing a [*fine structure effect*]{}. For the low mass states the level spacing is approximately constant (up to corrections of the order $\lambda^2$). For the high mass states, the changes are more drastic and they depend crucially on the sign of $\Lambda.$ A value $\Lambda<0$ causes the [*growing*]{} of the level spacing linearly with $N$ instead of being constant as in Minkowski space. Consequently, the density of states $\rho(m)$ grows with the exponential of $\sqrt{m}$ (instead of $m$ as in Minkowski space) discarding the existence of a Hagedorn temperature in AdS spacetime, and the possibility of a phase transition. In addition, another important feature of the high mass string spectrum in AdS spacetime is that it becomes independent of $\alpha'.$ The string scale for the high mass states is given by $\mid\Lambda\mid,$ instead of $1/\alpha'$ for the low mass states. We have found new classes of exact string solutions obtained either by the circular string ansatz, Eq.(2.16), or by the stationary string ansatz, Eq.(4.3), in a variety of curved backgrounds including Schwarzschild, de Sitter and anti de Sitter spacetimes. Many different types of solutions have been found: oscillating circular strings, closed stationary strings, infinitely long stationary strings, dynamical straight strings and multi-string solutions describing finitely or infinitely many stationary or dynamical strings. In all the cases we have obtained the exact solutions in terms of either elementary or elliptic functions. Furthermore, we have analyzed the physical properties (length, energy, pressure) of the string solutions. Finally, we solved the equations of motion and constraints for circular strings in static Robertson-Walker spacetimes. We computed the equations of state and found that there exists a self-consistent solution to the Einstein equations in the case of positive spatial curvature. The solutions were quantized semi-classically using the stationary phase approximation method, and the resulting spectra were analyzed and discussed. We have also found all stationary string configurations in these spacetimes and we computed the corresponding physical quantities, string length, energy and pressure.\ A.L. Larsen is supported by the Danish Natural Science Research Council under grant No. 11-1231-1SE [11]{} H.J. de Vega and N. Sánchez, Phys. Rev. D42 (1990) 3969.\ H.J. de Vega, M. Ramón-Medrano and N. Sánchez, Nucl. Phys. B374 (1992) 405. H.J. de Vega and N. Sánchez, Phys. Lett. B244 (1990) 215, Phys. Rev. Lett. 65C (1990) 1517, IJMP A7 (1992) 3043, Nucl. Phys. B317 (1989) 706 and ibid 731.\ D. Amati and C. Klimcik, Phys. Lett. B210 (1988) 92.\ M. Costa and H.J. de Vega, Ann. Phys. 211 (1991) 223 and ibid 235.\ C. Loustó and N. Sánchez, Phys. Rev. D46 (1992) 4520. V.E. Zakharov and A.V. Mikhailov, JETP 47 (1979) 1017.\ H. Eichenherr, in “Integrable Quantum Field Theories”, ed. J. Hietarinta and C. Montonen (Springer, Berlin, 1982). H.J. de Vega and N. Sánchez, Phys. Rev. D47 (1993) 3394. I. Bars and K. Sfetsos, Mod. Phys. Lett. A7 (1992) 1091. H.J. de Vega, J.R. Mittelbrunn, M.R. Medrano and N. Sánchez, Phys. Lett. B232 (1994) 133 and “The Two-Dimensional Stringy Black Hole: a new Approach and a Pathology”, PAR-LPTHE 93/14. H.J. de Vega and N. Sánchez, Phys. Lett. B197 (1987) 320. H.J. de Vega and N. Sánchez, Nucl. Phys. B299 (1988) 818. P.F. Mende, in “String Quantum Gravity and Physics at the Planck Energy Scale”. Proceedings of the Erice Workshop held in June 1992, ed. N. Sánchez (World Scientific, 1993). A.L. Larsen and V.P. Frolov, Nucl. Phys. B414 (1994) 129. H.J. de Vega, A.V. Mikhailov and N. Sánchez, Teor. Mat. Fiz. 94 (1993) 232, Mod. Phys. Lett. A29 (1994) 2745. F. Combes, H.J. de Vega, A.V. Mikhailov and N. Sánchez, Phys. Rev. D50 (1994) 2754. H.J. de Vega, A.L. Larsen and N. Sánchez, Nucl. Phys. B427 (1994) 643. I. Krichever, “Two-Dimensional Algebraic-Geometrical Operators with Self-Consistent Potentials”, Landau Institute Preprint, March, 1994, to appear in Funct. Analisis and Appl. H.J. de Vega and N. Sánchez, Nucl. Phys. B309 (1988) 552 and ibid, 577. C.O. Loustó and N. Sánchez, Phys. Rev. D47 (1993) 4498. H.J. de Vega and I.L. Egusquiza, Phys. Rev. D50 (1994) 763. A.L. Larsen, Phys. Rev. D50 (1994) 2623. A.L. Larsen and M. Axenides, Phys. Lett. B318 (1993) 47. H.J. de Vega and N. Sánchez, Phys. Rev. D50 (1994) 7202. N. Sánchez and G. Veneziano, Nucl. Phys. B333 (1990) 253.\ M. Gasperini, N. Sánchez and G. Veneziano, IJMP A6 (1991) 3853, Nucl. Phys. B364 (1991) 365. H.J. de Vega, M. Ramón-Medrano and N. Sánchez, Nucl. Phys. B374 (1992) 425. H.J. de Vega and N. Sánchez, Phys. Rev. D45 (1992) 2783.\ M. Ramón-Medrano and N. Sánchez, Class. Quantum Grav. 10 (1993) 2007. N. Sánchez, Phys. Lett. B195 (1987) 160. H.J. de Vega, M. Ramón-Medrano and N. Sánchez, Nucl. Phys. B351 (1991) 277. A.L. Larsen and N. Sánchez, Phys. Rev. D50 (1994) 7493. A.L. Larsen and N. Sánchez, “Mass Spectrum of Strings in Anti de Sitter Spacetime”, DEMIRM-Paris-94048, to appear in Phys. Rev. D. A.L. Larsen and N. Sánchez, “New Classes of Exact Multi-String Solutions in Curved Spacetimes”, DEMIRM-Paris-95003, to appear in Phys. Rev. D. A.L. Larsen and N. Sánchez, “The Effect of Spatial Curvature on the Classical and Quantum Strings”, DEMIRM-Paris-95004. M. Bañados, C. Teitelboim and J. Zanelli, Phys. Rev. Lett. 69 (1992) 1849. G.T. Horowitz and D.L. Welch, Phys. Rev. Lett. 71 (1993) 328. N. Kaloper, Phys. Rev. D48 (1993) 2598. M. Bañados, M. Henneaux, C. Teitelboim and J. Zanelli, Phys. Rev. D48 (1993) 1506. C. Farina, J. Gamboa and A.J. Seguí-Santonja, Class. Quantum Grav. 10 (1993) L193.\ N. Cruz, C. Martínez and L. Peña, Class. Quantum Grav. 11 (1994) 2731. A. Achúcarro and M.E. Ortiz, Phys. Rev. D48 (1993) 3600.\ D. Cangemi, M. Leblanc and R.B. Mann, Phys. Rev. D48 (1993) 3606. J. Horne and G.T. Horowitz, Nucl. Phys. B368 (1992) 444. T. Buscher, Phys. Lett. B201 (1988) 466, B194 (1987) 59. A. Vilenkin, Phys. Rev. D24 (1981) 2082. H.J. de Vega and N. Sánchez, Int. J. Mod. Phys. A7 (1992) 3043. V.P. Frolov, V.D. Skarzhinsky, A.I. Zelnikov and O. Heinrich, Phys. Lett. B224 (1989) 255. R. Dashen, B. Hasslacher and A. Neveu, Phys. Rev. D11 (1975) 3424. H.J. de Vega, A.L. Larsen and N. Sánchez, “Semi-Classical Quantization of Circular Strings in de Sitter and anti de Sitter Spacetimes”, DEMIRM-Paris-94049, submitted to Phys. Rev. D. A.L. Larsen, “Stable and Unstable Circular Strings in Inflationary Universes”, NORDITA-94/14-P, to appear in Phys. Rev. D. A.L. Larsen, Nucl. Phys. B412 (1994) 372. [**Figure Captions**]{} Figure 1. The potential $V(r),$ Eq.(2.21), for a circular string in the $2+1$ BH-AdS spacetime. In (a) we have $J^2>4E^2$ and a barrier between the inner horizon and $r=0$, while (b) represents a case where $J^2<4E^2$ and a string will always fall into $r=0$. The static limit is $r_{\mbox{erg}}=l$. Figure 2. The potential $V(r),$ Eq.(2.18), for a circular string in the equatorial plane of the four $3+1$ dimensional spacetimes: (a) Minkowski (Min), (b) Schwarzschild black hole (S), (c) anti de Sitter (AdS), and (d) Schwarzschild anti de Sitter space (S-AdS). Figure 3. The $(N,M)=(5,1)$ multi-string solution in anti de Sitter spacetime. The $(N,M)$ multi-string solutions describe $N$ stationary strings with $M$ windings in the Azimuthal angle $\phi,$ in anti de Sitter spacetime. Figure 4. The $(N,M)=(3,2)$ stationary string solution inside the horizon of de Sitter spacetime. Besides the circular string, this is the simplest stationary closed string configuration in de Sitter spacetime. Figure 5. Schematic representation of the time evolution of the $(N,M)=(5,1)$ dynamical multi-string solution, outside the horizon of de Sitter spacetime. Only one of the 5 strings is shown; the others are obtained by rotating the figure by the angles $2\pi/5,\;4\pi/5,\;6\pi/5$ and $8\pi/5.$ During the “scattering” at the horizon, the string collapses to a point and re-expands. Figure 6. The dynamical straight string inside the horizon of the Schwarzschild black hole. The string infalls non-radially towards the singularity with indefinetely growing length. Figure 7. The potential $V(r)$ introduced in Eqs.(5.3) for a circular string in the static Robertson-Walker spacetimes: (a) flat $(K=0),$ (b) closed $(K>0\;\;and\;\;bK\alpha'^2\leq 1),$ (c) closed $(K>0\;\;and\;\;bK\alpha'^2>1),$ (d) hyperbolic $(K<0).$ Figure 8. Parametric plot of $K\alpha'W$ as a function of $K\alpha'^2m^2$ in the three cases: (a) $K>0,\;bK\alpha'^2\leq 1$ (strings oscillating on one hemisphere, only finitely many states), (b) $K>0,\;bK\alpha'^2>1$ (strings oscillating on the full hypersphere, infinitely many states), (c) $K<0$ (strings oscillating on the hyperboloid, infinitely many states). Figure 9. The potential $V(r)$ introduced in Eqs.(5.10) for a stationary string in the static Robertson-Walker spacetimes: (a) flat $(K=0),$ (b) closed $(K>0\;\;and\;\;KL^2\leq 1),$ (c) hyperbolic $(K<0).$ [**Table Captions**]{} Table 1. String motion described by the string perturbation series approach in the 2+1 BH-AdS, ordinary black hole-AdS, de Sitter (dS) and black string backgrounds. Table 2. Circular exact string solutions in the indicated backgrounds. $S(\tau)$ is the invariant string size. Table 3. Characteristic features of the quantum string mass spectrum in anti de Sitter (AdS) and de Sitter (dS) spacetimes. Table 4. Short summary of the features of the string solutions found in this section. In anti de Sitter spacetime and outside the horizon of de Sitter and $2+1$ BH-AdS spacetimes, the solutions describe a [*finite*]{} number of strings provided a condition of the form $N\Delta\phi=2\pi M$ is fulfilled, where $\Delta\phi$ is the angle betwen the “arms” of the string and $(N,M)$ are integers. Table 5. Classical circular strings in the static Robertson-Walker spacetimes. Notice that a self-consistent solution to the Einstein equations, with the string back-reaction included, can be obtained only for $K>0.$ Table 6. Semi-classical quantization of the circular strings in the static Robertson-Walker spacetimes. Notice in particular the different behaviour of the high mass spectrum of strings in the three cases. Table 7. Stationary strings in the static Robertson-Walker spacetimes. Notice that the pressure densities are always negative in all three cases. [^1]: Research report to appear in [**“New developments in string gravity and physics at the Planck energy scale”**]{}. Edited by N. Sánchez (World Scientific, Singapore, 1995)

ArXiv

--- author: - Jürg Fröhlich and Zhou Gang title: ON THE THEORY OF SLOWING DOWN GRACEFULLY --- Abstract {#abstract .unnumbered} ======== *“A moving body will come to rest as soon as the force pushing it no longer acts on it in the manner necessary for its propulsion.”* (Aristotle) We discuss the transport of a tracer particle through the Bose Einstein condensate of a Bose gas. The particle interacts with the atoms in the Bose gas through two-body interactions. In the limiting regime where the particle is very heavy and the Bose gas is very dense, but very weakly interacting (“mean-field limit"), the dynamics of this system corresponds to classical Hamiltonian dynamics. We show that, in this limit, the particle is decelerated by emission of gapless modes into the condensate (Cerenkov radiation). For an ideal gas, the particle eventually comes to rest. In an interacting Bose gas, the particle is decelerated until its speed equals the propagation speed of the Goldstone modes of the condensate. This is a model of “Hamiltonian friction". It is also of interest in connection with the phenomenon of “decoherence" in quantum mechanics. It is based on work we have carried out in collaboration with D Egli, IM Sigal and A Soffer. Introduction ============ We review some recent results on the effective dynamics of particles moving through a dispersive wave medium, such as the Bose-Einstein condensate of a very dense quantum gas at very low temperature. The motion of the particles exhibits friction as long as their speed is larger than the propagation speed of waves in the medium. Friction is caused by the emission of Cerenkov radiation. This phenomenon provides an example of recent mathematical results in transport theory. A *key problem* in transport theory consists in deriving the effective dynamics of experimentally observable degrees of freedom interacting with unobserved degrees of freedom of a macroscopically large environment from an underlying fundamental (Hamiltonian or unitary quantum) dynamics of the entire system; as indicated in the table below. (n1) [**Fundamental Dynamics**]{}; [(n2) [**Effective Dynamics**]{};]{} -------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------- Hamiltonian dynamics of point particles and waves - celestial mechanics; Dissipative dynamics of particles moving through a medium; Schrödinger equation, mechanics of atom gases;... friction, diffusion, Brownian motion, (quantum) Boltzmann equation, dynamics of viscous fluids,... (n1.north) to \[bend left\] node\[above\][$?$]{} (n2.north); The key problem alluded to, above, concerns the passage from the left to the right column in this table. Typically, one might study a system consisting of an observable tracer particle, or a whole tribe of such particles, coupled to unobserved or only crudely observed degrees of freedom of a medium such as the electromagnetic field in an optically dense material, sound waves in a crystal or the Goldstone modes of an atom gas exhibiting Bose-Einstein condensation. At low temperatures, the emission of Cerenkov radiation by the particle(s) into the medium is the basic mechanism causing friction of the particles’ motion. When the medium is, initially, in thermal equilibrium at some temperature $T >0$, it exhibits thermal noise that causes a diffusive type of motion of the particle(s) around its (their) mean trajectory.\ The problem of making these remarks precise by identifying the correct effective dynamics of the particle(s) and analyzing its properties leads one to study a variety of interesting, novel phenomena in the realm of non-linear Hamiltonian evolution equations and quantum dynamics for systems with infinitely many degrees of freedom. Some simple model systems {#s1} ========================= Dynamics of a particle moving through a medium {#s1.1} ----------------------------------------------- We consider a classical particle moving in physical space $\mathbb{E}^{3}$. The Hamilton function of this system is given by $$\label{1} H(X,P)=\frac{P^{2}}{2M} + V(X)$$ with $X \in \mathbb{E}^{3}$ the position of the particle, $P \in \mathbb{R}^{3}$ its momentum, $M$ its mass, and $V$ the potential of an external force. An example of a potential is $$\label{2} V(X)=-F \cdot X$$ where $F \in \mathbb{R}^{3}$ is a constant external force pushing the particle; e.g., the gravitational force close to the surface of the earth.\ The particle is assumed to interact with a wave medium. We consider the example of a Bose gas exhibiting Bose-Einstein condensation. We would like to treat the entire system within the framework of classical Hamiltonian dynamics. This forces us to pass to a limiting regime of the system, the so-called *mean field limit*: The average density of the Bose gas is given by $\rho_N:=N \rho$, $\rho >0$, the strength of two-body interactions among the atoms of the Bose gas is $\lambda_N:=N^{-1} \lambda$, $\lambda >0$; we consider the limiting regime where $N \rightarrow \infty$, with $\rho$ and $\lambda$ kept fixed. In this limit, the state of the gas can be described by a Ginzburg-Landau order parameter, $\psi(x) \in \mathbb{C}$, where $N \vert \psi(x) \vert^{2}$ is the density of the gas at the point $x \in \mathbb{E}^{3}$. In microscopic units, the mass of a quantum-mechanical tracer particle moving through the Bose gas is given by $NM$ (i.e., it becomes very heavy, as $N \rightarrow \infty$) and the external potential is given by $N V(X)$, with $M$ and $V$ fixed. The limit $N \rightarrow \infty$ corresponds to a “classical limit” in which the quantum dynamics of the system approaches some classical Hamiltonian dynamics.\ The Ginzburg-Landau order parameter of the Bose gas, $\psi(x)$, and its complex conjugate, $\bar{\psi}(x)$, are interpreted as complex coordinates of an infinite-dimensional affine phase space, $\Gamma_{BG}:=H^{1}_{\rho}(\mathbb{R}^{3})$, (a certain complex Sobolev space). The standard symplectic structure on $\Gamma_{BG}$ leads to the following *Poisson brackets* : $$\label{3} \begin{split} \lbrace \psi(x), \psi(y)\rbrace&=\lbrace \bar{\psi}(x), \bar{\psi}(y) \rbrace=0\\ \lbrace \psi(x), \bar{\psi}(y)\rbrace&=i \delta(x-y) \end{split}$$ Quantum mechanically, $\psi$ and $\bar{\psi}$ would be interpreted as annihilation- and creation operators, and the Poisson brackets are replaced by commutators, $\left[\cdot, \cdot \right]$, with $\left[\cdot, \cdot \right] \sim -\frac{i}{N} \lbrace \cdot, \cdot \rbrace$. In the mean-field limit, $N \rightarrow \infty$, the following Hamilton functional is appropriate to describe a non-relativistic Bose gas: $$\begin{aligned} \label{4} \mathcal{H}_{BG}(\psi, \bar{\psi})=\int d^{3}x \text{ } \left[ \frac{1}{2m} \vert \nabla \psi(x) \vert^{2} + \lambda \int d^{3}y \text{ } ( \vert \psi(x) \vert^2 -\rho) \Phi(x-y) ( \vert \psi(y) \vert^2 -\rho) \right]\end{aligned}$$ where $\Phi$ is a two-body potential of short range and, for reasons of thermodynamic stability, of *positive type*, $\lambda >0$ is a coupling constant and $\rho>0$ is an average (rescaled) density. We consider a regime where the $U(1)$-symmetry $$\psi(x) \mapsto e^{i \theta} \psi(x), \text{ } \bar{\psi}(x) \mapsto e^{-i \theta} \bar{\psi}(x), \text{ } \theta \in \mathbb{R}$$ is spontaneously broken. It is convenient to impose symmetry-breaking boundary conditions $$\label{6} \overset{(-)}{\psi}(x)=\sqrt{\rho} + \overset{(-)}{\beta}(x)$$ with $$\beta(x) \rightarrow 0, \text{ as } \vert x \vert \rightarrow \infty$$\ The Poisson brackets between $\beta$ and $\bar{\beta}$ are the same as those in (\[3\]). In terms of the variables $\beta$ and $\bar{\beta}$, the Hamilton functional is given by $$\label{7} \begin{split} \mathcal{H}_{BG}(\beta, \bar{\beta})=&\int d^{3}x \text{ } \Bigl[ \frac{1}{2m} \vert \nabla \beta(x) \vert^{2} + \\ &\lambda \int d^{3}y \text{ } \left( \vert \beta(x) \vert^2 + 2 \sqrt{\rho} \text{ } \Re \beta(x) \right) \Phi(x-y) \left( \vert \beta(y) \vert^2 + 2 \sqrt{\rho} \text{ } \Re \beta(y) \right)\Bigr] \end{split}$$ The interaction between the tracer particle and the Bose gas is given by $$\label{8} \begin{split} \mathcal{H}_{I}(X; \beta,\bar{\beta}):=&g \int d^{3}x \text{ } W(X-x) \left( \vert \psi(x) \vert^{2} - \rho \right)\\ =&g \int d^{3}x \text{ } W(X-x) \left( \vert \beta(x) \vert^{2} + 2 \sqrt{\rho} \text{ }\Re \beta(x) \right) \end{split}$$ where $X$ is the position of the tracer particle, $W$ is a spherically symmetric two-body potential of short range, and $g >0$ is a coupling constant. The phase space of the coupled system is given by $$\label{9} \Gamma_{\text{tot}}:=\mathbb{R}^{6}_{\text{particle}} \times \Gamma_{BG}$$ and its Hamilton functional by $$\label{10} \mathcal{H}(X,P; \beta,\bar{\beta} )=H(X,P)+\mathcal{H}_{BG}(\beta,\bar{\beta} )+\mathcal{H}_{I}(X;\beta,\bar{\beta} )$$ where $H$ is as in (\[1\]), $\mathcal{H}_{BG}$ as in (\[7\]) and $\mathcal{H}_{I}$ as in (\[8\]). A list of models of decreasing difficulty {#s1.2} ----------------------------------------- \(1) : This is the model specified by (\[9\]) and (\[10\]), with $\rho < \infty$, $\lambda >0$, $g>0$.\ \ (2) : We consider the “*Bogoliubov limit*”, $$\label{11} \left.\begin{aligned} \rho&\rightarrow \infty, \text{ }\lambda \rightarrow 0,\text{ } \lambda \rho=\text{const}:=\frac{\kappa}{4} \\ g&\rightarrow 0,\text{ } g \sqrt{\rho}=: \nu, \text{ } \kappa \text{ and } \nu \text{ fixed}\end{aligned} \right\}$$ In this limit, $$\mathcal{H}(X,P; \beta,\bar{\beta})=H(X,P)+\mathcal{H}^{(E)}_{BG}(\beta,\bar{\beta})+\mathcal{H}^{(E)}_{I}(X;\beta,\bar{\beta} )$$ where $$\label{12} \begin{split} \mathcal{H}^{(E)}_{BG}(\beta,\bar{\beta})&=\int d^{3}x \text{ } \Bigl[ \frac{1}{2m} \vert \nabla \beta(x) \vert^{2} + \kappa \int d^{3}y \text{ } \Re \beta(x) \text{ } \Phi(x-y) \text{ } \Re \beta(y) \Bigr]\\ \mathcal{H}^{(E)}_{I}(X; \bar{\beta},\beta)&= 2 \nu \int d^{3} x \text{ }W(X-x) \text{ } \Re \beta(x) \end{split}$$ It is useful to introduce real coordinates, $$\beta(x)=\frac{1}{\sqrt{2}} \left(\varphi(x) +i \pi(x)\right)$$ with $$\lbrace \varphi , \varphi \rbrace=\lbrace \pi , \pi \rbrace=0, \text{ } \lbrace \varphi(x) , \pi(y) \rbrace=\delta(x-y)$$\ Then $$\mathcal{H}^{(E)}_{BG}=\int d^{3}x \text{ } \Bigl[ \frac{1}{4m} \left( \nabla \pi \right)^{2} (x) + \left( \nabla \varphi \right)^{2} (x) + \frac{\kappa}{2} \int d^{3}y \text{ } \varphi(x) \Phi(x-y) \varphi(y) \Bigr]$$\ and $$\mathcal{H}^{(E)}_{I}=\sqrt{ 2} \nu \int d^{3} x \text{ }W(X-x) \varphi(x)$$ The equation of motion for the Bose gas ($g=0$) are $$\begin{aligned} \dot{\varphi}&=&\lbrace\mathcal{H}, \varphi \rbrace=\frac{1}{2m} \Delta \pi\\ \dot{\pi}&=&\lbrace\mathcal{H}, \pi \rbrace=-\frac{1}{2m} \Delta \varphi + \kappa \Phi *\varphi\end{aligned}$$ It follows that $$\ddot{\varphi} =\frac{1}{2m} \Delta \dot{\pi}=-\frac{1}{4m^2} \Delta^{2} \varphi + \frac{\kappa}{2m} \Delta(\Phi) * \varphi$$ which is a wave equation for $\varphi$ with constant coefficients that can be solved by Fourier transformation.\ One concludes that the frequency, $\Omega$, of a plane wave with wave vector $k \in \mathbb{R}^{3}$ is given by $$\label{v} \begin{split} \Omega(k)&=\vert k\vert \sqrt{\left( \frac{k}{2m} \right)^2+ \frac{\kappa \hat{\Phi}(k)}{2m}}\\ &\approx \vert k\vert \sqrt{ \frac{\kappa \hat{\Phi}(0)}{2m}} \equiv v_{*} \vert k \vert, \text{ } k \approx 0. \end{split}$$ Here, $v_{*}=\sqrt{\frac{\kappa \hat{\Phi}(0)}{2m}}$ is the minimal propagation speed of a Goldstone mode (sound wave) in the Bose-Einstein condensate of the Bose gas. The equations of motion of the coupled system are given by $$\label{B} \begin{split} M \dot{X} &=P, \text{ } \dot{P}=- \nabla V(X)-\nu \int d^{3} x \text{ } \nabla W(X-x) \text{ } \Re \beta(x) \\ i \dot{\beta} (x)&= - \frac{1}{2m} \Delta \beta(x) + \frac{\kappa}{4}\left( \Phi * \Re \beta \right)(x) + \nu W^{X}(x) \end{split}$$ where $W^{X}(x):=W(x-X)$.\ \ (3) : $\lambda=0, \text{ } \rho<\infty$, $g>0$; see [@EG].\ \ (4) : $\lambda=0, \text{ } \rho \rightarrow \infty$, with $g\sqrt{\rho}=\nu$ fixed; see [@FGS1], [@FGS2]. The equations of motion in the B-model are those given in (\[B\]), with $\kappa=0$. Some special solutions of the equations of motion {#s2} ================================================= Static solutions: $\dot{X}=\dot{P}=\dot{\beta} \equiv 0$ {#s2.1} -------------------------------------------------------- Let $X_* \in \mathbb{E}^{3}$ be a critical point of the external potential (with $X_*$ arbitrary if $V \equiv 0$). Let $\beta_*$ be a critical point of the energy functional $$\begin{split} \mathcal{E} (X_*; \beta,\bar{\beta})&:=\int d^{3}x \text{ } \Bigl[ \frac{1}{2m} \vert \nabla \beta(x) \vert^{2} + g W(X_*-x) \left( \vert \beta(x) \vert^{2} + 2 \sqrt{\rho} \text{ }\Re \beta(x) \right) \\ &+ \lambda \int d^{3} y \text{ } \left( \vert \beta(x) \vert^{2} + 2 \sqrt{\rho} \text{ }\Re \beta(x) \right) \Phi(x-y) \left( \vert \beta(y) \vert^{2} + 2 \sqrt{\rho} \text{ }\Re \beta(y) \right) \Bigr] \end{split}$$\ A critical point $\beta_*$ can be constructed with the help of variational calculus. Then $\left(X \equiv X_*, P \equiv 0, \overset{(-)}{\beta}(x) \equiv \overset{(-)}{\beta_*}(x)\right)$ is a static solution of the equations of motion. If $X_*$ is a non-degenerate local minimum of $V$ one expects that there exist solutions of the equations of motion close to this static solution and converging to it, as time $t$ tends to $\infty$.\ In the E-model, $\beta_*$ is the real solution of the elliptic equation $$\label{coucou} \left( -\frac{\Delta}{2m} + \frac{\kappa}{4} \Phi *\right) \beta=-\nu W^{X_*},$$ which can be solved by Fourier transformation. If $\kappa \hat{\Phi}(0)=\kappa \int d^{3}x \text{ } \Phi(x) >0$ then the solution exhibits exponential decay in $x$. Traveling waves for $V \equiv 0$ {#s2.2} --------------------------------- $$\label{tr} \left.\begin{aligned} &X_t=vt+X_0, \text{ } P_t \equiv Mv, \text{ }v \in \mathbb{R}^{3}\\ &\beta_t (x)=\gamma_v(x-X_t)\end{aligned} \right\}$$ Let $W$ be smooth, spherically symmetric and of rapid decay at $\infty$. We consider the E-model, with $v_*=\sqrt{\frac{\kappa \hat{\Phi}(0)}{2m}}>0$; see (\[v\]). Then, for $\vert v \vert \leq v_c \apprle v_*$ the equations of motion (\[B\]) have a traveling wave solution of the form (\[tr\]), and $\gamma_v$ decays exponentially fast at $\infty$, for $\vert v \vert <v_c$; $\gamma_v$ has power-law decay at $\infty$, for $\vert v \vert =v_c$. As $\vert v \vert \rightarrow 0$, $\gamma_v$ converges to $\beta_*$ (see (\[coucou\])), with $X_*=X_0$. If $\Phi(x)=\delta(x)$ then $v_c=v_*$.\ The proofs of these claims follow from easy calculations, using Fourier-transformation.\ For $\vert v \vert >v_c \approx v_*$, the equation $$\frac{k^{2}}{2m}+ \frac{\kappa}{4} \hat{\Phi}(k)-v \cdot k=0$$ has real solutions, $k=k_v \in \mathbb{R}^{3}$, with $k_v \parallel v$. Then there do *not* exist any traveling wave solutions for $V \equiv 0$. If the ansatz (\[tr\]) is plugged into the equations of motion one finds that $\dot{P} \neq 0$, with $\dot{P}$ anti-parallel to $v$; i.e., a non-zero friction force, $F_v$, arises. This force is caused by emission of Cerenkov radiation by the particle, and its precise form can be inferred from *Fermi’s Golden Rules*.\ If $\kappa =0$ (i.e., $v_*=0$) the expression for $F_v$ is very simple: $$\label{Fv} F_v=-\text{const } v \frac{1}{\vert v \vert^{3}} \int_{0}^{(2 \pi v)^{2}} d \rho \text{ } \rho \vert \hat{W}(\sqrt{\rho}) \vert^{2},$$ where the constant is proportional to $\nu$. The behaviour of $\vert F_v \vert$ is indicated in the following figure. (0,0)(5,5) (-2,0)(-1,0.4)(0,1.35)(1.8,4)(2.3,3.98)(2.6,3.75)(3.25,2.5)(4,1.4)(5,0.8)(6,0.5)(7,0.25) (-2,0)[[(0,0)(0,4.5)]{}]{} (-2,0)[[(0,0)(9.5,0)]{}]{} \[90\](-2,4.4)[$\vert F_v \vert$]{} \[90\](-2.6,3.8)[$ F_{\max}$]{} \[0\](7.5,0)[$\vert v \vert$]{} \[0\](6,1)[$O(\vert v \vert^{-2})$]{} \[0\](-2,1.5)[$O(\vert v \vert^{2})$]{} (-2.1,4.035)[[(0,0)(0.2,0)]{}]{} (1.3,4.035)[[(0,0)(1.35,0)]{}]{} (-1.3,1.1)[[(0,0)(0.4,-0.4)]{}]{} (7,0.7)[[(0,0)(-0.3,-0.3)]{}]{} Our discussion suggests to proceed to study *forced* traveling waves, with $V(X)=-F \cdot X$, for a non-vanishing external force $F \in \mathbb{R}^{3}$. Forced traveling waves {#s2.3} ---------------------- We study the motion of a tracer particle under the influence of a constant external force $F \in \mathbb{R}^{3}$, with $V(X)=-F \cdot X$. For simplicity, we consider the B- or the E-model. [@EFGS] Suppose that $W$ is smooth, spherically symmetric and of rapid decay at infinity. Then there exists a positive constant $F_{\text{max}} < \infty$ such that \(1) if $\vert F \vert <F_{\text{max}}$ the equations of motion (\[B\]) of the B- or E-model have *two* traveling wave solutions of the form (\[tr\]) with velocity $v \parallel F$, $\vert v \vert >v_c \approx v_*$; \(2) if $\vert F \vert = F_{\text{max}} $ there exists a unique traveling wave solution with $v \parallel F$, $\vert v \vert >v_c$; \(3) if $\vert F \vert >F_{\text{max}}$ there do not exist any traveling wave solutions of the form (\[tr\]). . To prove this theorem we must look for solutions of the equations of motion of the form $X_t=vt+X_0$, $P_t \equiv Mv$, $\beta_t(x)=\gamma_v (x- X_t)$.\ The equations of motion (\[B\]) then imply that $$\label{mo} \begin{split} &- \nabla V(X) \equiv F =\nu \int d^{3} x \text{ }(\nabla W)(x) \text{ }\Re \gamma_v(x) \\ &-i v \cdot \nabla \gamma_v= -\Delta \gamma_v + \frac{\kappa}{4} \left( \Phi * \Re \gamma_v \right)(x)+ \nu W(x) \end{split}$$\ If $\vert v \vert >v_c$ the Fourier transform, $\hat{\gamma}_v$, of the second equation in (\[mo\]) is singular on a spherical surface in k-space, and it follows that the first equation in (\[mo\]) does not have any solutions unless $\vert F \vert>0$, as discussed in section \[s2.2\].\ The right side of the first equation in (\[mo\]) is a function of $\vert v \vert$ that vanishes, for $\vert v \vert <v_c$, tends to $0$, as $\vert v \vert \rightarrow \infty$, and has a unique maximum at some value, $v^{*}$, of $\vert v \vert$. It is quadratic in $W$. From these observations the proof of the theorem stated above can easily be inferred. For $\vert v \vert >v_c$, $\vert F \vert \neq 0$, the solution, $\gamma_v$, of the second equation describes a conical wave whose tip is at the position, $X_t$, of the particle. Significance of special solutions {#s2.4} --------------------------------- The construction of special solutions, as outlined in Sects \[s2.1\], \[s2.2\] and \[s2.3\] is elementary, and one may wonder why such solutions are of any interest, at all. The reason is that one expects that rather general classes of (time-dependent, non-stationary) solutions of the equations of motion approach the special solutions constructed above, as time $t$ tends to $\infty$, (in a mathematically precise sense). We briefly discuss this expectation in several situations. \(1) Let us suppose that $V(X)$ has a non-degenerate (quadratic) minimum, $V(X_*) <0$, at a point $X_* \in \mathbb{E}^{3}$, with $V(X) \rightarrow 0$, as $\vert X \vert \rightarrow \infty$. We consider the G-model and choose initial conditions, $(X_0, P_0, \beta_0)$, whose total energy, $\mathcal{H}(X_0,P_0; \beta_0,\bar{\beta}_0)$, is strictly *negative*. Then we expect that the solution, $(X_t, P_t,\beta_t)$, of the Hamiltonian equations of motion corresponding to the Hamilton functional $\mathcal{H} (P,X;\beta,\bar{\beta})$ in (\[10\]) approaches a static solution, $X_t \rightarrow X_*$, $P_t \rightarrow 0$, $\beta_t \rightarrow \beta_*$, as $t \rightarrow \infty$, where $\beta_*$ is a minimizer of the energy functional $\mathcal{E}(X_*; \beta,\bar{\beta})$ introduced in section \[s2.1\], provided the coupling constant $g$ is sufficiently small. (If the potential $V$ is as specified above, but $V(X) \rightarrow \infty$, as $\vert X \vert \rightarrow \infty$, then one expects that the static solution $(X_*, P=0,\beta_*)$ is asymptotically stable without further assumptions on initial conditions of finite total energy and on the coupling constant $g$.) \(2) We consider the B- or E-model, with $V(X) \equiv 0$. We choose initial conditions, ($X_0,P_0,\beta_0$), of energy $\mathcal{H}(X_0,P_0; \beta_0,\bar{\beta}_0)$ (see (\[12\])) not “much larger” than $\frac{M}{2} v^{2}_c$ (with $v_c=v_*=0$, for the B-model) and $g$ small enough. Then we expect that the solution, $(X_t,P_t,\beta_t)$, of the equations of motion (\[B\]), with the given initial conditions, approaches a traveling wave solution (\[tr\]), with $\vert v \vert \leq v_c$, as $t \rightarrow \infty$; (convergence of $\beta_t$ is understood locally, in arbitrary balls of finite radius around the position, $X_t$, of the tracer particle). \(3) We consider the B- or E-model, but with $V(X)=- F \cdot X$, $0<\vert F \vert < F_{\text{max}}$. Let $v_{<}>0$ and $v_>>v_<$ be the two speeds for which the equations in (\[mo\]) have solutions, with $\vert v \vert =v_<$ or $\vert v \vert =v_>$. We first discuss the stability of solutions with $\vert v \vert =v_<$. We expect that a solution of the equations of motion (\[B\]) corresponding to an initial condition $(X_0,P_0,\beta_0)$, with $X_0 \in \mathbb{E}^{3}$ arbitrary, $\vert P_0 \vert <<Mv_<$, $\mathcal{H}^{(E)}_{BG}( \beta_0,\bar{\beta}_0)$ small enough, converges to a traveling wave solution of the type discussed in the theorem of section \[s2.3\], with $\vert v \vert =v_<$, as time $t \rightarrow \infty$, provided $g$ is small enough. In contrast, if $\vert P_0 \vert \apprge Mv_>$, $\mathcal{H}^{(E)}_{BG}( \beta_0,\bar{\beta}_0)$ small, the solution $(X_t, P_t,\beta_t)$ of the equations of motion (\[B\]) will in general not approach a traveling wave solution with $\vert v \vert =v_>$. Rather, we expect that $\vert P_t \vert $ diverges, as $t \rightarrow \infty$. (Similar behaviour is expected if $\vert F \vert >F_{\text{max}}$.)\ Most of these expectations remain unproven, except for expectation (2) in the context of the B-model. A theorem deserving this name {#s3} ============================= In this section, we consider the B- and C-models, which are the only models for which non-trivial results have been established, so far; see [@FGS1], [@FGS2],[@EG]. We state our main result for the B-model; but a very similar result has also been proven for the C-model. We propose to study solutions of the equations of motion (\[B\]), for $\kappa=0$. We choose units such that $M=m=1$. The two-body potential $W$ is assumed to be smooth, spherically symmetric and of rapid decay at $\infty$, with $\vert \hat{W}(0) \vert=(2 \pi)^{-3/2} \vert \int d^{3} x \text{ } W(x) \vert=1$. The wave field $\beta$ is normalized appropriately. The external potential $V$ is assumed to vanish identically. Then the equations of motion (\[B\]) take the form $$\label{mo3} \begin{split} \dot{X}_t &=P_t, \text{ } \dot{P}_t=\nu \int d^{3} x \text{ } (\nabla W^{X_t})(x) \text{ }\Re \beta_t (x) \\ i \dot{\beta} _t (x)&= - \frac{1}{2} (\Delta \beta_t)(x) + W^{X_t}(x) \end{split}$$ where $ W^{X_t}(x):=W(x-X_t)$, $\nu=O(1)$ a constant. Note that the equations (\[mo3\]) have static solutions $$X_t \equiv X_*, \text{ } P_t \equiv 0, \text{ }\beta_t(x) \equiv \beta_*(x)=2(\Delta^{-1} W^{X_*})(x)$$ indexed by $X_*$, and we propose to analyze their asymptotic stability. Our main result is the following theorem. [@FGS2] Under the assumptions described above, there exists an interval $I \supseteq \left[0,0.66 \right) \supset \left[ 0, \frac{1}{2} \right]$ such that, for any $\delta \in I$, there is some $\epsilon_0 = \epsilon_{0}(\delta) >0$ with the property that if $$\vert P_0 \vert, \text{ } \vert \vert \langle x \rangle^{3} \beta_0 \vert \vert_2, \text{ } \vert \vert \nabla \beta_0 \vert \vert_2 <\epsilon_0$$ where $\langle x \rangle:=\sqrt{1+ \vert x \vert^{2}}$, then $$\vert P_t \vert \leq c \langle t \rangle^{-\frac{1}{2}- \delta}, \text{ as } t \rightarrow \infty$$ for some finite constant $c$, and $$\lim_{t \rightarrow \infty} \vert \vert \beta_t-2 \Delta^{-1} W^{X_t} \vert \vert_{\infty}=0$$ :\ (1) If $\delta$ is chosen to be larger than $\frac{1}{2}$ (note that I contains such values of $\delta$) then $$X_t =X_0+ \int_{0}^{t} P_s\text{ } ds \rightarrow X_*, \text{ as } t \rightarrow \infty$$ with $\vert X_* \vert < \infty$.\ \ (2) We suspect that $\vert P_t \vert \sim t^{-\frac{1}{2}- \delta}$, as $t \rightarrow \infty$, for an exponent $\delta > \frac{1}{2}$ *independent* of initial conditions, as long as $\vert P_0 \vert$, $\vert \vert \langle x \rangle^{3} \beta_0 \vert \vert_2$, $\vert \vert \nabla \beta_0 \vert \vert_2$, are finite. However, this has not been proven, so far. (If the condition that $\vert \vert \langle x \rangle^{3} \beta_0 \vert \vert_2 < \infty$ is dropped then $\delta$ may depend on initial conditions.\ For, consider the equation $$\label{fvv} \dot{v}_t=F_{v_t}$$ with $F_v$ as in (\[Fv\]). This corresponds to $\beta_t(x)=\gamma_{v_t} (x-X_t)$. The solution of (\[fvv\]) behaves like $\vert v_t \vert \sim \langle t \rangle^{-1}$, i.e., $\delta=\frac{1}{2}$. However, $\vert \vert \beta_t \vert \vert_2$ is divergent!)\ \ (3) The theorem stated above has interesting applications in the analysis of *decoherence* in systems of a heavy quantum-mechanical particle moving through a Bose-Einstein condensate. Although the initial moves in the proof of the theorem described above are quite natural and fairly standard, the technical details are suprisingly tedious and involve exhibiting very subtle cancellations among terms that do not, a priori, exhibit adequate decay, as time $t \rightarrow \infty$, in order to “close the estimates”. It would be of considerable interest to study systems of finitely many tracer particles moving through a Bose-Einstein condensate, or a gas of such particles. It is not very hard to come up with plenty of interesting and plausible looking conjectures concerning the behaviour of such systems. But it appears to be very difficult to rigorously prove these conjectures.\ \ : For further results on friction, see also [@bruneau; @caprino; @KM]. We thank D.Egli, I.M. Sigal and A. Soffer for many enlightening discussions. We are indebted to B.Schubnel for carefully reading the manuscript and much help with the typesetting. [1]{} L. Bruneau and S. De Bi[è]{}vre. A hamiltonian model for linear friction in a homogeneous medium. , 229(3):511–542, 2002. S. Caprino, C. Marchioro, and M. Pulvirenti. Approach to equilibrium in a microscopic model of friction. , 264(1):167–189, 2006. D. Egli, Fröhlich J., Zhou G., and Sigal I.M. Papers in preparation. D. Egli and G. Zhou. Some hamiltonian models of friction ii. , 2011. J. Fröhlich, Z. Gang, and A. Soffer. Friction in a model of hamiltonian dynamics. , 2011. J. Fr[ö]{}hlich, Z. Gang, and A. Soffer. Some hamiltonian models of friction. , 52:083508, 2011. DL Kovrizhin and LA Maksimov. Cherenkov radiation of a sound in a bose condensed gas. , 282(6):421–427, 2001.

ArXiv

--- abstract: 'An LDPC coded modulation scheme with probabilistic shaping, optimized interleavers and noniterative demapping is proposed. Full-field simulations show an increase in transmission distance by 8% compared to uniformly distributed input.' address: | ^(1)^Institute for Communications Engineering, Technische Universität München, 80333 Munich, Germany\ ^(2)^Optical Networks Group, University College London (UCL), London, WC1E 7JE, UK author: - 'Tobias Fehenberger^(1)^, Georg Böcherer^(1)^, Alex Alvarado^(2)^ and Norbert Hanik^(1)^' title: for Optical Fiber Systems --- Introduction ============ In the pursuit of higher data rates in optical communications, high-order modulation formats are an established technique to increase spectral efficiencies. As the optical channel is effectively power-limited by the nonlinearities, optimizing the signaling is necessary in order to further improve the spectral efficiency without increasing the launch power. Probabilistic shaping and geometric shaping are two possibilities to achieve this. The former uses constellations with nonuniform distributions on a regular grid while the latter uses nonequidistant spacings of equiprobable symbols. The main implementation advantage of probabilistic shaping over geometric shaping is that it does not require modifications of the digital-to-analog converters and the optical signal processing algorithms. Probabilistic shaping has been proposed for optical transmission, e.g. in [@smith2012trellisshaping; @beygi2014polarshaping; @yankov2014shaping]. In [@smith2012trellisshaping], a multi-level coded modulation scheme with trellis shaping and hard-decision decoding is found to operate close to a capacity estimate. Probabilistic shaping via shell-mapping improving a 4D coded modulation scheme is presented in [@beygi2014polarshaping]. In [@yankov2014shaping], probabilistic shaping is implemented for turbo codes by a many-to-one mapping. This type of mapping, however, requires information exchange between the demapper and the decoder at the receiver, known as iterative demapping. In this paper, we adapt the probabilistic shaping scheme introduced in [@boecherer2014ShapingLikeARealMan] to a multi-span wavelength-division multiplexing (WDM) optical fiber system with quadrature amplitude modulation (QAM). Off-the-shelf binary encoders and decoders are used and no iterative demapping is required. Using bit-interleaved coded modulation and a low-density parity-check (LDPC) code, the presented scheme increases the transmission distance by 8%. To the best of our knowledge, this is the first demonstration of such gains based on probabilistic shaping in optical communications.  \[fig:simsetup\] Mutual Information Analysis {#sec:MIsim} =========================== Optical Transmission Setup {#ssec:implementation} -------------------------- The setup of the optical coded modulation system is shown in Fig. \[fig:simsetup\]. The QAM symbols are either uniformly distributed or shaped as explained in Sec. \[ssec:shaping\]. We consider a dual-polarization multi-span WDM system with 15 co-propagating channels. The baud rate per channel is 28 GBaud, the pulses are root-raised-cosine (RRC) shaped with 5% roll-off, the WDM spacing is 30 GHz. We simulate 2^16^ symbols per polarization. The fiber is a standard single-mode fiber (SSMF) with $\alpha$=0.2 dB/km, $\gamma$=1.3 (W$\cdot$km)^-1^ and $D$=17 ps/nm/km. Each span of length 100 km is followed by an Erbium-doped fiber amplifier (EDFA) with a noise figure of 4 dB. Signal propagation is simulated using the split-step Fourier method with 32 samples per symbol and a step size of 100 m. At the receiver, the center WDM channel is bandpass-filtered and converted into the digital domain with a coherent receiver. Chromatic dispersion is compensated digitally and an RRC filter is applied again. Laser phase noise and polarization mode dispersion are not included in the simulations as perfect compensation is assumed. We compute nonparametric estimates of the mutual information (MI) between input and output symbols without considering memory (see the dashed blue box in Fig. \[fig:simsetup\]). In particular, Gaussian statistics are not assumed a priori for MI estimation. Optimized Input Distribution {#ssec:shaping} ---------------------------- We probabilistically shape 16-QAM and 64-QAM by assigning larger probabilities to the points with lower energy. At the same power, this increases the Euclidean distance between constellation points compared to uniform input. To find a suitable distribution for the shaped input, we first resort to the Gaussian noise (GN) model [@poggiolini2014GN] and calculate the signal-to-noise ratio (SNR) for every transmission distance. We then follow [@kschischang1993mbdistribution] and jointly optimize the Maxwell-Boltzmann distribution of the input symbols and the constellation scaling such that the MI is maximized for the additive white Gaussian noise (AWGN) channel under the power constraint imposed by the SNR. The obtained distribution is used for the shaped input of the optical fiber simulations. Insets b) and c) of Fig. \[fig:MIvsL\] illustrate the effect of uniform and shaped input on the received symbol distributions, respectively, with hotter colors representing a larger number of occurrences. The innermost points of the shaped input are sent more often and the Euclidean distance between constellation points is 18% larger than for uniform input. Mutual Information Results -------------------------- MI versus transmission distance is shown in Fig. \[fig:MIvsL\] for 16-QAM and 64-QAM, both with uniform and shaped input. For each overall transmission distance of 1 to 80 spans, we use an optimized shaped input and determine a candidate for the optimal launch power from the GN model. By testing different launch powers in our simulations, we found that the optimum is within 0.2 dB of $-1.6$ dBm for all simulated modulations and distances. For small distances, shaped and uniform QAM achieve similar MI because the impact of distortions is small in this region and the shaped distribution is close to the uniform distribution. Using shaped 64-QAM at 15 spans, however, gives 2 additional spans compared to uniform input, which is an increase of 15%. For longer distances, the improvement over uniform 64-QAM disappears, and thus, we did not simulate shaped 64-QAM for more than 50 spans. Once the signal propagates further and distortions accumulate, we observe that shaped 16-QAM performs better than uniform 16-QAM. From 50 spans onwards, the MI curves of shaped 16-QAM and uniform 64-QAM coincide. Inset a) of Fig. \[fig:MIvsL\] shows that shaped 16-QAM increases the transmission distance by 3 to 4 spans compared to uniform 16-QAM.  \[fig:MIvsL\] BER Analysis {#sec:implementation} ============ To verify our MI results, we implement an LDPC code as inner forward error correction (FEC) and compare the BERs of uniform and shaped 16-QAM (see the dotted red box in Fig. \[fig:simsetup\]). We operate at an information rate of 3 bits per symbol per polarization, without considering the coding overhead (OH) of the outer FEC. The LDPC codes of our coded modulation scheme are from the DVB-S2 standard and have a block length of 64800 bits. Implementation -------------- A total of four bits map to one 16-QAM symbol. For uniform input and a target information rate of 3 bits per symbol, we use 3 bits for data and one bit for redundancy. The rate 3/4 LDPC code is used to achieve this. For our shaped input, we effectively map 3 uniformly distributed data bits to 3.2 shaped bits. We use the remaining $4-3.2=0.8$ bits per QAM symbol for redundancy by encoding the shaped data with a rate $3.2/4=4/5$ LDPC code. Hence, the overall redundancy of one bit is successively added in two steps, i.e., by probabilistic shaping and by LDPC encoding. [r]{}[0.5]{}  \[fig:BERvsL\] The mapping from uniformly distributed data bits to shaped bits is done by a distribution matcher [@baur2014ADM] as shown in Fig. \[fig:simsetup\]. In our simulations, we emulated the distribution matcher output by directly generating the shaped bits. We optimize the LDPC codes by wrapping the encoder with two interleavers and the decoder with the respective inverse interleavers. The interleavers optimize the mapping of the coded bits to the QAM symbols and ensure that the distribution imposed on the data is preserved at the modulator output [@boecherer2014ShapingLikeARealMan Sec. V-D]. A staircase code with 6.7% OH [@smith2012staircase] is assumed as outer code. It must be placed between the distribution matcher and the inner FEC to avoid error propagation. A systematic encoder must be used to preserve the statistics generated by the distribution matcher. BER Results ----------- In Fig. \[fig:BERvsL\], the plot of BER versus distance shows that the shaping gain predicted by the MI in Fig. \[fig:MIvsL\] translates into an improvement in BER. At a BER of 1.3$\cdot$10^-3^, we observe a shaping gain of 4 spans compared to uniform 16-QAM. This increase in transmission distance by 8% is in excellent agreement with the prediction of the MI curves in Fig. \[fig:MIvsL\]. Conclusions =========== We have presented an LDPC coded modulation scheme with probabilistic shaping. Compared to previously suggested shaping schemes, the implementation has lower complexity because no iterative demapping is required at the receiver. We show in optical fiber simulations that 16-QAM and 64-QAM with shaped input allow for an increase in transmission distance by 8% and 15%, respectively, compared to uniformly distributed input. The results are promising and further gains are expected by optimizing the input distribution for the nonlinear optical channel. Also, four-dimensional constellations that are more robust against nonlinear distortions will be considered in future work. [99]{} B. P. Smith and F. R. Kschischang, “A pragmatic coded modulation scheme for high-spectral-efficiency fiber-optic communications,” *J. Lightw. Technol.* **30**(13), pp. 2047–2053, Jul. 2012. L. Beygi *et al.*, “Rate-adaptive coded modulation for fiber-optic communications,” *J. Lightw. Technol.* **32**(2), pp. 333–343, Jan. 2014. M. P. Yankov *et al.*, “Constellation shaping for fiber-optic channels with QAM and high spectral efficiency,” *Photon. Technol. Lett.* **26**(23), pp. 2407–2410, Dec. 2014. G. Böcherer, “Probabilistic signal shaping for bit-metric decoding,” in *Proc. Int. Symp. Inf. Theory (ISIT)*, pp. 431–435, Jul. 2014. P. Poggiolini *et al.*, “The GN-model of fiber non-linear propagation and its applications,” *J. Lightw. Technol.* **32**(4), pp. 694–721, Feb. 2014. F. R. Kschischang and S. Pasupathy, “Optimal nonuniform signaling for Gaussian channels,” *Trans. Inf. Theory* **39**(3), pp. 913–929, May 1993. S. Baur and G. Böcherer, “Arithmetic Distribution Matching,” \[Online\]. Available: <http://arxiv.org/abs/1408.3931>, Aug. 2014. B. P. Smith *et al.*, “Staircase codes: FEC for 100 Gb/s OTN,” *J. Lightw. Technol.* **30**(1), pp. 110–117, Jan. 2012.

ArXiv

--- abstract: 'Partial decay widths of lowest lying nucleon resonances $S_{11}(1535)$, $S_{11}(1650)$, $D_{13}(1520)$ and $D_{13}(1700)$ to the pseudoscalar mesons and octet baryons are studied within a chiral constituent quark model. Effects of the configurations mixing between the states $|N^{2}_{8}P_{M}\rangle$ and $|N^{4}_{8}P_{M}\rangle$ are considered, taking into account $SU(6)\otimes O(3)$ breaking effects. In addition, possible contributions of the strangeness components in the $S_{11}$ resonances are investigated. Experimental data for the partial decay widths of the $S_{11}$ and $D_{13}$ resonances are well reproduced. Predictions for coupling constants of the four nucleon resonances to pseudoscalar mesons and octet baryons, crucial issues in the photo- and hadron-induced meson production reactions, are reported. Contributions from five-quark components in the $S_{11}$ resonances are found crucial in reproducing the partial widths.' author: - 'C. S. An' - 'B. Saghai' title: 'Strong decay of low-lying [$S_{11}$]{} and [$D_{13}$]{} nucleon resonances to pseudoscalar mesons and octet baryons' --- Introduction {#intro} ============ Production of mesons with hidden or open strangeness via electromagnetic or hadronic probes, in the baryon resonance energy range, is subject to extensive experimental and theoretical investigations. In this realm, partial decay widths of resonances to meson-baryon final states, as well as the relevant coupling constants are crucial, but not well enough known [@Nakamura:2010zzi], ingredients in our understanding of the reaction mechanisms, and also of the nature of those resonances. Phenomenological approaches, dealing with the above ingredients, arise mainly from two families of formalisms: effective Lagrangians based on meson-baryon degrees of freedom [@Kaiser:1995cy; @Riska:2000gd; @Penner:2002ma; @Penner:2002md; @Inoue:2001ip; @Aznauryan:2003zg; @Liu:2005pm; @Chiang:2001pw; @Chiang:2004ye; @Nakayama:2004ek; @Nakayama:2005ts; @Nakayama:2008tg; @Vrana:1999nt; @Matsuyama:2006rp; @JuliaDiaz:2007kz; @Durand:2008es; @Cao:2008st; @Shyam:2003uz; @Shyam:2007iz; @Shklyar:2006xw; @Arndt:2005dg; @Anisovich:2005tf; @Sarantsev:2005tg; @Kaptari:2008ae; @Ceci:2006jj; @Geng:2008cv; @Xie:2010md; @Bruns:2010sv; @Gamermann:2011mq] and QCD based/inspired models  [@Koniuk:1979vy; @Capstick:1992th; @Capstick:1993kb; @Capstick:1998uh; @Capstick:2000qj; @Capstick:2004tb; @Kiswandhi:2003ca; @Golli:2011jk; @Kim:1997ym; @Yoshimoto:1999dr; @Zhu:1998qw; @Zhu:1998am; @Jido:1997yk; @Hyodo:2008xr; @Goity:2005rg; @Jayalath:2011uc]. Among the low-lying nucleon excitations, the $S_{11}(1535)$ resonance plays a special role due to its large $\eta N$ decay width [@Nakamura:2010zzi], though its mass is very close to the threshold of the decay. Moreover, in the $KY$ production reactions the importance of the $S_{11}(1650)$ is well established. For the two other first orbitally excited (quark model prediction) nucleon resonances, $D_{13}(1520)$ and $D_{13}(1700)$, the couplings to the pseudoscalar meson and octet baryons seem to be rather weak, but the first one is known to intervene significantly in the polarization asymmetries. The observables of interest in this paper are partial decay widths. Experimental values are available [@Nakamura:2010zzi] for all four resonances’ decay to $\pi N$ and $\eta N$ final states, as well as for the $S_{11}(1650)$ and $D_{13}(1700)$ resonances to $K \Lambda$, though with rather large uncertainties. However, in spite of extensive studies mentioned above, to our knowledge no single formalism has reproduced [*simultaneously*]{} those partial widths. The only exception here is a very recent comprehensive study [@Jayalath:2011uc] based on the $1/N_C$ expansion approach. Besides the fact that a large number of investigations concentrate on the $S_{11}$ resonances, recent copious photoproduction data have not yet been fully exploited by sophisticated coupled-channels phenomenological approaches. The main motivation of the present work is then to study those partial decay widths within a QCD inspired formalism, and shed light on the structure of those baryons. The theoretical frame of the present work is based on a chiral constituent quark model ($\chi$CQM), complemented with the $SU(6)\otimes O(3)$ symmetry breaking effects. The outcomes of those formalisms are compared to the known [@Nakamura:2010zzi] partial decay widths of the above mentioned resonances. This approach gives satisfactory results for the $D_{13}$ resonances, but misses partly the data for $S_{11}$. Attempting to cure the observed theory / experiment discrepancies, the $\chi$CQM is subsequently complemented with including contributions from higher Fock-components, namely, five-quark configurations. Actually, several authors [@Li:2005jn; @JuliaDiaz:2006av; @An:2005cj; @Li:2006nm; @Li:2005jb; @An:2008xk; @An:2008tz; @An:2010wb], have shown that contributions from the five-quark components are quite significant in describing the properties of baryons and their electromagnetic and strong decays, especially contributions from the $qqqq\bar{q} \to M(\gamma)+qqq$ transitions. For recent reviews on five-quark components in baryons, see Refs. [@Zou:2009zz; @Zou:2010tc; @riska]. The extended $\chi$CQM allows reproducing the known partial decay widths for both $S_{11}$ resonances. Following the successful results obtained for low-lying baryon resonances, we put forward predictions for the coupling constants of those resonances to seven meson-baryon final states, i.e. $\pi^{0} p$, $\pi^{+} n$, $\eta p$, $K^{+} \Lambda$, $K^{0}\Sigma^{+}$, $K^{+} \Sigma^{0}$, $\eta^{\prime} p$. The present manuscript is organized in the following way: in section \[theo\], we present the theoretical formalism which includes the wave functions, strong decays and the resulting transition amplitudes for the $S_{11}(1535)$, $S_{11}(1650)$, $D_{13}(1520)$ and $D_{13}(1700)$ to the pseudoscalar mesons and octet baryons. Numerical results are given in section \[num\], and finally section \[con\] contains summary and conclusions. Theoretical Formalism {#theo} ===================== In section \[wfm\], we present the wave functions of the nucleon resonances $S_{11}(1535)$, $S_{11}(1650)$, $D_{13}(1520)$ and $D_{13}(1700)$. Section \[form\] embodies a brief review of the formalism for the strong decay of the baryon resonances to meson-baryon in a $\chi$CQM, where we derive transition coupling amplitudes for the above four nucleon resonances to the $\pi N$, $\eta N$, $K\Lambda$, $K\Sigma$ and $\eta^{\prime}N$ channels. Wave functions {#wfm} -------------- In the $\chi$CQM, complemented with five-quark components, a baryon is a superposition of three- and five-quark mixture and the wave function can be written as $$|B\rangle=A_{3}|qqq\rangle+A_{5}|qqqq\bar{q}\rangle\,,$$ with $A_{3}$ and $A_{5}$ the probability amplitudes for the corresponding $qqq$ and $qqqq\bar{q}$ states, respectively. For the three-quark components, we employ the wave functions in traditional three-quark $\chi$CQM. In the $SU(6)\otimes O(3)$ conserved case, the general form for the wave functions of the octet baryons, $N(^{2}_{8}P_{M})_{S^{-}}$ and $N(^{4}_{8}P_{M})_{S^{-}}$ states, can be expressed as $$\begin{aligned} |B(^{2}_{8}S_{S})_{\frac{1}{2}^{+}},S_{z}\rangle&=&\frac{1}{\sqrt{2}}(|B\rangle_{\lambda}|\frac{1}{2},s_{z}\rangle_{\lambda} +|B\rangle_{\rho}|\frac{1}{2},s_{z}\rangle_{\rho})\varphi_{000}^{s}(\vec{\lambda},\vec{\rho})\,,\\ % \nonumber\\ |N(^{2}_{8}P_{M})_{S^{-}},S_{z}\rangle&=&\frac{1}{2}\sum_{m,s_{z}}C^{SS_{z}}_{1m,\frac{1}{2}s_{z}} [(|N\rangle_{\rho}|\frac{1}{2},s_{z}\rangle_{\lambda}+|N\rangle_{\lambda}|\frac{1}{2},s_{z}\rangle_{\rho}) \varphi^{\rho}_{11m}(\vec{\lambda},\vec{\rho})\nonumber\\ && +(|N\rangle_{\rho}|\frac{1}{2},s_{z}\rangle_{\rho}-|N\rangle_{\lambda}|\frac{1}{2},s_{z}\rangle_{\lambda}) \varphi^{\lambda}_{11m}(\vec{\lambda},\vec{\rho})]\,, % \nonumber\\ \label{suwfb} \\ % |N(^{4}_{8}P_{M})_{S^{-}},S_{z}\rangle&=&\frac{1}{\sqrt{2}}\sum_{m,s_{z}}C^{SS_{z}}_{1m,\frac{3}{2}s_{z}} [|N\rangle_{\rho}|\frac{3}{2},s_{z}\rangle \varphi^{\rho}_{11m}(\vec{\lambda},\vec{\rho}) +|N\rangle_{\lambda}|\frac{3}{2},s_{z}\rangle \varphi^{\lambda}_{11m}(\vec{\lambda},\vec{\rho})]\,, \label{suwfc}\end{aligned}$$ where $|B\rangle_{\rho(\lambda)}$ denotes the mixed symmetric flavor wave function of the three-quark system for the corresponding baryon. $|\frac{1}{2},s_{z}\rangle_{\rho(\lambda)}$ and $|\frac{3}{2},s_{z}\rangle$ are the mixed symmetric and symmetric spin wave functions of the three-quark system, respectively. $\varphi_{Nlm}(\vec{\lambda},\vec{\rho})$ is the harmonic oscillator basis orbital wave function for the three quarks with the subscripts $Nlm$ being the corresponding quantum numbers. Finally, $C^{SS_{z}}_{1m,ss_{z}}$ are the Clebsch-Gordan coefficients for the coupling of the orbital and spin of the three-quark system to form a baryon state with spin $S$ and z-component $S_{z}$. The explicit forms for all of the above flavor, spin, and orbital wave functions can be found in [@An:2010wb]. Taking into account the breakdown of $SU(6)\otimes O(3)$ symmetry due to either the color-magnetic [@De; @Rujula:1975ge] or flavor-magnetic [@Glozman:1995fu] hyperfine interactions between the quarks, one can express the wave functions of the $S_{11}$ and $D_{13}$ resonances in terms of the given $N(^{2}_{8}P_{M})_{S^{-}}$ and $N(^{4}_{8}P_{M})_{S^{-}}$ wave functions, Eqs. (\[suwfb\]) and (\[suwfc\]) , by introducing the configuration mixing angles $\theta _{S}$ and $\theta_{D}$ $$\begin{aligned} \pmatrix{ |S_{11}(1535)\rangle\cr |S_{11}(1650)\rangle\cr}&=& \pmatrix{ cos\theta_{S} & -sin\theta_{S} \cr sin\theta_{S} & cos\theta_{S} \cr} \pmatrix{ |N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle \cr |N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle \cr}\,, \label{ThetaS}\end{aligned}$$ $$\begin{aligned} \pmatrix{ |D_{13}(1520)\rangle\cr |D_{13}(1700)\rangle\cr}&=& \pmatrix{ cos\theta_{D} & -sin\theta_{D} \cr sin\theta_{D} & cos\theta_{D} \cr} \pmatrix{ |N(^{2}_{8}P_{M})_{\frac{3}{2}^{-}}\rangle \cr |N(^{4}_{8}P_{M})_{\frac{3}{2}^{-}}\rangle \cr}.\end{aligned}$$ For the octet baryons, other than the lowest lying $S_{11}$ and $D_{13}$, the configuration mixing effects are not so significant. So, for those baryons we take the wave functions within the exact $SU(6)\otimes O(3)$ symmetry. For the five-quark components of $S_{11}(1535)$, we use the wave functions given in Ref. [@An:2008tz], $$\begin{aligned} % \psi_{t,s}&=&\sum_{a,b,c}\sum_{Y,y,T_z,t_z}\sum_{S_z,s_z} C^{[1^4]}_{[31]_a[211]_a} C^{[31]_a}_{[F]_b [S]_c} [F]_{b,Y,T_z} [S]_{c,S_z} [211;C]_a (Y,T,T_z,y,\bar t,t_z|1,1/2,t)\nonumber\\ &&(S,S_z,1/2,s_z|1/2,s)\bar\chi_{y,t_z}\bar\xi_{s_z}\varphi_{[5]}\, . \label{wfc}\end{aligned}$$ In fact, this general wave function is appropriate for the five-quark components in all the low-lying nucleon resonances with $S^{p}=\frac{1}{2}^{-}$, albeit with different probabilities for five-quark components. As reported in Ref. [@An:2008tz], there are 5 different flavor-spin configurations which may form five-quark components in the resonances with negative parity. If the hyperfine interaction between the quarks is assumed to depend on flavor and spin, the energy of the second and third configurations should be about 80 MeV and 200 MeV higher than the first configuration, respectively. Since $S_{11}(1535)$ and $S_{11}(1650)$ are the first two orbital excitations of the nucleon with spin $1/2$, the configurations with low energies, namely the first two five-quark configurations should be the most appropriate ones to form higher Fock components in those two resonances. Moreover, the contribution of the second five-quark configuration is very similar to that of the first one, because of the same flavor structure, which rules out the five-quark components with light quark and anti-quark pairs in the $S_{11}$ resonances. Actually, the transition elements between all of the 5 five-quark configurations and the octet baryons differ just by constant factors. Therefore, the contributions from all the 5 configurations are similar, albeit with appropriate probability amplitudes. Consequently, the first configuration is enough for us to study the strong decays of $S_{11}(1535)$ and $S_{11}(1650)$. Then the wave functions for the five-quark components in $S_{11}(1535)$ and $S_{11}(1650)$ reduce to the following form: $$\psi_{5q}=\sum_{abc}C^{[1^4]}_{[31]_{a} [211]_{a}}C^{[31]_{a}}_{[211]_{b}[22]_{c}}[4]_{X}[211]_{F}(b)[22]_{S}(c)[211]_{C}(a) \bar \chi_{s_{z}}\varphi(\{\vec{\xi}_i\})\,,$$ the explicit form of which is given in Ref. [@An:2008xk]. Following Eq. (\[ThetaS\]), the introduction of five-quark wave functions leads to $$\begin{aligned} |S_{11}(1535)\rangle&=&A_{3}\left[cos\theta_{S}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle-sin\theta_{S} |N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle \right]+A_{5}\psi_{5q},\\ |S_{11}(1650)\rangle&=&A_{3}^{'}\left[sin\theta_{S}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle+cos\theta_{S} |N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle\right]+A_{5}^{\prime}\psi_{5q}\,. \label{wfc}\end{aligned}$$ The probability amplitude for the five-quark component in a baryon can be related to the coupling $_{5q}\langle\hat{V}_{cou} \rangle_{3q}$ between the $qqq$ and $qqqq\bar{q}$ configurations in the corresponding baryon $$\begin{aligned} A_{5q}=\frac{_{5q}\langle\hat{V}_{cou} \rangle_{3q}}{M-E_{5}}, \label{amp5}\end{aligned}$$ with $E_{5}$ the energy of the five-quark component. Given that the resonances considered here have negative parity, all of the quarks and anti-quark in the five-quark system should be in their ground states. Hence, we can take $\hat{V}_{cou}$ to be of the following form: $$\hat{V}_{cou}=3V(r_{34})\frac{\hat{\sigma}_{3}\cdot\vec{p}_{3}}{2m_{3}}\chi_{00}^{45}C_{00}^{45} F_{00}^{45}\varphi_{00}(\vec{p}_{4}-\vec{p}_{5})b_{4}^{\dagger}(\vec{p}_{4})d_{5}^{\dagger}(\vec{p}_{5})\,,$$ where $\chi_{00}^{45}$, $C_{00}^{45}$, $F_{00}^{45}$ and $\varphi_{00}(\vec{p}_{4}-\vec{p}_{5})$ denote the spin, flavor, color and orbital singlets of the quark and anti-quark pair, respectively. $b_{4}^{\dagger}(\vec{p}_{4})$ and $d_{5}^{\dagger}(\vec{p}_{5})$ are the creation operators for a quark and anti-quark pair with momentum $\vec{p}_{4}$ and $\vec{p}_{5}$, respectively. $V(r_{34})$ is the coupling potential which depends on the relative coordinate $|\vec{r}_{3}-\vec{r}_{4}|$. Then we obtain $$\frac{\langle\psi_{5q}|\hat{V}_{cou}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle} {\langle\psi_{5q}|\hat{V}_{cou}|N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}}\rangle}=-2,$$ and $$\frac{A_{5q}^{\prime}}{A_{5q}} = \frac{sin\theta_{S}-\frac{1}{2}cos\theta_{S}}{cos\theta_{S}+\frac{1}{2}sin\theta_{S}} %\times \frac{M_{S_{11}(1535)}-E_{5}}{M_{S_{11}(1650)}-E_{5}}. \label{ampratio}$$ Here we would like to emphasize that the considered $D_{13}$ resonances are not relevant for five-quark components issues. Actually, all of the quarks and anti-quark should be in their ground states (lowest energy) to form the negative parity. Then the spin configuration of four-quark subsystem is limited to be $[31]_{S}$, for which the total spin of the four-quark subsystems is $S=1$, in order to combine with the anti-quark to form the required total spin $3/2$. For the configurations with spin $[31]_{S}$, the flavor-spin overlap factors between such five-quark configurations and the $D_{13}$ states vanish. Therefore, the probabilities for these five-quark components in the $D_{13}$ resonances are $0$. Some additional five-quark configurations, other than those given in Ref. [@An:2008tz], could also be considered, for instance, the configurations with the anti-quark orbitally excited ($l_{\bar{q}}=2,4\cdots$), the ones in which the four-quark subsystem with spin symmetry $[4]_{S}$ ($S_{4}=2$), or the ones given in Ref. [@An:2005cj] with the four quark subsystem orbital symmetry $[31]_{X}$ and orbital momentum $L_{4}=2,4\cdots$. However, all those configurations have very high energies, far away from the lowest lying $D_{13}$ resonances masses. Finally, we do not consider the five-quark components in the ground states of octet baryons in this manuscript, because on the one hand their probabilities in the baryons are very small [@JuliaDiaz:2006av; @Li:2007hn], and on the other hand their contributions to electromagnetic and strong decays of nucleon resonances are negligible [@An:2008xk]. Actually, the five-quark configurations in the ground states of octet baryons cannot transit to three-quark components of the first orbitally excited baryon resonances due to the vanishing flavor-spin overlap factors. Formalism for strong decay {#form} -------------------------- It is well known that the pseudoscalar meson-quark coupling, in the tree level approximation, takes the form $$H_M=\sum_{j}\frac{g^{q}_{A}}{2f_{M}}\bar{\psi}_{j}\gamma^{j}_{\mu}\gamma_{5}^{j}\psi_{j}\partial^{\mu}\phi_{M}\,, \label{cou}$$ where $\psi_{j}$ and $\phi_{M}$ are the quark and pseudoscalar fields, respectively, and $g^{q}_{A}$ is the axial coupling constant for the constituent quarks, the value of which is in the range $0.7-1.26$ [@Goity:1998jr; @Riska:2000gd; @Lahde:2002fe]. $f_{M}$ denotes the decay constant of the corresponding meson; the empirical values for the decay constants of $\pi$, $K$, $\eta$ and $\eta^{\prime}$ are $f_{\pi}=93$ MeV, $f_{K}=113$ MeV, $f_{\eta}=1.2f_{\pi}$, $f_{\eta^{\prime}}=-0.58f_{\pi}$. In the framework of non-relativistic $qqq$ quark model, the coupling, Eq. (\[cou\]), takes the following form: $$H^{NR(3)}_{M}=\sum_{j}\frac{g^{q}_{A}}{2f_{M}}(\frac{\omega_{M}}{E_{f}+M_{f}}\sigma\cdot\vec{P}_{f}+ \frac{\omega_{M}}{E_{i}+M_{i}}\sigma\cdot\vec{P}_{i}-\sigma\cdot\vec{k}_{M} +\frac{\omega_{M}}{2\mu}\sigma\cdot\vec{p}_{j})X^{j}_{M}\exp\{-i\vec{k}_{M}\cdot\vec{r}_{j}\}\,. \label{op3}$$ Here, $\vec{k}_{M}$ and $\omega_{M}$ are the three momentum and energy of the final meson, $\vec{P}_{i(f)}$ and $M_{i(f)}$ denote the mass and three momentum of the initial (final) baryon, $\vec{p}_{j}$ and $\vec{r}_{j}$ the three momentum and coordinate of the $j^{th}$ quark, and $\mu$ is the reduced mass of the initial and final $j^{th}$ quark which emits the meson. Finally, $X^{j}_{M}$ is the flavor operator for emission of the meson from the corresponding $j^{th}$ quark, given by following expressions: $$\begin{aligned} & X^{j}_{\pi^{0}}=\lambda_{3}^{j}, ~X^{j}_{\pi^{\pm}}=\mp\frac{1}{\sqrt{2}}(\lambda_{1}^{j}\mp\lambda_{2}^{j}),\nonumber\\ & X^{j}_{K^{\pm}}=\mp\frac{1}{\sqrt{2}}(\lambda_{4}^{j}\mp\lambda_{5}^{j}),~X^{j}_{K^{0}}= \mp\frac{1}{\sqrt{2}}(\lambda_{6}^{j}\mp\lambda_{7}^{j}),\\ & X^{j}_{\eta}=cos\theta\lambda_{8}^{j}-sin\theta\sqrt{\frac{2}{3}}\mathcal{I}, ~X^{j}_{\eta^\prime}=sin\theta\lambda_{8}^{j}+cos\theta\sqrt{\frac{2}{3}}\mathcal{I}\nonumber\,,\end{aligned}$$ where $\lambda^{j}_{i}$ are the $SU(3)$ Gell-Mann matrices, and $\mathcal{I}$ the unit operator in the $SU(3)$ flavor space. $\theta$ denotes the mixing angle between $\eta_{1}$ and $\eta_{8}$, leading to the physical $\eta$ and $\eta ^\prime$ $$\begin{aligned} \eta &=& \eta _8 cos \theta - \eta _1 sin \theta \,,\\ % \eta ^\prime &=& \eta _8 sin \theta + \eta _1 cos \theta \,, \label{eq:etamix}\end{aligned}$$ it takes the value $\theta=-23$ [@Gobbi:1993au]. Taking into account the five-quark components in the resonances, we have to calculate the transition coupling amplitudes for $qqqq\bar{q}\to qqq+M$. The reduced form of the coupling in Eq. (\[cou\]) reads $$H_{M}^{NR(5)}=\sum_{j}\frac{g^{q}_{A}}{2f_{M}}C_{XFSC}^{j}(m_{i}+m_{f}) \bar{\chi}^{\dagger}_{z}\pmatrix{ 1&0\cr 0&1 \cr}\chi_{z}^{j}X_{M}^{j}\exp\{-i\vec{k}_{M}\cdot\vec{r}_{j}\}\,, \label{op5}$$ where $m_{i}$ and $m_{f}$ denote the constituent masses of the quark and anti-quark which combine to form a pseudoscalar meson, $C_{XFSC}^{j}$ denotes the overlap between the three-quark configuration of the final baryon and the residual orbital-flavor-spin-color configuration of the three-quark system that is left in the initial $qqqq\bar{q}$ after the combination of the $j^{th}$ quark with the anti-quark into a final meson. The transitions $qqqs\bar{s}\to B+M$ scheme is shown in Fig. \[fig\]. where three quarks of the five-quark system go as spectators to form the final three-quark baryon, and the fourth quark gets combined with the strange anti-quark to form a meson: $K$, $\eta$ or $\eta^{\prime}$. ![(Color online)Strangeness component transit in the $S_{11}$ resonances to $\eta p$ or $\eta^\prime p$ (a), $K^+ \Lambda$ or $K^+ \Sigma ^0$ (b) and $K^0 \Sigma^+$ (c). \[fig\]](fig.eps) Then, the transition coupling amplitude for a resonance to a pseudoscalar meson and a octet baryon is obtained by calculating the following matrix element: $$T^{MB}=\langle B(^{2}_{8}S_{S})_{\frac{1}{2}^{+}}|(H_{M}^{NR(3)}+H_{M}^{(5)})|N^{*}\rangle \equiv T^{MB}_{3}+T^{MB}_{5},$$ the resulting transition coupling amplitudes $T^{MB}_{3}$ and $T^{MB}_{5}$ for the $S_{11}$ and $D_{13}$ resonances to $\pi^{0}p$, $\pi^{+}n$, $\eta p$, $K^{+}\Lambda$, $K^{0}\Sigma^{+}$, $K^{+}\Sigma^{0}$ and $\eta^{'}p$ channels are shown in Tables \[a3qsd\] and \[a5q\], respectively. $S_{11}(1535)$ $S_{11}(1650)$ ------------------- ------------------------------------------------------------------------------------------------------------------------- -- -- ------------------------------------------------------------------------------------------------------------------------- $\pi^{0}p$ $\frac{\sqrt{2}}{9}(2cos\theta_{S}-sin\theta_{S})$ $\frac{\sqrt{2}}{9}(2sin\theta_{S}+cos\theta_{S})$ $\pi^{+}n$ $-\frac{2}{9}(2cos\theta_{S}-sin\theta_{S})$ $-\frac{2}{9}(2sin\theta_{S}+cos\theta_{S})$ $\eta p$ $\frac{\sqrt{2}}{3}(cos\theta_{S}+sin\theta_{S})(\frac{1}{\sqrt{3}}cos\theta-\sqrt{\frac{2}{3}}sin\theta)$ $\frac{\sqrt{2}}{3}(sin\theta_{S}-cos\theta_{S})(\frac{1}{\sqrt{3}}cos\theta-\sqrt{\frac{2}{3}}sin\theta)$ $K^{0}\Lambda$ $-\frac{1}{\sqrt{6}}cos\theta_{S}$ $-\frac{1}{\sqrt{6}}sin\theta_{S}$ $K^{0}\Sigma^{+}$ $-\frac{1}{9}(cos\theta_{S}+4sin\theta_{S})$ $-\frac{1}{9}(sin\theta_{S}-4cos\theta_{S})$ $K^{+}\Sigma^{0}$ $-\frac{1}{9\sqrt{2}}(cos\theta_{S}+4sin\theta_{S})$ $-\frac{1}{9\sqrt{2}}(sin\theta_{S}-4cos\theta_{S})$ $\eta^{\prime}p$ $\frac{\sqrt{2}}{3}(cos\theta_{S}+sin\theta_{S})(\frac{1}{\sqrt{3}}sin\theta+\sqrt{\frac{2}{3}}cos\theta)$ $\frac{\sqrt{2}}{3}(sin\theta_{S}-cos\theta_{S})(\frac{1}{\sqrt{3}}sin\theta+\sqrt{\frac{2}{3}}cos\theta)$ $D_{13}(1520)$ $D_{13}(1700)$ $\pi^{0}p$ $-\frac{2}{9}(2cos\theta_{D}-\frac{1}{\sqrt{10}}sin\theta_{D})$ $-\frac{2}{9}(2sin\theta_{D}+\frac{1}{\sqrt{10}}cos\theta_{D})$ $\pi^{+}n$ $\frac{2\sqrt{2}}{9}(2cos\theta_{D}-\frac{1}{\sqrt{10}}sin\theta_{D})$ $\frac{2\sqrt{2}}{9}(2sin\theta_{D}+\frac{1}{\sqrt{10}}cos\theta_{D})$ $\eta p$ $-\frac{2}{3}(cos\theta_{D}+\frac{1}{\sqrt{10}}sin\theta_{D})(\frac{1}{\sqrt{3}}cos\theta-\sqrt{\frac{2}{3}}sin\theta)$ $-\frac{2}{3}(sin\theta_{D}-\frac{1}{\sqrt{10}}cos\theta_{D})(\frac{1}{\sqrt{3}}cos\theta-\sqrt{\frac{2}{3}}sin\theta)$ $K^{0}\Lambda$ $\frac{1}{\sqrt{3}}cos\theta_{D}$ $\frac{1}{\sqrt{3}}sin\theta_{D}$ $K^{0}\Sigma^{+}$ $\frac{1}{9}(\sqrt{2}cos\theta_{D}+\frac{4}{\sqrt{5}}sin\theta_{D})$ $\frac{1}{9}(\sqrt{2}sin\theta_{D}-\frac{}{\sqrt{5}}4cos\theta_{D})$ $K^{+}\Sigma^{0}$ $-\frac{1}{9\sqrt{2}}(\sqrt{2}cos\theta_{D}+\frac{4}{\sqrt{5}}sin\theta_{D})$ $-\frac{1}{9\sqrt{2}}(\sqrt{2}sin\theta_{D}-\frac{4}{\sqrt{5}}cos\theta_{D})$ $\eta^{\prime}p$ $-\frac{2}{3}(cos\theta_{D}+\frac{1}{\sqrt{10}}sin\theta_{D})(\frac{1}{\sqrt{3}}sin\theta+\sqrt{\frac{2}{3}}cos\theta)$ $-\frac{2}{3}(sin\theta_{D}-\frac{1}{\sqrt{10}}cos\theta_{D})(\frac{1}{\sqrt{3}}sin\theta+\sqrt{\frac{2}{3}}cos\theta)$ : Transition coupling amplitudes $T^{MB}_{3}$ for the low-lying $S_{11}$ and $D_{13}$ resonances to meson-baryon final states. Note that the full amplitudes are obtained by multiplying each term by the following expressions: $\frac{g^{q}_{A}}{2f_{M}}\omega_{3}[(a_{M}-\frac{b_{M}}{3})\frac{k_{M}^{2}}{\omega_{3}^{2}}-3b_{M}] \exp\{-\frac{k^{2}_{M}}{6\omega_{3}^{2}}\}$ for $S_{11}$ and $\frac{g^{q}_{A}}{2f_{M}}(a_{M}-\frac{b_{M}}{3})\frac{k_{M}^{2}}{\omega_{3}} \exp\{-\frac{k^{2}_{M}}{6\omega_{3}^{2}}\}$ for $D_{13}$ resonances. Here, $\omega_{3}$ is the harmonic oscillator parameter for the three-quark components, $a_{M}=1+\frac{\omega_{M}}{E_{f}+M_{f}}$ and $b_{M}=\frac{\omega_{M}}{2\mu}$. \[a3qsd\] Notice that (Table \[a3qsd\]), within the exact $SU(6)\otimes O(3)$ symmetry, the matrix elements for transition $N(^{4}_{8}P_{M})_{S^{-}}\to K\Lambda$ vanish, and hence the decay widths of $S_{11}(1650)$ and $D_{13}(1700)$ to $K\Lambda$ are null. Moreover, Table \[a5q\], the transition elements for $5q\to MB$ do not vanish when $k_{M}=0$, and it may enhance or depress the transitions $S_{11}\to MB$ significantly near the meson-baryon threshold. Finally, the strangeness component does not transit to $\pi^{0}p$, since the matrix element of the flavor operator $X^{j}_{\pi^{0}}$ between the $s\bar{s}$ pair is $0$. $\pi^{0}p$ $\pi^{+}n$ $\eta p$ $K^{+}\Lambda$ $K^{0}\Sigma^{+}$ $K^{+}\Sigma^{0}$ $\eta^{\prime}p$ ------------ -- ------------ -- --------------------------------------------------------- -- ------------------------------- -- --------------------- -- ------------------- -- --------------------------------------------------------- $0$ $0$ $\frac{2}{\sqrt{3}}m_{s}(2cos\theta+\sqrt{2}sin\theta)$ $\frac{1}{\sqrt{3}}(m+m_{s})$ $\sqrt{2}(m+m_{s})$ -$(m+m_{s})$ $\frac{2}{\sqrt{3}}m_{s}(2sin\theta-\sqrt{2}cos\theta)$ : Transition coupling amplitudes $T^{MB}_{5}$. Note that the full amplitudes are obtained by multiplying each term by the following expression: $\frac{g^{q}_{A}}{2f_{M}}C_{35}\exp\{-\frac{3k^{2}_{M}}{20\omega_{5}^{2}}\}$, with $C_{35}$ related to the harmonic oscillator parameter for the three- and five-quark components as $C_{35}=(\frac{2\omega_{3}\omega_{5}}{\omega_{3}^{2}+\omega_{5}^{2}})^{3}$. \[a5q\] To obtain the relevant expressions for partial decay widths, we take the Lagrangian for $N^{*}MB$ coupling in hadronic level to be of the following form: $$\begin{aligned} \mathcal{L}_{S_{11}BM}&=&-ig_{S_{11}BM}\bar{\psi}_{B}\phi_{M}\psi_{S_{11}}+h.c.,\\ \mathcal{L}_{D_{13}BM}&=&\frac{1}{m_{M}} g_{D_{13}BM} \bar{\psi}_{B}\partial_{\mu}\phi_{M}\psi_{D_{13}}^{\mu}+h.c., \label{lag}\end{aligned}$$ where $\bar{\psi}_{B}$ and $\psi_{S_{11}}$ denote the Dirac spinor fields for the final baryon and the $S_{11}$ resonances, respectively, and $\phi_{M}$ is the scalar field for the final meson. For the $D_{13}$ resonances, with spin $3/2$, we employ the Rarita-Schwinger vector-spinor fields $\psi_{D_{13}}^{\mu}$ [@Rarita:1941mf; @Nath:1971wp], which are defined as $$\begin{aligned} \psi_{D_{13}}^{\mu}(S_{z})=\sum_{ms}C^{\frac{3}{2}S_{z}}_{1m,\frac{1}{2}s} \epsilon^{\mu}_{m}u_{s}. \end{aligned}$$ One can directly obtain the transition coupling amplitudes for $N^{*}\to MB$ in the hadronic level using the Lagrangian, Eq. (\[lag\]). Then, the coupling constants $g_{N^{*}MB}$ are extracted by comparing the transition coupling amplitudes $T^{MB}$ in the quark model to those in the hadronic model. With the resulting coupling constants, the strong decay widths for the $S_{11}$ and $D_{13}$ resonances to the pseudoscalar meson and octet baryon read $$\begin{aligned} \Gamma_{S_{11}\to MB}&=&\frac{1}{4\pi}g_{S_{11}MB}^{2}\frac{E_{f}+M_{f}}{M_{i}}|\vec{k}_{M}|\,, \label{widthS}\end{aligned}$$ $$\begin{aligned} \Gamma_{D_{13}\to MB}&=&\frac{1}{12\pi}\frac{1}{m_{M}^{2}} g_{D_{13}MB}^{2} \frac{E_{f}-M_{f}}{M_{i}}|\vec{k}_{M}|^{3}\,. \label{widthD}\end{aligned}$$ Note that in the center of mass frame of the initial resonance, $\vec{P}_{i}=0$, $\vec{k}_{M}$ and $E_{f}$ can be related to the masses of the initial and final hadrons as $$\begin{aligned} |\vec{k}_{M}|&=&|\vec{P}_{f}|=\frac{\sqrt{[M^{2}_{i}-(M_{f}+m_{M})^{2}][M^{2}_{i}-(M_{f} -m_{M})^{2}]}}{2M_{i}}\,,\\ \nonumber\\ E_{f}&=&\sqrt{|\vec{k}_{M}|^{2}+M_{f}^{2}}=\frac{M_{i}^{2}-m_{M}^{2}+M_{f}^{2}}{2M_{i}}. \label{kMEf}\end{aligned}$$ For decay channels with thresholds above the mass of the initial resonance, off-shell effects are taken into account by putting $|\vec{k}_{M}|=0$ and introducing the form factor [@Penner:2002ma] $$F=\frac{\Lambda^{4}}{\Lambda^{4}+(q_{N^{*}}^{2}-M_{N^{*}}^{2})^{2}}\,,$$ with the cutoff parameter $\Lambda=1$ GeV, and $q_{N^{*}}$ the threshold of the corresponding channel. In fact, this form factor affects mainly the $N^* \to \eta^{\prime}N$ process, since thresholds for all other channels are below or slightly above the masses of the four resonances. Numerical Results {#num} ================= In this section our results for partial decay widths $\Gamma_{N^* \to MB}$ and coupling constants $g_{N^* MB}$ are reported for the four investigated resonances, with $MB \equiv \pi^{0}p$, $\pi^{+}n$, $\eta p$, $K^{+}\Lambda$, $K^{0}\Sigma^{+}$, $K^{+}\Sigma^{0}$ and $\eta^{\prime}p$. The starting point, section  \[num3q\], is the standard $\chi$CQM. Then, in section \[num3qb\] we introduce $SU(6)\otimes O(3)$ breaking and finally, in section \[numtt\], the five-quark components are embodied for the $S_{11}$ resonances. For the partial decay widths, we compare our results to the experimental values reported in PDG [@Nakamura:2010zzi], and produce predictions for yet unmeasured channels. Pure $qqq$ configuration and exact $SU(6)\otimes O(3)$ symmetry {#num3q} --------------------------------------------------------------- Within this simplest configuration, there are three input parameters: quarks’ masses and harmonic oscillator parameter. For the constituent quarks’ masses, we use the traditional $qqq$ quark model values [@Glozman:1995fu; @Koniuk:1979vy; @Capstick:2000qj], namely, $m \equiv m_{u}=m_{d}=340$ MeV and $m_{s}=460$ MeV. The scale of the oscillator parameter, $\omega_{3}$, can be inferred from the empirical radius of the proton via $\omega_{3}=1/\sqrt{\langle r^{2}\rangle}$, which leads to $\omega_{3}\simeq250$ MeV, for $\sqrt{\langle r^{2}\rangle} \simeq 1$ fm. However, since the photon couples to $u$ and $d$ quarks through $\rho$ and $\omega$ mesons, the measured proton charge radius may reflect partly the vector meson propagator [@weisebook]. Moreover, pion cloud have some influence on the measured proton charge radius. Consequently, the intrinsic size of the proton still has some model dependence, and hence, the oscillator parameter $\omega_{3}$ might deviate from 250 MeV, within the range $100-400$ MeV [@Li:2006nm; @JuliaDiaz:2006av; @Koniuk:1979vy; @Capstick:2000qj]. ![(Color online) Partial decay widths of $S_{11}(1535)$ and $S_{11}(1650)$ to $\pi N$ and $\eta N$ channels as a function of the harmonic oscillator parameter $\omega_{3}$. Results of the present work are depicted in full and dashed curves for $S_{11} \to \pi N$ and $S_{11} \to \eta N$, respectively, without the $SU(6)\otimes O(3)$ breakdown effects. The horizontal lines are the bands given in PDG, for $S_{11} \to \pi N$ (dash-dotted) and $S_{11} \to \eta N$ (dash-dot-dotted). \[omega\]](3qo.eps) Figure \[omega\] shows the decay widths for $S_{11}(1535) \to \pi N,~\eta N$ (left panel) and $S_{11}(1650) \to \pi N,~\eta N$ (right panel) as a function of $\omega_{3}$. The full and dashed curves are our results and the horizontal lines give the bands reported in PDG [@Nakamura:2010zzi]. The width for $S_{11}(1535) \to \pi N$ (full curve) falls in the experimental range (dash-dotted lines) for $ 300 \lesssim \omega_{3} \lesssim 340$ MeV, while for $S_{11}(1535) \to \eta N$ the dashed curve and dash-dot-dotted lines lead to $ 300 \lesssim \omega_{3} \lesssim 380$ MeV. Accordingly, in the former range for $\omega_{3}$, the simple $qqq$ configuration allows reproducing the decay widths of $S_{11}(1535)$ in both $\pi N$ and $\eta N$ channels. The situation with respect to the second $S_{11}$ resonance is dramatically different. In the whole $\omega_{3}$ range, the calculated $S_{11}(1650) \to \pi N$ width (full curve), underestimates the experimental band (dash-dotted lines). For the $\eta N$ decay channel, predicted values (dashed curve) agree with experimental band (dash-dot-dotted lines) below $\omega_{3} \approx 200$ MeV, where $\Gamma_{S_{11}(1650) \to \pi N}$ turns out vanishing. It is also worthwhile mentioning that, within exact $SU(6)\otimes O(3)$ symmetry, $\Gamma_{S_{11}(1650) \to K \Lambda}=0$ and hence, disagrees with the experimental value [@Nakamura:2010zzi]: $4.8 \pm 0.7$ MeV. In summary the pure $qqq$ configuration, within exact $SU(6)\otimes O(3)$, is not appropriate in describing the $S_{11}(1650)$ resonance properties. Consequently, one has to consider the $SU(6)\otimes O(3)$ breakdown effects. Pure $qqq$ configuration and broken $SU(6)\otimes O(3)$ symmetry {#num3qb} ---------------------------------------------------------------- As discussed in section \[wfm\], $SU(6)\otimes O(3)$ symmetry breaking effects can be related to the mixing angles $\theta_{S}$ and $\theta_{D}$. Several predictions on those angles are available (for a recent review see e.g. Ref. [@Saghai:2009zz]). Here, we will extract ranges for both angles and discuss them with respect to the two most common approaches leading to $SU(6)\otimes O(3)$ symmetry breaking, namely, one-gluon-exchange (OGE) [@Isgur:1977ef; @Isgur:1978xb; @Isgur:1978xi; @Isgur:1978xj; @Isgur:1978wd; @Koniuk:1979vy] and one-boson-exchange models (OBE) [@Glozman:1995fu]. Those approaches have raised some controversy [@Isgur:1999jv; @Glozman:1999ms]. Given that both the sign and the magnitude of the mixing angles in those approaches are different (see e.g. Refs. [@Saghai:2009zz; @Chizma:2002qi]), and that even within a given approach, the sign depends on the convention used  [@Saghai:2009zz; @Koniuk:1979vy] or on the exchanged mesons included [@Glozman:1998wk], we give in Appendix \[apdx:mix\] values obtained within each approach in line with the de Swart [@de; @Swart:1963gc] convention for SU(3). In order to investigate the sign and range for $\theta_{S}$, in this section we report our numerical results for partial decay widths of $S_{11}(1535)$ and $S_{11}(1650)$ to ${\pi N}$ and $\eta N$ as a function of $\omega_3$ for six values of $\theta_{S}$, namely, $\pm 15^\circ,~\pm 30^\circ,~\pm 45^\circ$, and compare them to the data ranges. In Fig. \[minus\] the strong decay partial widths $\Gamma_{S_{11} \to \pi N}$ and $\Gamma_{S_{11} \to \eta N}$ for $S_{11}(1535)$ (upper panel) and $S_{11}(1650)$ (lower panel) are shown as a function of $\omega_{3}$, with negative values for $\theta_{S}$. Conventions for the curves are the same as in Fig. \[omega\], and due to $SU(6)\otimes O(3)$ symmetry breaking, $\Gamma_{S_{11}(1650) \to K\Lambda}$ gets non-vanishing values, depicted in dotted curves. The experimental bands for this latter width are not shown, because they are almost identical to those for $\Gamma_{S_{11}(1650) \to \eta N}$. At all the three mixing angles, our predictions for $\Gamma_{S_{11}(1535) \to \pi N}$ and $\Gamma_{S_{11}(1650) \to \eta N}$ fall in the experimental bands for $\omega_3 \approx$ 300 MeV, while the model underestimates very badly $\Gamma_{S_{11}(1535) \to \eta N}$ and $\Gamma_{S_{11}(1650) \to \pi N}$. Accordingly, within our approach, negative values for $\theta_{S}$ lead to unacceptable results compared to the data. ![(Color online) Decay widths of $S_{11}(1535)$ (upper panel) and $S_{11}(1650)$ (lower panel) as a function of $\omega_{3}$, with $\theta_{S}$ taken to be $- \frac{\pi}{4}$, $- \frac{\pi}{6}$ and $ - \frac{\pi}{12}$, respectively. The dotted curves are our results for $\Gamma_{S_{11}(1650) \to K\Lambda}$, and the other ones are as in Fig. \[omega\]. \[minus\]](minus.eps) ![(Color online) Same as Fig. \[minus\], but with $\theta_{S}$ taken to be $\frac{\pi}{12}$, $\frac{\pi}{6}$ and $\frac{\pi}{4}$. \[posit\]](posit.eps) In Fig. \[posit\] the strong decay partial widths $\Gamma_{S_{11} \to \pi N}$ and $\Gamma_{S_{11} \to \eta N}$ for both $S_{11}$ resonances, as well as $\Gamma_{S_{11}(1650) \to K\Lambda}$, are depicted as a function of $\omega_{3}$ with positive values for $\theta_{S}$. For the $S_{11}(1535)$ resonance, we obtain good agreement with data for $\theta_{S}= 15^\circ$ and $\omega_3 \approx 300$ MeV, for both ${\pi N}$ and ${\eta N}$ decay widths. This is also the case at all angles for $S_{11}(1650) \to {\pi N}$, but for $\omega_3 \approx 350$ MeV. To go further in our investigation, we fix the harmonic parameter at $\omega_3 = 340$ MeV and calculate partial widths and coupling constants for two extreme positive values of the mixing angle, $\theta_{S}= 15^\circ$ and $35^\circ$. Moreover, we extend our study to the $D_{13}(1520)$ and $D_{13}(1700)$ resonances, with the relevant mixing angle, also at two extreme values, $\theta_{D}= 0^\circ$ and $17.5^\circ$. Results obtained within this procedure are hereafter referred to as model A. [lccccccccccc]{} $N^*$ & $\Gamma_{tot}$ && $\pi N$ && $\eta N$ && $K \Lambda$ && Ref.\ $S_{11}(1535)$ & 150 $\pm$ 25 && 68 $\pm$ 15 && 79 $\pm 11$ && && PDG [@Nakamura:2010zzi]\ & && 51 $\pm$ 21 &&121 $\pm$ 15 && && Model A\ \ $S_{11}(1650)$ & 165 $\pm$ 20 && 128 $\pm$ 29 && 3.8 $\pm$ 3.6 && 4.8 $\pm$ 0.7 && PDG [@Nakamura:2010zzi]\ & && 81 $\pm$ 22 && 28 $\pm$ 22 && 9 $\pm$ 6 && Model A\ \ $D_{13}(1520)$ & 115 $\pm$ 15 && 69 $\pm$ 6 && 0.26 $\pm$ 0.05 && && PDG [@Nakamura:2010zzi]\ & && 66 $\pm$ 7 && 0.19 $\pm$ 0.01&& && Model A\ & && 72 $\pm$ 11 && 0.26 $\pm$ 0.07&& && Jayalath [*et al.*]{} [@Jayalath:2011uc]\ \ $D_{13}(1700)$ & 100 $\pm$ 50 && 10 $\pm$ 5 && 0.5 $\pm$ 0.5 && 1.5 $\pm 1.5$ && PDG [@Nakamura:2010zzi]\ & && 13 $\pm$ 10 && 0.5 $\pm$ 0.5 && 0.1 $\pm$ 0.1 && Model A\ & && 12 $\pm$ 13 && $\leq$ 0.15 && $\leq$ 0.03 && Jayalath [*et al.*]{} [@Jayalath:2011uc]\ \[BRD13\] In Table \[BRD13\], we present our results for the strong decay partial widths $\Gamma_{\pi N}$, $\Gamma_{\eta N}$ and $\Gamma_{K \Lambda}$ for the low lying $S_{11}$ and $D_{13}$ resonances studied here. Within model A, the reduced $\chi^2$ per data point is 10.3. However, this large value is due to $S_{11} \to \eta N,~ K \Lambda$ decay channels. It is worthwhile mentioning that for the five $D_{13}$ partial decay widths, we get $\chi^2_{d.p.}$ = 0.7. Here, $\Gamma_{S_{11}(1535) \to \pi N}$ is well reproduced, while $\Gamma_{S_{11}(1535) \to \eta N}$ is overestimated at the level of 3$\sigma$ and $\Gamma_{S_{11}(1650) \to \pi N}$ underestimated by roughly 2$\sigma$. For the remaining two other channels, large uncertainties on $\Gamma_{S_{11}(1650) \to \eta N}$ (both experiment and model), and on $\Gamma_{S_{11}(1650) \to K \Lambda}$ (mainly model) do not lead to reliable conclusions. Because of those undesirable features, we postpone to the next section the discussion on results from other sources, as well as the extraction of coupling constants. For both $D_{13}$ resonances, the model A allows reproducing satisfactorily enough (Table \[BRD13\]) the known partial widths, and agrees with values obtained within the $1/N_C$ expansion framework [@Jayalath:2011uc]. model A is hence appropriate to put forward predictions for $D_{13}$-meson-baryon coupling constants. In Table \[ccd13\], our predictions for $\Gamma_{D_{13}MB}$ for seven meson-baryon sets are reported. $N^*$ $\pi^{0}p$ $\pi^{+}n$ $\eta p$ $K^{+}\Lambda$ $K^{0}\Sigma^{+}$ $K^{+}\Sigma^{0}$ $\eta^{\prime}p$ ---------------- -- ------------------ ----------------- ------------------ ----------------- ------------------- ------------------- ------------------ -- $D_{13}(1520)$ -1.51 $\pm$ 0.07 2.13 $\pm$ 0.10 -8.33 $\pm$ 0.20 3.44 $\pm$ 0.08 0.99 $\pm$ 0.14 -0.69 $\pm$ 0.09 2.11 $\pm$ 0.05 $D_{13}(1700)$ -0.35 $\pm$ 0.17 0.50 $\pm$ 0.25 0.93 $\pm$ 0.91 1.43 $\pm$ 1.43 -2.80 $\pm$ 0.05 1.98 $\pm$ 0.04 1.67 $\pm$ 0.52 : Coupling constants for $D_{13}$ resonances to pseudoscalar meson and octet baryon within model A. \[ccd13\] To end this section, we summarize our main findings within a traditional $qqq$ $\chi$CQM, complemented with $SU(6)\otimes O(3)$ breakdown effects, and using following input values for adjustable parameters: $\omega _3$ = 340 MeV, $15^\circ \leq \theta_{S} \leq 35^\circ$ and $0^\circ\leq \theta_{D}\leq 17.5^\circ$. Model A is found appropriate for the $D_{13}$ resonances, given that the partial decay widths show from reasonable to good agreements with the PDG values. So, we do not push further our studies with respect to the $D_{13}(1520)$ and $D_{13}(1700)$. The main shortcomings of the model A concern: $\Gamma_{S_{11}(1535) \to \eta N}$ and the fact that for the $S_{11}(1650)$ resonance, central values for all three channels show significant discrepancies with those reported in PDG. This latter point remains problematic because of large uncertainties. Attempting to cure those disagreements with respect to the $S_{11}$ resonances, we proceed in the next section to considering possible contributions from higher Fock-components. Mixed $qqq$ and $qqqq \bar q$ configuration and broken $SU(6)\otimes O(3)$ symmetry {#numtt} ----------------------------------------------------------------------------------- To produce numerical results, seven input parameters are needed, the values of which are discussed below. [*a) Constituent quarks’ masses:*]{} due to the introduction of five-quark components, masses to be used are smaller than those we adopted in section \[num3q\], while dealing with pure three-quark states. In line with Ref. [@An:2008xk], we take $m=290$ MeV and $m_{s}=430$ MeV. [*b) Oscillator parameters:*]{} following results presented in section \[num3q\], we fix the oscillator parameter at $\omega_{3}=340$ MeV. For the five-quark components a commonly used value for the oscillator parameter, $\omega_{5}=600$ MeV, is adopted. [*c) Mixing angle:*]{} in Section \[num3qb\], we showed that to fit the decay widths , the mixing angle should be in the range $15^\circ \leq \theta_{S} \leq 35^\circ$. In the following, this angle is treated as adjustable parameter. [*d) Probabilities of five-quark components:*]{} the probabilities of the five-quark components in $S_{11}(1535)$ ($P_{5q}=A_{5q}^{2}$) and $S_{11}(1650)$ ($P^{\prime}_{5q}=A^{\prime ^2}_{5q}$) are also adjustable parameters in our model search. The latter three adjustable parameters have been extracted by mapping out the whole phase space defined by $15^\circ \leq \theta_{S} \leq 35^\circ$ and from 0 to 100% for five-quark probabilities in both $S_{11}(1535)$ and $S_{11}(1650)$. The calculated observables are: the partial decay widths of both $S_{11}$ resonances to $\pi N$ and $\eta N$, as well as $\Gamma_{S_{11}(1650) \to K \Lambda}$. Sets \[$\theta_{S},~P_{5q},~P^{\prime}_{5q}$\] leading [@An:2011nu] to decay widths within ranges reported in PDG have been singled out. Then, for each partial widths, extreme values for those parameters are retained as model ranges, namely, $$26.8^\circ \leq \theta_{S} \leq 29.8^\circ~;~ 21\% \leq P_{5q} \leq 30\%~;~ 11\% \leq P^\prime_{5q} \leq 18\%\,. % \label{Ranges}$$ The obtained model is hereafter called model B. As an example, Fig. \[prob\] illustrates how the known ranges for the partial decay widths allow determining ranges for the five-quark components’ probabilities. There, for each decay width intersections of the model curve with the horizontal bands taken from PDG, determine the extreme values for the relevant five-quark probability. ![(Color online) Partial decay widths (in MeV) for $S_{11}$ resonances as a function of five-quark components, $\theta _S=28^\circ$. Curves the same as in Fig. \[posit\]. \[prob\]](prob.eps) -1.0 cm Notice that the probability range for five-quark component in $S_{11}(1535)$ given above is compatible with previous results [@An:2008xk; @An:2008tz], obtained within $\chi CQM$ approaches. The latter one [@An:2008tz] puts an upper limit of $P_{5q} \leq 45\%$, based on the axial charge study of the resonance. While the former one [@An:2008xk], dedicated to the electromagnetic transition $\gamma ^* N \to S_{11}(1535)$, reports $25\% \leq P_{5q} \leq 65\%$. ### Partial decay widths $\Gamma_{S_{11} \to MB}$ The resulting numerical partial decay widths, within both models A and B, are reported in Table \[BR\] and compared with the PDG data [@Nakamura:2010zzi] as well as with results from other authors, based on various approaches [@Penner:2002ma; @Inoue:2001ip; @Aznauryan:2003zg; @Vrana:1999nt; @Shyam:2007iz; @Arndt:2005dg; @Ceci:2006jj; @Golli:2011jk; @Jayalath:2011uc]. Comparing results of the models A and B with the data for all five channels, shows clearly the superiority of the model B. The $\chi^2_{d.p.}$ is 0.15, instead of 19.9 in the case of model A. The most striking feature here is that $\Gamma_{S_{11}(1535) \to \eta N}$ is nicely reproduced, which was not the case with previous configurations, namely, pure $qqq$ without or with $SU(6) \otimes O(3)$ symmetry breaking. Moreover, $\Gamma_{S_{11}(1535) \to \pi N}$ agrees with PDG values within better than 1$\sigma$. The range for $\Gamma_{S_{11}(1650) \to \pi N}$ gets significantly reduced within the model B with respect to the model A result and is compatible with the PDG value within less than 1$\sigma$. Narrow experimental widths for $\Gamma_{S_{11}(1650) \to \eta N}$ and $\Gamma_{S_{11}(1650) \to K\Lambda}$ are well reproduced by the model B, with uncertainties comparable to those of the data. In the following, we proceed to comparisons with results from other sources. The most complete set of results comes from a very recent comprehensive study [@Jayalath:2011uc] of all known partial decay widths for sixteen baryon resonances, within the framework of the $1/N_C$ expansion in the next to leading order (NLO) approximation. Results for the $S_{11}(1535)$ decay channels from that work and model B are in excellent agreement. For the $S_{11}(1650)$, given that the authors of Ref. [@Jayalath:2011uc] use branching fractions data in PDG for $\eta N$ and $K \Lambda$ channels, rather than the branching ratios, we postpone the comparisons to sec. \[brbf\]. [lccccccccccc]{} $N^*$ & $\Gamma_{tot}$ && $\pi N$ && $\eta N$&& $K \Lambda$ && Approach && Ref.\ $S_{11}(1535)$ & 150 $\pm$ 25 && 68 $\pm$ 15 && 79 $\pm$ 11 && && && PDG [@Nakamura:2010zzi]\ & && 51 $\pm$ 21 &&121 $\pm$ 15 && && Model A && Present work\ & && 58 $\pm$ 5 && 79 $\pm$ 11 && && Model B && Present work\ & && 57 $\pm$19 && 73 $\pm$ 44 && && $1/N_C$-NLO && Jayalath [*et al.*]{} [@Jayalath:2011uc]\ &112 $\pm$ 19&& 39 $\pm$ 5 && 57 $\pm$ 6 && && Coupled-channel && Vrana [*et al.*]{} [@Vrana:1999nt]\ &129 $\pm$ 8&& 46 $\pm$ 1 && 68 $\pm$ 1 && && Coupled-channel && Penner-Mosel [@Penner:2002ma]\ &136 && 34.4 && 56.2 && && Coupled-channel && Shyam [@Shyam:2007iz]\ & && 42 $\pm$ 6 && 70 $\pm$ 10 && && PWA && Arndt [*et al.*]{} [@Arndt:2005dg]\ & && 21.3 && 65.7 && && Chiral Unitary && Inoue [*et al.*]{} [@Inoue:2001ip]\ & 95 && 42 && 51 && && Chiral quark model && Golli [*et al.*]{} [@Golli:2011jk]\ & 165 && 64 && 89 && && K-Matrix && Ceci [*et al.*]{} [@Ceci:2006jj]\ & 142 && && 71 && && Disp. Rel. && Aznauryan [@Aznauryan:2003zg]\ & 195 && && 97 && && Isobar && Aznauryan [@Aznauryan:2003zg]\ \ $S_{11}(1650)$ & 165 $\pm$ 20 &&128 $\pm$ 29&& 3.8 $\pm$ 3.6 && 4.8 $\pm$ 0.7 && && PDG [@Nakamura:2010zzi]\ & && 81 $\pm$ 22&& 28 $\pm$ 22 && 9 $\pm$ 6 && Model A && Present work\ & &&143 $\pm$ 5 && 4.5 $\pm$ 3.0 && 4.8 $\pm$ 0.7 && Model B && Present work\ &202 $\pm$ 40&& 149 $\pm$ 4 && 12 $\pm$ 2 && && Coupled-channel && Vrana [*et al.*]{} [@Vrana:1999nt]\ & 138 $\pm$ 7 && 90 $\pm$ 6 && 1.4 $\pm$ 0.8 && 3.7 $\pm$ 0.6 && Coupled-channel && Penner-Mosel [@Penner:2002ma]\ &133 && 71.9 && 2.5 && && Coupled-channel && Shyam [@Shyam:2007iz]\ & 144 && 86 && 1.4 && 13 && Chiral quark model && Golli [*et al.*]{} [@Golli:2011jk]\ & 233 && 149 && 37 && && K-Matrix && Ceci [*et al.*]{} [@Ceci:2006jj]\ & 85 && && 3.2 && && Disp. Rel. && Aznauryan [@Aznauryan:2003zg]\ & 125 && && 6.9 && && Isobar && Aznauryan [@Aznauryan:2003zg]\ \[BR\] The Pitt-ANL [@Vrana:1999nt] multichannel analysis of $\pi N \to \pi N,~\eta N$, produces rather small total widths for $S_{11}(1535)$ and large one for $S_{11}(1650)$. Those features lead to underestimate of $\Gamma_{S_{11}(1535) \to \pi N}$ and $\Gamma_{S_{11}(1535) \to \eta N}$, and overestimate of $\Gamma_{S_{11}(1650) \to \eta N}$. However, $\Gamma_{S_{11}(650) \to \pi N}$ comes out in agreement with PDG and model B results. An extensive coupled-channels analysis [@Penner:2002ma; @Penner:2002md] studied within an isobar approach all available data by year 2002 for following processes: $\gamma N \to \gamma N$, $\pi N$, $\pi \pi N$, $\eta N$, $K \Lambda$, $K \Sigma$, $\omega N$ and $\pi N \to \pi N$, $\eta N$, $K \Lambda$, $K \Sigma$, $\omega N$. That work describes successfully four out of the five decay channels, albeit with a few tens of free parameters, with the main shortcoming being the underestimate of $\Gamma_{S_{11}(1535) \to \pi N}$. Interpreting $p N \to p N \eta$ data, within an effective Lagrangian approach [@Shyam:2007iz], underestimates all partial decay widths, except $\Gamma_{S_{11}(1650) \to K \Lambda}$. The latest available results from SAID [@Arndt:2005dg], in 2005, analyzing $\pi N$ elastic scattering and $\eta N$ production data, give a smaller $\Gamma_{S_{11}(1535) \to \pi N}$ with respect to PDG, and compatible with PDG value for $\Gamma_{S_{11}(1535) \to \eta N}$. A chiral unitary approach [@Inoue:2001ip] dedicated to the $S$-wave meson-baryon interactions, reproduces well $\Gamma_{S_{11}(1535) \to \eta N}$, but underestimates $\Gamma_{S_{11}(1535) \to \pi N}$ by more than a factor of 2. A recent chiral quark model [@Golli:2011jk], concentrating on the meson scattering and $\pi$ and $\eta$ electroproduction amplitudes, leads to rather small total width for both resonances, underestimating all $\pi N$ and $\eta N$ partial decay widths by roughly 2$\sigma$, and overestimating $\Gamma_{S_{11}(1650) \to K \Lambda}$ by more than 10$\sigma$. The authors conclude however that the $S_{11}(1535)$ resonance is dominated by a genuine three-quark state. Results of a K-matrix approach [@Ceci:2006jj] for $\pi N$ and $\eta N$ final states provide realistic values for all considered partial widths, except for $\Gamma_{S_{11}(1650) \to \eta N}$. Finally, in Ref. [@Aznauryan:2003zg], studying the $\eta N$ final states, dispersion relations lead to values in agreement with data, while the isobar model tends to overestimate $\Gamma_{S_{11}(1535) \to \eta N}$. The ambitious EBAC [@EBAC] program offers a powerful frame to study the properties of baryons, including partial decay widths [@Kamano:2008gr], extraction of which requires non ambiguous determination of the poles positions [@Suzuki:2008rp]; a topic under extensive investigations [@Suzuki:2008rp; @Capstick:2007tv; @Doring:2009bi; @Doring:2009yv; @Doring:2009uc; @Oset:2009vf; @Ceci:2011ae; @Osmanovic:2011xn]. ### Coupling constants $g_{S_{11}MB}$ {#2cc} In Table \[cc\], predictions for the relevant resonance-meson-baryon coupling constants, $g_{S_{11}MB}$, from models A and B are given in particle basis. In order to emphasize the most sensitive decay channels to the five-quark components in $S_{11}(1535)$, we compare results from models A and B. For $K^{+}\Sigma^{0}$ and $K^{0}\Sigma^{+}$, we observe variations by a factor of 2 between the two models, with central values differing from each other by more than 4$\sigma$. Next come $K^{+}\Lambda$ and $\eta p$, with about 30% differences and 2$\sigma$. The other three channels ($\pi^{0}p$, $\pi^{+}n$, $\eta^{\prime}p$) show no significant sensitivities to the five-quark components. In the case of $S_{11}(1650)$, similar sensitivities are observed. However, the rather small branching ratios to those final states, require substantial experimental efforts and sophisticated phenomenological approaches, e.g. for $\gamma p \to K^{0}\Sigma^{+},~ K^{+}\Sigma^{0}$. In Table \[cc\], results from a chiral unitary approach [@Inoue:2001ip] are also reported, showing compatible values with those of model B for $K^{+}\Sigma^{0}$, $K^{0}\Sigma^{+}$ and $\eta p$. For the other three channels the two sets differ by roughly 60%. [lccccccccc]{} $N^*$ && $\pi^{0}p$ & $\pi^{+}n$ & $\eta p$ & $K^{+}\Lambda$ & $K^{0}\Sigma^{+}$ & $K^{+}\Sigma^{0}$ & $\eta^{\prime}p$ & Ref.\ $S_{11}(1535)$ && -0.58 $\pm$ 0.13 & 0.82 $\pm$ 0.18 & -2.57 $\pm$ 0.17 & 1.42 $\pm$ 0.11 & 0.95 $\pm$ 0.20 &-0.62 $\pm$ 0.09 & 3.09 $\pm$ 0.20 & Model A\ && -0.63 $\pm$ 0.03 & 0.89 $\pm$ 0.04 & -2.07 $\pm$ 0.15 & 1.76 $\pm$ 0.02 & 1.81 $\pm$ 0.06 & -1.28 $\pm$ 0.04 & 3.33 $\pm$ 0.10 & Model B\ && $\pm0.39$ & $\pm0.56$ & $\pm1.84$& $\pm0.92$& $\pm2.12$ &$\pm1.50$ && [@Inoue:2001ip]\ \ $S_{11}(1650)$ && -0.70 $\pm$ 0.10 & 0.94 $\pm$ 0.19 & 0.84 $\pm$ 0.40 & 0.67 $\pm$ 0.25 & -1.42 $\pm$ 0.21 & 0.95 $\pm$ 0.10& -1.61 $\pm$ 0.79 &Model A\ && -0.94 $\pm$ 0.02 & 1.33 $\pm$ 0.03 & 0.35 $\pm$ 0.12 & 0.51 $\pm$ 0.03 & -2.17 $\pm$ 0.05 &1.53 $\pm$ 0.04 & -1.62 $\pm$ 0.14 &Model B\ \[cc\] [lcccccccccccccc]{} $N^*$ && $\pi N$ && $\eta N$ && $K \Lambda$ && $K \Sigma$ && $\eta^\prime N$ && Approach && Ref.\ $S_{11}(1535)$ && -1.09 $\pm$ 0.05 && -2.07 $\pm$ 0.15 && 1.76 $\pm$ 0.02 && 2.21 $\pm$ 0.07 && 3.3 $\pm$ 0.1 && Model B && Present work\ &&$\pm$(0.62$\pm$0.32) &&$\pm$(0.97$\pm$0.45) && $\pm$(0.55$\pm$0.32) && $\pm$(0.55$\pm$0.32)&& && PWA && Sarantsev [*et al.*]{} [@Sarantsev:2005tg]\ && $\pm$0.6 && $\pm$2.1 && $\pm$1.7 && $\pm$2.4 && && Chiral Lagrangian && Gamermann [*et al.*]{} [@Gamermann:2011mq]\ \ $S_{11}(1650)$ && -1.64 $\pm$ 0.03 && 0.35 $\pm$ 0.14 && 0.53 $\pm$ 0.04 && -2.66 $\pm$ 0.06 && -1.62 $\pm$ 0.14 && Model B && Present work\ &&$\pm$(1.05$\pm$0.45) &&$\pm$(0.63$\pm$0.32) && $\pm$(0.32$\pm$0.32) && $\pm$(0.71$\pm$0.39)&& && PWA && Sarantsev [*et al.*]{} [@Sarantsev:2005tg]\ && $\pm$1.2 && $\pm$0.8 && $\pm$0.6 && $\pm$1.7 && && Chiral Lagrangian && Gamermann [*et al.*]{} [@Gamermann:2011mq]\ \[cci\] In Table \[cci\], predictions in isospin basis are reported for model B and other sources. Additional results reported in the literature and limited to fewer channels are also discussed below. Within an isobar approach [@Anisovich:2005tf], a combined analysis [@Sarantsev:2005tg] of the pseudoscalar mesons photoproduction data available by 2005 has extracted coupling constants in isospin basis, with around $\pm$60% uncertainties. The reported couplings $g_{S_{11}(1535) \pi N}$ and $g_{S_{11}(1535) \eta N}$ are compatible with the model B predictions within 2$\sigma$, while discrepancies between the two approaches for $g_{S_{11}(1535) K \Lambda}$ and $g_{S_{11}(1535) K \Sigma}$ reach factors 3 to 4 and 4$\sigma$. For the second resonance, results from the two calculations agree within 1$\sigma$ for $g_{S_{11}(1650) \pi N}$, $g_{S_{11}(1650) \eta N}$ and $g_{S_{11}(1650) K \Lambda}$, with only significant disagreement observed for $g_{S_{11}(1650) K \Sigma}$. Copious data released since then, if interpreted within the same approach might bring in new insights into the coupling constants. Results from a recent SU(6) extended chiral Lagrangian [@Gamermann:2011mq], embodying eleven meson-baryon final states, are also reported in Table \[cci\] and show consistent values between that approach and model B for $g_{S_{11}(1535) \eta N}$, $g_{S_{11}(1535) K \Lambda}$, $g_{S_{11}(1535) K \Sigma}$, and $g_{S_{11}(1650) K \Lambda}$. An effective Lagrangian focused on interpreting [@Shyam:2007iz] $\eta$ production data in $NN$ and $\pi N$ collisions, leads to $g_{S_{11}(1535) \eta N}$ = 2.2 and $g_{S_{11}(1650) \eta N}$=0.55, compatible with our values. Another effective Lagrangian approach [@Cao:2008st] studying $\eta$ and $\eta^\prime$ production data in the same reactions gives $g_{S_{11}(1535) \eta^{\prime}p}$ = 3.7, about only 10% higher than the value given by model B. Here, we wish to make a few comments with respect to the relative values of some of the coupling constants. [*i)*]{} While the $\eta N N$ coupling constant is known to be smaller than that of $\pi N N$, the ratio $|g_{S_{11}(1535) \eta N}/g_{S_{11}(1535)\pi N}|$ comes out significantly larger than 1. This result is in line with the finding [@Jido:1997yk] that, in the soft pion limit, $\pi N N^*$ coupling vanishes due to chiral symmetry, while that of $\eta N N^*$ remains finite. [*ii)*]{} The ratio $|g_{S_{11}(1535)K\Lambda}/g_{S_{11}(1535)\eta N}|$ takes the value $1.3\pm 0.3$, within an isobar model [@Liu:2005pm] interpreting $J/\psi\to \bar{p}p\eta$ and $\psi \to \bar{p}K^{+}\Lambda$ data, larger than the results reported in Table \[cci\]. Dressed versus bare mass considerations [@Ceci:2009zz], might affect the reported ratio in Ref. [@Liu:2005pm]. Investigation of the same reaction within a unitary chiral approach [@Inoue:2001ip; @Geng:2008cv] puts that ratio around 0.5 to 0.7, smaller than our result. [*iii)*]{} The ratio $|g_{S_{11}(1650) K \Sigma}/g_{S_{11}(1650) K \Lambda}|$ turns out to be around 5. Actually, $S_{11}(1650)$ is dominant by the state $N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}}$, which cannot transit to $K\Lambda$ channel. Moreover, there is a cancellation between the contributions from $qqq \to K \Lambda$ and $qqqq \bar q \to K \Lambda$, which leads also to a very small decay width $\Gamma_{S_{11}(1650) \to K \Lambda}$. In addition, the threshold for $S_{11}(1650) \to K\Sigma$ decay channel being very close to the mass of $S_{11}(1650)$, contributions from the five-quark component enhance significantly the coupling constant $g_{S_{11}(1650)K\Sigma}$. [*iv)*]{} It is worthy to be noticed that he coupling constants $g_{S_{11}\eta N}$, $g_{S_{11}K\Sigma}$ and $g_{S_{11}\eta^{\prime}N}$ for $S_{11}(1535)$ and $S_{11}(1650)$ have opposite signs. Moreover, the ratio $|g_{S_{11}(1535)K\Sigma}/g_{S_{11}(1650)K\Sigma}|$ is close to unity. Those features might lead to significant cancellations in the interference terms in $KY$ photo- and/or hadron-induced productions. [*v)*]{} In Tables \[cc\] and  \[cci\], one finds the following orderings for magnitudes of the coupling constants, predicted by model B, and in Refs.[@Bruns:2010sv; @Gamermann:2011mq], noted below as [**a)**]{}, [**b)**]{} and [**c)**]{}, respectively: [**- For**]{} $\bullet$ [*In particle basis*]{} $$\begin{aligned} {\bf (a)}&:&~|g_{S_{11}\pi^{0}p}| < |g_{S_{11}\pi^{+}n}| < |g_{S_{11}K^{+}\Sigma^{0}}| < |g_{S_{11}K^{+} \Lambda} | \approx |g_{S_{11}K^{0}\Sigma^{+}}| < |g_{S_{11}\eta p}| < |g_{S_{11}\eta^{\prime}p}|, \\ % {\bf (b)}&:&~|g_{S_{11}\pi^{0}p}| \approx |g_{S_{11}K^{+}\Sigma^{0}}| < |g_{S_{11}\pi^{+}n}| \approx |g_{S_{11}K^{0}\Sigma^{+}}| < |g_{S_{11}\eta p}| < |g_{S_{11}K^{+} \Lambda} |.\end{aligned}$$ The main feature of our results [**(a)**]{} is that the strongest couplings are found the hidden strangeness sector, while those for open strangeness channels come out in between $\pi N$ and $\eta N$ final states. Inequalities in [**(b)**]{} come from a recent unitarized chiral effective Lagrangian [@Bruns:2010sv], in which both $S_{11}(1535)$ and $S_{11}(1650)$ are dynamically generated. Within that model, the coupling to $K^{+}\Sigma^{0}$ is highly suppressed, and that to $K^{+} \Lambda$ turns out larger than coupling to $\eta p$. $\bullet$ [*In isospin basis*]{} $$\begin{aligned} {\bf (a^\prime)}&:&~|g_{S_{11}\pi N}| < |g_{S_{11}K \Lambda} | < |g_{S_{11}\eta N}| \approx |g_{S_{11}K \Sigma}| < |g_{S_{11}\eta^{\prime}N}|,\\ % {\bf (c^\prime)}&:&~|g_{S_{11}\pi N}| < |g_{S_{11}K \Lambda} | < |g_{S_{11}\eta N}| \approx |g_{S_{11}K \Sigma} |.\end{aligned}$$ Results from a chiral Lagrangian study [@Gamermann:2011mq], [**(c’)**]{}, give the same ordering for couplings as model B. It is also the case for results from a chiral unitary approach [@Inoue:2001ip], while another chiral unitary approach [@Hyodo:2008xr], distinguishing dynamically generated resonances from genuine quark states, leads to $$\begin{aligned} |g_{S_{11}\pi N}| < |g_{S_{11}K \Lambda} | < |g_{S_{11}\eta N}| < |g_{S_{11}K \Sigma}|.\end{aligned}$$ [**- For**]{} $\bullet$ [*In particle basis*]{} $$\begin{aligned} {\bf (a)}&:&~|g_{S_{11}\eta p}| < |g_{S_{11}K^{+} \Lambda} | < |g_{S_{11}\pi^{0}p}| < |g_{S_{11}\pi^{+}n}| < |g_{S_{11}K^{+}\Sigma^{0}}| < |g_{S_{11}\eta^{\prime}p}| < |g_{S_{11}K^{0}\Sigma^{+}}|, \\ % {\bf (b)}&:&~|g_{S_{11}K^{+} \Lambda} | < |g_{S_{11}\pi^{0}p}| < |g_{S_{11}\pi^{+}n}| \approx |g_{S_{11}K^{+}\Sigma^{0}}| < |g_{S_{11}\eta p}| < |g_{S_{11}K^{0}\Sigma^{+}}|.\end{aligned}$$ In our model, the ordering in strangeness sector is separated by $\pi N$, according to the fact that the relevant disintegration channel is above or below the resonance mass. The main differences between results from model B and those in Ref. [@Bruns:2010sv] concern couplings to $K^{+} \Lambda$ and $\eta p$. $\bullet$ [*In isospin basis*]{} $$\begin{aligned} {\bf (a^\prime)}&:&~ |g_{S_{11}K \Lambda} | < |g_{S_{11}\eta N}| < |g_{S_{11}\pi N}|\approx |g_{S_{11}\eta^{\prime}N}| < |g_{S_{11}K \Sigma} |, \\ % {\bf (c^\prime)}&:&~ |g_{S_{11}}\eta N | \lesssim |g_{S_{11}K \Lambda}| < |g_{S_{11}\pi N}| < |g_{S_{11}K \Sigma} |.\end{aligned}$$ Here again model B and Ref. [@Gamermann:2011mq] lead basically to identical orderings. To end this section, we would like to emphasize the following point, with respect to the importance of five-quark components. Our model leads to probability for the strangeness component in $S_{11}(1650)$ being smaller than that for the five-quark component in $S_{11}(1535)$. Moreover, the probability amplitude turns out to be positive for $S_{11}(1535)$, but negative for $S_{11}(1650)$. Taking the ranges determined for probabilities (Eq. (\[Ranges\])), one gets $-77.4 \leq A_{5q}/A_{5q}^\prime \leq -72.5$. This latter range and that for $\theta_{S}$, embodied in Eq. (\[ampratio\]), allow extracting values for the energy of the strangeness component, $1641.60 \leq E_{5} \leq 1649.99$ MeV. The coupling between $qqq$ and $qqqq\bar{q}$ in the corresponding baryon $_{5q}\langle\hat{V}_{cou} \rangle_{3q}$, Eq. (\[ampratio\]), turns out to be negative for both $S_{11}$ resonances. ### Branching fraction versus branching ratio considerations {#brbf} As mentioned earlier, in PDG [@Nakamura:2010zzi] estimates for both branching fractions (BF) to meson-baryon states and branching ratios (BR), ($\Gamma_{MB}/\Gamma_{total}$), are reported. In the case of the $S_{11}$ resonances considered here, those estimates are not identical for $S_{11}(1650) \to \eta N,~ K \Lambda$. In the present work we have used BR. However, a very recent work [@Jayalath:2011uc] has adopted BF. In order to compare the results of this latter work with those of model B, we have investigated the drawback of using BF instead of BR in our approach. Accordingly, a third model, hereafter called model C, was obtained. Though we extract simultaneously the partial decay widths for both $S_{11}$ resonances, the above changes in the data do not affect results for the $S_{11}(1535)$. In Table \[BR-BF\], results from PDG, Ref. [@Jayalath:2011uc] and our models B and C are given for $S_{11}(1650)$. The $\chi^2_{d.p.}$ for the three models are comparable, namely, 0.15 (model B), 0.25 (model C) and 0.19 (ref. [@Jayalath:2011uc]). Model C leads to results in agreement with the two other sets, within the uncertainties therein. Comparing models B and C, we observe that the most sensitive width is $\Gamma_{S_{11}(1650) \to K \Lambda}$ and to a lesser extent $\Gamma_{S_{11}(1650) \to \eta N}$, while $\Gamma_{S_{11}(1650) \to \pi N}$ increases very slightly. In Table \[cci-Gam\], results for coupling constant from models B and C are reported. We find of cours the same features as for partial decay widths. In addition, given the associated uncertainties, it turns out that $\Gamma_{S_{11}(1650) \to \eta^\prime N}$ and $\Gamma_{S_{11}(1650) \to K \Sigma}$ change very slightly within the two models. $\Gamma_{tot}$ $\pi N$ $\eta N$ $K \Lambda$ Approach Ref. ---------------- -- -------------- -- ----------------- -- ---------------- -- ------------- -- ------------------------------------------ 165 $\pm$ 20 128 $\pm$ 29 3.8 $\pm$ 3.6 4.8 $\pm$ 0.7 [**BR**]{} PDG [@Nakamura:2010zzi] 143 $\pm$ 5 4.5 $\pm$ 3.0 4.8 $\pm$ 0.7 Model B Present work 128 $\pm$ 29 10.7 $\pm$ 5.8 11.5 $\pm$ 6.6 [**BF**]{} PDG [@Nakamura:2010zzi] 148 $\pm$ 8 9.7 $\pm$ 6.7 7.9 $\pm$ 0.3 Model C Present work 133 $\pm$ 33 12.5 $\pm$ 11.0 11.5 $\pm$ 6.4 $1/N_C$-NLO Jayalath [*et al.*]{} [@Jayalath:2011uc] : Strong decay widths (in MeV) for $S_{11}(1650)$. \[BR-BF\] $\pi N$ $\eta N$ $K \Lambda$ $K \Sigma$ $\eta^\prime N$ Approach Ref. ------------------ -- ----------------- -- ----------------- -- ------------------ -- ------------------ -- ---------- -- -------------- -- -1.64 $\pm$ 0.03 0.35 $\pm$ 0.14 0.53 $\pm$ 0.04 -2.66 $\pm$ 0.06 -1.62 $\pm$ 0.14 Model B Present work -1.66 $\pm$ 0.05 0.55 $\pm$ 0.16 0.62 $\pm$ 0.09 -2.49 $\pm$ 0.16 -1.74 $\pm$ 0.24 Model C Present work : $S_{11}(1650)$-meson-baryon coupling constants ($g_{S_{11} MB}$) in isospin basis. \[cci-Gam\] $\pi^{0}p$ $\pi^{+}n$ $\eta p$ $K^{+}\Lambda$ $K^{0}\Sigma^{+}$ $K^{+}\Sigma^{0}$ $\eta^{\prime}p$ Ref. ------------------ ----------------- ----------------- ----------------- ------------------- ------------------- ------------------ --------- -- -0.94 $\pm$ 0.02 1.33 $\pm$ 0.03 0.35 $\pm$ 0.14 0.51 $\pm$ 0.03 -2.17 $\pm$ 0.05 1.53 $\pm$ 0.04 -1.62 $\pm$ 0.14 Model B -0.96 $\pm$ 0.03 1.36 $\pm$ 0.04 0.55 $\pm$ 0.16 0.62 $\pm$ 0.09 -2.03 $\pm$ 0.13 1.44 $\pm$ 0.09 -1.74 $\pm$ 0.24 Model C : $S_{11}(1650)$-meson-baryon coupling constants ($g_{S_{11} MB}$) in particle basis. \[cc-c\] Those trends are also present in the coupling constants given in particle basis (Table \[cc-c\]). Taking into account the associated uncertainties to the coupling constants, model C does not significantly modify the coupling constants ordering obtained in sec. \[2cc\] for model B. To end this section, we give the phase space defined by model C: $$%26.8^\circ \leq \theta_{S} \leq 29.8^\circ~;~21\% \leq P_{5q} \leq 30\%~;~11\% \leq P^\prime_{5q} \leq 18\%\,. 24.7^\circ \leq \theta_{S} \leq 30.0^\circ~;~19.8\% \leq P_{5q} \leq 31\%~;~3.0\% \leq P^\prime_{5q} \leq 12.6\%\,. % \label{Ranges-c}$$ Compared to model B, Eq. (\[Ranges\]), the ranges for $\theta_{S}$ and $P_{5q}$ get slightly increased. The most significant change concerns $P^\prime_{5q}$, which goes from $11\% \leq P^\prime_{5q} \leq 18$ down to $3\% \leq P^\prime_{5q} \leq 13$. This feature shows the sensitivity of $\Gamma_{S_{11}(1650) \to K \Lambda}$ and, to a lesser extent, that of $\Gamma_{S_{11}(1650) \to \eta N}$ to the five-quark components in $S_{11}(1650)$. Summary and Conclusions {#con} ======================= Within a constituent quark approach, we studied the properties of four low-lying baryon resonances with respect to their partial decay widths to seven meson-baryon channels and associated resonance-meson-baryon coupling constants. The starting point was the simplest chiral constituent quark model ($\chi$CQM). The second step consisted in introducing $SU(6) \otimes O(3)$ breaking effects. Finally, five-quark components in the $S_{11}$ resonances were implemented and investigated. The outcome of the present work is reported below, focusing on the considered nucleon resonances ($S_{11}(1535)$, $S_{11}(1650)$, $D_{13}(1520)$ and $D_{13}(1700)$) and their strong decays to $\pi N$, $\eta N$, $\eta^\prime N$, $K \Lambda$ and $K \Sigma$ final states. Within the $\chi$CQM, the only adjustable parameter ($\omega_3$) did not allow reproducing the partial widths of resonances. Introducing $SU(6) \otimes O(3)$ breaking, via configuration mixing angles $\theta_S$ and $\theta_D$, brought in significant improvements with respect to the decay widths of the $D_{13}$ resonances, but missed the data for the $S_{11}$ resonances partial decay widths. Nevertheless, this second step allowed fixing the value of $\omega_3$ and extracting ranges for the mixing angles, treated as free parameters. Trying to cure this unsatisfactory situation, possible roles due to five-quark component in the baryons’ wave functions were investigated. Given that the latter issue is irrelevant with respect to the $D_{13}$ resonances and the properties of which were well descried in the second step, the final phase of our study was devoted to the $S_{11}$ resonances. We calculated the partial decay widths $S_{11}(1535)\to\pi N$, $\eta N$ and $S_{11}(1650)\to \pi $, $\eta N$, $K \Lambda$ in the whole phase space defined by the mixing angle $\theta_S$ and the probability of five-quark components in each of the two resonances. Regions of the phase space allowing to reproduce the data for those widths were selected. Accordingly, that procedure allowed us extracting ranges for partial widths, with decay threshold below the relevant resonance mass, and resonance-meson-baryon coupling constants for the following meson-baryon combinations: $\pi^{0}p$, $\pi^{+}n$, $\eta p$, $K^{+}\Lambda$, $K^{0}\Sigma^{+}$, $K^{+}\Sigma^{0}$ and $\eta^{\prime}p$. The main findings of the present work are summarized below with respect to the approaches studied in describing the properties of the four low-lying nucleon resonances. - The chiral constituent quark approach in three-quark configuration and exact $SU(6)\otimes O(3)$ symmetry is not appropriate to reproduce the known partial decay widths. - Introducing symmetry breaking effects due to one-gluon-exchange mechanism, allows accounting for the partial decay width of the $D_{13}(1520)$ and $D_{13}(1700)$ resonances, but not for those of $S_{11}$ resonances. - Complementing the formalism with five-quark components in the $S_{11}$ resonances leads to satisfactory results with respect to all known partial decay widths investigated here. - The complete formalism puts ranges on the three adjustable parameters, namely, the mixing angle between configurations $|N^{2}_{8}P_{M}\rangle$ and $|N^{4}_{8}P_{M}\rangle$, and five-quark component probabilities in $S_{11}(1535)$ and $S_{11}(1650)$ resonances. - For $S_{11}(1535)$, the most sensitive entities to the five-quark component turn out to be $\Gamma_{S_{11}(1535) \to \eta N}$, $g_{S_{11}K^{+}\Sigma^{0}}$, $g_{S_{11}K^{0}\Sigma^{+}}$ and $g_{S_{11}\eta p}$, all with sizeable magnitudes. - For $S_{11}(1650)$, the same trends as for $S_{11}(1535)$ are observed. In addition $\Gamma_{S_{11}(1650) \to \pi N}$ undergoes significant change due to five-quark mixture. Here, $\eta N$ channel have smaller width and coupling constant compared to the $S_{11}(1535)$ case. To go further, interpretation of recent data, obtained using electromagnetic and/or hadronic probes, within approaches with reasonable number of free parameters is very desirable. Within the present extended $\chi$CQM approach, analysis of the $\gamma p \to \eta p$ data is underway [@next]. One of us (C. S. A.) thanks X. H. Liu and J. J. Xie for very helpful discussions. [$S_{11}(1535)$]{} and [$S_{11}(1650)$]{} resonances mixing angle in one-gluon-exchange and one-boson-exchange models {#apdx:mix} ===================================================================================================================== The mixing angle $\theta_{S}$ can be obtained by diagonalizing the following matrix: $$\begin{aligned} \pmatrix{ \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle, & \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}|N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle \cr \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle, & \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|h_{hyp}|N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle\cr}\,,\end{aligned}$$ where $H_{hyp}$ is the hyperfine interaction between the quarks. In the OGE [@De; @Rujula:1975ge] and OBE models [@Glozman:1995fu], the explicit forms of $H_{hyp}$ are $$\begin{aligned} H_{hyp}^{OGE}&=&\sum_{i<j}\frac{2\alpha_{s}}{3m_{i}m_{j}}\{ \frac{8\pi}{3}\vec{S}_{i}\cdot\vec{S}_{j}\delta^{3}(\vec{r}_{ij}) +\frac{1}{r_{ij}^{3}}[\frac{3\vec{S}_{i}\cdot\vec{r}_{ij}\vec{S}_{j}\cdot\vec{r}_{ij}}{r_{ij}^{2}}- \vec{S}_{i}\cdot\vec{S}_{j}]\}\\ H_{hyp}^{OBE}&=&\sum_{i<j}\sum_{F}\frac{g^{2}}{4\pi}\frac{1}{12m_{i}m_{j}} \vec{\lambda}_{i}^{F}\cdot\vec{\lambda}_{j}^{F}\{[\vec{\sigma}_{i}\cdot\vec{\sigma}_{j}(\frac{\mu^{2}e^{-\mu r_{ij}}}{r_{ij}}-4\pi\delta(\vec{r}_{ij}))]\nonumber\\ &&+ (\frac{3\vec{\sigma}_{i}\cdot\vec{r}_{ij}\vec{\sigma}_{j}\cdot\vec{r}_{ij}}{r_{ij}^{2}}-\vec{\sigma}_{i}\cdot\vec{\sigma}_{j}) \frac{\mu^{2}e^{-\mu r_{ij}}}{r_{ij}}(1+\frac{3}{\mu r_{ij}}+\frac{3}{\mu^{2}r_{ij}^{2}})\}\end{aligned}$$ One-Gluon-Exchange (OGE) model ------------------------------ The OGE hyperfine interaction leads to the following matrix elements: $$\begin{aligned} \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OGE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&-C,\\ \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OGE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&C,\\ \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OGE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&C,\\ \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OGE}|N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&0\,,\end{aligned}$$ with the constant $C=\frac{2\alpha_{s}}{m^{2}}\omega_{3}^{3}\pi^{-\frac{1}{2}}$, where $m$ and $\omega_{3}$ are the light quark mass and the harmonic oscillator parameter, respectively. Then, we obtain $\theta_{S}^{OGE}\simeq32$. Here a comment is in order with respect to the sign of $\theta_{S}$. As, reported in Ref. [@Saghai:2009zz], a non ambiguous entity with respect to that sign is the following ratio: $$\begin{aligned} \label{eq:MixR} {\cal {R}} = \frac {<N|H_m|N(^4P_M)_{{\frac 12}^-}>} {<N|H_m|N(^2P_M)_{{\frac 12}^-}>},\end{aligned}$$ with $H_m$ the pseudovector couplings at the tree level. The ratio ${\cal {R}}$ is a constant determined by $SU(6)\otimes O(3)$ symmetry. Notice that in the present work, we have adopted the convention introduced by Koniuk and Isgur [@Koniuk:1979vy], where wave functions are in line with the SU(3) conventions of de Swart [@de; @Swart:1963gc]. In this frame, the constant ${\cal {R}}$ gets a negative value, and the relevant mixing angle for the $S-$wave, $\theta_{S}$, turns out positive. However, in line with the Hey, Litchfield, and Cashmore [@Hey:1974nc] analysis, Isgur and Karl in their early works [@Isgur:1977ef; @Isgur:1978xi; @Isgur:1978xj; @Isgur:1978wd] used another convention, for which ${\cal {R}}$ = +1 and $\theta_{S} <$ 0. In the literature both conventions are being used, often without explicit mention of the utilized convention. One-Boson-Exchange (OBE) model ------------------------------ The OBE hyperfine interaction results in $$\begin{aligned} \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OBE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&5V_{11}-7V_{00},\\ \langle N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OBE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&-8T_{11},\\ \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OBE}|N(^{2}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&-8T_{11},\\ \langle N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}|H_{hyp}^{OBE}|N(^{4}_{8}P_{M})_{\frac{1}{2}^{-}},S_{z}\rangle&=&4V_{11}-2V_{00}+8T_{11}\,,\end{aligned}$$ where $V_{00}$, $V_{11}$ and $T_{11}$ are constants from the orbital integral $$\begin{aligned} V_{00}&=&\langle\varphi_{00}|\frac{g^{2}}{4\pi}\frac{1}{12m_{i}m_{j}} (\frac{\mu^{2}e^{-\mu r_{ij}}}{r_{ij}}-4\pi\delta(\vec{r}_{ij}))|\varphi_{00}\rangle,\\ V_{11}&=&\langle\varphi_{1m}|\frac{g^{2}}{4\pi}\frac{1}{12m_{i}m_{j}} (\frac{\mu^{2}e^{-\mu r_{ij}}}{r_{ij}}-4\pi\delta(\vec{r}_{ij}))|\varphi_{1m}\rangle,\\ T_{11}&=&\langle\varphi_{1m}|\frac{g^{2}}{4\pi}\frac{1}{12m_{i}m_{j}} \frac{\mu^{2}e^{-\mu r_{ij}}}{r_{ij}}(1+\frac{3}{\mu r_{ij}}+\frac{3}{\mu^{2}r_{ij}^{2}})|\varphi_{1m}\rangle.\end{aligned}$$ Taking the same values for the parameters as in Ref. [@Glozman:1995fu], we obtain $\theta_{S}=-13$. However, if one considers contributions from the vector meson exchanges, the absolute value of $\theta_{S}$ might be decreased, or even the sign might change [@Glozman:1999ms; @Glozman:1998wk]. Relevance of the OGE versus the OBE has been studied by several authors, see e.g. Refs. [@Capstick:2004tb; @Yoshimoto:1999dr; @He:2003vi; @Liu:2005wg], favoring OGE mechanism, endorsed by the present work, as the origin of the $SU(6)\otimes O(3)$ symmetry breakdown. [99]{} K. Nakamura [*et al.*]{} \[Particle Data Group\], J. Phys. G [**37**]{}, 075021 (2010). N. Kaiser, P. B. Siegel and W. Weise, Phys. Lett.  B [**362**]{}, 23 (1995). D. O. Riska and G. E. Brown, Nucl. Phys.  A [**679**]{}, 577 (2001). G. Penner and U. Mosel, Phys. Rev.  C [**66**]{}, 055211 (2002). G. Penner and U. Mosel, Phys. Rev.  C [**66**]{}, 055212 (2002). T. Inoue, E. Oset and M. J. Vicente Vacas, Phys. Rev.  C [**65**]{}, 035204 (2002). I. G. Aznauryan, Phys. Rev.  C [**68**]{}, 065204 (2003). B. C. Liu and B. S. Zou, Phys. Rev. Lett.  [**96**]{}, 042002 (2006). W. T. H. Chiang, F. Tabakin, T. S. H. Lee and B. Saghai, Phys. Lett.  B [**517**]{}, 101 (2001). W. T. Chiang, B. Saghai, F. Tabakin and T. S. H. Lee, Phys. Rev.  C [**69**]{}, 065208 (2004). K. Nakayama and H. Haberzettl, Phys. Rev.  C [**69**]{}, 065212 (2004). K. Nakayama and H. Haberzettl, Phys. Rev.  [**C73**]{}, 045211 (2006). K. Nakayama, Y. Oh and H. Haberzettl, arXiv:0803.3169 \[hep-ph\]. T. P. Vrana, S. A. Dytman and T. S. H. Lee, Phys. Rept.  [**328**]{}, 181 (2000). A. Matsuyama, T. Sato and T. S. Lee, Phys. Rept.  [**439**]{}, 193 (2007). B. Julia-Diaz, T. S. Lee, A. Matsuyama and T. Sato, Phys. Rev.  C [**76**]{}, 065201 (2007). J. Durand, B. Julia-Diaz, T. S. Lee, B. Saghai and T. Sato, Phys. Rev.  C [**78**]{}, 025204 (2008). X. Cao and X. G. Lee, Phys. Rev.  C [**78**]{}, 035207 (2008). R. Shyam, H. Lenske and U. Mosel, Phys. Rev.  C [**69**]{}, 065205 (2004). R. Shyam, Phys. Rev.  C [**75**]{}, 055201 (2007). V. Shklyar, H. Lenske and U. Mosel, Phys. Lett.  B [**650**]{}, 172 (2007). R. A. Arndt, W. J. Briscoe, T. W. Morrison, I. I. Strakovsky, R. L. Workman and A. B. Gridnev, Phys. Rev.  C [**72**]{}, 045202 (2005). A. V. Anisovich, A. Sarantsev, O. Bartholomy, E. Klempt, V. A. Nikonov and U. Thoma, Eur. Phys. J.  A [**25**]{}, 427 (2005). A. V. Sarantsev, V. A. Nikonov, A. V. Anisovich, E. Klempt and U. Thoma, Eur. Phys. J.  A [**25**]{}, 441 (2005). L. P. Kaptari and B. Kampfer, Eur. Phys. J.  [**A37**]{}, 69-80 (2008). S. Ceci, A. Svarc, B. Zauner, M. Manley and S. Capstick, Phys. Lett.  B [**659**]{}, 228 (2008). L. S. Geng, E. Oset, B. S. Zou and M. Doring, Phys. Rev.  C [**79**]{}, 025203 (2009). J. J. Xie and C. Wilkin, Phys. Rev.  C [**82**]{}, 025210 (2010). P. C. Bruns, M. Mai and U. G. Meissner, Phys. Lett.  B [**697**]{}, 254 (2011). D. Gamermann, C. Garcia-Recio, J. Nieves and L. L. Salcedo, arXiv:1104.2737 \[hep-ph\]. R. Koniuk and N. Isgur, Phys. Rev.  D [**21**]{}, 1868 (1980) \[Erratum-ibid.  D [**23**]{}, 818 (1981)\]. S. Capstick and W. Roberts, Phys. Rev.  D [**47**]{}, 1994 (1993). S. Capstick and W. Roberts, Phys. Rev.  D [**49**]{}, 4570 (1994). S. Capstick and W. Roberts, Phys. Rev.  D [**58**]{}, 074011 (1998). S. Capstick and W. Roberts, Prog. Part. Nucl. Phys.  [**45**]{}, S241 (2000). S. Capstick and W. Roberts, Fizika B [**13**]{}, 271 (2004). A. Kiswandhi, S. Capstick and S. Dytman, Phys. Rev.  C [**69**]{}, 025205 (2004). B. Golli and S. Sirca, Eur. Phys. J.  A [**47**]{}, 61 (2011). H. c. Kim and S. H. Lee, Phys. Rev.  D [**56**]{}, 4278 (1997). T. Yoshimoto, T. Sato, M. Arima and T. S. H. Lee, Phys. Rev.  C [**61**]{}, 065203 (2000). S. L. Zhu, Mod. Phys. Lett.  A [**13**]{}, 2763 (1998). S. L. Zhu, W. Y. P. Hwang and Y. B. P. Dai, Phys. Rev.  C [**59**]{}, 442 (1999). D. Jido, M. Oka and A. Hosaka, Phys. Rev. Lett.  [**80**]{}, 448 (1998). T. Hyodo, D. Jido and A. Hosaka, Phys. Rev.  C [**78**]{}, 025203 (2008). J. L. Goity and N. N. Scoccola, Phys. Rev.  D [**72**]{}, 034024 (2005). C. Jayalath, J. L. Goity, E. G. de Urreta and N. N. Scoccola, arXiv:1108.2042 \[nucl-th\]. Q. B. Li and D. O. Riska, Phys. Rev.  C [**73**]{}, 035201 (2006). B. Julia-Diaz and D. O. Riska, Nucl. Phys.  A [**780**]{}, 175 (2006). C. S. An, D. O. Riska and B. S. Zou Phys. Rev.  C [**73**]{}, 035207 (2006). Q. B. Li and D. O. Riska, Phys. Rev.  C [**74**]{}, 015202 (2006). Q. B. Li and D. O. Riska, Nucl. Phys.  A [**766**]{}, 172 (2006). C. S. An and B. S. Zou, Eur. Phys. J.  A [**39**]{}, 195 (2009). C. S. An and D. O. Riska, Eur. Phys. J.  A [**37**]{}, 263 (2008). C. S. An, B. Saghai, S. G. Yuan and J. He, Phys. Rev.  C [**81**]{}, 045203 (2010). B. S. Zou, Nucl. Phys.  A [**827**]{}, 333C (2009). B. S. Zou, Nucl. Phys.  A [**835**]{}, 199 (2010). D. O. Riska, Chin. Phys. C [**34**]{}, 9 (2010). A. De Rujula, H. Georgi and S. L. Glashow, Phys. Rev.  D [**12**]{}, 147 (1975). L. Y. Glozman and D. O. Riska, Phys. Rept.  [**268**]{}, 263 (1996). Q. B. Li and D. O. Riska, Nucl. Phys.  A [**791**]{}, 406 (2007). J. L. Goity and W. Roberts, Phys. Rev.  D [**60**]{}, 034001 (1999). T. A. Lahde and D. O. Riska, Nucl. Phys.  A [**710**]{}, 99 (2002). C. Gobbi, F. Iachello and D. Kusnezov, Phys. Rev.  D [**50**]{}, 2048 (1994). W. Rarita and J. Schwinger, Phys. Rev.  [**60**]{}, 61 (1941). L. M. Nath, B. Etemadi and J. D. Kimel, Phys. Rev.  D [**3**]{}, 2153 (1971). T.E.O. Ericson and W. Weise, [*Pions and Nuclei*]{} (Clarendon, Oxford, 1988). B. Saghai and Z. Li, Few Body Syst.  [**47**]{}, 105 (2010). N. Isgur and G. Karl, Phys. Lett.  B [**72**]{}, 109 (1977). N. Isgur, G. Karl and R. Koniuk, Phys. Rev. Lett.  [**41**]{}, 1269 (1978) \[Erratum-ibid.  [**45**]{}, 1738 (1980)\]. N. Isgur and G. Karl, Phys. Lett.  B [**74**]{}, 353 (1978). N. Isgur and G. Karl, Phys. Rev.  D [**18**]{}, 4187 (1978). N. Isgur and G. Karl, Phys. Rev.  D [**19**]{}, 2653 (1979) \[Erratum-ibid.  D [**23**]{}, 817 (1981)\]. N. Isgur, Phys. Rev.  D [**62**]{}, 054026 (2000). L. Y. Glozman, arXiv:nucl-th/9909021. J. Chizma and G. Karl, Phys. Rev.  D [**68**]{}, 054007 (2003). L. Y. Glozman, Surveys High Energ. Phys.  [**14**]{}, 109 (1999). J. J. de Swart, Rev. Mod. Phys.  [**35**]{}, 916 (1963) \[Erratum-ibid.  [**37**]{}, 326 (1965)\]. C. An and B. Saghai, arXiv:1107.5991 \[nucl-th\]. Excited Baryon Analysis Center: http://ebac-theory.jlab.org/. H. Kamano, B. Julia-Diaz, T. -S. H. Lee, A. Matsuyama and T. Sato, Phys. Rev.  [**C79**]{}, 025206 (2009). N. Suzuki, T. Sato and T. -S. H. Lee, Phys. Rev.  [**C79**]{}, 025205 (2009). S. Capstick [*et al.*]{}, Eur. Phys. J.  A [**35**]{}, 253 (2008). M. Doring, C. Hanhart, F. Huang, S. Krewald and U. G. Meissner, Phys. Lett.  B [**681**]{}, 26 (2009). M. Doring, C. Hanhart, F. Huang, S. Krewald and U. G. Meissner, Nucl. Phys.  A [**829**]{}, 170 (2009). M. Doring and K. Nakayama, Eur. Phys. J.  A [**43**]{}, 83 (2010). E. Oset and A. Ramos, Eur. Phys. J.  A [**44**]{}, 445 (2010). S. Ceci, M. Doring, C. Hanhart, S. Krewald, U. G. Meissner and A. Svarc, Phys. Rev.  C [**84**]{}, 015205 (2011). H. Osmanovic, S. Ceci, A. Svarc, M. Hadzimehmedovic and J. Stahov, arXiv:1103.2855 \[hep-ph\]. S. Ceci, A. Svarc and B. Zauner, Phys. Rev. Lett.  [**102**]{}, 209101 (2009). C. S. An, J. He and B. Saghai, [*in progress.*]{} A. J. G. Hey, P. J. Litchfield and R. J. Cashmore, Nucl. Phys.  B [**95**]{}, 516 (1975). J. He and Y. -B. Dong, Nucl. Phys.  [**A725**]{}, 201-210 (2003). J. Liu, J. He and Y. B. Dong, Phys. Rev.  [**D71**]{}, 094004 (2005).

ArXiv

--- abstract: 'In large-scale computation of physics problems, one often encounters the problem of determining a multi-dimensional function, which can be time-consuming when computing each point in this multi-dimensional space is already time-demanding. In the work, we propose that the active learning algorithm can speed up such calculations. The basic idea is to fit a multi-dimensional function by neural networks, and the key point is to make the query of labeled data economically by using a stratagem called “query by committee". We present the general protocol of this fitting scheme, as well as the procedure of how to further compute physical observables with the fitted functions. We show that this method can work well with two examples, which are quantum three-body problem in atomic physics and the anomalous Hall conductivity in condensed matter physics, respectively. In these examples, we show that one reaches an accuracy of few percent error for computing physical observables with less than $10\%$ of total data points compared with uniform sampling. With these two examples, we also visualize that by using the active learning algorithm, the required data are added mostly in the regime where the function varies most rapidly, which explains the mechanism for the efficiency of the algorithm. We expect broad applications of our method on various kind of computational physics problems.' author: - Juan Yao - Yadong Wu - Jahyun Koo - Binghai Yan - Hui Zhai title: Active Learning Algorithm for Computational Physics --- Background ========== Neural network (NN) based supervised learning methods has nowadays found broad applications in studying quantum physics of condensed matter materials and atomic, molecular and optical systems [@review1; @review2]. On the theoretical side, applications include, for example, finding orders and topological invariants in quantum phases [@Phase1; @Phase2; @Phase3; @Phase4; @Phase5; @Phase6; @Phase7], generating variational wave functions for quantum many-body states [@VariationalWF; @WFRBM; @WF; @WFFewBody; @TheoryTomography] and speeding up quantum Monte Carlo sampling [@MC1; @MC2]. On the experimental side, these methods can help both optimizing experimental protocols [@State; @Tomography] and analyzing experimental data [@STM; @ExpPhase; @ExpFHM]. Usually the supervised learning scheme requires a huge set of labelled data. However, in many physics applications, labelling data can be quite expensive, for instance, performing computation or experiments repeatedly can be time- and resources-demanding. Therefore, in many cases labelled data are not abundant, which is a challenge that have prevented many applications. The active learning is a scheme to solve this problem [@ALbook]. It starts from training a NN with a small initial dataset, and then actively queries the labelled data based on the prediction of the NN and iteratively improves the performance of the NN until the goal of the task is reached. With this approach, sampling the large parameter space can be made more efficiently, and the request of labelled data is usually much more economical than normal supervised learning methods. Recently a few works have applied the active learning algorithm to determine the inter-atomic potentials in quantum materials [@PotentialAL1; @PotentialAL2; @PotentialAL3] and to optimize control in quantum experiments [@ExpControl1; @ExpControl2], where labelled data have to be obtained either by *ab initio* calculation or by repeating experiments, which are both time consuming. In this work we focus on a class of general and common task in computational physics that is to numerically determining a multi-dimensional function, say, $\mathcal{F}(\alpha_1,\alpha_2,\dots,\alpha_n)$, where $\alpha_i$ are parameters. Considering a uniform discretization, suppose we discretize each parameter into $L$ points, there are totally $L^{n}$ number of data points that need to be calculated. In many cases, calculation of each point already takes quite some time, and thus the total computational cost is massive. Nevertheless, for most functions, there are regimes where the function varies smoothly and regimes where the function varies rapidly. Ideally, one should sample more points in the steep regimes and less points in the smooth regimes in order to obtain a good fitting in the entire parameter space efficiently. However, it seems to be paradox because one does not know which regimes the function varies more rapidly prior to computing the function. Here we show that this goal can actually be achieved by using the active learning algorithm and the “query by committee" stratagem [@Query_by_committee] to add data points iteratively. Below we will first introduce the general protocol and then demonstrate the algorithm in two concrete problems. One is the quantum three-body problem and the other is the anomalous Hall conductivity problem. These two are very representative examples in atomic and condensed matter physics, respectively. Using these two examples, we will illustrate how the active learning and “query by committee" guide adding data to steep regimes of the fitting function. We will also discuss how to compute physical observables from the fitted functions. ![Active learning protocol for fitting a multiply dimensional function. Here “NN" represents “neural network". []{data-label="protocal"}](ActiveLearning.pdf){width="45.00000%"} General Protocol. ================= The main procedure is summarized in Fig. \[protocal\] and explained as follows: 1\) We start with an initial dataset with the number of data $\mathcal{S}_0\ll L^n$, whose values of $\mathcal{F}$ have been computed exactly. We use this dataset to train a number of different NNs. 2\) We ask all NNs to make predictions of $\mathcal{F}$ on all $L^n$ number of data points, and for each point we compute the variance among predications made by different NNs. 3\) We select $\mathcal{S}_t$ number of data point with the largest variance, again with $\mathcal{S}_t\ll L^n$, and we query the accurate value of $\mathcal{F}$ of these points by numerical calculation. 4\) We add the $\mathcal{S}_t$ number of new data into the training set to train all NNs again, and then repeat from step 2). We repeat the procedure for $m$-epochs until the results from all NNs converge. 5\) We use NN to calculate the value of $\mathcal{F}$ on all $L^n$ data points. In this protocol, one only needs to query the value of $\mathcal{F}$ on $\mathcal{S}_0+m\mathcal{S}_t$ number of data points, and we should keep $\mathcal{S}_0+m\mathcal{S}_t \ll L^n$. The trade-off is that we need to train a number of different NNs and keep using all NN to make predications on all data points. It can save computational cost if the computational cost for training NN and making predication with NN is much less than the computational cost of $\mathcal{F}$, which is often the case in many applications. In this protocol, the key idea is “query by committee" [@Query_by_committee]. Here the committee is made of different NNs, which contain different number of layers, different number of nodes at each layer and different activation functions. Thanks to the great expressive power of NN, we do not need to assume a specific form of the fitting functions. In the regime where the function varies smoothly, different NNs can quickly reach a consensus and the variance will be small. On the other hand, in the regime where the function varies rapidly, it is hard for different NNs to converge and the variance will be large. Hence, we can use this variance to guide sampling data point. In fact, as we will see in the examples below, the data points added in later epoch are all added in the regime where the function changes rapidly. ![Schematic of the fully connected neural network used in these two examples.[]{data-label="network"}](network.pdf){width="35.00000%"} Three-body Problem ================== In the first example, we consider a quantum three-boson problem. These bosons interact with a two-body pairwise potential described by an $s$-wave scattering length $a_\text{s}$ and a high-energy cut-off $\Lambda$. When $a_\text{s}$ is positive, there exists a two-body bound dimer state, and it is important to compute the atom-dimer scattering length $a_\text{ad}$ by solving the three-body problem. Here a key quantity is the scattering kernel $\mathcal{U}({\bf k}, {\bf k^\prime})$. Focusing on the $s$-wave scattering only, $\mathcal{U}$ only depends on the amplitude of momentum $k=|{\bf k}|$ and $k^\prime=|{\bf k}^\prime|$, and we can simplify $\mathcal{U}$ as $\mathcal{U}(k,k^\prime)$. Once we know $\mathcal{U}(k,k^\prime)$, one can calculate $a_\text{ad}$ through the Skorniakov-Ter-Martirosion (STM) equation [@Braaten]. The details of how to compute $\mathcal{U}(k,k^\prime)$ and how to obtain $a_\text{ad}$ from $\mathcal{U}(k,k^\prime)$ are summarized in Appendix A. In the conventional method, we uniformly discretize both $k$ and $k^\prime$ into $L=100$ points between zero and $\Lambda$, and we need to compute $\mathcal{U}$ for all $10^4$ data points. We then solve the STM equation with these $U(k,k^\prime)$ to obtain $a^{\rm exact}_\text{ad}$, and we have checked that such a discretization can ensure reaching the convergence, which is referred as the exact result and is shown by the dash line as reference in the Fig. \[performance\]. Here we will follow the general procedure described above to fit $\mathcal{U}(k,k^\prime)$ using NNs. We will show that i) at most only $300$ number of data points are needed in order to obtain a reliable fitting, and ii) most of the queried data occupy the parameter regime where the function $\mathcal{U}(k,k^\prime)$ is steep, and iii) we use the trained NN to generate $\mathcal{U}(k,k^\prime)$ on all $10^4$ data points to solve STM equation, and we find that the error is only few percent compared with the exact result. ![The atom-dimer scattering length $a_\text{ad}$ calculated with the active learning method. $a_\text{ad}$ converges with the increasing number of queried dataset $\mathcal{S}$. The blue empty circles are the results averaged over different NNs at each step, and the yellow solid dots are results averaged over the adjacent five steps. The dashed line denotes the exact results obtained with all $10^4$ data points. Here we take the number of initial dataset $\mathcal{S}_0=100$ and at each step $\mathcal{S}_t=10$ data points are added. $a_\text{s}\Lambda=10$. []{data-label="performance"}](FigAad.pdf){width="45.00000%"} To be more concrete, we start with an initial dataset with uniformly sampled $S_0=100$ points. Here we design five different fully connected NNs, whose structures are schematically shown in Fig. \[network\]. The input of all NN are two numbers $k$ and $k^\prime$, and the output is $\mathcal{U}$. Each layer of a NN is characterized by the number of nodes $N_\alpha$ and an activation function $f_\alpha$, and we describe each NN with by $(N_1, f_1; N_2,f_2; \dots)$. The five different NN used in this work are $$\begin{aligned} 1:& ~(20, \tanh;20,\tanh;6,{\rm LS}) \\ 2:& ~(20, \tanh;20,{\rm LS};6,\tanh) \\ 3:& ~(30, \tanh;20,\tanh;6,{\rm LS}) \\ 4:& ~(30, \tanh;20,{\rm LS};4,\tanh) \\ 5:& ~(30, \tanh;20,\tanh;10,{\rm LS};4,\tanh), \label{EqFNNs} \end{aligned}$$ where ${\rm LS}$ denotes the logistic sigmoid activation function with $f(a)=1/(1+e^{-a})$. It is important to keep these NNs different but their detailed structures are not important for final results. For each NN, the training results also depend on the initialization, which is particularly so when the number of data points are not enough. Therefore, for each NN, we also consider $20$ different initialization. ![$a_\text{ad}$ as a function of $a_\text{s}\Lambda$. The divergence of $a_\text{ad}$ is the atom-dimer resonance due to the Efimov effect. The black dots are calculated by evaluating $\mathcal{U}$ with uniform sampling all data points, and the yellow diamonds are results from the active learning algorithm with self-averaging. Arrow mark the $a_\text{s}\Lambda$ where the training processes are demonstrated Fig. \[performance\]. The inset plots the relative error $\epsilon$  around the first atom-dimer resonant point.[]{data-label="Efimov"}](Efimov.pdf){width="47.00000%"} Therefore, at each iteration, we totally have $5\times 20$ trained NN to predict $\mathcal{U}$ on all $\{k,k^\prime\}$ points in the set $\mathcal{M}$. For each point $i \equiv \{k,k^\prime\}$, we can compute variance $\sigma_i$ for all $100$ predications. We will select $\mathcal{S}_\text{t}=10$ points with the largest variance to compute $\mathcal{U}$ at these points, and add them into the training set. At each iteration, we can compute the mean variance $\sigma_{\rm mean}=1/L^2\sum_{i\subset \mathcal{M}}\sigma_{i}$. When the mean variance $\sigma_{\rm mean}$ is small enough, we start to evaluate the physical quantity $a_{\rm ad}$. Here it comes to another important consideration of our method. On one hand, we have utilized the discrepancy between NNs to guide the query processes more efficiently, but on the other hand, when we start to compute observables, we need to properly average out these variances between different NNs to obtain a converged result. Hence, we compute $a_\text{ad}$ as follows: 1\) For each NN, since there are $20$ copies depending on different initializations, we first average the predications over these $20$ copies to obtain a mean value $\bar{\mathcal{U}}_{\rm t}({k,k^\prime})$ for each point. We then choose one of $\mathcal{U}_{\rm t }^{j}({k,k^\prime})$ that is the closest to $\bar{\mathcal{U}}_{\rm t}({k,k^\prime})$, and we use this $\mathcal{U}_{\rm t }({k,k^\prime})$ (ignoring the upper index) as the representative of the $t$-th NN. 2\) We solve the STM equation with $\mathcal{U}_{\rm t }({k,k^\prime})$ and obtain $a^t_\text{ad}$. Among all five $a^t_\text{ad}$, we discard the largest and the smallest one, and take an average over the rest three, which is taken as $a_\text{ad}$ predicted by the active learning approach. The results are shown in Fig. \[performance\] with blue circles. One can see that the results fluctuate around the exact value, and the fluctuation decreases as $\mathcal{S}$ increases. 3\) To further suppress the fluctuation, we can further take a self-average of $a_\text{ad}$, that is to average over $a_\text{ad}$ for five successive iterations. This result is denoted by $\bar{a}_\text{ad}$ and shown in Fig. \[performance\] with yellow solid dots. Indeed, one can see that the fluctuation is already suppressed strongly at $\mathcal{S}\sim 200$. When the result does not change with training epoch, we take a self-average over the last five iterations to obtain $\bar{a}_\text{ad}$, and we compare this result with $a^{\rm exact}_{\rm ad}$. For instance, when $a_s\Lambda=10$, the relative error $(\bar{a}_{\rm ad}-a^{\rm exact}_{\rm ad})/a^{\rm exact}_{\rm ad}$ is about $1\%$. We apply this method to scan different values of $a_\text{s}$. We stop the iteration at $\mathcal{S}=300$, a number much smaller than uniformly sampled $10^4$ data points, and we take a self-average over the last five iterations to obtain $\bar{a}_\text{ad}$. The results are shown in Fig. \[Efimov\]. Our calculation can obtain the right location for the atom-dimer resonances, and our active learning method can well reproduce the Efimov scaling law. By compared with the exact result, we define a relative error as $(\bar{a}_{\rm ad}-a^{\rm exact}_{\rm ad})/a^{\rm exact}_{\rm ad}\equiv\epsilon$. The $\epsilon$ around the first resonance is shown in the inset of Fig. \[Efimov\], which shows that the relative error does increases nearby the Efimov resonance, but overall it is kept within a few thousandths. To further reveal the mechanism of how our method works, we plot in Fig. \[function\_U\](a) the function $\mathcal{U}(k,k^\prime)$ generated by the uniform sampling of all $10^4$ points. All the data points added during iteration (with $100<\mathcal{S}\leqslant300$) are shown in Fig. \[performance\](b). It is very clear that nearly all points are added in the regime around $k\sim 0$ and $k^\prime \sim 0$, where $U(k,k^\prime)$ is the steepest as one can see from Fig. \[function\_U\](a) and the equal number contour in Fig. \[function\_U\](b). In Fig. \[function\_U\](c) and (d), we compare the function $\mathcal{U}(k,k^\prime)$ generated by NNs and the function generated by uniform sampling. Here the NN results are averaged over five different NNs and different initializations of different NNs. It shows that the fitting is perfect when the calculation of $a_\text{ad}$ converges. ![ (a) Profile of the two-dimensional function $\mathcal{U}(k,k^\prime)$ in the whole momentum space, plotted with all uniformly sampled $10^4$ data points. (b) Visualization of the dataset queried during the active learning iterations. The blue dots denote the queried dataset with $100< \mathcal{S} \leqslant 300$. The red lines are the contour-plotting of (b). (c-d) Difference between function $\mathcal{U}(k,k^\prime)$ generated by the NNs and $\mathcal{U}(k,k^\prime)$ generated by uniform sampling. Here we have averaged over all five different NNs and twenty different initializations. The labelled data $\mathcal{S}=120$ for (c) and $\mathcal{S}=320$ for (d). []{data-label="function_U"}](FigUAll.pdf){width="45.00000%"} ![ (a) Profile of the two-dimensional function $\sigma(k_x,k_z)$ in the whole momentum space, plotted with all uniformly sampled $10^4$ data points. (b) Visualization of the dataset queried during the active learning iterations. The blue dots denote the queried dataset with $100< \mathcal{S} \leqslant 900$. The red lines are the contour-plotting of (b). $\sigma_\text{H}$ is in unit of $\text{S/cm}$. (c-d) The function $\sigma(k_x,k_z)$ generated by the NNs, averaged over all five different NNs and five different initializations for each NN. The labelled data $\mathcal{S}=1200$ for (c) and $\mathcal{S}=4000$ for (d). []{data-label="function_sigma"}](FigsigmaAll.pdf){width="47.00000%"} Anomalous Hall Conductivity Problem =================================== In the second example, we consider the anomalous Hall conductivity of a magnetic Weyl semimetal Mn$_3$Ge [@Kubler2014], which has been extensively studied recently [@Nakatsuji2015; @Nayak2016]. The anomalous Hall conductivity ($\sigma_{\text{Hall}}$) is an intrinsic quantity induced by the Berry curvature of the band structure. We have computed $\sigma_{\text{H}}=\sum_{\mathbf{k}} \sigma(k_x,k_y,k_z)$ based on the $ab~initio$ band structure calculations. The model of this material and the details of computing the Berry curvature are shown in Appendix B. For each $k_x$ and $k_z$ point, we can obtain a value $\sigma(k_x,k_z)$ by performing an integration over $k_y$. Here without the active learning method, when we discretize both $k_x$ and $k_z$ into $100$ points, there are totally $10^4$ data points to be computed. The total Hall conductivity is obtained by summing over $k_x$ and $k_z$ as $\sigma_{\text{H}}=\sum_{k_x,k_z} \sigma(k_x, k_z)$. As one can see from Fig. \[function\_sigma\](a), the function $\sigma(k_x,k_z)$ is much more singular than $\mathcal{U}(k,k^\prime)$ in the previous case. There are four broad peaks and four narrow peaks located around $(0.5\pm0.1,0.5\pm0.4)$ and $(0.5\pm0.4,0.5\pm0.15)$. We follow the same active learning procedure discussed above. In this case we take five different initializations for each NN. Compared with the first example, computing physical observable is easier because we only need to average over all five different initializations for each NN. Similarly, to suppress the fluctuation, at each iteration, we also average over three different predictions of $\sigma_\text{H}$ by discarding the largest and the smallest results. We also take a self-average over five successive iterations. As we show in Fig. \[Hall\], the prediction of the active learning method approaches the exact value when $\mathcal{S}_{\rm max}>700$, which is less than $10\%$ of the total number of uniformly sampled data, and the relative error is about $0.8$. As we shown in Fig. \[function\_sigma\](b), the “query by committee" stratagem guides most of the queried date distributed in the areas of four broad peaks. In Fig. \[function\_sigma\](c) and (d), we show the function $\sigma(k_x,k_z)$ generated by NNs. For $\mathcal{S}=1200$ shown in Fig. \[function\_sigma\](c), the fitting has already captured the four main peaks, and since the contribution to the Hall conductance mainly comes from the four main peaks, the results of $\sigma_\text{H}$ already converges very well. Nevertheless, when we continue to add labelled data until $\mathcal{S}=4000$, as shown in Fig. \[function\_sigma\](d), one can see that these extra data are mainly added in the four narrow peaks such that these four small peaks can also be generated very well. ![The Hall conductivity $\sigma_\text{H}$ calculated with the active learning method. $\sigma_\text{H}$ converges with the increasing number of queried dataset $\mathcal{S}$. The blue empty circles are the results averaged over different NNs at each step, and the yellow solid dots are results averaged over the adjacent five steps. The dashed line denotes the exact results obtained with uniformly sampled all $10^4$ data points. Here we take the number of initial dataset $\mathcal{S}_0=100$ and at each step $\mathcal{S}_t=50$ data points are added. []{data-label="Hall"}](FigHall.pdf){width="47.00000%"} Conclusion and Outlook ====================== In summary, we have developed a neural network based machine learning method to determine a multi-dimensional function efficiently. The method combines the great expressivity of NN and the advantage of the active learning scheme to reduce the demand for labeled data to a minimum. There are two key ingredients of this method. On the one hand, we make use of the variances between NNs to guide the queried data to be sampled into the regime where the function varies rapidly, which makes the calculation more efficiently. On the other hand, we need to properly average out these variances when calculating physical observables using the fitted results. With two examples, we show our method can work remarkably well. Compared with uniform sampling, our method can achieve an accuracy of less than $1\%$ error with only less than $10\%$ of total data points, even when the function has multiple sharp peaks. We believe our method can find broad applications in many areas of computational physics. Acknowledge =========== We thank colleagues in Microsoft Research Asia for discussing active learning. This work is supported Beijing Outstanding Young Scientist Program (HZ), MOST under Grant No. 2016YFA0301600 (HZ) and NSFC Grant No. 11734010 (HZ), the Willner Family Leadership Institute for the Weizmann Institute of Science (BY), the Benoziyo Endowment Fund for the Advancement of Science (BY), and Ruth and Herman Albert Scholars Program for New Scientists (BY). Three-boson problem =================== The Hamiltonian of the three-boson system is given by $$\hat{H}=\sum\limits_{i=1}^{3}-\frac{\nabla^2_i}{2m}+\sum\limits_{i<j,=1}^3V(|{\bf r}_i-{\bf r}_j|),$$ where ${\bf r}_i$ ($i=1,2,3$) is the spatial coordinate of the $i$-th particle, and $V(r)$ is an isotropic short-range potential. Normally we focus on the $s$-wave scattering for ultracold low-energy atoms, and $V(r)$ can be described by the $s$-wave scattering length $a_\text{s}$ and a high-momentum cut-off $\Lambda$. Most features of this quantum three-body problem only depend on the dimensionless parameter $a_\text{s}\Lambda$ [@Braaten]. We consider the case that $a_\text{s}$ is large and positive, where $V(r)$ supports a low-energy bound state called a dimer. A key process to this three-body problem is the atom-dimer scattering illustrated in Fig. \[atom\_dimer\](a), and the quantity to describe this process is a scattering kernel, which is usually denoted by $\mathcal{U}({\bf k},{\bf k}^\prime)$, and it can be computed diagrammatically with the Feynman diagram shown in Fig. \[atom\_dimer\](b) [@Braaten]. When focusing on the $s$-wave scattering only, $\mathcal{U}$ only depends on the amplitude of $k=|{\bf k}|$ and $k^\prime=|{\bf k}^\prime|$, and we can simplify $\mathcal{U}$ as $\mathcal{U}(k,k^\prime)$. ![(a) Schematic of the atom-dimer scattering process. (b) Diagrammatical description of the scattering kernel $\mathcal{U}({\bf k},{\bf k}^\prime)$. The single line stands for the free single-particle Green’s function $G_0^{\rm A}({\bf k},E)$. The double line is the free dimer Green’s function $D_s({\bf k},E)$. The black dot is the $s$-wave interacting vertex. []{data-label="atom_dimer"}](atom_dimer.pdf){width="45.00000%"} Knowing $\mathcal{U}(k,k^\prime)$, one can calculate the few-body quantity, such as atom-dimer scattering length $a_\text{ad}$, through the celebrated Skorniakov-Ter-Martirosion (STM) equation, which can be written as $$\begin{aligned} \int\frac{d k^\prime k^{\prime2}}{2\pi^2}\mathcal{U}(k,k^\prime)D_s\left(k^\prime,E\right)\mathcal{A}(k^\prime)-\mathcal{A}(k)=\mathcal{U}(k,0), \label{EqSTM}\end{aligned}$$ where $$\begin{aligned} D_s^{-1}\left(k^\prime,E\right)=\frac{m}{2\pi a_\text{s}}-\frac{m}{2\pi}\sqrt{\frac{3k^{\prime 2}}{4}-mE}, \label{EqU}\end{aligned}$$ is the full propagator for $s$-wave dimer. Solving Eq. \[EqSTM\] with $E$ fixed at the dimer energy $-\frac{1}{ma_\text{s}^2}$, we can obtain atom-dimer scattering amplitude $\mathcal{A}(k)$, and $a_\text{ad}$ is given by [@Braaten] $$a_{\rm ad}=-\frac{8}{3ma_s}\mathcal{A}(k=0).$$ In practices, suppose that we discretize both $k$ and $k^\prime$ into $L=100$ points between zero and $\Lambda$, Eq. \[EqSTM\] becomes a matrix equation after the discretization, which can be solved by inverting the matrix. Let $\mathcal{M}$ denotes the set of total data points, and the number of data points in $\mathcal{M}$ is $10^4$. With uniform sampling, we need to compute $\mathcal{U}$ for all $10^4$ points, with which we solve Eq. \[EqSTM\] to obtain $a^{\rm exact}_\text{ad}$. This is referred as the exact results in the main text. Anomalous Hall conductivity =========================== The band structure of Mn$_3$Ge calculated with the density-functional theory from Vienna *ab-initio* simulation package [@kresse1996] in the framework of the generalized-gradient approximation. The calculated plane wave basis result were projected to atomic-orbital-like Wannier functions [@Mostofi2008] to get the tight-binding parameters ($t_{ij}$). Based on the tight-binding Hamiltonian, = \_[i,j]{}t\_[ij]{}c\^[+]{}\_i c\_j, we calculated the Berry curvature and the anomalous Hall conductivity in the clean limit. For the sufficient data set, uniform K point grid as $100\times100\times100$ employed to produce $10^6$ numbers of data points. We have evaluated the AHC ($\sigma_{zx}$) and Berry curvature ($\Omega_{zx}$) by the Kubo-formula approach in the linear response scheme [@Xiao2010], \_[xz]{}()&= -\_[BZ]{} \_[\_n&lt;]{}\^n\_[xz ]{}(**k**)\ \_[xz]{}\^n(k)&= i \_[mn ]{} . Here $n, m$ are the band index, $\epsilon_n$ is the eigenvalue of the $|n\rangle$ eigenstate of $H_k$ (Fourier transform of $\hat{H}$), and $\hat{v}_i=d H_k/(\hbar dk_i)$ ($i=x,z$) is the velocity operator, $\mu$ is Fermi level of the system. The calculated anomalous Hall conductivity value is 296 S/cm where S denoting unit Siemens. [10]{} P. Mehta, M. Bukov, C.-H. Wang, A. G. R. Day, C. Richardson, C. K. Fisher, and D. J. Schwab, Phyics Reports **810**,1 (2019) . G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, L. Vogt-Maranto, and L. Zdeborová, Rev. Mod. Phys. **91**, 045002 (2019). D. L. Deng, X. Li, and S. Das Sarma, Phys. Rev. B **96**, 1 (2017). Y. Zhang and E. A. Kim, Phys. Rev. Lett. **118**, 1 (2017). J. Carrasquilla and R. G. Melko, Nat. Phys. **13**, 431 (2017). E. P. L. Van Nieuwenburg, Y. H. Liu, and S. D. Huber, Nat. Phys. **13**, 435 (2017). P. Zhang, H. Shen, and H. Zhai, Phys. Rev. Lett. **120**, 066401 (2018). Y.-H. Liu and E. P. L. van Nieuwenburg, Phys. Rev. Lett. **120**, 176401 (2018). X.-Y. Dong, F. Pollmann, and X.-F. Zhang, Phys. Rev. B **99**, 121104 (2019). G. Carleo and M. Troyer, Science **355**, 602 (2017). X. Gao and L. M. Duan, Nat. Commun. **8**, 662 (2017). Z. Cai and J. Liu, Phys. Rev. B **97**, 035116 (2018). H. Saito, J. Phys. Soc. Japan **87**, 074002 (2018). G. Torlai, G. Mazzola, J. Carrasquilla, M. Troyer, R. Melko, and G. Carleo, Nat. Phys. **14**, 447 (2018). H. Shen, J. Liu, and L. Fu, Phys. Rev. B **97**, 205140 (2018). T. Song and H. Lee, Phys. Rev. B **100**, 045153 (2019). G. Torlai, B. Timar, E. P. L. van Nieuwenburg, H. Levine, A. Omran, A. Keesling, H. Bernien, M. Greiner, V. Vuletić, M. D. Lukin, R. G. Melko, and M. Endres, Phys. Rev. Lett. **123**, 230504 (2019). A. Macarone-Palmier, E. Kovlakov, F. Bianchi, D. Yudin, S. Straupe, J. Biamonte, and S. Kulik, arXiv: 1904.05902(2019). Y. Zhang, A. Mesaros, K. Fujita, S. D. Edkins, M. H. Hamidian, K. Ch’ng, H. Eisaki, S. Uchida, J. C. S. Davis, E. Khatami, and E.-A. Kim, Nature **570**, 484 (2019). B. S. Rem, N. Käming, M. Tarnowski, L. Asteria, N. Fläschner, C. Becker, K. Sengstock, and C. Weitenberg, Nature Physics **15**, 917 (2019). A. Bohrdt, C. S. Chiu, G. Ji, M. Xu, D. Greif, M. Greiner, E. Demler, F. Grusdt, and M. Knap, Nature Physics **15**, 921 (2019). Burr Settles, *Active learning: Synthesis Lectures on Artificial Intelligence and Machine Learning* (Morgan & Claypool Publishers, 2012). L. Zhang, D.-Y. Lin, H. Wang, R. Car, and W. E, Phys. Rev. Materials **3**, 023804 (2018). J. S. Smith, B. Nebgen, N. Lubbers, O. Isayev, and A. E. Roitberg, J. Chem. Phys. **148**, 241733 (2018). K. Gubaev, E. V. Podryabinkin, G. L. W. Hart, and A. V. Shapeev, Comput. Mater. Sci. **156**, 148 (2019). P. B. Wigley, P. J. Everitt, A. Van Den Hengel, J. W. Bastian, M. A. Sooriyabandara, G. D. Mcdonald, K. S. Hardman, C. D. Quinlivan, P. Manju, C. C. N. Kuhn, I. R. Petersen, A. N. Luiten, J. J. Hope, N. P. Robins, and M. R. Hush, Sci. Rep. **6**, 25890 (2016). B. M. Henson, D. K. Shin, K. F. Thomas, J. A. Ross, M. R. Hush, S. S. Hodgman, and A. G. Truscott, Proc. Natl. Acad. Sci. **115**, 13216 (2018). H. S. Seung, M.Pper and H. Sompolinsky, Proceedings of the fifth annual workshop on Computational learning theory, 287, 1992. E. Braaten and H.-W. Hammer, Phys. Rep. **428**, 259 (2006). H. Yang, Y. Sun, Y. Zhang, W. Shi, S. Parkin, B. Yan, New Journal of Physics **19**, 015008 (2017). S. Nakatsuji, N. Kiyohara, and T.Higo, Nature **527**, 212 (2015). A. K. Nayak, J. E. Fischer, Y. Sun, B. Yan, J. Karel, A. C. Komarek, C. Shekhar, N. Kumar, W. Schnelle, J. K[ü]{}bler,C. Felser, and S. S. P. Parkin Sci. Adv. **2**, e1501870 (2016). G. Kresse and J. Furthm[ü]{}ller, Phys. Rev. B **54**, 11169 (1996). A. A. Mostofi, J. R. Yates, Y.-S. Lee, I. Souza, D. Vanderbilt, and N. Marzari, Comput. Phys. Commun. **178**, 685 (2008). D. Xiao, M.-C. Chang, and Q. Niu, Rev. Mod. Phys. **82**, 1959 (2010).

ArXiv

--- author: - 'Filippos Koliopanos[^1]' - Georgios Vasilopoulos bibliography: - 'general.bib' date: 'Received ... / Accepted ...' title: 'Accreting, highly magnetized neutron stars at the Eddington limit: A study of the 2016 outburst of SMC X-3' --- [We study the temporal and spectral characteristics of [SMCX-3]{}during its recent (2016) outburst to probe accretion onto highly magnetized neutron stars (NSs) at the Eddington limit.]{} [We obtained [[*XMM-Newton*]{}]{}observations of [SMCX-3]{}and combined them with long-term observations by [[*Swift*]{}]{}. We performed a detailed analysis of the temporal and spectral behavior of the source, as well as its short- and long-term evolution. We have also constructed a simple toy-model (based on robust theoretical predictions) in order to gain insight into the complex emission pattern of [SMCX-3]{}.]{} [We confirm the pulse period of the system that has been derived by previous works and note that the pulse has a complex three-peak shape. We find that the pulsed emission is dominated by hard photons, while at energies below ${\sim}1$keV, the emission does not pulsate. We furthermore find that the shape of the pulse profile and the short- and long-term evolution of the source light-curve can be explained by invoking a combination of a “fan” and a “polar” beam. The results of our temporal study are supported by our spectroscopic analysis, which reveals a two-component emission, comprised of a hard power law and a soft thermal component. We find that the latter produces the bulk of the non-pulsating emission and is most likely the result of reprocessing the primary hard emission by optically thick material that partly obscures the central source. We also detect strong emission lines from highly ionized metals. The strength of the emission lines strongly depends on the phase.]{} [Our findings are in agreement with previous works. The energy and temporal evolution as well as the shape of the pulse profile and the long-term spectra evolution of the source are consistent with the expected emission pattern of the accretion column in the super-critical regime, while the large reprocessing region is consistent with the analysis of previously studied X-ray pulsars observed at high accretion rates. This reprocessing region is consistent with recently proposed theoretical and observational works that suggested that highly magnetized NSs occupy a considerable fraction of ultraluminous X-ray sources. ]{} Introduction {#sec-intro} ============ X-ray pulsars (XRPs) are comprised of a highly magnetized ($B{>}10^9$G) neutron star (NS) and a companion star that ranges from low-mass white dwarfs to massive B-type stars . XRPs are some of the most luminous (of-nuclear) Galactic X-ray point-sources [e.g., @2017Sci...355..817I]. The X-ray emission is the result of accretion of material from the star onto the NS, which in the case of XRPs is strongly affected by the NS magnetic field: The accretion disk formed by the in-falling matter from the companion star is truncated at approximately the NS magnetosphere. From this point, the accreted material flows toward the NS magnetic poles, following the magnetic field lines, forming an accretion column, inside which material is heated to high energies . The opacity inside the accretion column is dominated by scattering between photons and electrons. In the presence of a high magnetic field, the scattering cross-section is highly anisotropic [@1971PhRvD...3.2303C; @1974ApJ...190..141L], the hard X-ray photons from the accretion column are collimated in a narrow beam (so-called pencil beam), which is directed parallel to the magnetic field axis . The (possible) inclination between the pulsar beam and the rotational axis of the pulsar, combined with the NS spin, produce the characteristic pulsations of the X-ray light curve. During episodes of prolonged accretion, XRPs are known to reach and often exceed the Eddington limit for spherical accretion onto a NS (${\sim}1.8\times10^{38}$erg/s for a $1.4\,{\rm M_{\odot}}$ NS). This is further complicated by the fact that material is accreted onto a very small area on the surface of the NS, and as a result, the Eddington limit is significantly lower. Therefore, even X-ray pulsars with luminosities of ${{\text{about }}}$afew$10^{37}$erg/s persistently break the (local) Eddington limit. [@1976MNRAS.175..395B] demonstrated that while the accreting material is indeed impeded by the emerging radiation and the accretion column becomes opaque along the magnetic field axis, the X-ray photons can escape from the (optically thin) sides of the accretion funnel in a “fan-beam” pattern that is directed perpendicular to the magnetic field. More recent considerations also predict that a fraction of the fan-beam emission can be reflected off the surface of the NS, producing a secondary polar beam that is directed parallel to the magnetic field axis [@2013ApJ...764...49T; @2013ApJ...777..115P]. The recent discoveries of pulsating ultraluminous X-ray sources (PULXs) have established that XRPs can persistently emit at luminosities that are hundreds of times higher than the Eddington luminosity [e.g., @2014Natur.514..202B; @2016ApJ...831L..14F; @2017Sci...355..817I]. Refinements of the mechanism reported by [@1976MNRAS.175..395B] demonstrated that accreting highly magnetized NSs can facilitate luminosities exceeding $10^{40}$erg${\rm s^{-1}}$ [@2015MNRAS.454.2539M], while other recent studies propose that many (if not most) ULXs are accreting (highly magnetized) NSs, rather than black holes [@2016MNRAS.458L..10K; @2017MNRAS.468L..59K; @2017ApJ...836..113P; @2017MNRAS.467.1202M; @kolio2017]. Understanding the intricacies of the physical mechanisms that underly accretion onto high-B NSs requires the detailed spectral and timing analysis of numerous XRPs during high accretion episodes. The shape of the pulse profile in different energy bands and at different luminosities provides valuable insights into the shape of the emission pattern of the accretion column and may also shed light on the geometry and size of the accretion column itself . Furthermore, the shape of the spectral continuum, its variability, and the fractional variability of the individual components of which it is comprised, along with the presence (or absence), shape and variability of emission-like features are directly related to the radiative processes that are at play in the vicinity of the accretion column, and may also reveal the presence of material trapped inside or at the boundary of the magnetosphere (; ; @2002ApJ...564L..21S; ; @2012MNRAS.425..595R; @2016MNRAS.461.1875V; Vasilopoulos et al. 2017 in press). The emission of the accretion column of XRPs has a spectrum that is empirically described by a very hard power law (spectral index ${\lesssim}1.8$) with a low-energy (${\lesssim}10$keV) cutoff [e.g., @2012MmSAI..83..230C]. It has been demonstrated that the spectrum can be reproduced assuming bulk and thermal Comptonization of bremsstrahlung, blackbody, and cyclotron seed photons [@1981ApJ...251..288N; @1985ApJ...299..138M; @1991ApJ...367..575B; @2004ApJ...614..881H; @2007ApJ...654..435B]. Nevertheless, a comprehensive, self-consistent modeling of the emission of the accretion column has not been achieved so far. The spectra of XRPs often exhibit a distinctive spectral excess below ${\sim}1$keV, which is successfully modeled using a blackbody component at a temperature of ${{\sim}}$0.1-0.2keV [e.g., @2004ApJ...614..881H and references therein]. While some authors have argued that this “soft excess” is the result of Comptonization of seed photons from the truncated disk by low-energy ($kT_{e}{\lesssim}1$keV) electrons [e.g., @2001ApJ...553..375L], others have noted that the temperature of the emitting region is hotter and its size considerably larger than what is expected for the inner edge of a truncated accretion disk, and they attribute the feature to reprocessing of hard X-rays by optically thick material, trapped at the boundary of the magnetosphere. More recently, [@2017MNRAS.467.1202M] have extended these arguments to the super-Eddington regime, arguing that at high mass accretion rates (considerably above the Eddington limit), this optically thick material engulfs the entire magnetosphere and reprocesses most (or all) of the primary hard emission. The resulting accretion “curtain” is substantially hotter (${{\gtrsim}}1\,$keV) than that of the soft excess in the sub-Eddington sources. Be/X-ray binaries are a subclass of high-mass X-ray binaries (HMXB) that contains the majority of the known accreting X-ray pulsars with a typical magnetic field strength equal to or above $3\times10^{11}$ G [@2015MNRAS.454.3760I; @2016ApJ...829...30C]. In BeXRBs, material is provided by a non-supergiant Be-type donor star. This is a young stellar object that rotates with a near-critical rotation velocity, resulting in strong mass loss via an equatorial decretion disk . As BeXRBs are composed of young objects, their number within a galaxy correlates with the recent star formation [e.g., @2010ApJ...716L.140A; @2016MNRAS.459..528A]. By monitoring BeXRB outbursts, it is now generally accepted that normal or so-called Type I outbursts ($L_x\sim$[$10^{36}$ erg s$^{-1}$]{}) can occur as the NS passes close to the decretion disk, thus they appear to be correlated with the binary orbital period. Giant or Type II outbursts ($L_{x}\ge$[$10^{38}$ erg s$^{-1}$]{}) that can last for several orbits are associated with wrapped Be-disks [@2013PASJ...65...41O]. From an observational point of view, the former can be quite regular, appearing in each orbit [e.g., LXP38.55 and EXO2030+375; @2016MNRAS.461.1875V; @2008ApJ...678.1263W], while for other systems, they are rarer [@2015int..workE..78K]. On the other hand, major outbursts producing [$10^{38}$ erg s$^{-1}$]{} are rare events with only a hand-full detected every year. Moreover, it is only during these events that the pulsating NS can reach super-Eddington luminosities, thus providing a link between accreting NS and the emerging population of NS-ULXs [@2014Natur.514..202B; @2016ApJ...831L..14F; @2017Sci...355..817I; @2017MNRAS.466L..48I]. The spectroscopic study of BeXRB outbursts in our Galaxy is hampered by the strong Galactic absorption and the often large uncertainties in their distances. The Magellanic Clouds (MCs) offer a unique laboratory for studying BeXRB outbursts. The moderate and well-measured distances of $\sim$50 kpc for the Large Magellanic Cloud (LMC) [@2013Natur.495...76P] and $\sim$62 kpc for the Small Magellanic Cloud (SMC) [@2014ApJ...780...59G] as well as their low Galactic foreground absorption ($\sim$6$\times10^{20}$cm$^{-2}$) make them ideal targets for this task. The 2016 major outburst of [SMCX-3]{}offers a rare opportunity to study accretion physics onto a highly magnetized NS during an episode of high mass accretion. [SMCX-3]{}was one of the earliest X-ray systems to be discovered in the SMC in 1977 by the SAS 3 satellite at a luminosity level of about [$10^{38}$ erg s$^{-1}$]{} [@1978ApJ...221L..37C]. The pulsating nature of the system was revealed more than two decades later, when 7.78 s pulsations were measured [@2003HEAD....7.1730C] using data from the Proportional Counter Array on board the Rossi X-ray Timing Explorer [PCA, RXTE; @2006ApJS..163..401J], and a proper association was possible through an analysis of archival [[*Chandra*]{}]{}observations [@2004ATel..225....1E]. By investigating the long-term optical light-curve of the optical counterpart of [SMCX-3]{}, @2004AJ....128..709C reported a 44.86 d modulation that they interpreted as the orbital period of the system. @2008MNRAS.388.1198M have reported a spectral type of B1-1.5 IV-V for the donor star of the BeXRB. The 2016 major outburst of [SMCX-3]{}was first reported by MAXI [@2016ATel.9348....1N] and is estimated to have started around June 2016 [@2016ATel.9362....1K], while the system still remains active at the time of writing (June 2017). Lasting for more than seven binary orbital periods, the 2016 outburst can be classified as one of the longest ever observed for any BeXRB system. Since the outburst was reported, the system has been extensively monitored in the X-rays by [[*Swift*]{}]{}with short visits, while deeper observations have been performed by NuSTAR, [[*Chandra*]{}]{}, and [[*XMM-Newton*]{}]{}. @2017arXiv170102336T have reported a 44.918 d period as the true orbital period of the system by modeling the pulsar period evolution (see also @2017ApJ...843...69W) during the 2016 outburst with data obtained by [[*Swift*]{}]{}/XRT and by analyzing the optical light-curve of the system, obtained from the Optical Gravitational Lensing Experiment [OGLE, @2015AcA....65....1U]. The maximum luminosity obtained by [[*Swift*]{}]{}/XRT imposes an above-average maximum magnetic field strength at the NS surface ($>10^{13}$ G), as does the lack of any cyclotron resonance feature in the high-energy X-ray spectrum of the system [@2016ATel.9404....1P; @2017arXiv170200966T]. On the other hand, the lack of a transition to the propeller regime during the evolution of the 2016 outburst suggests a much weaker magnetic field at the magnetospheric radius. According to @2017arXiv170200966T, this apparent contradiction could be resolved if there were a significant non-bipolar component of the magnetic field close to the NS surface. Following the evolution of the outburst, we requested a “non anticipated” [[*XMM-Newton*]{}]{}ToO observation (PI: F. Koliopanos) in order to perform a detailed study of the soft X-ray spectral characteristics of the system. In Sect. \[sec-observations\] we describe the spectral and temporal analysis of the [[*XMM-Newton*]{}]{}data. In Sect. \[sec:irr\] we introduce a toy model constructed to phenomenologically explain our findings, and in Sects. \[discussion\] and \[conclusion\] we discuss our findings and interpret them in the context of highly accreting X-ray pulsars in (or at the threshold of) the ultraluminous regime. Observations and data analysis {#sec-observations} ============================== [[*XMM-Newton*]{}]{}data extraction ----------------------------------- [[*XMM-Newton*]{}]{}observed SMC X-3 on October 14, 2016 for a duration of $\sim37$ks. All onboard instruments were operational, with MOS2 and pn operating in Timing Mode and MOS1 in Large Window Mode. All detectors had the Medium optical blocking filter on. In Timing mode, data are registered in one dimension, along the column axis. This results in considerably shorter CCD readout time, which increases the spectral resolution, but also protects observations of bright sources from pile-up[^2]. Spectra from all detectors were extracted using the latest XMM-Newton Data Analysis software SAS, version 15.0.0., and using the calibration files released[^3] on May 12, 2016. The observations were inspected for high background flaring activity by extracting high-energy light curves (E$>$10keV for MOS and 10$<$E$<$12keV for pn) with a 100s bin size. The examination revealed $\sim$2.5ks of the pn data that where contaminated by high-energy flares. The contaminated intervals where subsequently removed. The spectra were extracted using the SAS task `evselect`, with the standard filtering flags (`#XMMEA_EP && PATTERN<=4` for pn and `#XMMEA_EM && PATTERN<=12` for MOS). The SAS tasks `rmfgen` and `arfgen` were used to create the redistribution matrix and ancillary file. The MOS1 data suffered from pile-up and were therefore rejected. For simplicity and self-consistency, we opted to only use the EPIC-pn data in our spectral analysis. The pn spectra are optimally calibrated for spectral fitting and had $\sim$1.3$\times10^6$ registered counts, allowing us to accurately constrain emission features and minute spectral variations between spectra extracted during different pulse phases. While an official estimate is not provided by the [[*XMM-Newton*]{}]{}SOC, the pn data in timing mode are expected to suffer from systematic uncertainties in the 1-2% range (e.g., see Appendix A in ). Accordingly, we quadratically added 1% systematic errors to each energy bin of the pn data. The spectra from the Reflection Grating Spectrometer (RGS, ) were extracted using SAS task `rgsproc`. The resulting RGS1 and RG2 spectra were combined using [`r`gscombine]{}. [[*Swift*]{}]{}data extraction ------------------------------ To study the long-term behavior of the source hardness, pulse fraction, and evolution throughout the 2016 outburst, we also analyzed the available [[*Swift*]{}]{}/XRT data up to December 31, 2016. The data were analyzed following the instructions described in the [[*Swift*]{}]{}data analysis guide[^4] . We used [xrtpipeline]{} to generate the [[*Swift*]{}]{}/XRT products, while events were extracted by using the command line interface [xselect]{} available through HEASoft FTOOLS [@1995ASPC...77..367B][^5]. Source events were extracted from a 45region, while background was extracted from an annulus between 90and 120. X-ray timing analysis {#time} --------------------- We searched for a periodic signal in the barycentric corrected EPIC-pn event arrival time-series by using an epoch-folding technique . The EF method uses a series of trial periods within an appropriate range to phase-fold the detected event arrival times and performs a $\chi^2$ minimization test of a constant signal hypothesis. Folding a periodic light curve with an arbitrary period smears out the signal, and the folded profile is expected to be nearly flat. Thus a high value of the $\chi^2$ (i.e., a bad fit) indirectly supports the presence of a periodic signal. A disadvantage of this method is that it lacks a proper determination of the period uncertainty. In many cases, the FWHM of the $\chi^2$ distribution is used as an estimate for the uncertainty, but this value only improves with the length of the time series and not with the number of events [@1996ApJ...473.1059G] and does not have the meaning of the statistical uncertainty. In order to have a correct estimate for both the period and its error, we applied the Gregory-Loredo method of Bayesian periodic signal detection [@1996ApJ...473.1059G], and constrained the search around the true period derived from the epoch-folding method. To further test the accuracy in the derived pulsed period, we performed 100 repetitions of the above method while bootstrapping the event arrival times and by selecting different energy bands with 1.0 keV width (e.g., 2.0-3.0 keV and 3.0-4.0 keV). The derived period deviation between the above samples was never above $10^{-5}$ sec. From the above treatment, we derived a best-fit pulse period of 7.77200(6) s. The pulse profile corresponding to the best-fit period is plotted in Fig. \[fig:pp\]. The accurate calculation of the pulse profile enabled us to compute the pulsed fraction (PF), which is defined as the ratio between the difference and the sum of the maximum and minimum count rates over the pulse profile, i.e., $$PF=(F_{max}-F_{min})/(F_{max}+F_{min}). \label{pfeq}$$ For the [[*XMM-Newton*]{}]{}EPIC-pn pulse profile (Fig. \[fig:pp\]) we calculated a pulsed fraction of $PF=0.328\pm0.005$. We also estimated the PF for five energy bands, henceforth noted with the letter $i$, with $i=1,2,3,4,5$ and $1\rightarrow(0.2-0.5)$ keV, $2\rightarrow(0.5-1.0)$ keV, $3\rightarrow(1.0-2.0)$ keV, $4\rightarrow(2.0-4.5)$ keV, $5\rightarrow(4.5-10.0)$ keV . The resulting values for the energy-resolved PF are $PF_1=0.208$, $PF_2=0.249$, $PF_3=0.341$, $PF_4=0.402$ and $PF_5=0.410$. In Fig. \[fig:HR\] we present the phase-folded light curves for the five energy bands introduced above, together with the four phase-resolved hardness ratios. The hardness ratios $HR_i$ ($i=1,2,3,4$) are defined as $$HR_i = \frac{R_{i+1} - R_{i}}{R_{i+1} + R_{i}} ,$$ with R$_{\rm i}$ denoting the background-subtracted count rate in the i$^{\rm th}$ energy band. The plots revealed that the prominence of the three major peaks are strongly dependent on the energy range: the pulsed emission appears to be dominated by high-energy photons (i.e., $>2$keV), while in the softer bands, the pulse shape is less pronounced. To better visualize the spectral dependence of the pulse phase, we produced a 2D histogram for the number of events as function of energy and phase (top panel of Fig. \[fig:phase\_heat\]). The histogram of the events covers a wide dynamic range because the efficiency of the camera and telescope varies strongly with energy. To produce an image where patterns and features (such as the peak of the pulse) are more easily recognized, we normalized the histogram by the average effective area of the detector for the given energy bin and then normalized it a second time by the phase-averaged count rate in each energy bin (middle panel, Fig. \[fig:phase\_heat\]). Last, in the lower bin, each phase-energy bin was divided by the average soft X-ray flux (0.2-0.6 keV) of the same phase bin (bottom panel, Fig. \[fig:phase\_heat\]). This energy range was chosen because it displays minimum phase variability (Fig. \[fig:HR\]). X-ray spectral analysis ----------------------- We performed phase-averaged and phase-resolved spectroscopy of the pn data and a separate analysis of the RGS data, which was performed using the latest version of the [XSPEC]{} fitting package, Version 12.9.1 [@1996ASPC..101...17A]. The spectral continuum in both phase-averaged and phased-resolved spectroscopy was modeled using a combination of an absorbed multicolor disk blackbody (MCD) and a power-law component. The interstellar absorption was modeled using the [`t`bnew]{} code, which is a new and improved version of the X-ray absorption model [`t`babs]{} [@2011_in_prep]. The atomic cross-sections were adopted from [@1996ApJ...465..487V]. We considered a combination of Galactic foreground absorption and an additional column density accounting for both the interstellar medium of the SMC and the intrinsic absorption of the source. For the Galactic photoelectric absorption, we considered a fixed column density of nH$_{\rm GAL}$ = 7.06$\times10^{20}$cm$^{−2}$ , with abundances taken from [@2000ApJ...542..914W]. ### Phase-averaged spectroscopy Both thermal and nonthermal emission components were required to successfully fit the spectral continuum. More specifically, when modeling the continuum using a simple power law, there was a pronounced residual structure that strongly indicated thermal emission below 1keV. On the other hand, nonthermal emission dominated the spectrum at higher energies. The thermal emission was modeled by an MCD model (for the rationale of this choice, see the discussion in Section \[discussion\]), which we modeled using the [XSPEC]{} model [`diskbb`]{}. The continuum emission models alone could not fit the spectrum successfully, and strong emission-like residuals were noted at ${\sim}$0.51, ${\sim}$0.97, and ${\sim}$6.63keV and yielded a reduced $\chi^2$ value of 1.93 for 198 dof. (see ratio plot in Fig. \[fig:spec\_av\]). The corresponding emission features where modeled using Gaussian curves, yielding a $\chi^2$ value of 1.01 for 190 dof. The three emission lines are centered at ${\rm E_{1}}{\sim}0.51$keV, ${\rm E_{2}}{\sim}0.96$keV, and ${\rm E_{3}}{\sim}6.64$keV and have equivalent width values of 6.6$_{-4.1}^{+5.0}$eV, 19$_{-5.1}^{+7.0}$eV, and 72$_{-15}^{+17}$eV, respectively. The ${\rm E_{2}}$ and ${\rm E_{3}}$ lines have a width of $83.8_{-2.6}^{3.0}$eV and 361$_{-80.1}^{+89.7}$eV, respectively, and the ${\rm E_{1}}$ line was not resolved. The three emission-like features also appear in the two MOS detectors (see Fig. \[fig:spec\_av\]) and the RGS data. The emission-like features in the MOS spectra increase the robustness of their detection, but the MOS data are not considered in this analysis. The MOS1 detector was heavily piled up, and since we used the pn detector for the phase-resolved spectroscopy (as it yields more than three times the number of counts of the MOS2 detector), we also opted to exclude the MOS2 data and only use the pn data for the phase-averaged spectroscopy in order to maintain consistency. For display purposes, we show the MOS1, MOS2, and pn spectra and the data-to-model vs energy plots in Fig. \[fig:spec\_av\_MOS\]. For the purposes of this plot, we fit the spectra of the three detectors using the [`diskbb`]{} plus [`powerlaw`]{} model, with all their parameters left free to vary. This choice allows for the visual detection of the line-like features while compensating for the considerable artificial hardening of the piled-up MOS1 detector. The MOS data indicate that the line-like features at ${\sim}0.5\,$keV and ${\sim}1\,$keV are composed of more emission lines that are unresolved by pn. This finding is confirmed by the RGS data (see Sect. \[sec:RGS\] for more details). Narrow residual features are still present in the 1.7-2.5kev range. They stem most likely from incorrect modeling of the Si and Au absorption in the CCD detectors by the EPIC pn calibration, which often results in emission and/or absorption features at ${\sim}$1.84keV and ${\sim}$2.28keV (M$\beta$) and 2.4keV (M$\gamma$). Since the features are not too pronounced and did not affect the quality of the fit or the values of the best-fit parameters, they were ignored, but their respective energy channels were still included in our fits. The best-fit parameters for the modeling of the phase-averaged pn data are presented in Tables \[tab:cont\_phase\] and \[tab:lines\], with the normalization parameters of the two continuum components given in terms of their relevant fluxes (calculated using the multiplicative [XSPEC]{} model [`cflux`]{}). The phase-averaged pn spectrum along with the data-to-model ratio plot (without the Gaussian emission lines) is presented in Figure \[fig:spec\_av\]. The phase-averaged spectrum was also modeled using a single-temperature blackbody instead of the MCD component. The fit was qualitatively similar to the MCD fit, with a k${\rm T_{BB}}$ temperature of 0.19$\pm0.01$keV, a size of 168$_{-16.4}^{+22.1}$km, and a power-law spectral index of $0.98\pm0.01$. ### Phase-resolved spectroscopy {#phase-res} We created separate spectra from 20 phase intervals of equal duration. We studied the resulting phase-resolved spectra by fitting each individual spectrum separately using the same model as in the phase-averaged spectrum. We noted the behavior of the two spectral components and the three emission lines, and monitored the variation of their best-fit parameters as the pulse phase evolved. The resulting best-fit values for the continuum are presented in Table \[tab:cont\_phase\], and those of the emission lines in Table \[tab:lines\]. In Figure \[fig:phace\_spec\_res\] we present the evolution of the different spectral parameters and the source count rate with the pulse phase. For purposes of presentation and to further note the persistent nature of the soft emission, we also fit all 20 spectra simultaneously and tied the parameters of the thermal component, while the parameters of the power-law component were left free to vary. The results of this analysis were qualitatively similar to the independent spectral fitting (see Fig. \[fig:spec\_res\], where the spectral evolution of the source is illustrated). Nevertheless, the decision to freeze the temperature and normalization of the soft component at the phase-averaged best-fit value introduces bias to our fit, and we therefore only tabulate and study the best-fit values for the independent modeling of the phase-resolved spectra. ### RGS spectroscopy {#sec:RGS} The RGS spectra were not grouped and were fitted employing Cash statistics [@1979ApJ...228..939C] and with the same continuum model as the phased-averaged pn spectra. The model yielded an acceptable fit, with a reduced $\chi^2$ value of 1.05 for 1875 dof. The RGS spectra were also fitted simultaneously with the pn phase-averaged spectra (with the addition of a normalization constant, which was left free to vary) yielding best-fit parameter values consistent within 1${\sigma}$ error bars with the pn continuum model. We therefore froze the continuum parameters to the values tabulated in the first row of Table \[tab:cont\_phase\]. We note that a simple absorbed power-law model still describes the continuum with sufficient accuracy, and the use of a second component only improves the fit by $\sim$2.7$\sigma,$ as indicated by performing an ftest[^6]. We searched the RGS spectra for emission and absorption lines by a blind search for Gaussian features with a fixed width. The spectrum was searched by adding the Gaussian to the continuum and fitting the spectrum with a step of 10eV, and the resulting $\Delta$C was estimated for each step. Emission and absorption lines are detected with a significance higher than 3${\sigma.}$ To estimate the values for the $\Delta$C that correspond to the 3$\sigma$ significance, we followed a similar approach as [@2018MNRAS.473.5680K]. We simulated 1000 spectra based on the continuum model and repeated the process of line-search using the fixed-width Gaussian. We plot the probability distribution of the $\Delta$C values and indicate the range of $\Delta$C values within which lies the 99.7% of the trials. $\Delta$C values outside this interval (see Fig. \[fig:spec\_RGS\], dotted lines) are defined as having a 3$\sigma$ significance. In Fig. \[fig:spec\_RGS\] we plot the $\Delta$C improvement of emission and absorption lines detected in the RGS data assuming Gaussian lines with a fixed width of 2eV (black solid line) and 10eV (orange dashed line). The line significance is affected by the width of the Gaussian, and a more thorough approach should include a grid of different line widths in order to achieve the best fit. We here used the emission lines with a significance higher than 3$\sigma$ (dotted line) assuming a 2eV fixed width, which were then fit manually with the width as a free parameter. The best-fit values are presented in Table \[tab:rgs\]. We also note a strong absorption line at more than 5$\sigma$ significance and several tentative others. The most prominent absorption line is centered at ${\approx}1.585\,$keV. When the line is associated with MgXII absorption, it appears to be blueshifted by 7%. The apparent blueshift could be an indication of outflows, but he velocity appears to be considerable lower than the velocities inferred for ULXs [e.g., @2016Natur.533...64P]. A more thorough study of the RGS data using an extended grid of all line parameters will be part of a separate publication. \[tab:cont\_phase\] \[tab:lines\] \[tab:rgs\] X-ray spectral and temporal evolution during the 2016 outburst -------------------------------------------------------------- We report on the long-term evolution of the values of the PF and HR throughout the 2016 outburst, using the data from [[*Swift*]{}]{}observatory. Using the method described in Sect. \[time\], we searched for the pulse period of [SMCX-3]{}for all the analyzed [[*Swift*]{}]{}/XRT observations using the barycenter corrected events within the 0.5-10.0 keV energy band. The resulting pulse periods are presented in Fig. \[fig:t\_spin\]. Using the derived best-fit periods, we estimated the PF for all the [[*Swift*]{}]{}/XRT observations (see Fig. \[fig:cr\_pf\]). From the extracted event files, we calculated the phase-average hardness ratios for all [[*Swift*]{}]{}/XRT observations. For simplicity and in order to increase statistics, we used three energy bands to estimate two HR values: HR [ soft,]{} using the 0.5-2.0 keV and 2.0-4.5 keV bands, and HR [ hard,]{} using the 2.0-4.5 keV and 4.5-10.0 keV bands. We performed the same calculations for the phase-resolved spectrum of [SMCX-3]{}as measured by the [[*XMM-Newton*]{}]{}ToO. In particular, we estimated the above two HR values for 40 phase intervals. The evolution of the spectral state of [SMCX-3]{}with time and pulse phase as depicted by these two HR values is shown in Fig. \[fig:HR\_L\] (left). For comparison, we have plot in parallel (Fig. \[fig:HR\_L\], right panel) the evolution of the HR with NS spin phase during the [[*XMM-Newton*]{}]{}observation. We note that although quantitatively different HR values are expected for different instruments (i.e., [[*Swift*]{}]{}/XRT and [[*XMM-Newton*]{}]{}/EPIC-pn), a qualitative comparison of the HR evolution can be drawn. More specifically, the left-hand plot probes the long-term HR evolution versus the mass accretion rate (inferred from the registered source count-rate), and the right-hand side shows the HR evolution with pulse phase. Emission pattern, pulse reprocessing, and the observed PF {#sec:irr} ========================================================= To investigate the evolution of the PF with luminosity and to investigate the origin of the soft excess as determined from the phase-resolved analysis, we constructed a toy model for the beamed emission. The model assumes that the primary beamed emission of the accretion column is emitted perpendicularly to the magnetic field axis, in a fan-beam pattern [@1976MNRAS.175..395B]. Furthermore, it is assumed that all the emission originates in a point source at a height of ${\sim}2-3\,$km from the surface of the NS. A fraction of the fan emission will also be beamed toward the NS surface, off of which it will be reflected, resulting in a secondary polar beam that is directed perpendicularly to the fan beam [this setup is described in @2013ApJ...764...49T see also their Fig. 4 ]. Partial beaming of the primary fan emission toward the NS surface is expected to occur primarily as a result of scattering by fast electrons at the edge of accretion column [@1976SvA....20..436K; @1988SvAL...14..390L; @2013ApJ...777..115P], and to a lesser degree as a result of gravitational bending of the fan-beam emission. The latter may also result in the emission of the far-away pole entering the observer line of sight. In the model, gravitational bending is accounted for according to the predictions of [@2002ApJ...566L..85B]. The size of the accretion column (and hence the height of the origin of the fan-beam emission) is a function of the NS surface magnetic field strength and the mass accretion rate [see @2015MNRAS.454.2539M and references within]. As it changes, it affects the size of the illuminated region on the NS surface and therefore the strength of the reflected polar emission. If the accretion rate drops below the critical limit, the accretion column becomes optically thin in the direction parallel to the magnetic field axis, and the primary emission is then described by the pencil-beam pattern . In principle, the polar and pencil-beam emission patterns can be phenomenologically described by the same algebraic model. We also modeled the irradiation of any slab or structure (e.g., the inner part of the accretion disk that is located close to the NS) by the combined beams. Last, while arc-like accretion columns and hot-spots have been found in 3D magnetohydrodynamic simulations [@roma2004] of accreting pulsars, we did not include them in our treatment as they would merely introduce an anisotropy in the pulse profile and not significantly affect our results. Pulsed fraction evolution ------------------------- The beaming functions of the polar and fan beams are given by $f\sin^m{\phi}$ and $p\cos^k{\phi}$, respectively. Here $\phi$ is the angle between the magnetic field axis and the photon propagation and is thus a function of the angle between the magnetic and rotation axis ($\theta$), the angle of the NS rotation axis, the observing angle ($i$), and the NS spin phase. The exponent values $m$ and $k$ can be as low as 1, but in the general case of the accretion column emission, they can have significantly higher values [@2013ApJ...764...49T]. We constructed various pulse profiles for different combinations of fan- and polar-beam emission patterns with different relative intensities of the fan and polar beams (i.e., $F_{\rm Polar}/F_{\rm Fan}$ ranging from $\sim$0.01 to ${\sim}30$). We note that these values exceed any realistic configuration between the fan beam and the reflected polar beam. More specifically, since the polar beam is a result of reflection of the fan beam, it is only for a very limited range of observing angles that its contribution will (seemingly) exceed that of the fan beam, and thus a value of $F_{\rm Polar}/F_{\rm Fan}\gtrsim2$, is unlikely. On the other hand, the increasing height of the emitting region, which would result in a smaller fraction of the fan beam being reflected by the NS surface, is limited to $\lesssim10\,km$ [@2015MNRAS.454.2539M], thus limiting the $F_{\rm Polar}/F_{\rm Fan}$ ratio to values greater than 0.1 [see, e.g., Fig 2. of @2013ApJ...777..115P]. Despite these physical limitations, we have extended our estimates to these exaggerated values in order to better illustrate the contribution of each component (fan and polar) to the PF. Moreover, a value of $F_{\rm Polar}/F_{\rm Fan}$ exceeding 10 qualitatively describes the pencil-beam regime, which is expected at lower accretion rates. We estimated the PF for a range of observer angles and for different combinations of angles between the magnetic and rotation axis. An indicative result for a 45$^o$ angle between the NS magnetic field and rotation axis and for an $r_G/r_{NS}=0.25$ is plotted in Fig \[fig:PF\_evol\]. Reprocessed radiation --------------------- To investigate the origin of the soft excess (i.e., the thermal component in the spectral fits) and to assess its contribution to the pulsed emission, we studied the pattern of the reprocessed emission from irradiated optically thick material in the vicinity of the magnetosphere. More specifically, we considered a cylindrical reprocessing region that extends symmetrically above and below the equatorial plane. The radius of the cylinder is taken to be equal to the magnetospheric radius, where the accretion disk is interrupted. The latitudinal size of the reprocessing region corresponds to an angular size of 10$^{\rm o}$, as viewed from the center of the accretor [see, e.g., @2004ApJ...614..881H and our Fig. \[fig:cher\] for an illustration]. The model takes into account irradiation of this region by the combined polar- and fan-beam emission. If [**d**]{} is the unit vector toward the observer and [**n**]{} is the disk surface normal, the total radiative flux that the observer sees would be estimated from their angle [**d${\cdot}$n**]{}. We furthermor assume that the observer sees only half of the reprocessing region (and disk, i.e., $\pi$). Light-travel effects of the scattered/reprocessed radiation [e.g., equation 3 of @2016MNRAS.455.4426V] should have a negligible effect in altering the phase of the reprocessed emission, as the light-crossing time of the disk is less than 1% (0.05% $=(1000 km /c)/P_{NS} *100\%$) of the NS spin period. Thus, any time lag that is expected should be similar in timescale to the reverberation lag observed in LMXB systems [e.g., H1743-322: @2016ApJ...826...70D]. To determine the fraction of the pulsed emission that illuminates the reprocessing region, we integrated the emission over the 10$^{\rm o}$ angular size of the reprocessing region and estimated the PF for an inclination angle between the rotational and the magnetic filed axes ($\theta$) that ranges between 0$^{\rm o}$ and 90$^{\rm o}$ (Fig. \[fig:irrad\], dotted line). We also estimate the PF originating from the reprocessing region assuming a ratio of $F_{\rm Polar}/F_{\rm Fan}=$0.25 (Fig. \[fig:irrad\], solid line). To further probe the contribution of the two beam components to the reprocessed emission, we estimated the PF for two extreme cases in which all the primary emission is emitted in the polar beam (Fig. \[fig:irrad\], green dashed line) or only the fan beam (Fig. \[fig:irrad\], blue dashed line). As we discussed, this scenario is unrealistic, since the polar beam is the result of reflection of the fan beam. However, this setup would adequately describe the pencil-beam emission, expected at lower accretion rates . Discussion ========== The high quality of the [[*XMM-Newton*]{}]{}spectrum of SMC-X3 and the very large number of total registered counts have allowed us to perform a detailed time-resolved spectroscopy where we robustly constrained the different emission components and monitored their evolution with pulse phase. Furthermore, the high time-resolution of the [[*XMM-Newton*]{}]{}detectors, complimented with the numerous Swift/XRT observations (conducted throughout the outburst), have allowed for a detailed study of the long- and short-term temporal behavior of the source. Below, we present our findings and their implications with regard to the underlying physical mechanisms that are responsible for the observed characteristics. Emission pattern of the accretion column and its evolution ---------------------------------------------------------- The 0.01-12keV luminosity of SMC X-3 during the [[*XMM-Newton*]{}]{}observation is estimated at 1.45$\pm0.01\times10^{38}$erg/s for a distance of ${\sim}$62kpc. The corresponding mass accretion rate, assuming an efficiency of $\xi=0.21$ [e.g., @2000AstL...26..699S], is ${\dot M}{\sim}1.2\times10^{-8}\,{M_ \odot }/{\rm yr}$, which places the source well within the fan-beam regime [@1976MNRAS.175..395B; @1981ApJ...251..288N]. Furthermore, the shape of the pulse profile (Fig. \[fig:pp\]) indicates a more complex emission pattern, comprised of the primary fan beam and a secondary reflected polar beam, as was proposed in [@2013ApJ...764...49T] and briefly presented in Section \[sec:irr\]. The energy-resolved pulse profiles and the phase-resolved hardness ratios presented in Fig. \[fig:HR\] strongly indicate that the pulsed emission is dominated by hard-energy photons, while the smooth single-color appearance of the 0.2-0.6keV energy range in the lower bin heat map presented in Fig. \[fig:phase\_heat\] indicates a soft spectral component that does not pulsate. Using our toy-model for the emission of the accretion column, we were able to explore the evolution of the PF with the $F_{\rm Polar}/F_{\rm Fan}$ ratio, assuming a broad range of observer viewing angles. Interestingly, the resulting PF vs $F_{\rm Polar}/F_{\rm Fan}$ ratio (Fig. \[fig:PF\_evol\]) qualitatively resembles the evolution of the PF, with increasing source count-rate, as presented in Fig. \[fig:cr\_pf\], using Swift/XRT data. Following the paradigm of and [@1976MNRAS.175..395B], we expect the emission pattern of the accretion column to shift from a pencil- to a fan-beam pattern as the accretion rate increases. Below $\sim$1cts/sec, the emission is dominated by the pencil beam (which is qualitatively similar to the polar beam). This regime is defined by $F_{\rm Polar}/F_{\rm Fan}>2$. As the luminosity increases, the radiation of the accretion column starts to be emitted in a fan beam, which is accompanied by the secondary polar beam, corresponding to $0.7{\lesssim}F_{\rm Polar}/F_{\rm Fan}{\lesssim}2$ (this only refers to the X-axis of Fig. \[fig:cr\_pf\] and does not imply chronological order, which is indicated with arrows in Fig. \[fig:HR\_L\]). This is the era during which the [[*XMM-Newton*]{}]{}observation was conducted. As the accretion rate continued to increase, the accretion column grew larger, the height of the emitting region increased, and the emission started to become dominated by the fan beam as the reflected component decreases, yielding a lower value of the $F_{\rm Polar}/F_{\rm Fan}$ ratio. Nevertheless, it is evident from Fig. \[fig:cr\_pf\] that the PF does not readily increase as the fan beam becomes more predominant (i.e., above ${\sim}7$cts/s). This indicates that the height of the accretion column is limited to moderate values [e.g., as argued by @2013ApJ...777..115P; @2015MNRAS.454.2539M], and therefore a considerable contribution from the polar beam is always present. The presence of the polar beam is also indicated by the shape of the pulse profile and its evolution during different energy bands and HR values (Figures \[fig:pp\] and \[fig:HR\]). The pulse profile has three distinct peaks that we label A, B, and C from left to right in Fig. \[fig:pp\]. The energy and HR-resolved pulse profiles presented in Fig. \[fig:HR\] reveal that the pulse peaks are dominated by harder emission. A closer inspection indicates that peak B is softer than the other two peaks and is still present (although much weaker) in the lower energy bands (Fig. \[fig:HR\], lower left bin). This finding is further supported by our study of the HR evolution of the sources throughout the [[*XMM-Newton*]{}]{}observation, presented in Fig. \[fig:HR\_L\] (right panel). This representation clearly shows that the source is harder during high-flux intervals (the peaks of the pulse profile), but also that peak B is distinctly softer than peaks A and C. We argue that these findings indicate a different origin of the three peaks. More specifically, we surmise that the softer and more prominent peak B originates mostly from the fan beam, while the two harder peaks (A and C) are dominated by the polar-beam emission, which is expected to be harder than the fan beam. This is because only the harder incident photons (of the fan beam) will be reflected (i.e., backscattered) off the NS atmosphere, while the softer photons will be absorbed [e.g., @2013ApJ...764...49T; @2015MNRAS.452.1601P]. The analysis of the [[*Swift*]{}]{}data taken throughout the outburst allowed us to also probe the long-term spectral evolution of the source. In the left panel of Fig. \[fig:HR\_L\] we present the HR ratio evolution of the source throughout the 2016 outburst. Unlike the short-term evolution of the source, during which the emission is harder during the high-flux peaks of the pulses (see the data from the [[*XMM-Newton*]{}]{}observation presented in the right panel of Fig. \[fig:HR\_L\]), the emission of SMC X-3 becomes softer in the long term as the luminosity of the source increases (color-coded count rate in Fig. \[fig:HR\_L\]). This behavior is also evident in Fig. \[fig:HR\_swift\], where the HR ratio of the 1-3keV to the 3-10keV band is plotted versus the source count-rate. The source becomes clearly softer as the accretion rate (which corresponds to the observed flux) increases. Long-term observations of numerous XRPs suggests that they can be divided into two groups, based on their long-term spectral variability. Sources in group 1 (e.g., 4U 0115+63) exhibit a positive correlation between source hardness and its flux, while sources in group 2 (e.g., Her X-1 or this source) a negative . It can be argued that the apparent two populations are just the result of observing the XRPs in two different regimes of mass accretion rates. At low accretion rates, deceleration of the in-falling particles in the accretion column occurs primarily via Coulomb interactions, while in the high-rate regime, the mass is decelerated through pressure from the strong radiation field . As the mass accretion rate increases, so does the height of the accretion column, and therefore the reflected fraction (i.e., the polar beam) starts to decrease. As the polar emission is harder than the fan beam, the total registered spectrum stops hardening and even becomes moderately softer [@2015MNRAS.452.1601P]. Observationally, the positive correlation is expected at luminosities below a critical value and the negative above it. This critical value is theoretically predicted at ${\sim}1-7\times10^{37}$erg/s , which is in agreement with the earlier prediction of [@1976MNRAS.175..395B]. Observations of XRPs at different luminosities indeed show that a single source can exhibit the behavior of both groups when monitored above or below this critical value [e.g., @2015MNRAS.452.1601P]. The high-luminosity regime, which is the only regime probed by the Swift data points in Fig. \[fig:HR\_swift\], is essentially the diagonal branch of the hardness-intensity diagrams presented by . The source behavior, presented in Fig. \[fig:HR\_swift\], is consistent with the theoretical predictions for sources accreting above the critical luminosity of ${\sim}10^{37}$erg/s. However, while the previous observations by and [@2015MNRAS.452.1601P] showed the pause in the hardening of the source with increasing luminosity, and (in some sources) the slight decrease in hardness, this drop is clearly demonstrated here thanks to the numerous high-quality observations provided by the [ Swift]{} telescope. Furthermore, our analysis indicates the presence of perhaps a third branch that appears when the source luminosity exceeds ${\sim}4{\times}L_{\rm Edd}$ , in which the source luminosity increases but the emission does not become softer. If this additional branch is indeed real, it may have eluded detection by simply because the sources studied there never reached such high luminosities. It is plausible that the stabilization of the source hardness is a manifestation of the predicted physical limitations imposed on the maximum height of the accretion column [e.g., @2013ApJ...777..115P; @2015MNRAS.454.2539M]. Origin of the soft excess and the optically thin emission --------------------------------------------------------- The phase-averaged spectral continuum of SMC X-3 during the [[*XMM-Newton*]{}]{}observation is described by a combination of a emission in the shape of a hard power-law and a softer thermal component with a temperature of ${\sim}0.24$keV. The nonthermal component is consistent with emission from the accretion column [e.g., @2007ApJ...654..435B], but the origin of the soft thermal emission is less clear. Assuming that the accretion disk is truncated approximately at the magnetosphere, and for an estimated magnetic field of ${\sim}2-3\times10^{12}$G [@2014MNRAS.437.3863K; @2017arXiv170200966T], the maximum effective temperature of the accretion disk would lie in the far-UV range (20-30eV) for ${\dot M}{\sim}1.2\times10^{-8}\,{M_ \odot }/{\rm yr}$, as inferred from the observed luminosity and for $R_{\rm in}=0.5-1\,R_{\rm mag}$. Although we can exclude a typical thin disk as the origin of the emission, an inflated disk that is irradiated from the NS might be the origin of the thermal emission, because such a disk can reach higher temperatures [see, e.g., @2017arXiv170307005C and the discussion below]. The emission is also unlikely to originate in hot plasma on the surface of the NS, as the size of a blackbody-emitting sphere that successfully fits the data is an order of magnitude larger (${\sim}170$km) than the ${\sim12}$km radius of a 1.4${M_ \odot }$ NS. The observed thermal emission could be the result of reprocessing of the primary emission by optically thick material that lies at a distance from the NS and is large enough to partially cover the primary emitting region (i.e., the accretion column). If a significant fraction of the primary emission is reprocessed, then this optically thick region will emit thermal radiation at the observed temperatures. A detailed description of this configuration is presented in [@2000PASJ...52..223E] and [@2004ApJ...614..881H]. This optically thick material is composed of accreted material trapped inside the magnetosphere and also the inside edge of the accretion disk, which at these accretion rates will start to become radiation dominated and geometrically thick . The reprocessing region is expected to have a latitudinal temperature gradient, and its emission can be modeled as a multi-temperature blackbody [e.g., @2017MNRAS.467.1202M], which we modeled using the [XSPEC]{} model [`diskbb`]{}. Nevertheless, we stress that the soft emission most likely does not originate in an accretion disk heated through viscous dissipation, but is rather the result of illumination of an optically thick region at the disk-magnetosphere boundary. Following @2004ApJ...614..881H and assuming that the reprocessing region subtends a solid angle $\Omega$ (as viewed from the source of the primary emission) and has a luminosity given by $L_{\rm soft}$ = ($\Omega$/4$\pi$) $L_{\rm X}$, where $L_{\rm X}$ is the observed luminosity and further assuming that the reprocessed emission is emitted isotropically and follows a thermal distribution (i.e. $L_{\rm soft} = \Omega R^2 \sigma T^4_{\rm soft}$), we can estimate the distance between the primary emitting region and the reprocessing region from the relation $R^2 = L_{\rm X}$/($4\pi \sigma T^4_{\rm soft})$ [@2004ApJ...614..881H]. For our best-fit parameters and for a spectral hardening factor of $\sim1.5-1.7$ [@1995ApJ...445..780S], the relation yields a value of $1.5{\pm}0.3{\times}10^{8}$cm, which is in very good agreement with the estimated magnetospheric radius of $R_{\rm m}=1.6{\pm}0.4{\times}10^{8}$cm that corresponds to the expected (i.e., ${\sim}2-3{\times}10^{12}$G) magnetic field strength of SMC X-3 and for $R_{\rm m} {\sim}2 \times 10^7{\alpha}{\dot{M}_{15}}^{-2/7}{B_{9}}^{4/7}{M_{1.4}}^{-1/7}R_6^{12/7}$ cm [@1977ApJ...217..578G; @1979ApJ...232..259G; @1979ApJ...234..296G; @1992xbfb.work..487G], where ${\alpha}$ is a constant that depends on the geometry of the accretion flow, with ${\alpha}$=0.5 the commonly used value for disk accretion, ${\dot{M}_{15}}$ is the mass accretion rate in units of $10^{15}$g/s (estimated for the observed luminosity of ${\sim}$1.5${\times}10^{38}$erg/s and assuming an efficiency of 0.2), $B_{9}$ is the NS magnetic field in units of $10^{9}$ G, $M_{1.4}$ is the NS mass in units of 1.4 times the solar mass, and $R_6$ is the NS radius in units of $10^6$ cm. Three prominent broad emission lines are featured on top of the broadband continuum. Centered at ${\sim}$0.51keV ${\sim}$0.97keV and ${\sim}$6.64keV, the lines are consistent with emission from hot ionized plasma. The 6.6keV line is most likely due to K-shell fluorescence from highly ionized iron, while the 0.5keV line is most likely the K${\alpha}$ line of the - Ly$\alpha$ (500) ion. The 1keV line is most likely the result of combined Ne K${\alpha}$ and Fe L-shell emission lines. Both the 1keV and 6.6keV lines appear to be broadened, although as we discuss below, the broadening is most likely artificial and is the result of blending of unresolved emission lines from relevant atoms at different ionization states. Both the 1kev and the 6.6keV lines are often detected in the spectra of XRBs and appear to be particularly pronounced in the spectra of X-ray pulsars (e.g., ; ; ; @2014MNRAS.437..316K). While in many XRBs the emission lines (and particularly the iron K${\alpha}$ line) are attributed to reflection of the primary emission from the optically thick accretion disk [e.g., @2010LNP...794...17G and references therein], in the case of X-ray pulsars (and in this source), the emission line origin is more consistent with optically thin emission from rarefied hot plasma that is trapped in the Alfv[é]{}n shell of the highly magnetized NS [e.g., @1978ApJ...223..268B]. The (apparent) line broadening is consistent with microscopic processes, i.e., Compton scattering for a scattering optical depth of 0.5-1 [e.g., equation 30 in @1978ApJ...223..268B], rather than macroscopic motions, such as the rotation of the accretion disk or the surface of the donor star. The measured width of the two lines corresponds to line-of-sight velocities of ${\sim}1.6\times10^{4}$km/sec. For the disk inclination of $\lesssim45^o$, expected for SMC X-3 [@2017arXiv170102336T], this corresponds to a 3D velocity of $\gtrsim2.3\times10^{4}$km/sec. Assuming Keplerian rotation, these velocities would correspond to a distance of $\lesssim350$km from the NS, which is almost an order of magnitude smaller than the estimated magnetospheric radius. Furthermore, the phase-resolved pn spectra and the RGS spectra (discussed below) indicate that the line broadening may be an artifact, resulting from the blending of unresolved thin emission lines at different centroid energies. Short-term evolution of the spectral continuum and emission lines. ------------------------------------------------------------------- The timing analysis of the source emission has revealed that the pulsed emission is dominated by hard photons (Fig. \[fig:phase\_heat\]), and the phase-folded light-curves additionally revealed a non-pulsating component of the total emission at energies below $\sim1$keV (see Fig. \[fig:HR\] and Fig. \[fig:phase\_heat\], bottom). The soft persistent component of the source light-curve appears to coincide with the thermal spectral component. To investigate this further, we divided the pn spectra into 20 equally timed phase intervals, which were modeled independently. We find that while the contribution from the soft thermal component is more or less stable (Fig. \[fig:L\_BB\_PL\]), the contribution of the power-law component strongly correlates with the pulse profile (Fig. \[fig:phace\_spec\_res\]). We note, however, that the use of a simple power law in our spectral fittings fails to model the low-energy roll-over expected from a photon distribution, which is the result of multiple Compton upscattering of soft, thermal seed photons. Nevertheless, adding a parameter for the low-energy turn-over would only add to the uncertainties of the soft component parameters, but would not qualitatively affect these findings. The variation of the power-law component is also illustrated in Fig. \[fig:spec\_res\], in which the phase-resolved spectra were unfolded from the phase-averaged model. This behavior further supports the presence of a large, optically thick reprocessing region located at a considerable distance (i.e., the magnetospheric boundary). As demonstrated in Fig. \[fig:irrad\], the reprocessed emission contributes a very small fraction of the pulsed emission. This finding contradicts the conclusions of regarding another known Be-XRB, XMMU J054134.7-682550. also noted thermal emission in the spectrum of the source (they simultaneously analyzed [[*XMM-Newton*]{}]{}and RXTE data), which they also attributed to reprocessing of the primary emission by optically thick material. However, contrary to this work, concluded that the reprocessed emission also pulsates. They furthermore argued that this is the result of highly beamed emission from the pulsar that is emitted toward the inner disk border. The authors reached this conclusion by noting the (tentative) presence of a sinusoidal-like shape of the pulse profile in the $\lesssim$1keV range. They then concluded that this is mostly due to the contribution of the reprocessed emission. However, the apparent pulsations of the soft component may be due to the fact that did not distinguish between the 0.5-1keV and ${<}0.5$keV energy range, as the total number of counts in their observation did not allow it. It is highly likely that the authors were probing the lower energy part of the accretion column emission and not the reprocessed thermal emission. In our Fig. \[fig:HR\] (left), peak B (and to a lesser extent, peak A) is still detected in the 0.5-1keV range and it is only below ${\sim}0.5$keV (where the thermal component dominates the emission) that the pulses disappear. More importantly, the high quality of our observation allowed us to perform the detailed pulse-resolved analysis that demonstrated the invariance of the thermal component. Furthermore, the fan-beam emission is not expected to be extremely narrow and while optically thick material in the disk/magnetosphere boundary is expected to become illuminated by the primary emission (as both this work and propose), there is no reason to expect that the beam will be preferentially (and entirely) aimed toward the disk inner edge (especially when one accounts for the gravitational bending). Of course, XMMU J054134.7-682550 could be a special case in which such an arrangement took place. However, we conclude that the higher quality of the data available for [SMCX-3]{}most likely allowed for a more detailed analysis that revealed the non-pulsating nature of the thermal emission. This finding highlights the importance of long-exposure high-resolution observations of X-ray pulsars. The emission lines noted in the phase-averaged spectra are also detected in (most of) the phase-resolved spectra. Their apparent strength and centroid energy variation between different phases (when the lines are detected, see Table \[tab:lines\]) lies within the 90% confidence range. However, the lines completely disappear during some phase intervals. While there is no discerning pattern between the presence (or absence) of the lines and the pulse phase, this variability is striking. Furthermore, it is strongly indicated (from the RGS data) that the seemingly broadened lines of the phase-averaged pn spectrum are the result of blending of narrow emission features at different centroid energies. The presence and variability of multiple emission lines indicates optically thin material that is located close to the central source and has a complex shape. As the different parts of this structure are illuminated by the central source, their ionization state varies with phase and position, resulting in emission lines with moderately different energies and different strengths. Such a complex structure is also supported by the predictions of [@roma2004] for accretion of hot plasma onto highly magnetized NSs. Based on our line identification in the RGS data, no significant blueshift can be established for the emission lines listed in Table \[tab:rgs\]. We note that similar lines have been reported in the X-ray spectra of other BeXRBs during outbursts [e.g., SMCX-2, SXP2.16 @2016MNRAS.458L..74L; @vas2017]. For SMCX-2, which was observed by [[*XMM-Newton*]{}]{}during a luminosity of $\sim$1.4[$\times 10^{38}$ erg s$^{-1}$]{} (0.3-12.0 keV band), it has been proposed that these lines could be associated with the circum-source photoionized material in the reprocessing region at the inner disk [@2016MNRAS.458L..74L]. However, if the apparent emission line variability noted in out analysis is real, this would indicate a complex structure of optically thin plasma that lies closer to the source of the primary emission and is illuminated by it at different time intervals. Photoionized material at the boundary of the magnetosphere would not exhibit such rapid variability. SMC X-3 in the context of ULX pulsars -------------------------------------- It is interesting to note that all the emission lines resolved in [SMCX-3]{}have also been identified in ultraluminous X-ray sources [NGC 1313 X-1 and NGC 5408 X-1, @2016Natur.533...64P; @2017AN....338..234P] and ultraluminous soft X-ray sources [NGC55ULX, @2017MNRAS.468.2865P]. Moreover, there are striking similarities between the 1keV emission feature of [SMCX-3]{}and the same feature as described for NGC55ULX within the above studies; “emission peak at 1 keV and absorption-like features on both sides” [@2017MNRAS.468.2865P]. However, we note that for [SMCX-3]{}and other BeXRB systems, to our best knowledge, no significant blueshift of these lines has been measured. This might mean that these lines are produced by the same mechanism both in BeXRB systems (observed at luminosities around the Eddington limit) and in ULXs. However, there is no ultrafast wind/outflow in BeXRBs, as in the case of ULX systems. The findings of this work regarding the continuum and line emission and their temporal variations construct a picture of X-ray pulsars in which the accretion disk is truncated at a large distance from the NS (i.e., the magnetosphere), after which accretion is governed by the magnetic field. As material is trapped by the magnetic field, an optically thick structure is formed inside the pulsar magnetosphere, which (for a range of viewing angles) covers the primary source, partially reprocessing its radiation. The optically thick region is expected to have an angular size on the order of${\text{}}$ a few tens of degrees and to be centered at the latitudinal position of the accretion disk. At higher latitudes and closer to the accretion column, the trapped plasma remains optically thin, producing the observed emission lines. This configuration is schematically represented in Fig. \[fig:cher\] and is in agreement with similar schemes proposed by [@2000PASJ...52..223E] and [@2004ApJ...614..881H]. At high accretion rates, the size of the reprocessing region becomes large enough to reprocess a measurable fraction of the primary emission. This is due to the increase in the accreting material trapped by the pulsar magnetosphere and the increase in the thickness of the accretion disk, which in turn is due to the increase in radiation pressure . This optically thick reprocessing region has been noted in numerous X-ray pulsars (e.g., Cen X-3: @2000ApJ...530..429B; Her X-1: @2002MNRAS.337.1185R; LXP 8.04: Vasilopoulos et al. in prep; SMC X-1: @2005ApJ...633.1064H; SMC X-2: @2016MNRAS.458L..74L). Its presence in sub- or moderately super-Eddington accreting sources becomes especially pertinent in light of recent publications that postulated that at even higher accretion rates (corresponding to luminosities exceeding $10^{39}$erg/s, i.e., in the ULX regime), the entire magnetosphere becomes dominated by optically thick material, which reprocesses all the primary photons of the accretion column [@2017MNRAS.467.1202M], essentially washing out the pulsation information and the spectral characteristics of the accretion column emission. The prediction of this model has also been used to argue that a considerable fraction of ULXs may in fact be accreting, highly magnetized NSs and not black holes [@kolio2017]. In light of these considerations, [SMCX-3]{}is of particular significance, since it stands right at the threshold between sub-Eddington X-ray pulsars and ULXs, and its spectro-temporal characteristics support the notion of a reprocessing region that is already of considerable size. Conclusion ========== We have analyzed the high-quality [[*XMM-Newton*]{}]{}observation of [SMCX-3]{}during its recent outburst. The [[*XMM-Newton*]{}]{}data where complemented with [[*Swift*]{}]{}/XRT observations, which were used to study the long-term behavior of the source. By carrying out a detailed temporal and spectral analysis (including phase-resolved spectroscopy) of the source emission, we found that its behavior and temporal and spectral characteristics fit the theoretical expectations and the previously noted observational traits of accreting highly magnetized NSs at high accretion rates. More specifically, we found indications of a complex emission pattern of the primary pulsed radiation, which most likely involves a combination of a fan-beam emission component directed perpendicularly to the magnetic field axis and a secondary polar-beam component reflected off the NS surface and directed perpendicularly to the primary fan beam (as discussed in, e.g., @1976MNRAS.175..395B; and more recently by @2013ApJ...764...49T). The spectroscopic analysis of the source further reveals optically thick material located at approximately the boundary of the magnetosphere. The reprocessing region has an angular size (as viewed from the NS) that is large enough to reprocess a considerable fraction of the primary beamed emission, which is remitted in the form of a soft thermal-like component that contributes very little to the pulsed emission. These findings are in agreement with previous works on X-ray pulsars [e.g., @2000PASJ...52..223E; @2004ApJ...614..881H], but also with the theoretical predictions for highly super-Eddington accretion onto highly magnetized NSs [e.g., @2017MNRAS.468L..59K; @2017MNRAS.467.1202M; @kolio2017], where it has been argued that this reprocessing region grows to the point that it may reprocess the entire pulsar emission. Acknowledgements ================ The authors thank Maria Petropoulou for valuable advice and stimulating discussion and Olivier Godet for contributing to the RGS analysis. The authors also extend their warm gratitude to Cheryl Woynarski and woyadesign.com for designing the source schematic. Finally we extend our gratitude to the anonymous referee, whose keen observations significantly improved our manuscript. [^1]: [^2]: For more information on pile-up, see http://xmm2.esac.esa.int/docs/documents/CAL-TN-0050-1-0.ps.gz [^3]: XMM-Newton CCF Release Note: XMM-CCF-REL-334 [^4]: <http://www.swift.ac.uk/analysis/xrt/> [^5]: <http://heasarc.gsfc.nasa.gov/ftools/> [^6]: Algorithms taken from the CEPHES special function library by Stephen; Moshier. (<http://www.netlib.org/cephes/>)

ArXiv

--- abstract: | We prove the well-posedness of a system of balance laws inspired by [@HoldenRisebro], describing macroscopically the traffic flow on a multi-lane road network. Motivated by real applications, we allow for the the presence of space discontinuities both in the speed law and in the number of lanes. This allows to describe a number of realistic situations. Existence of solutions follows from compactness results on a sequence of Godunov’s approximations, while ${\mathbf{L^1}}$-stability is obtained by the doubling of variables technique. Some numerical simulations illustrate the behaviour of solutions in sample cases. *2010 Mathematics Subject Classification:* 35L65, 90B20, 82B21. *Keywords:* macroscopic multi-lane traffic flow model on networks; Godunov scheme; well-posedness. author: - Paola Goatin - Elena Rossi bibliography: - 'multilane.bib' title: 'A multi-lane macroscopic traffic flow model for simple networks' --- Introduction ============ Macroscopic traffic flow models consisting of hyperbolic balance laws have been developed in the scientific literature starting from the celebrated Lighthill-Whitham-Richards (LWR) model [@LW1955; @Richards1956]. Despite its simplicity, the LWR model is able to capture the basic features of road traffic dynamics, such as congestion formation and propagation. Nevertheless, it cannot describe many aspects of road traffic complexity. To this end, several improved models accounting for specific flow characteristics have subsequently been introduced: second-order models accounting for a momentum equation (see e.g. [@AwRascle]), multi-population models distinguishing between different classes of vehicles (e.g. [@BenzoniColombo]), etc. In this paper, we are interested in describing carefully the traffic dynamics on road networks with several lanes, allowing for lane change and overtaking. Multi-lane models for vehicular traffic have been proposed in [@CC2006; @HoldenRisebro; @KW1; @KW2]. In the macroscopic setting, these models consist in a system of balance laws in which the transport is expressed by a LWR equation for each lane, and the source term accounts for the lane change rate. In particular, the equations of the system are coupled in the source term only.\ Aiming to describe realistic situations in detail, we allow for the speed laws and the number of lane to change along the road. In the study, for sake of simplicity, we consider the model proposed in [@HoldenRisebro], but more general source terms could be taken into account. We consider an infinite road described by the real line. Let ${{\mathcal{M}}}_\ell\subset {{\mathbb N}}^+$ be the set of indexes of the [*active*]{} lanes on $]-\infty, 0[$, with $M_\ell := {{\left|{{\mathcal{M}}}_\ell\right|}}\geq 1$ its cardinality, and ${{\mathcal{M}}}_r\subset {{\mathbb N}}^+$ be the set of indexes of the [ *active*]{} lanes on $]0,+\infty[$, with $M_r := {{\left|{{\mathcal{M}}}_r\right|}}\geq 1$. Let us consider $M\geq \max\{M_\ell, M_r\}$, its choice depending on the specific situation under study. To cast the problem in a general setting, we extend the road considering the same number of lanes $M$ on the left and on the right of $x=0$. More precisely, we assume that there are $M - M_\ell$ and $M - M_r$ additional empty lanes on $]-\infty,0[$, respectively $]0,+\infty[$. Moreover, we prevent vehicles from passing from the active to the fictive lanes added, see condition  below. In the same way, we can consider multiple separate roads, thus accounting for network nodes. The problem under consideration is then the following: for $x \in {{\mathbb{R}}}$ and $t>0$, the vehicle density $\rho_j=\rho_j(t,x)$ on lane $j$ solves the Cauchy problem $$\label{eq:M} \left\{ \begin{array}{l@{\quad}l} \partial_t \rho_j + \partial_x f_j (x,\rho_j) = S_{j-1}(x,\rho_{j-1}, \rho_j) - S_j (x, \rho_j, \rho_{j+1}) & j=1, \ldots, M, \\ \rho_j(0,x) = \rho_{o,j} (x) & j=1,\ldots, M, \end{array} \right.$$ with $$\begin{aligned} \label{eq:vj} v_j (x,u) = \ & H (x) \, {v_{r,j}} (u) + (1 - H (x)) \, {v_{\ell,j}} (u), \\ \label{eq:flr} f_{\ell,j}(u) = \ & u \, {v_{\ell,j}} (u), \quad f_{r,j} (u) = \ u \, {v_{r,j}} (u), \\ \label{eq:3} f_j(x,u) = \ & u \, v_j (x,u) =\ H (x) f_{r,j} (u) + (1- H (x)) f_{\ell,j} (u),\end{aligned}$$ for $j=1, \ldots, M$, where $H$ is the Heaviside function. The velocities $v_{d,j}$, for $d= \ell, r$ and $j=1,\ldots, M$, are strictly decreasing positive functions such that $v_{d,j} (1) = 0$. We assume that each map $f_{d,j} (u) = u \, v_{d,j} (u)$ admits a unique global maximum in the interval $[0,1]$, attained at $u={\vartheta}^j_d$. We set $$\label{eq:theta} {\vartheta}^j (x) = H(x) \, {\vartheta}^j_r+ (1- H(x))\, {\vartheta}^j_\ell.$$ Moreover, we set $\rho_{o,j}:{\mathbb R}\to [0,1]$ for $j=1,\ldots, M,$ and $$\begin{aligned} \label{eq:idM} &\rho_{o,j}(x) = 0 \quad \mbox{ for } x\in\, ]-\infty,0[ \mbox{ and } j\not\in{{\mathcal{M}}}_\ell,\\ \label{eq:idM1} &\rho_{o,j}(x) = 1 \quad \mbox{ for } x\in\, ]0,+\infty[ \mbox{ and } j\not\in{{\mathcal{M}}}_r.\end{aligned}$$ Concerning the source terms, accounting for the flow rate across lanes, we define, as in [@HoldenRisebro], $$\label{eq:2} \begin{aligned} S_{d,j} (\rho_j, \rho_{j+1}) = \ & \left[ \left( v_{d,j+1} (\rho_{j+1}) - v_{d,j} (\rho_j) \right) ^+ \rho_j - \left( v_{d,j+1} (\rho_{j+1}) - v_{d,j} (\rho_j) \right) ^- \rho_{j+1} \right] \\ = \ & \left( v_{d,j+1} (\rho_{j+1}) - v_{d,j} (\rho_j) \right) \left\{ \begin{array}{l@{\quad}l} \rho_j & v_{d,j+1} (\rho_{j+1}) \geq v_{d,j} (\rho_j) , \\ \rho_{j+1} & v_{d,j+1} (\rho_{j+1}) < v_{d,j} (\rho_j) , \end{array} \right. \end{aligned}$$ for $d=\ell,r$ and $j=1, \ldots, M-1$, where $(a)^+=\max\left\{a,0\right\}$ and $a^-=-\min\{a,0\}$. To account for separate lanes, such as different roads or fictive lanes, we set $$\label{eq:sghost} S_{d,j_d} (u,w) = 0 \qquad \mbox{ for some } j_d \in \left\{1,\ldots,M-1\right\},~ d=\ell,r.$$ The functions appearing in the source term are then defined as follows $$\begin{aligned} \label{eq:4} S_j (x,u,w) = \ & H (x) \, {S_{r,j}} (u,w) + (1 - H (x)) \, {S_{\ell,j}} (u,w) & \mbox{ for } j= \ & 1, \ldots, M-1, \\ \label{eq:sbordo} S_0 (x,u,w) = \ & S_M (x,u,w) = 0. &\end{aligned}$$ For the sake of shortness, introduce the notation ${\boldsymbol{\rho}}= (\rho_1, \ldots, \rho_M)$, so that the initial data associated to problem –– read ${\boldsymbol{\rho}}(0,x) = {\boldsymbol{\rho}}_o (x)$. For simplicity, and with slight abuse of notation, we consider ${\boldsymbol{\rho}}={\boldsymbol{\rho}}(t,x)$ for $t>0$, $x\in\, {{\mathbb{R}}}$. However, we will show that, by , and , there holds $\rho_j(t,x) = 0$ for all $t>0$, $x\in\, ]-\infty,0[$ and $j\not\in{{\mathcal{M}}}_\ell$, respectively $\rho_j(t,x) = 1$ for all $t>0$, $x\in\, ]0,+\infty[$ and $j\not\in{{\mathcal{M}}}_r$. Following [@KRT2003 Definition 5.1], see also [@KRT2002 Definition 2.1 and Formula (5.8)] and [@HoldenRisebroBook2015 § 8.3], we recall the definition of *weak entropy solution* for ––. \[def:sol\] A map ${\boldsymbol{\rho}}= (\rho_1,\ldots,\rho_M) \in {\mathbf{L^\infty}} ([0,T]\times {{\mathbb{R}}}; [0,1]^M) $ is a *weak entropy solution* to the initial value problem  if 1. \[it:weak\] for any ${\varphi}\in {\mathbf{C_c^{1}}} ([0,T[ \times {{\mathbb{R}}}; {{\mathbb{R}}})$ and for all $j=1,\ldots, M$, $$\begin{aligned} \int_0^T\int_{{{\mathbb{R}}}} \left(\rho_j \, \partial_t {\varphi}+ f_j (x,\rho_j) \, \partial_x {\varphi}+ \left(S_{j-1} (x,\rho_{j-1}, \rho_j) - S_j (x, \rho_j, \rho_{j+1})\right) {\varphi}\right) {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}} & \\ + \int_{{\mathbb{R}}}\rho_{o,j} \, {\varphi}(0,x) {\mathinner{\mathrm{d}{x}}} & = 0. \end{aligned}$$ 2. \[it:entropy\] for any ${\varphi}\in {\mathbf{C_c^{1}}} ([0,T[ \times {{\mathbb{R}}}; {{\mathbb{R}}}^+)$, for any $c\in[0,1]$ and for all $j=1,\ldots, M$ $$\begin{aligned} \int_0^T\int_{{{\mathbb{R}}}} \left\{{{\left|\rho_j-c\right|}} \, \partial_t {\varphi}+ \operatorname{sgn}(\rho_j - c) \left(f_j (x,\rho_j) - f_j (x, c)\right) \partial_x {\varphi}\right. & \\ \left. \qquad + \operatorname{sgn}(\rho_j - c) \left(S_{j-1} (x,\rho_{j-1}, \rho_j) - S_j (x, \rho_j, \rho_{j+1})\right) {\varphi}\right\} {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}}& \\ + \int_0^T{{\left|f_{r,j} (c) - f_{\ell,j} (c)\right|}} \, {\varphi}(t,0) {\mathinner{\mathrm{d}{t}}} + \int_{{\mathbb{R}}}{{\left|\rho_{o,j} -c\right|}} {\varphi}(0,x) {\mathinner{\mathrm{d}{x}}} & \geq\ 0. \end{aligned}$$ The rest of the paper is organised as follows. In Section \[sec:existence\] we construct a sequence of approximate solutions based on Godunov finite volume scheme and we prove its convergence towards a solution of . We then provide a ${\mathbf{L^1}}$-stability estimate with respect to the initial data, which implies the uniqueness of solutions. Specific situations and the corresponding numerical simulations are discussed in Section \[sec:num\]. Well-posedness {#sec:existence} ============== We define the map ${\boldsymbol{v}}: [0,1] \to {{\mathbb{R}}}^{2M}$ by setting ${\boldsymbol{v}}_j = {v_{\ell,j}}$ and $ {\boldsymbol{v}}_{M+j}= {v_{r,j}}$, for $j=1, \ldots, M$. Moreover we define $$\label{eq:normav} \begin{aligned} V_{\max} = \ & {{\left\|{\boldsymbol{v}}\right\|}}_{{\mathbf{C^{0}}}([0,1];{{\mathbb{R}}}^{2 M})} = \max_{\substack{j=1,\dots,M\\d=\ell,r}} {{\left\|v_{d,j}\right\|}}_{{\mathbf{L^\infty}} ([0,1];{{\mathbb{R}}})}, \\ \mathcal{V} = \ & {{\left\|{\boldsymbol{v}}\right\|}}_{{\mathbf{C^{1}}}([0,1];{{\mathbb{R}}}^{2M})} = \max_{\substack{j=1,\dots,M\\d=\ell,r}} {{\left\|v_{d,j}\right\|}}_{{\mathbf{L^\infty}} ([0,1];{{\mathbb{R}}})} + \max_{\substack{j=1,\dots,M\\d=\ell,r}} {{\left\|v'_{d,j}\right\|}}_{{\mathbf{L^\infty}} ([0,1];{{\mathbb{R}}})}. \end{aligned}$$ We introduce the following quantity, which corresponds to the ${\mathbf{L^1}}$–norm of the vector ${\boldsymbol{\rho}}$ computed on *active* lanes: $$\label{eq:norma1} {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}}\|\hspace{-1pt}|}} = \sum_{j\in{{\mathcal{M}}}_\ell} {{\left\|\rho_j\right\|}}_{{\mathbf{L^1}} (]-\infty,0[)} + \sum_{j\in{{\mathcal{M}}}_r} {{\left\|\rho_j\right\|}}_{{\mathbf{L^1}} (]0,+\infty[)}.$$ Introduce a uniform space mesh of width ${{\Delta x}}$ and a time step ${{\Delta t}}$, subject to a CFL condition, to be detailed later on. For $k\in {{\mathbb{Z}}}$ set $$\begin{aligned} x_k = \ & \left(k+\frac12\right) {{\Delta x}}, & x_{k - 1/2} = \ & k {{\Delta x}}, \end{aligned}$$ where $x_k$ denotes the centre of the cell, while $x_{k\pm 1/2}$ its interfaces. Observe that $x=0$ corresponds to $x_{-1/2}$, so that non negative integers denote the cells on the positive part of the $x$-axis. Set $N_T= \lfloor T/{{\Delta t}}\rfloor$ and let $t^n= n \, {{\Delta t}}$, for $n=0, \ldots, N_T$, be the time mesh. Set $\lambda = {{\Delta t}}/ {{\Delta x}}$. Approximate the initial data in the following way: for $j=1,\ldots,M$, for $k\in {{\mathbb{Z}}}$ $$\rho_{j,k}^0 = \frac1{{\Delta x}}\int_{x_{k-1/2}}^{x_{k+1/2}}\rho_{o,j} (x) {\mathinner{\mathrm{d}{x}}}.$$ Define a piece-wise constant solution ${\boldsymbol{\rho}}_\Delta$ to  as, for $j=1,\ldots,M$, $$\label{eq:6} \rho_{j,\Delta} (t,x) = \rho_{j,k}^n \quad \mbox{ for } \quad \left\{ \begin{array}{l} t \in [t^n, t^{n+1}[,\\ x \in [x_{k-1/2}, x_{k+1/2}[, \end{array} \right. \quad \mbox{ where } \quad \begin{array}{l} n= 0, \ldots, N_t -1,\\ k \in {{\mathbb{Z}}}, \end{array}$$ through a Godunov type scheme (see [@AJVG2004]) together with operator splitting, to account for the source terms: \[alg:1\] $$\begin{aligned} \label{eq:fx} & F_j (x, u, w) = \left\{ \begin{array}{l@{\quad \mbox{ if }}l} \min\left\{f_j \left(x, \min\{u,{\vartheta}^j (x)\}\right), f_j \left(x, \max\{w, {\vartheta}^j (x)\}\right) \right\} & x\neq 0, \\ \min\left\{f_{\ell,j} \left(\min\{u, {\vartheta}^j_\ell\}\right), f_{r,j} \left(\max\{w, {\vartheta}^j_r \}\right) \right\} & x=0, \end{array} \right. \\ &\texttt{for } n=0,\ldots, N_T-1 \nonumber \\ & \quad \texttt{for } j=1,\ldots, M, \texttt{for } k \in {{\mathbb{Z}}}\nonumber \\ \label{eq:scheme} & \quad \quad \quad \rho_{j,k}^{n+1/2} = {\rho^{n}_{j,k}} - \lambda \left[ F_j (x_{k+1/2}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - F_j (x_{k-1/2}, {\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right] \\ \nonumber & \quad \texttt{end} \\ & \quad \texttt{for } j=1,\ldots, M, \texttt{for } k \in {{\mathbb{Z}}}\nonumber \\ \label{eq:scheme2} &\quad\quad\quad \rho_{j,k}^{n+1} = \rho_{j,k}^{n+1/2} + {{\Delta t}}\, S_{j-1} (x_k, \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - {{\Delta t}}\, S_j (x_k,\rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k}) \\ \nonumber & \quad \texttt{end} \\ \nonumber &\texttt{end} \end{aligned}$$ \[rem:resta0\] Observe that, under hypotheses –, for all $n=0, \dots, N_T-1$ and $k\leq -1$ (corresponding to $x<0$), it holds ${\rho^{n}_{j,k}}=0$ for all $j\not\in{{\mathcal{M}}}_\ell$. In particular, no wave can move backward into the segment $]-\infty,0[$ for $j\not\in{{\mathcal{M}}}_\ell$. Similarly, for all $n=0, \dots, N_T-1$ and $k\geq 0$ (corresponding to $x>0$), it holds ${\rho^{n}_{j,k}}=1$ for all $j\not\in{{\mathcal{M}}}_r$. In particular, no wave can move forward into the segment $]0,+\infty[$ for $j\not\in{{\mathcal{M}}}_r$. Positivity and upper bound {#sec:pos} -------------------------- We prove that, under a suitable CFL condition, if the initial data take values in the interval $[0,1]$, then also the approximate solution constructed via Algorithm \[alg:1\] attains values in the same interval $[0,1]$. \[lem:pos\] Let ${\boldsymbol{\rho}}_o \in {\mathbf{L^\infty}} ({{\mathbb{R}}}; [0,1]^M)$. Assume that $$\label{eq:cfl} \lambda \, \mathcal{V}\leq \frac12,$$ with $\mathcal{V}$ as in . Then, for all $t>0$ and $x \in {{\mathbb{R}}}$, the piece-wise constant approximate solution ${\boldsymbol{\rho}}_\Delta$ constructed through Algorithm \[alg:1\] is such that $0 \leq \rho_{j, \Delta} (t,x) \leq 1$, for all $j=1,\dots, M$. By induction, assume that $0 \leq {\rho^{n}_{j,k}} \leq 1$ for all $k \in {{\mathbb{Z}}}$ and $j=1,\dots, M$. Consider : it is well known that, for a Godunov type scheme with discontinuous flux function, it holds $0 \leq \rho_{j,k}^{n+1/2} \leq 1$, see [@AJVG2004 Lemma 4.3]. We now focus on the remaining step, involving the source term. In particular, fix $k\geq 0$, corresponding to $x>0$, the other case being entirely similar. Exploiting , equation  reads $$\rho_{j,k}^{n+1} = \rho_{j,k}^{n+1/2} + {{\Delta t}}\, {S_{r,j-1}} ( \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - {{\Delta t}}\, {S_{r,j}} (\rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k}).$$ To improve readability, in what follows we omit the index $n+1/2$. Moreover, we take into account a *complete* case, in which the source term contains the contributions from both the previous and the subsequent lane. Without loss of generality, we take $j=2$ and we assume both ${S_{r,1}} ( \rho_{1,k}, \rho_{2,k}) \not= 0$ and ${S_{r,2}} (\rho_{2,k}, \rho_{3,k}) \not= 0$. By  and  we obtain $$\begin{aligned} \label{eq:7} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\, {S_{r,1}} ( \rho_{1,k}, \rho_{2,k}) - {{\Delta t}}\, {S_{r,2}} (\rho_{2,k}, \rho_{3,k}) \\ \nonumber = \ & \rho_{2,k} + {{\Delta t}}\left[ \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) ^+ \rho_{1,k} - \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) ^- \rho_{2,k}\right] \\ \nonumber &\qquad \!\! - {{\Delta t}}\left[ \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) ^+ \rho_{2,k} - \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) ^- \rho_{3,k} \right]. \end{aligned}$$ There are four possibilities: $\boldsymbol{v_{r,2} (\rho_{2,k}) \geq v_{r,1} (\rho_{1,k})}$ $\boldsymbol{v_{r,2} (\rho_{2,k}) < v_{r,1} (\rho_{1,k})}$ ---------------------------------------------------------------- --------------------------------------------------------------- ------------------------------------------------------------ $\boldsymbol{v_{r,3} (\rho_{3,k}) \geq v_{r,2} (\rho_{2,k}) }$ Case \[item:++\] Case \[item:-+\] $\boldsymbol{v_{r,3} (\rho_{3,k}) < v_{r,2} (\rho_{2,k}) }$ Case \[item:+-\] Case \[item:–\] We analyse them in details. 1. \[item:++\] Equation  reads $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{1,k} - {{\Delta t}}\left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{2,k} \\ \geq \ & \rho_{2,k} - {{\Delta t}}\left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{2,k} \\ \geq \ & \rho_{2,k} \left( 1 - {{\Delta t}}\, v_{r,3} (\rho_{3,k}) \right) \\ \geq \ & \rho_{2,k} \left( 1 - {{\Delta t}}\, V_{\max}\right) \\ \geq \ & 0, \end{aligned}$$ by the CFL condition , since ${{\Delta x}}<1$. Moreover, since $v_{r,2}(1) =0$ and $\rho_{2,k}\leq 1$, $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{1,k} - {{\Delta t}}\left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{2,k} \\ \leq \ & \rho_{2,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{1,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{2,k} \\ = \ & \rho_{2,k} + {{\Delta t}}\, {v_{r,2}}' (\sigma) \left(\rho_{2,k} - 1\right)\left(\rho_{1,k} + \rho_{2,k}\right) \\ = \ & \rho_{2,k} \left( 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{2,k}\right) \right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{2,k}\right) \\ \leq \ & 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{2,k}\right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{2,k}\right) \\ = \ & 1, \end{aligned}$$ with $\sigma \in \,]\rho_{2,k},1[$ and we exploit the fact that $1 + {{\Delta t}}\,{v_{r,2}}' \left(\rho_{1,k} + \rho_{2,k}\right) \geq 0$, due to the CFL condition . 2. \[item:+-\] By equation  and the hypotheses on the signs, it follows immediately that $$\rho_{2,k}^{n+1} = \rho_{2,k} + {{\Delta t}}\left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{1,k} - {{\Delta t}}\left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{3,k} \geq 0.$$ Moreover, since $v_{d,2} (1)=0$ and $\rho_{2,k}\leq 1$, we get $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{1,k} - {{\Delta t}}\left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{3,k} \\ \leq \ & \rho_{2,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{1,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{3,k} \\ = \ & \rho_{2,k} + {{\Delta t}}\, {v_{r,2}}' (\sigma) \left(\rho_{2,k} - 1\right)\left(\rho_{1,k} + \rho_{3,k}\right) \\ = \ & \rho_{2,k} \left( 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{3,k}\right) \right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{3,k}\right) \\ \leq \ & 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{3,k}\right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{3,k}\right) \\ = \ & 1, \end{aligned}$$ where $\sigma \in\, ]\rho_{2,k},1[$. 3. \[item:-+\] By equation  and the hypotheses on the sign, we get $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left[ \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{2,k} - \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{2,k} \right] \\ \geq \ & \rho_{2,k} \left(1 - {{\Delta t}}\, v_{r,1} (\rho_{1,k}) - {{\Delta t}}\, v_{r,3} (\rho_{3,k}) \right) \\ \geq & \rho_{2,k} (1 - 2 \, {{\Delta t}}\, V_{\max}) \\ \geq \ & 0, \end{aligned}$$ by the CFL condition , since ${{\Delta x}}<1$. Moreover, since $v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) < 0$ and $v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k})\geq 0$, we get $$\rho_{2,k}^{n+1} = \rho_{2,k} + {{\Delta t}}\left[ \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{2,k} - \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{2,k} \right] \leq \rho_{2,k} \leq 1.$$ 4. \[item:–\] By equation  and the CFL condition  we obtain $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left[ \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{2,k} - \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{3,k} \right] \\ \geq \ & \rho_{2,k} + {{\Delta t}}\left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{2,k} \\ \geq \ & \rho_{2,k} \left( 1 - {{\Delta t}}\, v_{r,1} (\rho_{1,k})\right) \\ \geq \ & \rho_{2,k} (1 - {{\Delta t}}\, V_{\max}) \\ \geq \ & 0. \end{aligned}$$ Moreover, since ${v_{r,2}}(1)=0$ and $\rho_{2,k}\leq 1$, $$\begin{aligned} \rho_{2,k}^{n+1} = \ & \rho_{2,k} + {{\Delta t}}\left[ \left( v_{r,2} (\rho_{2,k}) - v_{r,1} (\rho_{1,k}) \right) \rho_{2,k} - \left( v_{r,3} (\rho_{3,k}) - v_{r,2} (\rho_{2,k}) \right) \rho_{3,k} \right] \\ \leq \ & \rho_{2,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{2,k} + {{\Delta t}}\, v_{r,2} (\rho_{2,k}) \, \rho_{3,k} \\ = \ & \rho_{2,k} + {{\Delta t}}\, {v_{r,2}}' (\sigma) \left(\rho_{2,k} - 1\right)\left(\rho_{2,k} + \rho_{3,k}\right) \\ = \ & \rho_{2,k} \left( 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{2,k} + \rho_{3,k}\right) \right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{1,k} + \rho_{3,k}\right) \\ \leq \ & 1 + {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{2,k} + \rho_{3,k}\right) - {{\Delta t}}\,{v_{r,2}}' (\sigma) \left(\rho_{2,k} + \rho_{3,k}\right) \\ = \ & 1, \end{aligned}$$ where $\sigma \in \,]\rho_{2,k},1[$. Hence, we conclude that $\rho_{j,k}^{n+1} \in[0,1]$ for all $j=1,\dots,M$ and $k \in {{\mathbb{Z}}}$. L1–bound {#sec:l1} -------- The following Lemma shows that, if the initial datum ${\boldsymbol{\rho}}_o$ satisfies ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$, i.e. it is in ${\mathbf{L^1}}$ on the *active* lanes, the same holds for the corresponding solution. Moreover, the ${\mathbf{L^1}}$–norm is constant, thus the total number of vehicles is preserved over time. \[lem:l1\] Let ${\boldsymbol{\rho}}_o \in ({\mathbf{L^1}} \cap {\mathbf{L^\infty}}) ({{\mathbb{R}}}; [0,1]^M)$. Let ${\boldsymbol{\rho}}_o \in {\mathbf{L^\infty}} ({{\mathbb{R}}}; [0,1]^M)$, with ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$ . Under the CFL condition , the piece-wise approximate solution ${\boldsymbol{\rho}}_\Delta$ constructed through Algorithm \[alg:1\] is such that, for all $t>0$, $$\label{eq:l1} {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_\Delta (t)}\|\hspace{-1pt}|}}= {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}.$$ By induction, assume that  holds for $t^n = n \, {{\Delta t}}$. The Godunov type scheme  is conservative, see [@AJVG2004], hence $${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}^{n+1/2}}\|\hspace{-1pt}|}} = {{\Delta x}}\sum_{j\in{{\mathcal{M}}}_\ell}\sum_{k\leq -1} {{\left|\rho^{n+1/2}_{j,k}\right|}} + {{\Delta x}}\!\!\!\sum_{j\in{{\mathcal{M}}}_r} \sum _{k\geq 0}{{\left|\rho^{n+1/2}_{j,k}\right|}} = |\hspace{-1pt}{{\left\|{\boldsymbol{\rho}}_o\right\|}}\hspace{-1pt}|.$$ Pass now to : by the positivity of ${\boldsymbol{\rho}}_\Delta$, see Lemma \[lem:pos\], and the assumptions on the source terms , it follows immediately that $ {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}^{n+1}}\|\hspace{-1pt}|}}= {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}^{n+1/2}}\|\hspace{-1pt}|}}= {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}$. L1 continuity in time {#sec:l1cont-time} --------------------- Following the idea introduced in [@KRT2002 Lemma 3.3], we now prove the ${\mathbf{L^1}}$-continuity in time of the numerical approximation, constructed through Algorithm \[alg:1\]. The result is of key importance in the subsequent analysis. \[prop:l1cont-time\] Let ${\boldsymbol{\rho}}_o \in {\mathbf{BV}}({{\mathbb{R}}}; [0,1]^M)$ with ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$ . Assume that the CFL condition  holds. Then, for $n=0, \ldots, N_T-1$ $$\label{eq:8} {{\Delta x}}\, \sum_{j=1}^M\sum_{k \in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1} - \rho_{j,k}^n\right|}} \leq 2 \, e^{4 \, \mathcal{V}\, T} \, {{\Delta t}}\Bigl( \mathcal{V} \, \sum_{j=1}^M {\mathinner{\rm TV}}(\rho^0_j) + M \, V_{\max} + 2 \, V_{\max} \, {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}} \Bigr),$$ with $V_{\max}$ and $\mathcal{V}$ as in . \[rem:somma\] Observe that, by Remark \[rem:resta0\], the sums appearing in  are actually sums over the *active* lanes only, the terms corresponding to fictive lanes being equal to 0. For example $$\sum_{j=1}^M\sum_{k \in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1} - \rho_{j,k}^n\right|}} = \sum_{j\in\mathcal{M}_\ell}\sum_{k \leq -1} {{\left|\rho_{j,k}^{n+1} - \rho_{j,k}^n\right|}} + \sum_{j\in \mathcal{M}_r}\sum_{k \geq 0} {{\left|\rho_{j,k}^{n+1} - \rho_{j,k}^n\right|}}.$$ However, for the sake of shortness, we keep the first notation throughout the proof. Fix $k\in {{\mathbb{Z}}}$ and $j \in \{1,\dots,M\}$. By  we have: $$\begin{aligned} \nonumber \rho_{j,k}^{n+1} - {\rho^{n}_{j,k}} = \ & \rho_{j,k}^{n+1/2} - \rho_{j,k}^{n-1/2} \\ \label{eq:10} & +{{\Delta t}}\, S_{j-1}(x_k, \rho_{j-1,k}^{n+1/2}, \rho_{j,k}^{n+1/2}) - {{\Delta t}}\, S_{j-1}(x_k, \rho_{j-1,k}^{n-1/2}, \rho_{j,k}^{n-1/2}) \\ \nonumber & -{{\Delta t}}\, \, S_{j}(x_k, \rho_{j,k}^{n+1/2}, \rho_{j+1,k}^{n+1/2}) +{{\Delta t}}\, \, S_{j}(x_k, \rho_{j,k}^{n-1/2}, \rho_{j+1,k}^{n-1/2}). \end{aligned}$$ Observe that, by , terms of type $ {{\Delta t}}\left( S_{j}(x_k, \rho_{j,k}^{n+1/2}, \rho_{j+1,k}^{n+1/2}) - S_{j}(x_k, \rho_{j,k}^{n-1/2}, \rho_{j+1,k}^{n-1/2})\right)$ are non zero for $j=1,\ldots,M-1$. For $x\in{{\mathbb{R}}}$ and $j=1,\dots, M-1$, the function $(u,w) \mapsto S_{j}(x, u, w)$ defined in , together with  and , is Lipschitz in both variables, with Lipschitz constant $$K_{j} = \max\left\{ {{\left\|v'_{j} (x)\right\|}}_{{\mathbf{L^\infty}} ([0,1])} + v_{j+1} (x, 0), \, {{\left\|v'_{j+1} (x)\right\|}}_{{\mathbf{L^\infty}} ([0,1])} + v_{j} (x,0) \right\} \leq \mathcal{V},$$ with $\mathcal{V}$ as in . Hence, for $j=1,\dots,M-1$, we get $$\begin{aligned} & {{\Delta t}}\, {{\left| S_{j}(x_k, \rho_{j,k}^{n+1/2}, \rho_{j+1,k}^{n+1/2}) - S_{j}(x_k, \rho_{j,k}^{n-1/2}, \rho_{j+1,k}^{n-1/2})\right|}} \\ \leq\ & {{\Delta t}}\, \mathcal{V} \, \left( {{\left|\rho_{j,k}^{n+1/2} - \rho_{j,k}^{n-1/2}\right|}} + {{\left|\rho_{j+1,k}^{n+1/2} - \rho_{j+1,k}^{n-1/2}\right|}} \right). \end{aligned}$$ By , taking into account also , we conclude $$\begin{aligned} \nonumber & \sum_{j=1}^M {{\left|\rho_{j,k}^{n+1} - {\rho^{n}_{j,k}}\right|}} \\ \nonumber \leq \ & \sum_{j=1}^M {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}} \\ \nonumber & + 2 \, \mathcal{V} \, {{\Delta t}}\left( {{\left|\rho_{1,k}^{n+1/2} - \rho^{n-1/2}_{1,k}\right|}} + 2 \sum_{j=2}^{M-1} {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}} + {{\left|\rho_{M,k}^{n+1/2} - \rho^{n-1/2}_{M,k}\right|}} \right) \\ \nonumber \leq \ & (1+ 4 \, \mathcal{V}\, {{\Delta t}}) \sum_{j=1}^M {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}} \\ \label{eq:9} \leq \ & e^{4 \, \mathcal{V}\, {{\Delta t}}} \sum_{j=1}^M {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}}. \end{aligned}$$ Exploit now : we have, for fixed $j \in \{1,\dots,M\}$ and $k \in {{\mathbb{Z}}}$, $$\begin{aligned} \nonumber \rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k} = \ & {\rho^{n}_{j,k}} - \rho_{j,k}^{n-1} - \lambda \left[ F_j (x_{k+1/2}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - F_j (x_{k-1/2}, {\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right] \\ \label{eq:12} & + \lambda \left[ F_j (x_{k+1/2}, \rho^{n-1}_{j,k}, \rho^{n-1}_{j,k+1}) - F_j (x_{k-1/2}, \rho^{n-1}_{j,k-1}, \rho^{n-1}_{j,k}) \right]. \end{aligned}$$ We closely follow the proof of [@KRT2002 Lemma 3.3]. In  add and subtract $\lambda \, F_j (x_{k+1/2}, {\rho^{n}_{j,k}}, \rho^{n-1}_{j,k+1})$ and $ \lambda \, F_j (x_{k-1/2}, {\rho^{n}_{j,k-1}}, \rho^{n-1}_{j,k})$ and, setting $$\begin{aligned} \label{eq:alfa} \alpha_{j,k}^n = \ & \left\{ \begin{array}{l@{\quad \mbox{ if }}l} - \lambda \, \dfrac{F_j (x_{k-1/2}, {\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) - F_j (x_{k-1/2}, {\rho^{n}_{j,k-1}}, \rho^{n-1}_{j,k})}{{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}} & {\rho^{n}_{j,k}} \neq \rho^{n-1}_{j,k}, \\ 0 & {\rho^{n}_{j,k}} = \rho^{n-1}_{j,k}, \end{array} \right. \\ \label{eq:beta} \beta_{j,k}^n = \ & \left\{ \begin{array}{l@{\qquad \mbox{ if }}l} \lambda \, \dfrac{F_j (x_{k+1/2}, {\rho^{n}_{j,k}}, \rho^{n-1}_{j,k+1}) - F_j (x_{k+1/2}, \rho^{n-1}_{j,k}, \rho^{n-1}_{j,k+1})}{{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}} & {\rho^{n}_{j,k}} \neq \rho^{n-1}_{j,k}, \\ 0 & {\rho^{n}_{j,k}} = \rho^{n-1}_{j,k}, \end{array} \right. \end{aligned}$$ rearrange the resulting expression to obtain $$\label{eq:11} \begin{aligned} \rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k} = \ & \left({\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}\right) \left( 1 - \alpha_{j,k}^n - \beta_{j,k}^n\right) \\ & + \alpha_{j,k+1}^n \left({\rho^{n}_{j,k+1}} - \rho^{n-1}_{j,k+1}\right) + \beta_{j,k-1}^n \left({\rho^{n}_{j,k-1}} - \rho^{n-1}_{j,k-1}\right). \end{aligned}$$ Since the numerical flux $F_j$ defined in  is non decreasing in the second variable and non increasing in the third, we get $\alpha_{j,k}^n, \, \beta_{j,k}^n \geq 0$ for all $j=1,\dots,M$ and $k \in {{\mathbb{Z}}}$. Moreover, $F_j(x, \cdot, \cdot)$ is Lipschitz in both arguments, for $x\in {{\mathbb{R}}}$, with Lipschitz constant bounded by $\mathcal{V}$ as in . Therefore, $$\begin{aligned} \beta_{j,k}^n = \ & \frac{\lambda}{{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}} \left( F_j (x_{k+1/2}, {\rho^{n}_{j,k}}, \rho^{n-1}_{j,k+1}) - F_j (x_{k+1/2}, \rho^{n-1}_{j,k}, \rho^{n-1}_{j,k+1}) \right) \\ \leq \ & \frac{\lambda}{{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}} \, \mathcal{V} \, {{\left|{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}\right|}} = \lambda \, \mathcal{V} \leq \frac12, \end{aligned}$$ by the CFL condition . A similar argument applies to $\alpha^n_{j,k}$. As a consequence, $1 - \alpha^n_{j,k} - \beta^n_{j,k} \geq 0$, thus all the coefficients appearing in  are positive and so $$\begin{aligned} \nonumber \sum_{k\in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}} \leq \ & \sum_{k\in {{\mathbb{Z}}}} {{\left|{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}\right|}} \left( 1 - \alpha_{j,k}^n - \beta_{j,k}^n\right) \\ \nonumber & + \sum_{k\in {{\mathbb{Z}}}} \alpha_{j,k+1}^n \, {{\left|{\rho^{n}_{j,k+1}} - \rho^{n-1}_{j,k+1}\right|}} + \sum_{k\in {{\mathbb{Z}}}} \beta_{j,k-1}^n \, {{\left|{\rho^{n}_{j,k-1}} - \rho^{n-1}_{j,k-1}\right|}} \\ \label{eq:13} = \ & \sum_{k\in {{\mathbb{Z}}}} {{\left|{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}\right|}}. \end{aligned}$$ Collecting together  and  leads to $$\sum_{j=1}^M \sum_{k\in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1} - {\rho^{n}_{j,k}}\right|}} \leq e^{4 \, \mathcal{V}\, {{\Delta t}}} \sum_{j=1}^M \sum_{k\in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1/2} - \rho^{n-1/2}_{j,k}\right|}} \leq e^{4 \, \mathcal{V}\, {{\Delta t}}} \sum_{j=1}^M \sum_{k\in {{\mathbb{Z}}}} {{\left|{\rho^{n}_{j,k}} - \rho^{n-1}_{j,k}\right|}},$$ which applied recursively yields $$\label{eq:14} {{\Delta x}}\sum_{j=1}^M \sum_{k\in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{n+1} - {\rho^{n}_{j,k}}\right|}} \leq e^{4\, \mathcal{V} \, T } {{\Delta x}}\sum_{j=1}^M \sum_{k\in {{\mathbb{Z}}}} {{\left|\rho_{j,k}^{1} - \rho^0_{j,k}\right|}},$$ where we also multiplied both sides of the inequality by ${{\Delta x}}$. Using  and , compute $$\begin{aligned} \nonumber \rho^{1}_{j,k} - \rho^0_{j,k} = \ & \rho^{1/2}_{j,k} - \rho^0_{j,k} + {{\Delta t}}\, S_{j-1} (x_{k}, \rho^{1/2}_{j-1,k}, \rho^{1/2}_{j,k}) - {{\Delta t}}\, S_j (x_k, \rho^{1/2}_{j,k}, \rho^{1/2}_{j+1,k}) \\ \label{eq:A} = \ & -\lambda \, \left[ F_j (x_{k+1/2}, \rho^0_{j,k}, \rho^0_{j,k+1}) - F_j (x_{k-1/2}, \rho^0_{j,k-1}, \rho^0_{j,k}) \right] \\ \label{eq:B} & + {{\Delta t}}\, S_{j-1} (x_{k}, \rho^{1/2}_{j-1,k}, \rho^{1/2}_{j,k}) - {{\Delta t}}\, S_j (x_k, \rho^{1/2}_{j,k}, \rho^{1/2}_{j+1,k}). \end{aligned}$$ Focus first on : by the definition of $S_j$ ––, for $j=1,\dots,M-1$ we have $$\label{eq:boundS} {{\left|S_j (x_k, \rho^{1/2}_{j,k}, \rho^{1/2}_{j+1,k})\right|}} \leq V_{\max} \left( \rho^{1/2}_{j,k} + \rho^{1/2}_{j+1,k}\right).$$ Therefore, recalling Remark \[rem:somma\], with slight abuse of notation $$\begin{aligned} \nonumber {{\Delta x}}\sum_{j=1}^M\sum_{k \in {{\mathbb{Z}}}} {{\Delta t}}{{\left|S_{j-1} (x_{k}, \rho^{1/2}_{j-1,k}, \rho^{1/2}_{j,k}) - S_j (x_k, \rho^{1/2}_{j,k}, \rho^{1/2}_{j+1,k})\right|}} \leq \ & {{\Delta x}}\, {{\Delta t}}\, V_{\max} \sum_{j=1}^M\sum_{k \in {{\mathbb{Z}}}} 4 \, \rho^{1/2}_{j,k} \\ \nonumber = \ & 4 \, {{\Delta t}}\, V_{\max} \, {{|\hspace{-1pt}\|{\rho^{1/2}}\|\hspace{-1pt}|}} \\ \label{eq:15} = \ & 4 \, {{\Delta t}}\, V_{\max} \, {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}, \end{aligned}$$ where we use Lemma \[lem:l1\]. Pass now to . Since we are interested in the sum over $k \in {{\mathbb{Z}}}$, we distinguish among four cases: $k<-1$, $k>0$, $k=-1$ and $k=0$. The first case, $k<-1$, amounts to $x_{k-1/2}<x_{k+1/2}<0$. Thus, by the definition of $F_j$ , together with , the numerical flux does not depend on the variable $x$, namely $$\mbox{for } x<0: \quad F_j (x, u, w) = \min\bigl\{f_{\ell,j} \bigl(\min\{u, {\vartheta}_\ell^j\}\bigr), f_{\ell,j} \left(\max\{w, {\vartheta}_\ell^j\}\right) \bigr\},$$ and the function above is clearly Lipschitz in both $u$ and $w$, with Lipschitz constant $\mathcal{V}$ as in , leading to $$\label{eq:xneg} \begin{aligned} & \sum_{k < -1} {{\left| F_j (x_{k+1/2}, \rho^0_{j,k}, \rho^0_{j,k+1}) - F_j (x_{k-1/2}, \rho^0_{j,k-1}, \rho^0_{j,k})\right|}} \\ \leq \ & \mathcal{V} \sum_{k<-1} \left({{\left|\rho^0_{j,k} - \rho^0_{j,k-1}\right|}} + {{\left|\rho^0_{j,k+1} - \rho^0_{j,k}\right|}} \right). \end{aligned}$$ The case $k>0$ can be treated analogously, leading to $$\label{eq:xpos} \begin{aligned} & \sum_{k >0} {{\left| F_j (x_{k+1/2}, \rho^0_{j,k}, \rho^0_{j,k+1}) - F_j (x_{k-1/2}, \rho^0_{j,k-1}, \rho^0_{j,k})\right|}} \\ \leq \ & \mathcal{V} \, \sum_{k >0} \left({{\left|\rho^0_{j,k} - \rho^0_{j,k-1}\right|}} + {{\left|\rho^0_{j,k+1} - \rho^0_{j,k}\right|}} \right). \end{aligned}$$ Pass now to $k=0$. Recall that $x_{-1/2}=0$. By the definition of $F_j$ , together with , we have $$\begin{aligned} F_j (x_{1/2}, \rho^0_{j,0}, \rho^0_{j,1}) - F_j (x_{-1/2}, \rho^0_{j,-1}, \rho^0_{j,0}) = \ & \min\bigl\{f_{r,j} \left(\min\{\rho^0_{j,0}, {\vartheta}_r^j\}\right), f_{r,j} \left(\max\{\rho^0_{j,1}, {\vartheta}_r^j\}\right) \bigr\} \\ & - \min\bigl\{f_{\ell,j} \left(\min\{\rho^0_{j,-1}, {\vartheta}_\ell^j\}\right), f_{r,j} \left(\max\{\rho^0_{j,0}, {\vartheta}_r^j\}\right) \bigr\}. \end{aligned}$$ We immediately get $$\label{eq:x0} {{\left|F_j (x_{1/2}, \rho^0_{j,0}, \rho^0_{j,1})\! - \! F_j (x_{-1/2}, \rho^0_{j,-1}, \rho^0_{j,0})\right|}} \leq {{\left\|f_j\right\|}}_{{\mathbf{L^\infty}}} \leq V_{\max},$$ with $V_{\max}$ as in . The case $k=-1$ follows analogously. Hence, collecting together , and  and using the fact that $\lambda \, {{\Delta x}}= {{\Delta t}}$, we obtain $$\begin{aligned} \nonumber & {{\Delta x}}\sum_{k \in {{\mathbb{Z}}}} \lambda \, {{\left| F_j (x_{k+1/2}, \rho^0_{j,k}, \rho^0_{j,k+1}) - F_j (x_{k-1/2}, \rho^0_{j,k-1}, \rho^0_{j,k})\right|}} \\ \nonumber \leq \ & \mathcal{V} \, {{\Delta t}}\sum_{k \in {{\mathbb{Z}}}} \left({{\left|\rho^0_{j,k} - \rho^0_{j,k-1}\right|}} + {{\left|\rho^0_{j,k+1} - \rho^0_{j,k}\right|}} \right) + 2 \, {{\Delta t}}\, V_{\max} \\ \label{eq:16} \leq \ & 2 \, \mathcal{V} \, {{\Delta t}}\sum_{k \in {{\mathbb{Z}}}}{{\left|\rho^0_{j,k} - \rho^0_{j,k-1}\right|}} +2 \, {{\Delta t}}\, V_{\max}. \end{aligned}$$ By –, insert  and  into : $$\begin{aligned} {{\Delta x}}\sum_{j=1}^M \sum_{k \in {{\mathbb{Z}}}} {{\left|\rho^{n+1}_{j,k} - \rho^n_{j,k}\right|}} \leq \ & 2 \, e^{4 \, \mathcal{V} \, T} \, {{\Delta t}}\Bigl( \mathcal{V} \, \sum_{j=1}^M {\mathinner{\rm TV}}(\rho^0_j) + M \, V_{\max} + 2 \, V_{\max} \, {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}} \Bigr), \end{aligned}$$ concluding the proof. Spatial BV bound {#sec:bv} ----------------- We follow the idea of [@BGKT2008 Lemma 4.2] of providing a *local* spatial ${\mathbf{BV}}$ bound, in the sense that the estimate in  below blows up if one of the endpoints of the interval $[a,b]$ approaches $x=0$. \[lem:bv\] Let ${\boldsymbol{\rho}}_o \in {\mathbf{BV}}({{\mathbb{R}}}; [0,1]^M)$ with ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$. Assume that the CFL condition  holds. For any interval $[a,b] \subseteq {{\mathbb{R}}}$ such that $0 \notin [a,b] $, fix $s>0$ such that $2 \, s < \min\{{{\left|a\right|}}, {{\left|b\right|}}\}$ and $s>{{\Delta x}}$. Then, for any $n=1, \ldots, N_T-1$ the following estimate holds: $$\label{eq:bv} \sum_{j=1}^M \sum_{k \in \mathbf{K}_a^b} {{\left|{\rho^{n}_{j,k+1}} - {\rho^{n}_{j,k}}\right|}} \leq e^{4 \, \mathcal{V} \, T} \biggl( \sum_{j=1}^M {\mathinner{\rm TV}}(\rho_{o,j}) + 8\, M \, V_{\max}\, T + \frac{2 \, C}{s}\biggr),$$ with $\mathbf{K}_a^b = \left\{ k \in {{\mathbb{Z}}}\colon a \leq x_{k} \leq b\right\}$, $V_{\max}$ and $\mathcal{V}$ as in  and $C$ independent of ${{\Delta x}}$ and ${{\Delta t}}$. Let $$\begin{aligned} \mathcal{A}_\Delta = \ & \left\{k \in {{\mathbb{Z}}}\colon x_{k-1/2} \in [a-s-{{\Delta x}}, a]\right\}, & \mathcal{B}_\Delta = \ & \left\{k \in {{\mathbb{Z}}}\colon x_{k+1/2} \in [b, b+s+{{\Delta x}}]\right\}. \end{aligned}$$ By the assumptions on $s$, observe that there are at least 2 elements in each of the sets above, i.e. ${{\left|\mathcal{A}_\Delta\right|}}, \, {{\left|\mathcal{B}_\Delta\right|}} \geq 2$. Moreover, ${{\left|\mathcal{A}_\Delta\right|}} \, {{\Delta x}}\geq s$ and ${{\left|\mathcal{B}_\Delta\right|}} {{\Delta x}}\geq s$. Furthermore, notice that - if $0<a<b$: it holds $x_{k-1/2}>0$ for any $k\in\mathcal{A}_\Delta$; - if $a<b<0$: it holds $x_{k+1/2}<0$ for any $k \in \mathcal{B}_\Delta$. By Proposition \[prop:l1cont-time\], there exists a constant $C$ such that $${{\Delta x}}\sum_{n=0}^{N_T -1} \sum_{j=1}^M \, \sum_{k \in {{\mathbb{Z}}}} {{\left|\rho^{n+1}_{j,k} - {\rho^{n}_{j,k}}\right|}} \leq C,$$ with $C = 2 \, T \, e^{4 \, \mathcal{V} \, T} \left( \mathcal{V} \, {\mathinner{\rm TV}}({\boldsymbol{\rho}}_o) + M \, V_{\max} + 2 V_{\max} \, {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}\right)$. Hence, when restricting the sum over $k$ in the set $\mathcal{A}_\Delta$, respectively $\mathcal{B}_\Delta$, it clearly follows that $$\begin{aligned} \label{eq:19} {{\Delta x}}\sum_{n=0}^{N_T -1} \sum_{j=1}^M \, \sum_{k \in \mathcal{A}_\Delta} {{\left|\rho^{n+1}_{j,k} - {\rho^{n}_{j,k}}\right|}} \leq \ & C, & {{\Delta x}}\sum_{n=0}^{N_T -1} \sum_{j=1}^M \, \sum_{k \in \mathcal{B}_\Delta} {{\left|\rho^{n+1}_{j,k} - {\rho^{n}_{j,k}}\right|}} \leq \ & C. \end{aligned}$$ Choose $k_a \in \mathcal{A}_\Delta$ and $k_b$ with $k_b+1 \in \mathcal{B}_\Delta$ such that $$\begin{aligned} \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_a} - {\rho^{n}_{j,k_a}}\right|}} = \ & \min_{k \in \mathcal{A}_\Delta} \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k} - {\rho^{n}_{j,k}}\right|}}, \\ \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_b+1} - {\rho^{n}_{j,k_b+1}}\right|}} = \ & \min_{k \in \mathcal{B}_\Delta} \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k} - {\rho^{n}_{j,k}}\right|}},. \end{aligned}$$ Thus, by , $$\label{eq:20} \begin{aligned} \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_a} - {\rho^{n}_{j,k_a}}\right|}} \leq \ & \frac{C}{{{\left|\mathcal{A}_\Delta\right|}} \, {{\Delta x}}} \leq \frac{C}{s}, \\ \sum_{n=0}^{N_T -1} \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_b+1} - {\rho^{n}_{j,k_b+1}}\right|}} \leq \ & \frac{C}{{{\left|\mathcal{B}_\Delta\right|}} \, {{\Delta x}}} \leq \frac{C}{s}. \end{aligned}$$ In view of the next steps, observe that $$\label{eq:decomp} \sum_{k=k_a}^{k_b}{{\left|\rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k}\right|}} = {{\left|\rho^{n+1}_{j,k_a+1} - \rho^{n+1}_{j,k_a}\right|}} + \sum_{k=k_a+1}^{k_b-1}{{\left|\rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k}\right|}} + {{\left|\rho^{n+1}_{j,k_b+1} - \rho^{n+1}_{j,k_b}\right|}}.$$ Focus on the central sum on the right hand side of . By , for $k_a < k < k_b$ and $j=1,\dots,M$, we have $$\begin{aligned} \rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k} = \ & \rho^{n+1/2}_{j,k+1} - \rho^{n+1/2}_{j,k} \\ & + {{\Delta t}}\left( S_{j-1} (x_k, \rho^{n+1}_{j-1,k+1} - \rho^{n+1}_{j,k+1}) - S_{j-1} (x_k, \rho^{n+1}_{j-1,k} - \rho^{n+1}_{j,k})\right. \\ & \qquad\quad \left. - S_{j} (x_k, \rho^{n+1}_{j,k+1} - \rho^{n+1}_{j+1,k+1}) + S_{j} (x_k, \rho^{n+1}_{j,k} - \rho^{n+1}_{j+1,k}) \right). \end{aligned}$$ By the Lipschitz continuity of the map $(u,w) \mapsto S_j (x,u,w)$ for $x\in {{\mathbb{R}}}$ and $j=1,\dots, M-1$, we get $$\begin{aligned} \label{eq:17} \sum_{j=1}^M{{\left| \rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k}\right|}} \leq \ & (1 + 4 \, \mathcal{V} \, {{\Delta t}}) \sum_{j=1}^M {{\left| \rho^{n+1/2}_{j,k+1} - \rho^{n+1/2}_{j,k}\right|}} . \end{aligned}$$ Fix now $j\in \{1,\dots,M\}$. Recall that for all $k_a \leq k \leq k_b$ either $x_{k-1/2}>0$ or $x_{k+1/2}<0$. Therefore, when applying , observe that the numerical flux $F_j$  is never computed at $x=0$, leading to $$\label{eq:schemeG} \rho^{n+1/2}_{j,k} = {\rho^{n}_{j,k}} - \lambda \left[ G_{d,j} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - G_{d,j} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right],$$ for $d=\ell, r$, with $$\label{eq:Gdj} G_{d,j} (u,w) = \min\bigl\{ f_{d,j} \left(\min\{u, {\vartheta}_d^j\}\right), \, f_{d,j}\left(\max\{w, {\vartheta}_d^j\}\right) \bigr\}.$$ Clearly, it is $d=\ell$ whenever $a<b<0$ and $d=r$ whenever $0<a<b$. Adding and subtracting $\lambda \, G_{d,j} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k}}) = \lambda \, f_{d,j} ({\rho^{n}_{j,k}})$ into and setting $$\begin{aligned} \label{eq:gamma} \gamma_{d,j,k}^n = \ & \left\{ \begin{array}{l@{\quad \mbox{ if }}l} -\lambda \, \dfrac{G_{d,j} ({\rho^{n}_{j,k+1}},{\rho^{n}_{j,k}}) - G_{d,j} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k}})}{{\rho^{n}_{j,k+1}}- {\rho^{n}_{j,k}}} & {\rho^{n}_{j,k+1}} \neq {\rho^{n}_{j,k}}, \\ 0 & {\rho^{n}_{j,k+1}} = {\rho^{n}_{j,k}}, \end{array} \right. \\ \label{eq:delta} \delta_{d,j,k}^n = \ & \left\{ \begin{array}{l@{\qquad \mbox{ if }}l} \lambda \, \dfrac{G_{d,j} ({\rho^{n}_{j,k}},{\rho^{n}_{j,k}}) - G_{d,j} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}})}{{\rho^{n}_{j,k}}- {\rho^{n}_{j,k-1}}} & {\rho^{n}_{j,k}} \neq {\rho^{n}_{j,k-1}}, \\ 0 & {\rho^{n}_{j,k}} = {\rho^{n}_{j,k-1}}, \end{array} \right. \end{aligned}$$ we can rearrange  to get $$\label{eq:scomp} \rho^{n+1/2}_{j,k} = {\rho^{n}_{j,k}} + \gamma_{d,j,k}^n \left({\rho^{n}_{j,k+1}} - {\rho^{n}_{j,k}}\right) - \delta_{d,j,k}^n \left({\rho^{n}_{j,k}} - {\rho^{n}_{j,k-1}}\right).$$ The function $G_{d,j}$ is non decreasing in the first argument and non increasing in the second, so that we easily get $\gamma_{d,j,k}^n , \, \delta_{d,j,k}^n \geq 0$. Furthermore, $G_{d,j}$ is Lipschitz continuous in both variables, with the same Lipschitz constant $\mathcal{V}$  as $F_j$: by the CFL condition  $$\begin{aligned} \gamma_{d,j,k}^n\leq \ & \lambda \, \mathcal{V} \leq \frac12, & \delta_{d,j,k}^n\leq \ & \lambda \, \mathcal{V} \leq \frac12, \end{aligned}$$ and hence $\gamma_{d,j,k}^n + \delta_{d,j,k+1}^n\leq 1$. Therefore, for $k_a< k < k_b$ $$\label{eq:21} \begin{aligned} \rho^{n+1/2}_{j,k+1} - \rho^{n+1/2}_{j,k} = \ & \left({\rho^{n}_{j,k+1}} - {\rho^{n}_{j,k}}\right) \left(1 - \gamma_{d,j,k}^n - \delta_{d,j,k+1}^n\right) \\ & + \gamma_{d,j, k+1}^n \left({\rho^{n}_{j,k+2}} - {\rho^{n}_{j,k+1}}\right) + \delta_{d,j,k}^n \left({\rho^{n}_{j,k}} - {\rho^{n}_{j,k-1}}\right). \end{aligned}$$ We are left with the boundary terms in . Fix $j \in \{1,\dots, M\}$. For $k=k_a$, applying first  then , in the form of , we have $$\begin{aligned} & \rho^{n+1}_{j,k_a+1} - \rho^{n+1}_{j,k_a} \\ = \ & \rho^{n+1/2}_{j,k_a+1} + {{\Delta t}}\, S_{j-1} (x_{k_a+1}, \rho^{n+1/2}_{j-1,k_a+1}, \rho^{n+1/2}_{j,k_a+1}) - {{\Delta t}}\, S_{j} (x_{k_a+1}, \rho^{n+1/2}_{j,k_a+1}, \rho^{n+1/2}_{j+1,k_a+1}) - \rho^{n+1}_{j,k_a} \\ = \ & {\rho^{n}_{j,k_a+1}} + \gamma^n_{d,j,k_a+1} \left({\rho^{n}_{j,k_a+2}} - {\rho^{n}_{j,k_a+1}}\right) - \delta^n_{d,j,k_a+1}\left({\rho^{n}_{j,k_a+1}} - {\rho^{n}_{j,k_a}}\right) \\ & + {{\Delta t}}\, S_{j-1} (x_{k_a+1}, \rho^{n+1/2}_{j-1,k_a+1}, \rho^{n+1/2}_{j,k_a+1}) - {{\Delta t}}\, S_{j} (x_{k_a+1}, \rho^{n+1/2}_{j,k_a+1}, \rho^{n+1/2}_{j+1,k_a+1}) - \rho^{n+1}_{j,k_a}. \end{aligned}$$ Add and subtract ${\rho^{n}_{j,k_a}}$, then take the absolute value and sum over $j=1,\dots,M$: exploiting  leads to $$\label{eq:ka} \begin{aligned} \sum_{j=1}^M {{\left| \rho^{n+1}_{j,k_a+1} - \rho^{n+1}_{j,k_a} \right|}} \leq \ & \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_a} -\ {\rho^{n}_{j,k_a}} \right|}} + \sum_{j=1}^M (1-\delta^n_{d,j,k_a+1}) {{\left|{\rho^{n}_{j,k_a+1}} -{\rho^{n}_{j,k_a}}\right|}} \\ & + \sum_{j=1}^M \gamma^n_{d,j,k_a+1} {{\left|{\rho^{n}_{j,k_a+2}} - {\rho^{n}_{j,k_a+1}}\right|}} + 4 \, {{\Delta t}}\, V_{\max} \sum_{j=1}^M \rho^{n+1/2}_{j,k_a+1}. \end{aligned}$$ Proceed similarly for $k=k_b$: $$\begin{aligned} & \rho^{n+1}_{j,k_b+1} - \rho^{n+1}_{j,k_b} \\ = \ & \rho^{n+1}_{j,k_b+1} - \rho^{n+1/2}_{j,k_b} - {{\Delta t}}\, S_{j-1} (x_{k_b}, \rho^{n+1/2}_{j-1,k_b}, \rho^{n+1/2}_{j,k_b}) + {{\Delta t}}\, S_{j} (x_{k_b}, \rho^{n+1/2}_{j,k_b}, \rho^{n+1/2}_{j+1,k_b}) \\ = \ & \rho^{n+1}_{j,k_b+1} - {\rho^{n}_{j,k_b}} - \gamma^n_{d,j,k_b} \left({\rho^{n}_{j,k_b+1}} - {\rho^{n}_{j,k_b}}\right) + \delta^n_{d,j,k_b}\left({\rho^{n}_{j,k_b}} - {\rho^{n}_{j,k_b-1}}\right) \\ & - {{\Delta t}}\, S_{j-1} (x_{k_b}, \rho^{n+1/2}_{j-1,k_b}, \rho^{n+1/2}_{j,k_b}) + {{\Delta t}}\, S_{j} (x_{k_b}, \rho^{n+1/2}_{j,k_b}, \rho^{n+1/2}_{j+1,k_b}). \end{aligned}$$ Now add and subtract ${\rho^{n}_{j,k_b+1}}$, take the absolute value and sum over $j=1,\dots,M$: $$\label{eq:kb} \begin{aligned} \sum_{j=1}^M {{\left| \rho^{n+1}_{j,k_b+1} - \rho^{n+1}_{j,k_b} \right|}} \leq \ & \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_b+1} -\ {\rho^{n}_{j,k_b+1}} \right|}} + \sum_{j=1}^M (1-\gamma^n_{d,j,k_b}) {{\left|{\rho^{n}_{j,k_b+1}} -{\rho^{n}_{j,k_b}}\right|}} \\ & + \sum_{j=1}^M \delta^n_{d,j,k_b} {{\left|{\rho^{n}_{j,k_b+1}} - {\rho^{n}_{j,k_b}}\right|}} + 4 \, {{\Delta t}}\, V_{\max} \sum_{j=1}^M \rho^{n+1/2}_{j,k_b}. \end{aligned}$$ By , collect together , , and : since all the coefficients appearing there are positive, we obtain $$\begin{aligned} & \sum_{j=1}^M \, \sum_{k=k_a}^{k_b} {{\left|\rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k}\right|}} \\ \leq \ & \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_a} -\ {\rho^{n}_{j,k_a}} \right|}} + \sum_{j=1}^M (1-\delta^n_{d,j,k_a+1}) {{\left|{\rho^{n}_{j,k_a+1}} -{\rho^{n}_{j,k_a}}\right|}} + \sum_{j=1}^M \gamma^n_{d,j,k_a+1} {{\left|{\rho^{n}_{j,k_a+2}} - {\rho^{n}_{j,k_a+1}}\right|}} \\ & + 4 \, {{\Delta t}}\, V_{\max} \sum_{j=1}^M \rho^{n+1/2}_{j,k_a+1} + e^{4\, \mathcal{V} \, {{\Delta t}}} \sum_{j=1}^M \, \sum_{k=k_a+1}^{k_b-1} \left( 1 - \gamma_{d,j,k}^n - \delta_{d,j,k+1}^n \right) {{\left|{\rho^{n}_{j,k+1}} - {\rho^{n}_{j,k}}\right|}} \\ & + e^{4 \, \mathcal{V}\, {{\Delta t}}} \sum_{j=1}^M \, \sum_{k=k_a+1}^{k_b-1} \gamma_{d,j,k+1}^n {{\left|{\rho^{n}_{j,k+2}} - {\rho^{n}_{j,k+1}}\right|}} + e^{4 \, \mathcal{V} \, {{\Delta t}}} \sum_{j=1}^M \, \sum_{k=k_a+1}^{k_b-1} \delta_{d,j,k}^n {{\left|{\rho^{n}_{j,k}} - {\rho^{n}_{j,k-1}}\right|}} \\ & + \sum_{j=1}^M {{\left|\rho^{n+1}_{j,k_b+1} -\ {\rho^{n}_{j,k_b+1}} \right|}} + \sum_{j=1}^M (1-\gamma^n_{d,j,k_b}) {{\left|{\rho^{n}_{j,k_b+1}} -{\rho^{n}_{j,k_b}}\right|}} + \sum_{j=1}^M \delta^n_{d,j,k_b} {{\left|{\rho^{n}_{j,k_b+1}} - {\rho^{n}_{j,k_b}}\right|}} \\ & + 4 \, {{\Delta t}}\, V_{\max} \sum_{j=1}^M \rho^{n+1/2}_{j,k_b} \\ \leq \ & \sum_{j=1}^M \left( e^{4 \mathcal{V} \, {{\Delta t}}} \sum_{k=k_a}^{k_b} {{\left|{\rho^{n}_{j,k+1}} - {\rho^{n}_{j,k}}\right|}} + {{\left|\rho^{n+1}_{j,k_a} -\ {\rho^{n}_{j,k_a}} \right|}} + {{\left|\rho^{n+1}_{j,k_b+1} -\ {\rho^{n}_{j,k_b+1}}\right|}}\right) +8 \, M \, V_{max} \, {{\Delta t}}, \end{aligned}$$ where we exploit also Lemma \[lem:pos\]. Proceeding recursively we finally get, for $1\leq n < N_T-1$, $$\begin{aligned} \sum_{j=1}^M \, \sum_{k=k_a}^{k_b} {{\left|\rho^{n+1}_{j,k+1} - \rho^{n+1}_{j,k}\right|}} \leq \ & e^{4 \, \mathcal{V} \, (n+1) \, {{\Delta t}}} \sum_{j=1}^M {\mathinner{\rm TV}}(\rho_{o,j}) + e^{4\, \mathcal{V} \, n \, {{\Delta t}}} \, 8 \, M \, V_{\max} (n+1)\, {{\Delta t}}\\ & + e^{4\, \mathcal{V} \, n \, {{\Delta t}}}\sum_{m=0}^n\sum_{j=1}^M \left( {{\left|\rho^{m+1}_{j,k_a} - \rho^m_{j,k_a}\right|}} + {{\left|\rho^{m+1}_{j,k_b+1} - \rho^m_{j,k_b+1}\right|}} \right) \\ \leq \ & e^{4 \, \mathcal{V} \, T} \left( \sum_{j=1}^M {\mathinner{\rm TV}}(\rho_{o,j}) + 8 \, M \, V_{\max}\, T + \frac{2 \, C}{s}\right), \end{aligned}$$ where we used also . Noticing that $[a,b] \subseteq [x_{k_a}, x_{k_b+1}]$ completes the proof. Discrete Entropy Inequality {#sec:entropyIneq} --------------------------- We follow the idea of [@KRT2002 Lemma 5.1]. \[lem:entropyIneq\] Let ${\boldsymbol{\rho}}_o \in {\mathbf{BV}}({{\mathbb{R}}}; [0,1]^M)$ with ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$. Assume that the CFL condition  holds. Then the approximate solution ${\boldsymbol{\rho}}_\Delta$ defined by  through Algorithm \[alg:1\] satisfies the following discrete entropy inequality: for all $j=1,\dots,M$, for $k\in {{\mathbb{Z}}}$, for $n=0,\ldots, N_T-1$ and for any $c \in [0,1]$ $$\begin{aligned} \nonumber {{\left|\rho^{n+1}_{j,k} - c\right|}} - {{\left|{\rho^{n}_{j,k}} - c\right|}} + \lambda \left( \mathscr{F}^{c}_{j,k+1/2} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - \mathscr{F}^{c}_{j, k-1/2} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right) & \\ \label{eq:die} - \lambda \, {{\left|F_j (x_{k+1/2}, c,c) - F_j (x_{k-1/2},c, c)\right|}} & \\ \nonumber - {{\Delta t}}\, \operatorname{sgn}(\rho^{n+1}_{j,k} - c) \left(S_{j-1} (x_{k}, \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - S_j (x_k, \rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k})\right) & \leq \ 0, \end{aligned}$$ with $$\mathscr{F}^{c}_{j,k+1/2} (u,w) = F_j (x_{k+1/2},u \vee c, w \vee c) - F_j (x_{k+1/2},u \wedge c, w \wedge c),$$ where $a \vee b = \max\{a,b\}$, $a \wedge b = \min\{a,b\}$. Fix $j \in \{1, \dots, M\}$ and $k \in {{\mathbb{Z}}}$. Let $$\mathscr{G}_{j,k} (u,w,z) = w - \lambda \left[ F_j (x_{k+1/2}, w, z) - F_j (x_{k-1/2}, u,w) \right].$$ Clearly $\rho^{n+1/2}_{j,k} = \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}},{\rho^{n}_{j,k+1}})$. Set $$\Delta_k F_j ^c = F_j (x_{k+1/2}, c, c) - F_j (x_{k-1/2}, c,c),$$ so that $ \mathscr{G}_{j,k} (c,c,c) = c - \lambda \, \Delta_k F_j^c$. By the properties of the numerical flux $F_j$, the map $\mathscr{G}_{j,k}$ is non decreasing in all its arguments. Therefore, $$\begin{aligned} \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}} \vee c, {\rho^{n}_{j,k}} \vee c, {\rho^{n}_{j,k+1}} \vee c) \geq \ & \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) \vee \mathscr{G}_{j,k} (c,c,c), \\ - \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}} \wedge c, {\rho^{n}_{j,k}} \wedge c, {\rho^{n}_{j,k+1}} \wedge c) \geq \ & - \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) \wedge \mathscr{G}_{j,k} (c,c,c). \end{aligned}$$ Sum the two inequalities above: since $a \vee b - a \wedge b = {{\left|a-b\right|}}$, observe that $$\begin{aligned} & \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}} \vee c, {\rho^{n}_{j,k}} \vee c, {\rho^{n}_{j,k+1}} \vee c) - \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}} \wedge c, {\rho^{n}_{j,k}} \wedge c, {\rho^{n}_{j,k+1}} \wedge c) \\ = \ & {{\left|\rho^{n}_{j,k} - c\right|}} -\lambda \left( \mathscr{F}^c_{j,k+1/2} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - \mathscr{F}^c_{j,k-1/2} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right), \end{aligned}$$ and $$\begin{aligned} & \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) \vee \mathscr{G}_{j,k} (c,c,c) - \mathscr{G}_{j,k} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) \wedge \mathscr{G}_{j,k} (c,c,c) \\ = \ & {{\left|\rho^{n+1/2}_{j,k} - c + \lambda \, \Delta_k F_j^c\right|}} \\ = \ & {{\left|\rho^{n+1}_{j,k} - c + \lambda \, \Delta_k F_j^c -{{\Delta t}}\left( S_{j-1} (x_k, \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - S_{j} (x_k, \rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k}) \right) \right|}} \\ \geq \ & {{\left|\rho^{n+1}_{j,k} - c\right|}} - \lambda \, {{\left| \Delta_k F_j^c\right|}} \\ & - {{\Delta t}}\, \operatorname{sgn}\!\left(\rho^{n+1}_{j,k} - c\right) \left( S_{j-1} (x_k, \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - S_{j} (x_k, \rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k}) \right), \end{aligned}$$ where we used also  and the inequality ${{\left|a+b\right|}} \geq {{\left|a\right|}} + \operatorname{sgn}(a) \, b$. The thesis immediately follows. Convergence {#sec:convergence} ----------- \[thm:main\] Let ${\boldsymbol{\rho}}_o \in {\mathbf{BV}}({{\mathbb{R}}}; [0,1]^M)$ with ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}< +\infty$. Let ${{\Delta x}}\to 0$ with $\lambda = {{\Delta x}}/ {{\Delta t}}$ constant and satisfying the CFL condition . The sequence of approximate solutions ${\boldsymbol{\rho}}_\Delta$ constructed through Algorithm \[alg:1\] converges in ${\mathbf{L^{1}_{\mathbf{loc}}}}$ to a function ${\boldsymbol{\rho}}\in {\mathbf{L^\infty}}([0,T] \times {{\mathbb{R}}}; [0,1]^M)$ such that ${{|\hspace{-1pt}\|{{\boldsymbol{\rho}}(t)}\|\hspace{-1pt}|}}= {{|\hspace{-1pt}\|{{\boldsymbol{\rho}}_o}\|\hspace{-1pt}|}}$ for $t\in [0,T]$. This limit function ${\boldsymbol{\rho}}$ is a weak entropy solution to problem –– in the sense of Definition \[def:sol\]. We follow [@BKT2009 Theorem 5.1] and [@KRT2002 Theorem 5.1]. Lemma \[lem:pos\] ensures that the sequence of approximate solutions ${\boldsymbol{\rho}}_\Delta$ is bounded in ${\mathbf{L^\infty}}$, in particular $\rho_{j,\Delta} (t,x) \in [0,1]$, for all $t>0$, $x \in {{\mathbb{R}}}$ and $j=1,\dots,M$. Proposition \[prop:l1cont-time\] proves the ${\mathbf{L^1}}$-continuity in time of the sequence ${\boldsymbol{\rho}}_\Delta$, while Lemma \[lem:bv\] guarantees a bound on the spatial total variation in any interval $[a,b]$ not containing $x=0$. Standard compactness results imply that, for any interval $[a,b]$ not containing $x=0$, there exists a subsequence, still denoted by ${\boldsymbol{\rho}}_\Delta$, converging in ${\mathbf{L^1}} ([0,T] \times [a,b]; [0,1]^M)$. Take now a countable set of intervals $[a_i,b_i]$ such that $\bigcup_i [a_i,b_i] = {{\mathbb{R}}}\setminus \{0\}$: by a standard diagonal process, we can extract a subsequence, still denoted by ${\boldsymbol{\rho}}_\Delta$, converging in ${\mathbf{L^{1}_{\mathbf{loc}}}} ([0,T]\times {{\mathbb{R}}}; [0,1]^M)$, and almost everywhere in $[0,T] \times {{\mathbb{R}}}$, to a function ${\boldsymbol{\rho}}\in {\mathbf{L^\infty}} ([0,T]\times {{\mathbb{R}}}; [0,1]^M)$. Moreover, Proposition \[prop:l1cont-time\], and in particular formula , implies that this limit function is such that ${\boldsymbol{\rho}}\in {\mathbf{C^{0}}} ([0,T]; {\mathbf{L^1}} ({{\mathbb{R}}}; [0,1]^M))$, with slight abuse of notation concerning the ${\mathbf{L^1}}$-norm. It remains to show that the limit function ${\boldsymbol{\rho}}$ satisfies the integral inequalities in Definition \[def:sol\]. Concerning point \[it:weak\], i.e. the weak formulation, it suffices to apply a Lax-Wendroff-type calculation, similarly to what has been done in [@KRT2002 Theorem 3.1]. Notice that the presence of the source terms does not add any difficulties in the proof. As for point \[it:entropy\] in Definition \[def:sol\], i.e. the entropy inequality, we follow [@KRT2002 Theorem 5.1]. Fix $j \in \{1,\dots,M\}$. Let ${\varphi}\in {\mathbf{C_c^{1}}} ([0,T[ \times {{\mathbb{R}}}; {{\mathbb{R}}}^+)$. Multiply the inequality  by ${{\Delta x}}\, {\varphi}_k^n = {{\Delta x}}\, {\varphi}(t^n, x_k)$, then sum over $k \in {{\mathbb{Z}}}$ and $n=0, \ldots, N_T -1$: $$\begin{aligned} \label{eq:p1} 0 \geq \ & {{\Delta x}}\sum_{n=0}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} \left[ {{\left|\rho^{n+1}_{j,k} - c\right|}} - {{\left|{\rho^{n}_{j,k}} - c\right|}} \right] {\varphi}^n_k \\ \label{eq:p2} & + {{\Delta t}}\sum_{n=0}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} \left[ \mathscr{F}^c_{j,k+1/2} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) - \mathscr{F}^c_{j,k-1/2} ({\rho^{n}_{j,k-1}}, {\rho^{n}_{j,k}}) \right] {\varphi}^n_k \\ \label{eq:p3} & - {{\Delta t}}\sum_{n=0}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} {{\left| F_j (x_{k+1/2}, c, c) - F_j (x_{k-1/2}, c, c)\right|}} \, {\varphi}^n_k \\ \label{eq:p4} & - {{\Delta t}}\, {{\Delta x}}\sum_{n=0}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} \operatorname{sgn}(\rho^{n+1}_{j,k} - c) \left[ S_{j-1} (x_k, \rho^{n+1/2}_{j-1,k}, \rho^{n+1/2}_{j,k}) - S_j (x_k, \rho^{n+1/2}_{j,k}, \rho^{n+1/2}_{j+1,k}) \right] {\varphi}^n_k. \end{aligned}$$ Take into account each term separately. Summing by parts and letting ${{\Delta x}}\to 0^+$, the Dominated Convergence Theorem yields $$\label{eq:p1ok} \begin{aligned} [\eqref{eq:p1}] = \ & - {{\Delta x}}\sum_{k \in {{\mathbb{Z}}}} {{\left|\rho^0_{j,k} - c\right|}} \, {\varphi}^0_k - {{\Delta x}}\, {{\Delta t}}\sum_{n=1}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} {{\left|{\rho^{n}_{j,k}} - c\right|}} \, \frac{{\varphi}^n_k - {\varphi}^{n-1}_k}{{{\Delta t}}} \\ \underset{{{\Delta x}}\to 0^+}{\longrightarrow} \ & - \int_{{{\mathbb{R}}}} {{\left|\rho_{o,j} - c\right|}} \, {\varphi}(0,x) {\mathinner{\mathrm{d}{x}}} - \int_0^T \int_{{\mathbb{R}}}{{\left|\rho_j (t,x) - c\right|}} \, \partial_t {\varphi}(t,x) {\mathinner{\mathrm{d}{x}}}{\mathinner{\mathrm{d}{t}}}, \end{aligned}$$ and $$\label{eq:p2ok} \begin{aligned} [\eqref{eq:p2}] = \ & - {{\Delta x}}\, {{\Delta t}}\sum_{n=0}^{N_T -1} \sum_{k \in {{\mathbb{Z}}}} \mathscr{F}^c_{j,k+1/2} ({\rho^{n}_{j,k}}, {\rho^{n}_{j,k+1}}) \frac{{\varphi}^n_k - {\varphi}^n_{k-1}}{{{\Delta x}}} \\ \underset{{{\Delta x}}\to 0^+}{\longrightarrow} \ & - \int_0^T \int_{{\mathbb{R}}}\operatorname{sgn}(\rho_j (t,x) - c) \left(f_j (x,\rho_j (t,x)) - f_j (x,c)\right) \, \partial_x {\varphi}(t,x) {\mathinner{\mathrm{d}{x}}}{\mathinner{\mathrm{d}{t}}}. \end{aligned}$$ Pass now to . Observe that, by the definition of the numerical flux , when $x \neq 0$ it holds $F_j (x,c,c) = f_{d,j} (c)$, with $d=\ell$ if $x <0$ and $d=r$ if $x>0$. Therefore  gives a contribution only for $k=-1$ and $k=0$: $$\begin{aligned} [\eqref{eq:p3}] =\ & - {{\Delta t}}\sum_{n=0}^{N_T -1} \sum_{k=-1}^0 {{\left|F_j (x_{k+1/2},c,c) - F_j (x_{k-1/2}, c, c)\right|}} \, {\varphi}_{k}^n \\ \underset{{{\Delta x}}\to 0^+}{\longrightarrow} \ & - \int_0^T\left({{\left|F_j (0,c,c) - f_{\ell,j} (c)\right|}} + {{\left|f_{r,j} (c) - F_j (0,c,c)\right|}}\right) \, {\varphi}(t,0) {\mathinner{\mathrm{d}{t}}} \end{aligned}$$ A careful analysis of all the possible cases yields $${{\left|F_j (0,c,c) - f_{\ell,j} (c)\right|}} + {{\left|f_{r,j} (c) - F_j (0,c,c)\right|}} = {{\left|f_{r,j} (c) - f_{\ell,j} (c)\right|}},$$ so that $$\label{eq:p3ok} [\eqref{eq:p3}] \underset{{{\Delta x}}\to 0^+}{\longrightarrow} - \int_0^T {{\left|f_{r,j} (c) - f_{\ell,j} (c)\right|}} \, {\varphi}(t,0) {\mathinner{\mathrm{d}{t}}}.$$ Focus now on the last term : by the Dominated Convergence Theorem $$\label{eq:p4ok} [\eqref{eq:p4}] \underset{{{\Delta x}}\to 0^+}{\longrightarrow} - \int_0^T \int_{{\mathbb{R}}}\operatorname{sgn}(\rho_j- c) \left( S_{j-1} (x, \rho_{j-1} , \rho_j ) - S_j (x, \rho_j , \rho_{j+1} ) \right) {\varphi}(t,x) {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}}.$$ Collecting together , , and  completes the proof. L1-Stability and uniqueness {#sec:unique} --------------------------- The following Theorem ensures that the solution to –– depends ${\mathbf{L^1}}$-Lipschitz continuously on the initial data, thus guaranteeing the uniqueness of solutions. \[thm:unique\] Let ${\boldsymbol{\rho}}, \, {\boldsymbol{\sigma{}}}$ be two weak entropy solutions, in the sense of Definition \[def:sol\], to problem –– with initial data ${\boldsymbol{\rho}}_o, \, {\boldsymbol{\sigma{}}}_o \in {\mathbf{L^\infty}} ({{\mathbb{R}}}; [0,1]^M)$ and such that ${\boldsymbol{\rho}}_o - {\boldsymbol{\sigma{}}}_o \in {\mathbf{L^1}} ({{\mathbb{R}}};[0,1]^M)$. Then, for a.e. $t \in [0,T]$, $$\label{eq:22} \sum_{j=1}^M {{\left\|\rho_j (t) - \sigma_j (t)\right\|}}_{{\mathbf{L^1}} ({{\mathbb{R}}})} \leq \sum_{j=1}^M {{\left\|\rho_{o,j} - \sigma_{o,j}\right\|}}_{{\mathbf{L^1}} ({{\mathbb{R}}})}.$$ Notice that the sums appearing in  are actually sums over the *active* lanes only, the terms corresponding to fictive lanes being equal to $0$. The idea is to combine together the results contained in [@KRT2003 § 2 and § 5], in particular [@KRT2003 Theorem 5.1], and in [@BKT2009 Theorem 3.1], and then adapt [@HoldenRisebro Theorem 3.3]. Indeed, fix $j\in\{1, \dots, M\}$. Following [@KRT2003 Theorem A.1 and Formula (2.22)], it is possible to derive the following inequality for any ${\varphi}\in {\mathbf{C_c^{1}}} (\,]0,T[ \times {{\mathbb{R}}}\setminus\{0\}; {{\mathbb{R}}}^+)$ $$\label{eq:1} \begin{aligned} -\int_0^T \int_{{\mathbb{R}}}\bigl\{ {{\left|\rho_j - \sigma_j\right|}} \partial_t {\varphi}+ \operatorname{sgn}(\rho_j - \sigma_j) \left( f_j (x,\rho_j) - f_j (x,\sigma_j)\right) \partial_x {\varphi}& \\ + \operatorname{sgn}(\rho_j - \sigma_j)\left(S (x, {\boldsymbol{\rho}}, j) - S (x, {\boldsymbol{\sigma{}}}, j)\right) {\varphi}& \bigr\} {\mathinner{\mathrm{d}{x}}}{\mathinner{\mathrm{d}{t}}} \leq 0, \end{aligned}$$ where, for the sake of simplicity, we set $$\label{eq:S} S (x, \boldsymbol{u}, j) = S_{j-1} (x,u_{j-1}, u_j) - S_j (x, u_j, u_{j+1}).$$ Inspired by [@HoldenRisebro Theorem 3.3], since $\rho_j$, respectively $\sigma_j$, satisfies Point \[it:weak\] in Definition \[def:sol\], we subtract to the above inequality the equation for $\rho_j$ and add the equation for $\sigma_j$, arriving at $$\begin{aligned} -\int_0^T\int_{{\mathbb{R}}}\bigl\{\left(\rho_j - \sigma_j\right)^+\partial_t {\varphi}+ H (\rho_j - \sigma_j) \left(f_j (x,\rho_j) - f_j (x, \sigma_j)\right) \partial_x{\varphi}& \\ + H (\rho_j - \sigma_j) \left(S (x,{\boldsymbol{\rho}},j) - S(x,{\boldsymbol{\sigma{}}},j)\right){\varphi}& \bigr\} {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}}\leq 0, \end{aligned}$$ for ${\varphi}\in {\mathbf{C_c^{1}}} (\,]0,T[ \times {{\mathbb{R}}}\setminus\{0\}; {{\mathbb{R}}}^+)$. Now, we extend the above inequality to $\Phi \in {\mathbf{C_c^{1}}} (\,]0,T[ \times {{\mathbb{R}}}; {{\mathbb{R}}}^+)$. The procedure is similar to that in [@KRT2003 Theorem 2.1] and it leads to $$\label{eq:5} \begin{aligned} -\int_0^T\int_{{\mathbb{R}}}\bigl\{\left(\rho_j - \sigma_j\right)^+ \partial_t \Phi + H (\rho_j - \sigma_j) \left(f_j (x,\rho_j) - f_j (x, \sigma_j)\right) \partial_x \Phi & \\ + H (\rho_j - \sigma_j) \left(S (x,{\boldsymbol{\rho}},j) - S(x,{\boldsymbol{\sigma{}}},j)\right) \Phi & \bigr\} {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}} \leq E, \end{aligned}$$ for all $\Phi \in {\mathbf{C_c^{1}}} (\,]0,T[ \times {{\mathbb{R}}}; {{\mathbb{R}}}^+)$, where $$E = \int_0^T\left[ H(\rho_j - \sigma_j) \left( f_j (x,\rho_j) - f_j (x,\sigma_j)\right) \right]_{x=0^-}^{x=0^+} \Phi (t,0) {\mathinner{\mathrm{d}{t}}}.$$ Analogously to [@KRT2003 Theorem 2.1] and [@BKT2009 Theorem 3.1], it can be proven that $E\leq 0$. Following again [@HoldenRisebro Theorem 3.3] and choosing $\Phi \approx \boldsymbol{1}_{[0,\tau]}$, for $\tau \in \,]0,T]$, we get $$\label{eq:18} \begin{aligned} \int_{{\mathbb{R}}}\left(\rho_j (\tau) - \sigma_j (\tau)\right)^+ {\mathinner{\mathrm{d}{x}}} \leq \ & \int_{{\mathbb{R}}}\left(\rho_j (0) - \sigma_j (0)\right)^+{\mathinner{\mathrm{d}{x}}} \\ & + \int_0^\tau \int_{{\mathbb{R}}}H (\rho_j - \sigma_j) \left(S (x,{\boldsymbol{\rho}},j) - S(x,{\boldsymbol{\sigma{}}},j)\right) {\mathinner{\mathrm{d}{x}}} {\mathinner{\mathrm{d}{t}}}. \end{aligned}$$ It is easy to verify that, for fixed $x$, the map $S_j (x,u,w)$ defined in , together with , is non decreasing in the second argument and non increasing in the third: setting for the sake of convenience $\Delta_+ v_j =v_{j+1} (x,w) - v_{j} (x,u)$ we obtain $$\begin{aligned} \partial_u S_j = \ & (\Delta_+v_j )^+ - v'_j (x,u) \, w - H (\Delta_+ v_j) \, v'_j (x,u) \, (u-w) \geq 0, \\ \partial_w S_j = \ & - (\Delta_+v_j)^- + v'_{j+1} (x,w) \, w + H (\Delta_+v_j) \, v'_{j+1} (x,w) \, (u-w) \leq 0. \end{aligned}$$ Hence, if $\rho_j > \sigma_j$ we have $$\begin{aligned} & S (x,{\boldsymbol{\rho}},j) - S(x,{\boldsymbol{\sigma{}}},j) \\ = \ & S_{j-1} (x,\rho_{j-1}, \rho_j)-S_{j-1} (x,\sigma_{j-1}, \sigma_j) - S_j (x, \rho_j, \rho_{j+1}) + S_j (x, \sigma_j, \sigma_{j+1}) \\ \leq \ & S_{j-1} (x,\rho_{j-1}, \sigma_j)-S_{j-1} (x,\sigma_{j-1}, \sigma_j) - S_j (x, \rho_j, \rho_{j+1}) + S_j (x, \rho_j, \sigma_{j+1}) \\ = \ & \partial_u S_{j-1} (x, \pi_{j-1}, \sigma_j)\, (\rho_{j-1} - \sigma_{j-1}) - \partial_w S_{j} (x,\sigma_j, \pi_{j+1}) \,(\rho_{j+1}- \sigma_{j+1}) \\ \leq \ & \mathcal{V} \left((\rho_{j-1} - \sigma_{j-1})^+ + (\rho_{j+1}- \sigma_{j+1})^+\right), \end{aligned}$$ with $\pi_{j\pm1}$ in the interval between $\rho_{j\pm1}$ and $\sigma_{j\pm1}$ respectively and $\mathcal{V}$ as in . Thus, $$\label{eq:25} \sum_{j=1}^M H (\rho_j - \sigma_j) \left(S (x,{\boldsymbol{\rho}},j) - S(x,{\boldsymbol{\sigma{}}},j)\right) \leq 2 \, \mathcal{V} \, \sum_{j=1}^M (\rho_j - \sigma_j)^+.$$ Define $$\Theta (t) = \sum_{j=1}^M \int_{{\mathbb{R}}}\left(\rho_j (t,x) - \sigma_j (t,x)\right)^+ {\mathinner{\mathrm{d}{x}}}.$$ By  and  it follows that $$\Theta (\tau) \leq \Theta (0) + 2 \, \mathcal{V} \int_0^\tau \Theta (t) {\mathinner{\mathrm{d}{t}}}.$$ Gronwall’s inequality then implies that $\Theta (t) \leq \Theta (0) \, \exp \left( 2 \, \mathcal{V} \, t\right)$. Therefore, if $\Theta (0)=0$, i.e. $\rho_{o,j}(x)\leq \sigma_{o,j} (x)$ a.e. in ${{\mathbb{R}}}$ and for all $j$, then $\Theta (t)=0$ for $t>0$, i.e. $\rho_{j}(t,x)\leq \sigma_{j} (t,x)$ a.e. in ${{\mathbb{R}}}$ and for all $j$. An application of the Crandall–Tartar Lemma [@HoldenRisebroBook2015 Lemma 2.13] concludes the proof of the ${\mathbf{L^1}}$–contractivity. Numerical experiments {#sec:num} ===================== We present some applications of our result in test cases describing realistic road junction examples. The study is not exhaustive: in particular, specific cases of diverging junctions could be handled adding some information on drivers’ routing preferences upstream the junction. Yet, these situations go beyond the scope of this paper. In all the numerical experiments, we choose $$v_{d,j} (u) = V_d (1-u) \quad \hbox{for}~d=\ell, r~ \hbox{and}~j=1,\ldots,M,$$ thus the maximal speed is the same for all the lanes before, respectively after, $x=0$. In particular, in each situation we consider two cases, $V_\ell < V_r$ and $V_\ell > V_r$. 1-to-1 junction: from 2 to 3 lanes {#sec:incr} ---------------------------------- We consider problem ––, with ${{\mathcal{M}}}_\ell=\left\{1,2\right\}$, ${{\mathcal{M}}}_r=\left\{1,2,3\right\}$ and $S_{\ell,2}(u,w)=0$. (0,0) – (6,0) node\[pos=0.25, align=center, above\][lane 1]{} node\[pos=0.5, align=center, below\] [$x=0$]{}; (0,1) – (6,1) node\[pos=0.25, align=center, above\][lane 2]{}; (0,2) – (3,2) (3,2) |- (6,3); (3,2) – (6,2) node\[pos=0.5, align=center, above\][lane 3]{}; (3,0) – (3,2); The initial data are chosen as follows: $$\begin{aligned} \label{eq:27} \rho_{o,1} (x) = \ & 0.7, & \rho_{o,2} (x) = \ & 0.6, & \rho_{o,3} (x) = \ & 0.5 * {{\chi_{\strut[0,+\infty[}}} (x).\end{aligned}$$ Moreover, we choose $ V_\ell = 1.5$, and $V_{r}=1$ or $2$ respectively. Figure \[fig:2to3\] displays the solutions in both cases at time $t= 1$: on the right the maximal speed decreases, on the left it is increasing. We notice the effect of the flow between neighbouring lanes: all along the $x$-axis vehicles moves from lane 1 to lane 2, for $x>0$ vehicles pass also from lane 2 to lane 3, and this is particularly evident near $x=0$. ![Solutions to ––, with ${{\mathcal{M}}}_\ell=\left\{1,2\right\}$, ${{\mathcal{M}}}_r=\left\{1,2,3\right\}$ and initial data  at time $t=1$. $V_\ell = 1.5$: left $V_r=1$, right $V_r=2$.[]{data-label="fig:2to3"}](i2-50.png "fig:"){width="48.00000%"}\ 1-to-1 junction: from 3 to 2 lanes {#sec:decr2} ---------------------------------- We consider problem ––, with ${{\mathcal{M}}}_\ell=\left\{1,2,3\right\}$, ${{\mathcal{M}}}_r=\left\{1,2\right\}$ and $S_{r,2}(u,w)=0$. (0,0) – (6,0) node\[pos=0.25, align=center, above\][lane 1]{} node\[pos=0.5, align=center, below\] [$x=0$]{}; (0,1) – (6,1) node\[pos=0.25, align=center, above\][lane 2]{}; (0,3) -| (3,2) (3,2) – (6,2); (0,2) – (3,2) node\[pos=0.5, align=center, above\][lane 3]{}; (3,0) – (3,2); The initial data are chosen as follows: $$\begin{aligned} \label{eq:26} \rho_{o,1} (x) = \ & 0.7, & \rho_{o,2} (x) = \ & 0.6, & \rho_{o,3} (x) = \ & 0.5 \, {{\chi_{\strut]-\infty,0]}}} (x)+ 1 \,{{\chi_{\strut]0,+\infty[}}} (x).\end{aligned}$$ We choose $ V_\ell = 1.5$, and $V_{r}=1$ or $2$ respectively. Figure \[fig:3to2\] displays the solutions in both cases at time $t= 1$: on the right the maximal speed decreases, on the left it is increasing. We display the solution also for the positive part of the third lane: it is constantly equal to the maximal density $1$. As in the case of an increasing number of lanes, we notice the effect of the flow between neighbouring lanes. Observe that no vehicle passes from lane 3 to lane 2 for $x>0$: indeed, lane 3 for $x>0$ is a fictive lane and we impose  (${S_{r,2}} (u,w)=0$). Focus on the queue forming before $x=0$ and compare the two cases, $V_r<V_\ell$ and $V_r>V_\ell$. When the maximal speed diminishes, the queue is longer and the number of vehicles in the queue is greater with respect to the case of increasing maximal speed: for $x<0$, in the former case it is more difficult for vehicles in lane 3 to pass in lane 2, since here the decrease in the maximal speed diminishes the flow at $x=0$. 2-to-1 junction : from 3 to 2 lanes {#sec:32plud} ----------------------------------- We consider the same setting of Section \[sec:decr2\], thus problem ––, with ${{\mathcal{M}}}_\ell=\left\{1,2,3\right\}$, ${{\mathcal{M}}}_r=\left\{1,2\right\}$ and initial data , with the additional assumption that there is no flow of vehicles between the first and the second lane on $]-\infty,0[$, i.e. ${S_{\ell,1}} (u,w)=0$ (we keep $S_{r,2}(u,w)=0$): (0,0) – (6,0) node\[pos=0.25, align=center, above\][lane 1]{} node\[pos=0.5, align=center, below\] [$x=0$]{}; (0,1) – (3,1) node\[pos=0.5, align=center, above\][lane 2]{}; (0,3) -| (3,2) (3,2) – (6,2); (3,1) –(6,1) (0,2) – (3,2) node\[pos=0.5, align=center, above\][lane 3]{}; (3,0) – (3,2); We choose $V_\ell = 1.5$ and $V_r \in \{1, \, 1.5, \, 2\}$. Figure \[fig:3to2Plus\] displays the solution in the three cases at time $t=0.5$: on the right the maximal speed decreases, in the centre it stays constant, on the left it increases. As before, we display the solution also for the positive part of the third lane, where it is constantly equal to $1$. 2-to-1 junction: from 4 to 2 lanes {#sec:42} ---------------------------------- We consider the problem ––, with ${{\mathcal{M}}}_\ell=\left\{1,2,3,4 \right\}$, ${{\mathcal{M}}}_r=\left\{2,3\right\}$ and initial data $$\label{eq:28} \begin{aligned} \rho_{o,1} (x) = \ & 0.7 \, {{\chi_{\strut]-\infty,0]}}} (x)+ 1 \,{{\chi_{\strut]0,+\infty[}}} (x), & \rho_{o,2} (x) = \ & 0.5, \\ \rho_{o,3} (x) = \ & 0.6, & \rho_{o,4} (x) = \ & 0.4 \, {{\chi_{\strut]-\infty,0]}}} (x)+ 1 \,{{\chi_{\strut]0,+\infty[}}} (x), \end{aligned}$$ with the additional assumption that there is no flow of vehicles between the second and the third lane on $]-\infty,0[$, i.e. ${S_{\ell,2}} (u,w)=0$ (we also impose $S_{r,1}(u,w)=S_{r,3}(u,w)=0$). The situation under consideration looks as follows: (0,0) -| (3,1) (3,1) – (6,1) (0,2) – (3,2) node\[pos=0.5, align=center, below\][lane 2]{} node\[pos=0.5, align=center, above\][lane 3]{} (3,3) – (6,3) (0,4) -| (3,3); (0,1) – (3,1) node\[pos=0.5, align=center, below\][lane 1]{} (3,2) –(6,2) (0,3) – (3,3) node\[pos=0.5, align=center, above\][lane 4]{}; (3,1) – (3,3); (3,0) node\[below\] [$x=0$]{}; We choose $V_\ell = 1.5$ and $V_r \in \{1, \, 1.5, \, 2\}$. Figure \[fig:4to2\] displays the solution in the three cases at time $t=1$: on the right the maximal speed decreases, in the centre it stays constant, on the left it increases. As before, we display the solution also for the positive part of the first and fourth lane, where it is constantly $1$. **Acknowledgement:** The authors are grateful to Rinaldo M. Colombo for stimulating discussions.

ArXiv

--- abstract: 'We consider two types of fractional integral moduli of smoothness, which are widely used in theory of functions and approximation theory. In particular, we obtain new equivalences between these moduli of smoothness and the classical moduli of smoothness. It turns out that for fractional integral moduli of smoothness some pathological effects arise.' author: - 'Yurii Kolomoitsev$^{1}$' title: | On moduli of smoothness and\ \[3pt\] averaged differences of fractional order --- [^1] [^2] Introduction {#sec:1} ============ Let $\T\cong[0,2\pi)$ be the circle. As usual, the space $L_p(\T)$, $0<p<\infty$, consists of measurable functions which are $2\pi$-periodic and $$\Vert f\Vert_p=\bigg(\frac1{2\pi}\int_{\T}|f(x)|^p \mathrm{d}x\bigg)^{\frac 1p}<\infty.$$ For simplicity, by $L_\infty(\T)$ we denote the space of all $2\pi$-periodic continuous functions on $\T$ which is equipped with the norm $$\Vert f\Vert_\infty=\max_{x\in\T}|f(x)|.$$ The classical (fractional) modulus of smoothness of a function $f\in L_p(\T)$, $0< p\le\infty$, of order $\b>0$ and step $h>0$ is defined by $$\label{eqmod1} \w_\b(f,h)_p=\sup_{0<\d<h} \Vert \D_\d^\b f\Vert_p,$$ where $$\D_\d^\b f(x)=\sum_{\nu=0}^\infty\binom{\b}{\nu}(-1)^\nu f(x+\nu\d),$$ $ \binom{\b}{\nu}=\frac{\b (\b-1)\dots (\b-\nu+1)}{\nu!},\quad \binom{\b}{0}=1. $ For solving particular problems, the usage of the classical modulus of smoothness may be technically very difficult or its application cannot give sharp and meaningful results. Therefore, it arises the necessity to employ modifications of the classical moduli of smoothness. Thus, in the papers [@DHI], [@K11], [@K], [@Pu], [@Rup], [@T2], [@Tumb] different means of averaging of finite differences and their modifications have been studied and applied for solving several problems of approximation theory. As a rule, to construct such modifications (special moduli of smoothness) one replaces the shift operator $\tau_h f(x)=f(x+h)$ by some smoothing operator, for example, by the Steklov means. In this paper, we consider two types of special moduli of smoothness given by the following formulas: $$\label{eqmod2} {\rm w}_\b(f,h)_p=\(\frac 1h\int_0^h \Vert \D_\d^\b f\Vert_p^{p_1} {\rm d}\d\)^\frac1{p_1}$$ and $$\label{eqmod3} \widetilde{\w}_\b(f,h)_p=\bigg\Vert\frac 1h \int_0^h \D_\d^\b f(\cdot) {\rm d}\d \bigg\Vert_p,$$ where $\b>0$, $h>0$, and $p_1=\min(1,p)$. Sometimes, the moduli  and  are called the integral moduli of smoothness or averaged differences. The modulus ${\rm w}_\b(f,h)_p$ is well known and it has been often applied for solving different problems of approximation theory, see e.g. [@DL Ch. 6, §5], [@DHI], [@Pu], [@Rup]. The modulus $\widetilde{\w}_\b(f,h)_p$ has been introduced and studied in [@Tvinity], see also [@Tumb] and [@TB Ch.8], in which it is also called the linearized modulus of smoothness. Some applications of the modulus $\widetilde{\w}_\b(f,h)_p$ as well as some of its modifications can be found in [@Ak], [@KS], [@K], [@KT], and [@T2]. Note that has sense for all function $f\in L_p(\T)$ with $0<p\le \infty$, while can be defined only for integrable functions $f$. At the same time, has some advantages. One of them concerns the direct application of the method of Fourier multipliers. Recall that a numerical sequence $\{\l_k\}_{k\in\Z}$ is a Fourier multiplier in $L_p(\T)$, if for all functions $f\in L_p(\T)$, $1\le p\le\infty$, the series $$\sum_{k\in\Z} \l_k \widehat{f}_k e^{\i kx},\quad \widehat{f}_k=\frac1{2\pi}\int_0^{2\pi} f(x)e^{-\i kx}{\rm d}x,$$ is the Fourier series of a function $\Lambda f\in L_p(\T)$ and $$\Vert\{\l_k\} \Vert_{M_p}=\Vert \Lambda\Vert_{L_p\mapsto L_p}=\sup_{\Vert f\Vert_p\le 1}\Vert \Lambda f\Vert_p < \infty$$ (see, e.g., [@SW Ch. I and Ch. VI]). Let us illustrate how the method of Fourier multipliers can be used in relation to the modulus $\widetilde{\w}_\b(f,h)_p$. It is easy to see that the Fourier series of the averaged difference $\frac 1h \int_0^h \D_\d^\b f(x){\rm d}\d$ can be written as follows $$\label{mainf0} \sum_{k\in\Z} \psi_\b(k h)\widehat{f}_k e^{\i kx},$$ where the function $\psi_\b$ is defined by $$%\label{mainf} \psi_\b(t)=\int_0^1 (1-e^{\i t\vp})^\b {\rm d}\vp$$ (here and throughout, we use the principal branch of the logarithm). Thus, if we want to obtain, for example, an inequality of the form $$\bigg\Vert \sum_{k\in\Z} \l_k(h) \widehat{f}_k e^{\i kx} \bigg\Vert_p\le C(p,\b)\widetilde{\w}_\b(f,h)_p,$$ we need only to verify that the sequence $\{\l_k(h)/\psi_\b(hk)\}_{k\in\Z}$ is a Fourier multiplier in $L_p(\T)$ and $$\sup_{h>0} \bigg\Vert \bigg\{\frac{\l_k(h)}{\psi_\b(kh)}\bigg\}\bigg\Vert_{M_p}\le C(p,\b),$$ see also Lemma \[lem3\] below. This method has been used, e.g., in [@K] and [@T2], see also [@TB Ch. 8]. In approximation theory, it is important to ascertain whether a special modulus of smoothness is equivalent to the classical modulus of smoothness $\w_\b(f,h)_p$. For the moduli of smoothness and with $\b\in \N$, this problem is well studied. In particular, for all $f\in L_p(\T)$, $1\le p\le \infty$, $\b\in \N$, and $h>0$, we have $$\label{equivmmm} {\rm w}_\b(f,h)_p\asymp {\w}_\b(f,h)_p\asymp\widetilde{\w}_\b(f,h)_p,$$ where $\asymp$ is a two-sided inequality with positive constants independent of $f$ and $h$. The equivalence ${\rm w}_\b(f,h)_p\asymp {\w}_\b(f,h)_p$, which also holds in the case $0<p<1$, was proved in [@DL p. 185], see also [@Pu Theorem 1] and [@Rup Theorem 3.1]. The proof of the second equivalence in  can be found in [@Tumb] (see also [@TB Theorem 8.4.1] for similar results in the Hardy spaces $H_p$). The main purpose of this paper is to investigate relations  for positive $\b\not \in \N$. Is easy to see that for any $\b>0$ and $1\le p\le \infty$, by Minkovsky’s inequality and trivial estimates, we have $$\label{lemmm} \widetilde{\w}_\b(f,h)_p\le {\rm w}_\b(f,h)_p\le {\w}_\b(f,h)_p.$$ Concerning the inverse inequalities, we have an unexpected result. In Theorem \[th3\] below, we show that there exist $f_0(x)\not\equiv\const$, $\b_0>0$, and $h_0>0$ such that $$\label{0} \widetilde{\w}_{\b_0}(f_0,h_0)_p=0.$$ Since $\w_\b(f,h)_p>0$ for all $f(x)\not\equiv\const$ and $h>0$,  implies that for any $C>0$ and particular $\b>0$ the inequality $$%\label{1} {\w}_\b(f,h)_p\le C\widetilde{\w}_\b(f,h)_p$$ does not hold in general. At the same time, we have a standard situation for ${\rm w}_\b(f,h)_p$: for all $f\in L_p(\T)$, $0<p\le \infty$, $\b>0$, and $h>0$ $${\w}_\b(f,h)_p\le C(p,\b){\rm w}_\b(f,h)_p$$ (see Theorem \[th1\] below). In the paper, we propose several ways to overcome the pathological property . In particular, we show that the following modification of  can be used instead of the modulus $\widetilde{\w}_\b(f,h)_p$, $$\w_\b^*(f,h)_p=\bigg\Vert\frac 1{h^2} \int_0^h \int_0^h \D_{\d_1}^{\lfloor\b\rfloor} \D_{\d_2}^{\{\b\}} f(\cdot) {\rm d}\d_1\, {\rm d}\d_2\bigg\Vert_p,$$ where $\lfloor\b\rfloor$ and $\{\b\}$ are the floor and the fractional part functions of $\b$, respectively. In Theorem \[th5\] below, we prove that $\w_\b^*(f,h)_p$ is equivalent to the classical modulus of smoothness for all $\b>0$. At the same time, $\w_\b^*(f,h)_p$ is a convenient modulus in the sense of applications of Fourier multipliers. Finally, we note that property  seems to be very unnatural for moduli of smoothness. However, even in the study of the approximation of functions by some classical methods, for example by Bernstein-Stechkin polynomials, it has been arisen the necessity to construct special moduli of smoothness for which a condition of type holds (see [@T2]). One has a similar situation for some non-classical methods of approximation (see [@KT]). The paper is organized as follows. In Section 2, we present the main results. In Section 3, we formulate and prove auxiliary results. In Section 4, we prove the main results of the paper. We denote by $C$ some positive constant depending on the indicated parameters. As usual, $A(f, h)\asymp B(f, h)$ means that there exists a positive constant $C$ such that $C^{-1} A(f,h)\le B(f,h)\le C A(f,h)$ for all $f$ and $h$. Main results {#sec:2} ============ We start from the modulus ${\rm w}_\b(f,h)_p$, for which we have a quite standard result in the following theorem. \[th1\] Let $f\in L_p(\T)$, $0<p\le\infty$, $\b\in \N\cup (1/p_1-1,\infty)$, and $h\in (0,1)$. Then $$%\label{eqthEqmods1} {\rm w}_\b(f,h)_p\asymp {\w}_\b(f,h)_p\,.$$ The next theorem is the key result of the paper. \[th3\] There exists a set $\{\b_k\}_{k=0}^\infty$ such that $\b_0\in (4,5)$, $\b_k\to\infty$ as $k\to\infty$, and the following assertions hold: $(i)$ for each function $f\in L_p(\T)$, $1\le p\le\infty$, and for all $\b\in (0,\b_0)\cup \N$ and $h\in (0,1)$ we have $$\widetilde{\w}_\b(f,h)_p\asymp \w_\b(f,h)_p\,;$$ $(ii)$ for each $k\in \Z_+$ there exists $t_k>2\pi$ such that for any $n\in \Z\setminus \{0\}$ we have $$\frac 1h\int_0^{h} \D_\d^{\b_k} e_n(x){\rm d}\d=0,\quad x\in \T,$$ where $e_n(x)=e^{inx}$ and $h=t_k/|n|$. In particular, for all $0<p\le \infty$, we have $$\label{02} \widetilde{\w}_{\b_k}(e_n,h)_p=0\,.$$ Theorem \[th3\] implies that, for small values of $\b$, the modulus  has the same properties as the classical modulus of smoothness , but for some $\b >4$ the modulus  has a pathological behaviour. In the next results, we show several ways to overcome the effect of . In particular problems of approximation theory, it is enough to know the behaviour of moduli of smoothness on a set of trigonometric polynomials. It turns out that one can reduce the influence of property  in such situation. Let $\mathcal{T}_n$ be the set of all trigonometric polynomials of order at most $n$, $$\mathcal{T}_n=\bigg\{T(x)=\sum_{\nu=-n}^n c_\nu {\rm e}^{\i\nu x}~:~ c_\nu \in \C\bigg\}.$$ \[th4\] Let $1\le p\le\infty$, $\b>0$, and $n\in\N$. Then for each $T_n\in \mathcal{T}_n$ and for each $h\in (0,1/n)$ we have $$\widetilde{\w}_\b(T_n,h)_p\asymp \w_\b(T_n,h)_p.$$ Theorem \[th4\] is also true in the case $0<p<1$. This follows from the proof of Theorem \[th4\] presented below and the corresponding results in [@K07]. Another way to reduce the effect of  is adding to $\widetilde{\w}_\b(f,h)_p$ a quantity that has a better behaviour than the classical modulus of smoothness. For this purpose, one may use the error of the best approximation. As usual, the error of the best approximation of a function $f$ in $L_p$ by trigonometric polynomials of order at most $n$ is given by $$E_n(f)_p=\inf_{T\in \mathcal{T}_n}\Vert f-T\Vert_p.$$ Recall the well-known Jackson inequality: [*for all $f\in L_p$, $0< p\le\infty$, $r\in \N$, and $n\in\N$ we have $$\label{eqI12} E_n(f)_p\le C\w_r\(f,\frac 1n\)_p,$$ where $C$ is a constant independent of $f$ and $n$*]{} (see [@St51] for $1\le p\le\infty$ and [@SO] for $0<p<1$; see also [@I], [@Rup], and [@SKO]). \[thbest\] Let $f\in L_p(\T)$, $1\le p\le\infty$, $\b>0$, and $h\in (0,1)$. Then $$%\label{eqthEqmods2} {\w}_\b(f,h)_p\asymp \widetilde{\w}_\b(f,h)_p+E_{[1/h]}(f)_p\,.$$ Finally, let us consider a modification of . Let $\b>\a>0$. Denote $$%\label{eqModMod} \w_{\b;\, \a}^*(f,h)_p=\bigg\Vert\frac 1{h^2} \int_0^h \int_0^h \D_{\d_1}^{\b-\a} \D_{\d_2}^{\a} f(\cdot) {\rm d}\d_1 {\rm d}\d_2\bigg\Vert_p.$$ It is easy to see that $\b$ is the main parameter in the above modulus of smoothness. \[th5\] Let $f\in L_p(\T)$, $1\le p\le\infty$, and $\b>0$ and $\a \in (0,4]$ be such that $\b-\a\in \Z_+$. Then for all $h\in (0,1)$ we have $$%\label{ab0} \w_{\b;\, \a}^*(f,h)_p\asymp \w_\b(f,h)_p\,.$$ Auxiliary results ================= Properties of the differences and moduli of smoothness. ------------------------------------------------------- Let us recall several basic properties of the differences and moduli of smoothness of fractional order (see [@But; @DL; @samko; @tabeR]). For $f,f_1,f_2\in L_p(\T)$, $0< p\le \infty$, $h>0$, and $\a,\b \in\N\cup \(1/{p_1}-1,\infty\)$, we have - $\D_h^\b(f_1+f_2)(x) = \D_h^{\b} f_1(x) + \D_h^{\b} f_2(x)$; - $\D_h^\alpha(\D_h^\beta f)(x) = \D_h^{\alpha+\beta} f(x)$; - $\|\D_h^{\b} f\|_{p} \le C(\b,p)\|f\|_{p}$; <!-- --> - $ \omega_\beta(f,\delta)_p$ is a non-negative non-decreasing function of $\delta$ such that $\lim_{\delta\to 0+} \omega_\beta(f,\delta)_p=0;$ - $ \omega_\beta(f_1+f_2,\delta)_p\le 2^{\frac1{p_1}-1} \big( \omega_\beta(f_1,\delta)_p+\omega_\beta(f_2,\delta)_p\big); $ - $ \omega_{\alpha+\beta}(f,\delta)_p\le C(\alpha,p) \omega_{\beta}(f,\delta)_p$; - for $\lambda\ge 1$ and $\b\in\N$, $$\omega_{\beta}(f,\lambda \delta)_p\le C(\beta,p) \lambda^{\beta+\frac1{p_1}-1} \omega_{\beta}(f,\delta)_p.$$ We will use the following Boas type inequality. \[lemNS\] Let $0<p\le\infty$, $\beta>0$, $n\in\mathbb{N}$, and $0<\delta, h\le {\pi}/{n}$. Then for each $T_n\in\mathcal{T}_n$ we have $$%\label{ineqNS3} h^{-\beta}\Vert\Delta_h^\beta T_n\Vert_p\asymp\delta^{-\beta}\Vert\Delta_\delta^\beta T_n\Vert_p,$$ where $\asymp$ is a two-sided inequality with absolute constants independent of $T_n$, $h$, and $\d$. In the case $1\le p \le\infty$, Lemma \[lemNS\] follows from [@timan 4.8.6 and 4.12.18] (for integer $\beta$), and [@tabeR] (for any positive $\beta$). In the case $0<p<1$, it follows from [@DHI] (for integer $\beta$) and from [@K07] (for any positive $\beta$). Properties of the function $\psi_\b(t)$. ---------------------------------------- Everywhere below we set $$z_\b(t)=t\psi_\b(t)=\int_0^t (1-e^{\i\vp})^\b {\rm d}\vp.$$ The next lemma is a key result for proving Theorem \[th3\] $(i)$. \[lem0\] [(See [@Tumb]).]{} Let $\b>0$ and $\psi_\b(t)\neq 0$ for $t\in\R\setminus \{0\}$. Then for each function $f\in L_p(\T)$, $1\le p\le\infty$, and $h>0$ we have $${\w}_\b(f,h)_p\le C(\b)\widetilde{{\w}}_\b(f,h)_p.$$ The following result was proved in [@TB 8.3.5 b)] by using Lindemann’s classical theorem about the transcendence of values of the exponential function. \[lemtran\] If $\b\in\N$, then $\psi_\b(t)\neq 0$ for all $t\in \R\setminus \{0\}$. Thus, combining Lemma \[lemtran\] and Lemma \[lem0\], we get the proof of Theorem \[th3\] $(i)$ for $\b\in \N$. Below, we obtain some unexpected properties of $\psi_\b$ for non-integer $\b$. The following lemma is the main auxiliary result for proving Theorem \[th3\] in the case $\b\not\in \N$. \[lem1\] There exists a set $\{\b_k\}_{k=0}^\infty$ such that $\b_0\in (4,5)$, $\b_k\to\infty$ as $k\to\infty$, and - for all $\b\in (0,\b_0)$ we have $$\label{eq0.lem1} z_\b(t)\neq 0\quad\text{for}\quad t\in \R\setminus \{0\};$$ - for each $k\in \Z_+$ there exists $t_k>2\pi$ such that $z_{\b_k}(t_k)=0$; - for all $\b\in (0,\infty)$ we have $$%\label{eq0.lem1 iii} z_\b(t)\neq 0\quad\text{for}\quad 0<|t|<\pi.$$ *[The proof of $(i)$.]{}* First let us derive basic properties of the function $z_\b$. By simple calculation, we get $$\label{eq.zv} x_\b(t)=\Re z_\b(t)=t+\sum_{\nu=1}^\infty\binom{\b}{\nu} (-1)^\nu\frac{\sin \nu t}\nu$$ and $$y_\b(t)=\Im z_\b(t)=\sum_{\nu=1}^\infty\binom{\b}{\nu} (-1)^\nu\frac{1-\cos \nu t}\nu.$$ These equalities imply that $$\label{+++} x_\b(-t)=-x_\b(t)\quad\text{and}\quad y_\b(-t)=y_\b(t),\quad t\in\R.$$ Thus, to prove the lemma, it is enough to consider only the case $t>0$. It is easy to see that for $t\in [0,2\pi]$ we have $$\label{eq1.lem1} x_\b(t)=2\pi-x_\b(2\pi-t)=x_\b(2\pi+t)-2\pi,$$ $$\label{eq2.lem1} y_\b(t)=y_\b(2\pi-t)=y_\b(2\pi+t),$$ and $$%\label{eq3.lem1} x_\b(\pi k)=\pi k,\quad y_\b(2\pi k)=0,\quad k\in\Z_+.$$ Denote $$\g_{\b,k}=\{z_\b(t),\, 2\pi k\le t< 2\pi (k+1)\}\quad\text{and}\quad \g_{\b}=\bigcup_{k=0}^\infty \g_{\b,k}.$$ Equalities (\[eq1.lem1\]) and (\[eq2.lem1\]) imply that the curve $\g_{\b,0}$ is symmetric with respect to the line $x=\pi$ on the complex plane $\C$. We also have that $$\label{eq4.lem1} \g_{\b,k+1}=\g_{\b,k}+2\pi,\quad k\in\Z_+.$$ See on Figure 1 the form of the curves $\g_{\b,k}$ in the cases $\b=4$ and $k=0,1,2$. Note that for $t\in [0,2\pi]$ the following equalities hold $$\label{eq5.lem1} x_\b(t)=\frac 2\b\int_{-\b\pi/2}^{\b(t-\pi)/2} \cos\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp$$ and $$\label{eq6.lem1} y_\b(t)=\frac 2\b\int_{-\b\pi/2}^{\b(t-\pi)/2} \sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp.$$ Therefore, $$x_\b'(t)=\cos\frac{\b(t-\pi)}{2}\left(2\sin\frac t2\right)^\b \quad \text{and} \quad y_\b'(t)=\sin\frac{\b(t-\pi)}{2}\left(2\sin\frac t2\right)^\b.$$ Below, these two equalities will be often used to indicate intervals of monotonicity of the functions $x_\b$ and $y_\b$. By Lemma \[lemtran\], we have that if $\b\in\N$, then $z_\b(t)\neq 0$ for all $t\neq 0$. Now, we show that $z_\b(t)\neq 0$ for $\b\in (0,4)\setminus \N$ and $t>0$. We split the proof of this fact into several cases. The simplest case is $\b \in(0,1)$. Indeed, by (\[eq5.lem1\]) and (\[eq4.lem1\]), we have that $x_\b(t)>0$ for $t>0$. This obviously implies . The next case $\b\in (1,2)$ is also simple. In this case, by (\[eq6.lem1\]), (\[eq2.lem1\]), and (\[eq4.lem1\]), we get that $y_{\b}(t)<0$ for all $t\in\R_+\setminus \{2\pi k\}_{k\in\Z_+}$, and $x_{\b}(2\pi k)=2\pi k$, $k\in\Z_+$. Hence, $z_\b(t)\neq 0$ for all $t>0$. In what follows we deal only with the case $\b\in (3,4)$. The proof of the lemma in the case $\b\in (2,3)$ is similar to the arguments presented below. First, let us consider the curve $\g_{\b,0}$ for $\b\in (3,4)$. Let $$I_1=(0,\pi(1-3/\b)]$$ and $$I_j=(\pi(1-(5-j)/\b),\pi(1-(4-j)/\b)],\quad j=2,3,4.$$ We are going to show that for each $j\in\{1,2,3,4\}$ $$\label{eq7.lem1} z_\b(t)\neq 0,\,\,\, \text{ $t\in I_j$}.$$ *1) [The case $j=1$.]{}* In this case, (\[eq7.lem1\]) easily follows from the fact that the functions $x_\b$ and $y_\b$ are strictly increasing and positive on $I_1$. *2) [The case $j=2$.]{}* We have $$%\label{eq10.lem1} \begin{split} x_\b\(\pi\(1-\frac2\b\)\)&= \frac 2\b\int_{-\b\pi/2}^{-\pi} \cos\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp \\ &<\frac 2\b \left(2\cos \frac{3\pi}{2\b}\right)^\b \sin \frac{\pi\b}2<0. \end{split}$$ Thus, the function $x_\b$ is strictly decreasing and changes the sign from “$+$” to “$-$” on $I_2$. At the same time, the function $y_\b$ is strictly increasing and positive on $I_2$. The last fact implies (\[eq7.lem1\]) for $j=2$. *3) [The case $j=3$.]{}* Let us show that $ %\label{eq11.lem1} y_\b\(\pi\(1-\frac1\b\)\)<0. $ We have $$\begin{split} \frac\b 2 y_\b&\(\pi\(1-\frac1\b\)\)=\int_{-\b\pi/2}^{-\pi/2} \sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &=\left\{\int_{-\b\pi/2}^{-3\pi/2}+\int_{-3\pi/2}^{-\pi}+\int_{-\pi}^{-\pi/2}\right\}\sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &=S_1+S_2+S_3. \end{split}$$ One can estimate the integrals $S_i$, $i=1,2,3$, by the following way: $$S_1<\left(2\cos \frac {3\pi}{2\b}\right)^\b \cos \frac {\pi\b}{2}<\left(2\cos \frac {3\pi}{8}\right)^3=(2-\sqrt{2})^{3/2},$$ $$\begin{split} S_2&=\left\{\int_{-3\pi/2}^{-5\pi/4}+\int_{-5\pi/4}^{-\pi}\right\}\sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &<\left(2\cos \frac {5\pi}{16}\right)^4\frac1{\sqrt 2}+\left(2\cos \frac {\pi}{\b}\right)^\b(1-1/{\sqrt 2}), \end{split}$$ and $$\begin{split} S_3&=\left\{\int_{-\pi}^{-3\pi/4}+\int_{-3\pi/4}^{-\pi/2}\right\}\sin\vp \left(2\cos \frac\vp\b\right)^\b{\rm d}\vp\\ &<\left(2\cos \frac {\pi}{\b}\right)^\b(1/{\sqrt 2}-1)-\left(2\cos \frac {\pi}{4}\right)^3\frac1{\sqrt 2}. \end{split}$$ Combining the above estimates for $S_i$, we get $$y_\b\(\pi\(1-\frac1\b\)\)<\frac 23\bigg[(2-\sqrt{2})^{3/2}+\frac 1{\sqrt{2}}\bigg(\left(2\cos \frac {5\pi}{16}\right)^4-2^{3/2}\bigg)\bigg]<0.$$ We have that the functions $x_\b$ and $y_\b$ are strictly decreasing, the function $y_\b$ changes the sign from “$+$” to “$-$”, but $x_\b$ is negative on $I_3$. Combining these facts, we get that (\[eq7.lem1\]) holds for $j=3$. *4) [The case $j=4$.]{}* We have that the function $y_\b$ is strictly decreasing and negative on $I_4$, the function $x_\b$ is strictly increasing and changes the sign from “$-$” to “$+$” on this interval. Therefore, $z_\b(t)\neq 0$ for all $t\in I_4$. Thus, we have shown that $z_\b(t)\neq 0$ for all $t\in (0,\pi]$. Taking into account that the curve $\g_{\b,0}$ is symmetric with respect to the line $x=\pi$, we get that $z_\b(t)\neq 0$ for all $t\in (\pi,2\pi]$, too. Now let us consider the curves $\g_{\b,k}$ for $k\in\N$. In view of , to finish the proof of part $(i)$ we need to show that $$\label{eq12.lem1} \g_{\b,1}\in \{z\in\C\,:\, \Re z>0\}.$$ Note that $\g_{\b,1}=\g_{\b,0}+2\pi$. Thus, to verify (\[eq12.lem1\]) we only need to investigate the extremal points of the function $x_\b(t)$ for all $\b\in (3,4)$ and $2\pi\le t\le 4\pi$. From the results obtained above, it follows that the points $\tau_0=\pi(3-3/\b)$, $\tau_1=\pi(3-1/\b)$, $\tau_2=\pi(3+1/\b)$, and $\tau_3=\pi(3+3/\b)$ are extremal for $x_\b(t)$. It is easy to see that it is enough to consider $x_\b$ at the points $\tau_1$ and $\tau_3$ (see Figure 1). {width="9cm"} [**Figure 1.**]{} Simple estimates show that $x_\b(\tau_3)=4\pi-x_\b(\pi(1-3/\b))>0$. We also have $$\label{eq13.lem1} x_\b(\tau_1)=2\pi+x_\b\(\pi\(1-\frac1\b\)\).$$ Using , we get $$\label{dopp} \begin{split} x_\b\(\pi\(1-\frac1\b\)\)&=\left\{\int_{-\pi\b/2}^{-3\pi/2}+\int_{-3\pi/2}^{-\pi/2}\right\}\frac 2\b\cos\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &=S_1+S_2. \end{split}$$ It is evident that $S_1>0$. The integral $S_2$ can be estimated by the following way $$\label{eq14.lem1} \begin{split} S_2&=\frac 2\b \sum_{k=2}^5 \int_{-\pi(k+1)/4}^{-\pi k/4}\cos \vp\left( 2\cos\frac\vp\b\right)^\b {\rm d}\vp\\ &>\frac 2\b \sum_{k=2}^5 \left( 2\cos\frac{\pi k}{16}\right)^\b\int_{-\pi(k+1)/4}^{-\pi k/4}\cos \vp {\rm d}\vp\\ &=-\frac 2\b \bigg[\bigg(1-\frac1{\sqrt 2}\bigg)\bigg\{\left( 2\cos\frac{5\pi}{16}\right)^\b+\left( 2\cos\frac{\pi }{8}\right)^\b\bigg\}\\ &\quad\quad\quad\quad\quad\quad\quad\quad\quad+ \frac1{\sqrt 2}\bigg\{\left( 2\cos\frac{\pi}{4}\right)^\b+\left( 2\cos\frac{3\pi }{16}\right)^\b\bigg\}\bigg]. \end{split}$$ From , taking into account that the function $f(x)={a^x}/{x}$ is increasing for $a>1$ and $x>1$, we get $$\begin{split} S_2>-\frac 12 \bigg[\bigg(1-\frac1{\sqrt 2}\bigg)\bigg\{&\left( 2\cos\frac{5\pi}{16}\right)^4+(\sqrt{2}+2)^2\bigg\}\\ &+ \frac1{\sqrt 2}\bigg\{2^2+\left( 2\cos\frac{3\pi }{16}\right)^4\bigg\}\bigg]>-2\pi. \end{split}$$ The last estimate together with (\[dopp\]) and (\[eq13.lem1\]) implies that $x_\b(\tau_1)>0$. Therefore, we have (\[eq12.lem1\]). Finally, combining (\[eq12.lem1\]) and (\[eq4.lem1\]), we prove the first part of Lemma \[lem1\]. *[The proof of $(ii)$.]{}* To prove the second part of the lemma, it is enough to investigate the curve $$\Gamma_{\b,1}=\{z_\b(t)\,:\, \pi(3-2/\b)\le t\le 3\pi\}\quad\text{for}\quad \b\in[4,5]$$ (see Figure 2 below). By analogy with the proof of the first part of the lemma, taking into account the equality $$\Gamma_{\b,1}=\{z_\b(t)\,:\, \pi(1-2/\b)\le t\le \pi\}+2\pi,$$ we get $$\label{eqDop.lem1} y_\b(\pi(3-2/\b))>0\quad\text{and}\quad y_\b(3\pi)<0\quad\text{for all}\quad \b\in[4,5].$$ Moreover, the functions $x_\b$ and $y_\b$ are strictly decreasing on $[\pi(3-2/\b),\\\pi(3-1/\b))$. At the same time, we see that the function $x_\b$ is strictly increasing and $y_\b$ is strictly decreasing on $[\pi(3-1/\b),3\pi]$. Combining these facts, we have that the curve $\Gamma_{\b,1}$ intersects the line $y=0$ only once. Thus, one can define the function $\vt:[4,5]\mapsto (\pi(3-2/\b),3\pi)$ by the rule $$\label{*} {y}_\b|_{[\pi(3-2/\b),3\pi]}(\vt(\b))=0,$$ where ${y}_\b|_A$ denotes a restriction of the function $y_\b$ on some set $A$. Let us consider the function $$F(\b)=x_\b(\vt(\b)).$$ We need to verify that $F(\b)$ is a continuous function on the interval $(4,5)$. First, let us show that the function $\vt(\b)$ is continuous on $(4,5)$. Let $4\le\b'<\b''\le 5$. Without loss of generality, we can suppose that $\vt(\b')<\vt(\b'')$. Using , , and , we obtain $$\label{**} \begin{split} y_{\b'}(\vt(\b''))-y_{\b''}(\vt(\b''))&=y_{\b'}(\vt(\b''))\\ &=y_{\b'}(\vt(\b''))-y_{\b'}(\vt(\b'))\\ &=y_{\b'}(\vt(\b'')-2\pi)-y_{\b'}(\vt(\b')-2\pi)\\ &=\frac2{\b'}\int_{\frac{\b'}2(\vt(\b')-3\pi)}^{\frac{\b'}2(\vt(\b'')-3\pi)}\sin \vp \left(2\cos\frac{\vp}{\b'}\right)^{\b'}{\rm d}\vp. \end{split}$$ By  and the mean value theorem, there exists a point $\xi\in({\b'}(\vt(\b')$ $-3\pi)/2, {\b'}(\vt(\b'')-3\pi)/2)$ such that $$\label{oo} y_{\b'}(\vt(\b''))-y_{\b''}(\vt(\b''))=\sin \xi \left(2\cos\frac{\xi}{\b'}\right)^{\b'}\(\vt(\b'')-\vt(\b')\).$$ By  and continuity of $y_\b(t)$, there exists a sufficiently small $\d>0$ such that $ %\label{o} \pi\(3-2/\b\)+\d<\vt(\b)<3\pi-\d. $ Therefore, we get $$\label{o} -\pi+2\d<-\pi+\frac{\b' \d}{2}<\xi<-\frac{\b' \d}{2}<-2\d.$$ Taking into account these inequalities, we obtain $$\label{ooo} \left| \sin \xi \left(2\cos\frac{\xi}{\b'}\right)^{\b'} \right|> \sin(2\d)\left(2\cos\(\frac\pi4-\frac\d2\)\right)^{4}=C(\d)>0.$$ Thus, by  and , we have $$\label{oooo} |\vt(\b'')-\vt(\b')|<\frac1{C(\d)}|y_{\b'}(\vt(\b''))-y_{\b''}(\vt(\b''))|.$$ Since $y_\b(t)$ is a continuous function of $\b$, from , we get that the function $\vt(\b)$ is also continuous on $(4,5)$. Now, taking into account that $x_\b(t)$ is continuous on $(4,5)\times (2\pi,3\pi)$, we get that the function $F(\b)$ is continuous on $(4,5)$. Next, from , we get that $F(4)>0$. Thus, if we show that $F(5)<0$, then, by the intermediate value theorem, we can find $\b_0\in (4,5)$ such that $F(\b_0)=0$. This and  will imply that $z_{\b_0}(f(\b_0))=0$. Indeed, it is easy to see that the functions $x_5$ and $y_5$ are strictly decreasing on $(13\pi/5,3\pi)$. Thus, to prove the existence of a point $t_0\in (13\pi/5,3\pi)$ such that $$%\label{eq2.lem2} x_{5}(t_0)<0\quad\text{and}\quad y_{5}(t_0)=0,$$ it is enough to verify that the origin of $\Gamma_{5,1}$ lies in the first quadrant, the extremal point of the curve (relating to the extremal point of the function $x_5$) lies in the third quadrant, and there exists an intermediate point that lies in the second quadrant. {width="7cm"} **Figure 2.** Performing simple calculation, we can verify that the origin of the curve $\Gamma_{5,1}$ takes the value $ z_5\left({13\pi}/5\right)\approx 3.622+{\rm i}2.327, $ that is it belongs to the first quadrant; since $ z_5\left({14\pi}/5\right)\approx -2.803-{\rm i}5.632, $ the point $z_5(14\pi/5)$, which is the extremal point of the curve $\Gamma_{5,1}$, belongs to the third quadrant; since $ z_5\left({27\pi}/{10}\right)\approx -0.413+{\rm i}0.504, $ the intermediate point $z_5(27\pi/10)$ belongs to the second quadrant (see also Figure 2 in which $4<\b<4.5$ and $\b_0\approx 4.85$). Now, let us prove assertion $(ii)$ for any $k\in \N$. First, we show that for each $\b>4$ $$\label{verif} y_\b\(\pi \(1-\frac 2\b\)\)>0\quad\text{and}\quad y_\b\(\pi\(1-\frac1\b\)\)<0.$$ Let us verify the first inequality. The second one can be proved similarly. Let $k_0$ be the smallest natural number such that ${\b \pi}/2\ge (2k_0+1)\pi$. Then, using , we have $$\begin{split} &\frac \b 2 y_\b\(\pi \(1-\frac 2\b\)\)\\ &=\left\{\int_{-\frac{ \b \pi}{2}}^{-\pi(2k_0+1)}+\sum_{k=k_0}^1 \int_{-(2k+1)\pi}^{-(2k-1)\pi} \right\} \sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &=S_1+S_2. \end{split}$$ It is evident that $S_1\ge 0$. Thus, to verify , it is enough to show that $S_2>0$. But this easily follows from the fact that for each $k\in\N$ we have $$\begin{split} \int_{-(2k+1)\pi}^{-(2k-1)\pi}&\sin\vp \left(2\cos \frac \vp\b\right)^\b {\rm d}\vp\\ &> \(2 \cos\frac{2\pi k}{\b}\)^\b \left\{\int_{-(2k+1)\pi}^{-2k \pi}+\int_{-2k\pi}^{-(2k-1) \pi}\right\}\sin\vp {\rm d}\vp=0. \end{split}$$ By analogy with the proof of part $(i)$, we see that the function $y_\b$ is strictly decreasing and changes the sign from “$+$” to “$-$” on $(\pi(1-2/\b),\pi)$ for all $\b>5$. Therefore, as above, for any $\b>5$ one can define a function $\t : [5,\infty)\mapsto (\pi\(1-2/\b\),\pi\(1-1/\b\))$ by the rule $$\label{*1} {y}_\b|_{(\pi\(1-2/\b\),\pi\(1-1/\b\))}(\t(\b))=0.$$ Let us show that for $n\in\N$ one has $$\label{eqkvad} x_{2n}\(\pi\(1-\frac 1{n}\)\)\to -\infty\quad\text{as}\quad n\to\infty.$$ Indeed, (\[eq.zv\]) implies that $$\label{eq.prN1} x_{2n}\(\pi\(1-\frac 1{n}\)\)=\pi\(1-\frac 1{n}\)+\sum_{\nu=1}^{2n} \binom {2n}{\nu} \frac{(-1)^\nu}{\nu} \sin \pi \(1-\frac 1n\)\nu.$$ Using the formula $ \binom{2n}{\nu}=\binom{2n}{n-\nu}=\binom{2n}{\nu+n}, $ we derive $$\begin{split} \sum_{\nu=1}^{2n} \binom {2n}\nu \frac{(-1)^\nu}{\nu} {\sin \pi \(1-\frac 1n\)}{\nu}&=-\sum_{\nu=1}^{2n} \binom {2n} \nu\frac 1\nu \sin \frac{\pi \nu}{n}\\ &=-\sum_{\nu=1}^{n}\binom {2n}\nu \(\frac 1\nu-\frac 1{\nu+n}\)\sin\frac{\pi \nu}{n}. \end{split}$$ Combining these equalities and the well-known asymptotic of the binomial coefficients $$\binom{2n}{n}\sim \frac{4^n}{\sqrt {\pi n}}\quad\text{as}\quad n\to\infty,$$ we obtain that (\[eqkvad\]) is fulfilled. Now, the decreasing of the function $x_\b(t)$ on $(\pi\(1-2/\b\),\pi\(1-1/\b\))$ and $(\ref{eqkvad})$ imply that for any $N_0\in 2\N$ one can find $N_1\in 2\N$ such that $$x_{N_1}(\t({N_1}))<x_{N_1}\(\pi\(1-\frac 2{N_1}\)\)<x_{N_0}(\t({N_0})).$$ At that, choosing $N_1$ to be sufficiently large number, we obtain that $$x_{N_0}(\t({N_0}))-x_{N_1}(\t({N_1}))>2\pi.$$ Then, taking into account , one can choose $k_1\in \N$ such that $$\label{oo1} x_{N_1}(\t({N_1})+2\pi k_1)<0\quad\text{and}\quad x_{N_0}(\t({N_0})+2\pi k_1)>0.$$ Denote $$F_1(\b)=x_\b(\t(\b)+2\pi k_1).$$ By analogy with the above proof of $(ii)$ in the case $k=0$, we can show that the function $F_1(\b)=x_\b(\t(\b)+2\pi k_1)$ is continuous. Then, taking into account  and applying the intermediate value theorem to $F_1$, we can find the points $t_1>\pi (1- 2/{N_0})+2\pi k_1$ and $\b_1>\b_0$ such that $z_{\b_{1}}(t_1)=0$. Repeating this scheme, we obtain infinite sets of points $\{\b_k\}_{k\in\Z_+}$ and $\{t_k\}_{k\in\Z_+}$, for which assertion $(ii)$ holds. *[The proof of $(iii)$.]{}* Suppose to the contrary that there exists $t_0\in (0,\pi)$ such that for some $\b>4$ one has $$\label{eqlem2iii.3} \psi_\b(t_0)=0.$$ Then from (\[eq5.lem1\]) and (\[eq6.lem1\]), we get $$\label{eqlem2iii.4} \int_0^{\frac \b 2 t_0}e^{\i \vp}\(\sin \frac\vp\b\)^\b {\rm d}\vp=0.$$ This equality implies that $$\int_0^{\frac \b 2 t_0}\sin(\vp-\vp_0) \(\sin \frac\vp\b\)^\b {\rm d}\vp =0$$ and, therefore, one has $$\label{eqlem2iii.5} \int_{\vp_0}^{2\pi l_0}\sin \vp\(\sin \frac{\vp-\vp_0}\b\)^\b {\rm d}\vp=0,$$ where we put $ \vp_0=2\pi l_0-\b t_0/2 $ and $l_0=\min\{l\in \N\,:\, \b t_0/2\le 2\pi l\}$. From , we obtain $$\label{eqlem2iii.6} \begin{split} \int_{\vp_0}^{2\pi}&\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp\\ &+\sum_{j=1}^{l_0-1}\int_{2\pi j}^{2\pi (j+1)}\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp=0 \end{split}$$ (if $l_0=1$, then this sum contains only one term). It is easy to see that all summands in the above sum are negative. Indeed, it is obvious that for $\vp_0 \in [\pi,2\pi)$ we have $$\int_{\vp_0}^{2\pi}\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp<0.$$ At the same time, if $\vp_0 \in [0,\pi)$, then $$\begin{split} \int_{\vp_0}^{2\pi}\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp=\left\{\int_{\vp_0}^{\pi}+\int_{\pi}^{2\pi}\right\}\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp\\ <\(\sin \frac{\pi-\vp_0}\b\)^\b\int_{\vp_0}^{\pi}\sin\vp {\rm d}\vp+\(\sin \frac{\pi-\vp_0}\b\)^\b \int_{\pi}^{2\pi}\sin\vp {\rm d}\vp<0. \end{split}$$ By analogy, we can prove that for any $l_0>1$ and $j=1,\dots,l_0-1$ one has $$\int_{2\pi j}^{2\pi (j+1)}\(\sin \frac{\vp-\vp_0}\b\)^\b \sin \vp {\rm d}\vp<0.$$ Thus, the sum in (\[eqlem2iii.6\]) is negative, which is a contradiction to (\[eqlem2iii.3\]). Properties of Fourier multipliers --------------------------------- To prove Theorems \[th4\] and \[th5\], we need some facts about Fourier multipliers $\{\l_k\}_{k\in \Z}$. In the case $\l_k=g(\e k)$, $g\in C(\R)$, $\e>0$, it is important to ascertain whether the function $g$ lies in the Banach algebra $$B(\R)=\bigg\{g\,:\,g(t)=\int_{-\infty}^\infty e^{-\i t u}{\rm d}\mu(u),\quad \Vert g\Vert_B={\rm var} \mu<\infty\bigg\},$$ where $\mu$ is a complex Borel measure which is finite on $\R$, and $\var\mu$ is the total variation of $\mu$. The point is that for any $1\le p\le\infty$, we have $$\sup_{\e>0}\Vert \{g(\e k)\}\Vert_{M_p}\le\sup_{\e>0}\Vert \{g(\e k)\}\Vert_{M_1}=\Vert g\Vert_B$$ (see [@SW] or [@TB Ch. 7]). In what follows, we will use the following comparison principle (see [@TB 7.1.11 and 7.1.14]). \[lem3\] Let $\vp$ and $\psi$ belong to $C(\R)$ and $$\Phi_\e (f;x)\sim\sum_{k\in\Z} \vp(\e k) \widehat{f}_k e^{\i kx},\quad \Psi_\e(f;x)\sim\sum_{k\in\Z} \psi(\e k) \widehat{f}_k e^{\i kx},\quad f\in L_1(\T),$$ where the series on the right are Fourier series. If $g=\vp/\psi\in B(\R)$, then for all $f\in L_p(\T)$, $1\le p\le\infty$, and $\e>0$ we have $$\Vert \Phi_\e(f)\Vert_p \le \Vert g\Vert_B \Vert \Psi_\e(f)\Vert_p.$$ To verify that the function $g$ belongs to $B(\R)$, one can use the following Beurling’s sufficient condition (see [@TB 6.4.2] or the survey [@LST]). \[lem4\] Let a function $g$ be locally absolutely continuous on $\R$, $g\in L_2(\R)$, and $g'\in L_2(\T)$. Then $$\Vert g\Vert_B\le C\(\Vert g\Vert_{L_2(\R)}+\Vert g'\Vert_{L_2(\R)}\).$$ Proofs of the main results ========================== The inequality ${\rm w}_\b(f,h)_p\le C{\w}_\b(f,h)_p$ is obvious. Let us prove the converse inequality. Let $n\in \N$ be such that $n=[1/h]$ and let $T_n\in\mathcal{T}_n$ be polynomials of the best approximation of $f$ in $L_p(\T)$, that is $\Vert f-T_n\Vert_p=E_n(f)_p$. By properties ${(e)}$ and ${(f)}$ of the modulus of smoothness and Lemma \[lemNS\] for $\d\in (h/2,h)$, we obtain $$\label{pro1} \begin{split} \w_\b(f,h)_p^{p_1}&\le \w_\b(f-T_n,h)_p^{p_1}+\w_\b(T_n,h)_p^{p_1}\\ &\le C\(\Vert f-T_n\Vert_p^{p_1}+\(\frac h\d\)^{\b p_1}\Vert \D_\d^\b T_n\Vert_p^{p_1}\)\\ &\le C\(\Vert f-T_n\Vert_p^{p_1}+\Vert \D_\d^\b f\Vert_p^{p_1}\). \end{split}$$ Integrating inequality  by $\d$ over $(h/2,h)$ and applying the Jackson inequality  and ${(g)}$, we get $$\label{pro2} \begin{split} \w_\b(f,h)_p^{p_1}&\le C\(E_n(f)_p^{p_1}+\frac1h\int_{h/2}^h\Vert \D_\d^\b f\Vert_p^{p_1}{\rm d}\d\)\\ &\le C\(\w_r\(f,\frac1n\)_p^{p_1}+\frac1h\int_{0}^h\Vert \D_\d^\b f\Vert_p^{p_1}{\rm d}\d\)\\ &\le C\(\w_r\(f,h\)_p^{p_1}+\frac1h\int_{0}^h\Vert \D_\d^\b f\Vert_p^{p_1}{\rm d}\d\), \end{split}$$ where we take $r=\b+\a\in \N$ with $\a>1/p_1-1$. Next, using the first equivalence from  for all $0<p\le\infty$ and ${(c)}$, we obtain $$\label{pro3} \begin{split} \w_r\(f,h\)_p^{p_1}&\le \frac Ch\int_{0}^h\Vert \D_\d^r f\Vert_p^{p_1}{\rm d}\d=\frac Ch\int_{0}^h\Vert \D_\d^\a\D_\d^\b f\Vert_p^{p_1}{\rm d}\d\\ &\le \frac Ch\int_{0}^h\Vert \D_\d^\b f\Vert_p^{p_1}{\rm d}\d. \end{split}$$ Finally, combining  and , we get ${\w}_\b(f,h)_p\le C{\rm w}_\b(f,h)_p$. *[The proof of $(i)$]{}* easily follows from Lemma \[lem0\], Lemma \[lem1\] $(i)$, and . *[The proof of $(ii)$]{}.* By , we have $$\frac 1h\int_0^{h} \D_\d^{\b_k} e_n(x){\rm d}\d=\psi_{\b_k}({\rm sign\,}n\cdot t_k)e_n(x).$$ It remains only to take into account that by Lemma \[lem1\] $(ii)$ and equalities  one has $\psi_{\b_k}(\pm t_k)=0$ . It is obvious that $ \widetilde{\w}_\b(T_n,h)_p \le \w_\b(T_n,h)_p$. Let us show $$\label{eqprth3.1} \w_\b(T_n,h)_p\le C \widetilde{\w}_\b(T_n,h)_p\quad\text{for all} \quad h\in (0,1/n).$$ Denote $$g_\t(t)=\frac{(1-e^{\i \t t})^\b v(t)}{\psi_\b(t)},\quad \t\in (0,1),$$ where $v\in C^\infty(\R)$, $v(t)=1$ for $|t|\le 1$ and $v(t)=0$ for $|t|> 2$. Note that $$\label{form} \D_\d^\b f(x)\sim\sum_{k\in\Z}(1-e^{\i k\d})^\b \widehat{f}_k e^{\i kx},\quad f\in L_1(\T).$$ Thus, by Lemma \[lem3\], to prove (\[eqprth3.1\]) it is enough to verify $$\label{eqprth3.2} \sup_{\t \in (0,1)}\Vert g_\t\Vert_B<\infty.$$ By Lemma \[lem1\] $(iii)$ we have that $\psi_\b(t)\neq 0$ for $0<|t|<\pi$. Moreover, it is easy to see that $g_\t \in C^\infty(\R)$ and $\sup_{\t \in (0,1)}\Vert g_\t^{(k)}\Vert_{L_\infty(\R)}<\infty$ for each $k\in \Z_+$. Thus, using Lemma \[lem4\] we see that (\[eqprth3.2\]) holds. In view of , we only need to verify that $$\label{preqthEqmods1} {\w}_\b(f,h)_p\le C\(\widetilde{\w}_\b(f,h)_p+E_{[1/h]}(f)_p\).$$ Let $n=[1/h]$ and let $T_n\in\mathcal{T}_n$ be such that $\Vert f-T_n\Vert_p=E_n(f)_p$. Using properties $(e)$ and $(f)$ and Theorem \[th4\], we obtain $$%\label{thpro1} \begin{split} \w_\b(f,h)_p&\le \w_\b(f-T_n,h)_p+\w_\b(T_n,h)_p\\ &\le C\Vert f-T_n\Vert_p+\widetilde{\w}_\b(T_n,h)_p\\ &\le C\Vert f-T_n\Vert_p+\widetilde{\w}_\b(f,h)_p\\ &\le C\(\widetilde{\w}_\b(f,h)_p+E_{[1/h]}(f)_p\). \end{split}$$ Thus, we have . First let us prove that $$\label{ab1} \w_{\b;\, \a}^*(f,h)_p\le C\w_\b(f,h)_p.$$ Let $n=[1/h]$ and let $T_n\in\mathcal{T}_n$ be such that $\Vert f-T_n\Vert_p=E_n(f)_p$. Then, by properties $(a), (b), (c)$, and $(g)$, and the Jackson inequality , we derive $$\label{ab2} \begin{split} \w_{\b;\, \a}^*(f,h)_p &\le \w_{\b;\, \a}^*(f-T_n,h)_p+\w_{\b;\, \a}^*(T_n,h)_p\\ &\le C\Vert f-T_n\Vert_p+\w_{\b;\, \a}^*(T_n,h)_p\\ &\le C\w_\b(f,h)_p+\w_{\b;\, \a}^*(T_n,h)_p. \end{split}$$ Thus, it remains to show $$\label{ab3} \w_{\b;\, \a}^*(T_n,h)_p\le C\w_{\b}(f,h)_p.$$ By Lemma \[lemNS\] with $0<\d_1, \d_2<h$ and by $(b)$ and $(c)$, we obtain $$\label{ab4} \begin{split} \Vert \D_{\d_1}^{\b-\a}\D_{\d_2}^{\a}T_n\Vert_p&\le C \Vert \D_{h}^{\b-\a}\D_{\d_1}^{\a}T_n\Vert_p= C \Vert \D_{\d_1}^{\a}\D_{h}^{\b-\a}T_n\Vert_p\\ &\le C\Vert \D_{h}^{\b}T_n\Vert_p\le C\w_{\b}(T_n,h)_p. \end{split}$$ At the same time, as above, by $(e)$, $(f)$, and , we get $$\label{ab5} \w_{\b}(T_n,h)_p\le C\Vert f-T_n\Vert_p+\w_{\b}(f,h)_p\le C\w_{\b}(f,h)_p.$$ Now, by the definition of $\w_{\b;\, \a}^*(f,h)_p$, , and , inequality  easily follows. Combining   and , we get . To prove the converse inequality $$\label{abc1} \w_\b(f,h)_p\le C\w_{\b;\, \a}^*(f,h)_p,$$ we use the de la Vallée-Poussin means of $f$ given by $$V_h(f;x)=\sum_{k\in \Z} v\(kh\)\widehat{f}_k e^{\i kx},$$ where the function $v$ is defined in the proof of Theorem \[thbest\] (in addition we suppose that $0\le v(t)\le 1$ for all $t\in\R$). By property $(e)$, we have $$%\label{abc2} \w_\b(f,h)_p\le \w_\b(V_h(f),h)_p+\w_\b(f-V_h(f),h)_p.$$ Therefore, the proof of  will follows from the following two inequalities $$%\label{abc3} \w_\b(V_h(f),h)_p\le C \w_{\b;\, \a}^*(f,h)_p$$ and $$\label{abc4} \w_\b(f-V_h(f),h)_p\le C \w_{\b;\, \a}^*(f,h)_p.$$ The first inequality can be verified by repeating the proof of Theorem \[th4\] with the function $$g_{1,\t}(t)=\frac{ (1-e^{\i\t t})^\b v(t)}{\psi_{\b-\a}(t)\psi_\a(t)}$$ instead of $g_\t$. Let us verify that holds. By Lemma \[lem3\] and it is enough to check that $$%\label{abc5} g_{2,\t}(t)=\frac{(1-e^{\i\t t})^\b(1-v(t))}{\psi_{\b-\a}(t)\psi_\a(t)}\in B(\R)$$ and $$\label{abc6} \sup_{\t \in (0,1)}\Vert g_{2,\t}\Vert_B<\infty.$$ Indeed, it is easy to see that $$\label{abc7} \Vert (1-e^{\i\t (\cdot)})^\b\Vert_B\le 2^\b.$$ At the same time we have that $\psi_\a\in B(\R)$, $\psi_\a(t)\neq 0$ for $t\in \R\setminus\{0\}$, and $\lim_{|t|\to\infty}\psi_\a(t)=1$. Thus, by the Wiener-Lévy theorem (see [@LST Theorem 4.4]), we get $$\label{abc8} \frac{(1-v(t))^{1/2}}{\psi_\a(t)}\in B(\R).$$ By analogy, we have $$\label{abc9} \frac{(1-v(t))^{1/2}}{\psi_{\b-\a}(t)}\in B(\R).$$ Finally, combining – and taking into account that $B(\R)$ is the Banach algebra, we derive and, therefore, . Theorem \[th5\] is proved. Acknowledgements {#acknowledgements .unnumbered} ================ The author thanks to R.M. Trigub for the proposed problem and to A.A. Dovgoshey for the valuable discussions. This research has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 704030. [99]{} [^1]: $^1$Universität zu Lübeck, Institut für Mathematik, Ratzeburger Allee 160, 23562 Lübeck [^2]: E-mail address: [email protected]

ArXiv

--- abstract: 'The geometry of spaces with indefinite inner product, known also as Krein spaces, is a basic tool for developing Operator Theory therein. In the present paper we establish a link between this geometry and the algebraic theory of $*$-semigroups. It goes via the positive definite functions and related to them reproducing kernel Hilbert spaces. Our concern is in describing properties of elements of the semigroup which determine shift operators which serve as Pontryagin fundamental symmetries.' address: | F. H. Szafraniec, Instytut Matematyki, Uniwersytet Jagielloński, ul. Łojasiewicza 6, 30 348 Kraków, Poland M. Wojtylak, Instytut Matematyki, Uniwersytet Jagielloński, ul. Łojasiewicza 6, 30 348 Kraków, Poland\ VU University Amsterdam, Department of Mathematics, Faculty of Exact Sciences, De Boelelaan 1081 a, 1081 HV Amsterdam\ author: - Franciszek Hugon Szafraniec - Michał Wojtylak title: 'Pontryagin space structure in reproducing kernel Hilbert spaces over $*$–semigroups' --- [^1] Introduction {#introduction .unnumbered} ============ There are two ways of looking at $*$–semigroups and positive definite functions defined on them. The first consists in intense analysis of the algebraic structure of a semigroup so as to establish conditions on it, which ensure prescribed properties to hold for positive definite functions. One of the properties frequently considered is representing of the positive definite functions as moments of a positive measure. This attitude has been successfully undertaken by Bisgaard resulting in a considerable number of papers, see in particular [@biss-scmat; @biss-two; @biss-sep; @biss-ext; @biss-CC; @biss-fact] and references therein, some of them we are going to exploit here. The other way is to determine of the positive definite functions posses desired properties; a typical example within this category is to detect multidimensional moment sequences. In the present paper we are going to develop the first thread putting forward the following problem. Impose necessary and sufficient conditions on a distinguished element $u$ of a $*$–semigroup $S$ to generate a Pontryagin fundamental symmetry of reproducing kernel Hilbert space over the semigroup in question; the precise formulation is exposed as ([P]{}), p. . Surprisingly, our problem has found a simple algebraic solution. In the context of $*$-separative commutative semigroups the condition on $u$ is: $u=u^*$, $u+u=0$ and $u+s\neq s$ for only a finite number of $s\in S$ (see Proposition \[krein\], Theorem \[main\]). Let us mention that $*$-semigroups and positive definite functions on them have been originated by Sz.-Nagy in his famous Appendix [@nagy]. He uses the reproducing kernel Hilbert space factorization to prove his “general dilation theorem” for operator valued functions. Since then the RKHS technique has been used from time to time for proving results with dilation flavour behind the screen. We are going to follow suit here. Shift operators connected with positive definite functions – formulation of the problem ======================================================================================= By a $*$-semigroup we understand a commutative semigroup with an involution, not necessarily having the neutral element. The involution is always denoted by the symbol “$*$” and the semigroup operation is always written in an additive way. In the case when the $*$–semigroup $S$ has a neutral element $0$ we say that $\phi:S\to{\mathbb{C}}$ is [*positive definite*]{} (we write $\phi\in{\frak{P}(S)}$) if for every $N\in{\mathbb{N}}:={\left\{0,1,\dots\right\}}$ and every $s_0{,\dots,}s_N\in S$, $\xi_0{,\dots,}\xi_N\in{\mathbb{C}}$ we have $\sum_{i,j=0}^N\xi_i\bar\xi_j \phi(s_j^*+s_i)\geq 0$. With such $\phi$ we link a reproducing kernel Hilbert space ${\mathcal{H}}^\phi{\subseteq}{\mathbb{C}}^S$ with a reproducing kernel defined by $K^\phi(s,t):=\phi(t^*+s)$, $t,s\in S$ (see e.g.  [@BChR p.81]). For $s\in S$ we set $K^\phi_s:=K(\cdot,s)$ ($s\in S$), it is known that $K^\phi_s\in{\mathcal{H}}^\phi$ and $f(s)={\left<f,K^\phi_s\right>}$ for every $f\in{\mathcal{H}}^\phi$. The set ${\mathcal{D}}^\phi:={\textrm{lin}}{\left\{K^\phi_s:s\in S\right\}}$ is dense in ${\mathcal{H}}^\phi$. For an element $u\in S$ we define [*the shift operator*]{} (sometimes called also the translation operator) $ A(u,\phi):{\mathcal{D}}^\phi\to{\mathcal{D}}^\phi$, by $$A(u,\phi)\left(\sum_j\xi_jK_{s_j}^\phi\right):=\sum_j\xi_jK_{s_j+u}^\phi.$$ It can be shown that $A( u,\phi)$ is well defined, linear and closable (see e.g. [@szaf-bdd Proposition, p.253], [@BChR p.90]). Our main object will be the closure of the above operator, denoted below by $u_\phi$. It is a matter of simple verification that ${\mathcal{D}}^\phi$ is contained in the domain of $(u_\phi)^*$ (the adjoint of the operator $u_\phi$) and $(u_\phi)^*f=(u^*)_\phi f$ for $f\in{\mathcal{D}}^\phi$. Under the above circumstances we can state our problem as follows. 1. Provide necessary and sufficient conditions on the element $u\in S$ for the operator $u_\phi$ to be a fundamental symmetry of a Pontryagin space, i.e. to satisfy $$u_\phi=u_\phi^*,\quad u_\phi^2=I_{\mathcal{H}^\phi},\quad \dim\ker(u_\phi+I_{\mathcal{H}^\phi})<\infty$$ $\phi\in{\frak{P}(S)}$. Obviously, the condition $u_\phi=u_\phi^*$ together with $u_\phi^2=I_{\mathcal{H}^\phi}$ imply that $u_\phi$ must be bounded on ${\mathcal{H}}$. We continue with some basic definitions and notations concerning $*$-semigroups. Let $S$ and $T$ be $*$-semigroups, a mapping $\chi:S\to T$ is called $*$-[*homomorphism*]{} if $\chi(s+t)=\chi(s)+\chi(t)$ for all $s,t\in S$, $\chi(s^*)=(\chi(s))^*$ for all $s\in S$. A [*character*]{} on $S$ is a nonzero $*$-homomorphism $\chi:S\to{\mathbb{C}}$ where the latter set is understood as a semigroup with multiplication as the operation and the conjugation as involution. It is obvious that if $S$ has a neutral element $0$ then $\chi(0)=1$ for all $\chi\in S^*$. Let ${\mathcal{A}}(S^*)$ be the least $\sigma$-algebra of subsets of $S^*$ rendering measurable all the functions $$\label{shat} \hat s:S^*\ni\chi\mapsto\chi(s)\in{\mathbb{C}},\quad s\in S.$$ If $\mu$ is a positive measure on $S^*$ such that all the functions listed in (\[shat\]) are square-integrable we define a function ${\mathcal{L}}(\mu):S\to{\mathbb{C}}$ by ${\mathcal{L}}(\mu)(s):=\int_{S^*}\sigma(s)d\mu(\sigma)$ ($s\in S$). We call $\phi:S\to{\mathbb{C}}$ a [*moment function*]{} ($\phi\in{\frak{M}(S)} $) if $\phi={\mathcal{L}}(\mu)$ for some measure $\mu$ on $S^*$. It is easy to verify that ${\frak{M}(S)} {\subseteq}{\frak{P}(S)}$, if the latter inclusion is an equality then we call $S$ [*semiperfect*]{}. For examples of semiperfect and non-semiperfect $*$-semigroups and more general concepts of semiperfectness see  [@biss-scmat]. We call $S$ [*$*$-separative*]{} if the characters separate points in $S$. The [*greatest $*$-homomorphic $*$-separative image of $S$*]{} is the semigroup ${S/_\sim}$, where the equivalence relation $\sim$ on $S$ is defined by the condition that $s\sim t$ if and only if $\sigma(s)=\sigma(t)$ for all $\sigma\in S^*$; addition and involution in ${S/_\sim}$ are those that make the quotient mapping a $*$-homomorphism. The elements of ${S/_\sim}$ will be denoted as equivalence classes $[s]$ ($s\in S$). If $S$ has a zero, then we will use the symbol $0$ for the neutral element of both $S$ and ${S/_\sim}$, instead of using $[0]$. Since every character on $S$ generates a character on ${S/_\sim}$, the latter semigroup is in fact $*$-separative. The following simple proposition gives answer to the question when the operator $ u_\phi$ defined above is a fundamental symmetry of a Krein space, i.e. when $( u_\phi)^2= u_\phi$ and $ u_\phi=( u_\phi)^*$. \[krein\] Let $S$ be a commutative $*$-semigroup with zero. For each element $u\in S$ the following conditions are equivalent: - [$[2u]= 0$ and $[u]=[u^*]$;]{} - [for every $\phi\in{\frak{M}(S)} $ we have $(u_\phi)^2= I_{{\mathcal{H}}^\phi}$ and $(u_\phi)^*= u_\phi$;]{} - [for every $\phi\in{\frak{P}(S)}$ we have $(u_\phi)^2 = I_{{\mathcal{H}}^\phi}$ and $(u_\phi)^*= u_\phi$;]{} Moreover, [(i)]{} implies that $ u_\phi$ is a bounded, selfadjoint operator on ${\mathcal{H}}^\phi$ for every $\phi\in{\frak{P}(S)}$. (i)${ \Rightarrow}$(iii) Suppose that (i) holds and let $\phi\in{\frak{P}(S)}$. For every $\sigma\in S^*$ we have $\sigma(u)=\sigma(u^*)$ and $\sigma(2u)=\sigma(0)=1$. Consequently, for every $\sigma\in S^*$ and every $s\in S$ $$\label{sigma2} \sigma(s+u)=\sigma(s+u^*),\quad\sigma(t+2u)=\sigma(t),\quad \sigma(s^*+u^*+u+s)=\sigma(s^*+s).$$ By [@biss-sep Thm.2] we have $$\label{sigma3-1} \phi(s+u)=\phi(s+u^*),\quad s\in S,$$ $$\label{sigma3-2} \phi(s+2u)=\phi(s),\quad s\in S,$$ $$\label{sigma3-3} \phi(s^*+u^*+u+s)=\phi(s^*+s),\quad s\in S,$$ The last of these three equalities, together with [@szaf-bdd Cor.1], implies that the operator $u_\phi$ is in ${\bold{B}}({\mathcal{H}}^\phi)$. It is also selfadjoint, since for $f=\sum_i \xi_i K_{s_i}^\phi\in{\mathcal{D}}^\phi$ we have $${\left<u_\phi f,f\right>}= \sum_{i,j}\xi_i\bar\xi_j{\left<K_{s_i+u}^\phi,K_{s_j}^\phi\right>}=\sum_{i,j}\xi_i\bar\xi_j\phi(s_i+u+s_j^*) \stackrel{(\ref{sigma3-1})}=$$ $$\sum_{i,j}\xi_i\bar\xi_j\phi(s_i+u^*+s_j^*)= {\left<f, u_\phi f\right>}.$$ The fact that $(u_\phi)^2=I_{\mathcal{H}}^\phi$ can be obtained similarly as selfadjointness of $u_\phi$, with the use of (\[sigma3-2\]) instead of (\[sigma3-1\]). This finishes the proof of (iii) and of the ‘Moreover’ part of the proposition. The implication (iii)${ \Rightarrow}$(ii) is trivial. The proof (ii)${ \Rightarrow}$(i) goes by contraposition. Suppose first that $[2u]\neq 0$, i.e. there exists a character $\sigma$ such that $\sigma(u)^2=\sigma(2u)\neq\sigma(0)=1$. We put $\phi:={\mathcal{L}}(\delta_\sigma)$, where $\delta_\sigma$ stands for the Dirack measure on $S^*$ concentrated in $\sigma$. Let ${\left<\,\cdot\,,-\right>}$ denote the scalar product on ${\mathcal{H}}^\phi$. Observe that $${\left<K_0^\phi,K_0^\phi\right>}=\phi(0)=\sigma(0)\neq\sigma(2u)=\phi(2u)={\left<K_{2u}^\phi,K_0^\phi\right>}={\left<( u_\phi)^2K_0^\phi,K_0^\phi\right>}.$$ In consequence, $I_{{\mathcal{H}}^\phi}\neq ( u_\phi)^2$. Similarly, if $\tau(u)\neq\tau(u^*)$ for some $\tau\in S^*$ then for $\psi:={\mathcal{L}}(\delta_\tau)$ the operator $u_\psi$ is not symmetric in ${\mathcal{H}}^\psi$. Let us consider now a situation when $S$ and $T$ are $*$-semigroups with zeros and $h$ is a $*$-homomorphism from $S$ into $T$ satisfying $h(0)=0$. Note that if an element $u\in S$ is such that $[2u]=0$, $[u]=[u^*]$ then $[2h(u)]=0$, $[h(u)]=[h(u)^*]$. This comes from the fact that for every character $\sigma$ on $T$ the function $\sigma\circ h$ is a character on $S$. Observe also that $\phi\circ h\in{\frak{P}(S)}$ for every $\phi\in{\frak{P}(T)}$. \[ST\] Assume that $S$, $T$ and $h$ are as above and that $h$ is additionally onto. Let $u\in S$ be such that $[2u]=0$, $[u]=[u^*]$ and let $\phi\in{\frak{P}(T)}$. Then the operators $u_{\phi\circ h}$ in ${\mathcal{H}}^{\phi\circ h}$ and $h(u)_{\phi}$ in ${\mathcal{H}}^\phi$ are unitarily equivalent. Let ${\left<\cdot,-\right>}_\phi$ and ${\left<\cdot,-\right>}_{\phi\circ h}$ denote the scalar products on ${\mathcal{H}}^\phi$ and ${\mathcal{H}}^{\phi\circ h}$ respectively. Since $$\begin{aligned} {\left<\sum_{j=1}^N \xi_j K^{\phi\circ h} _{s_j} ,\sum_{j=1}^N \xi_j K^{\phi\circ h} _{s_j}\right>}_{\!\!\phi\circ h} &=& \sum_{i,j=1}^N \xi_i\bar\xi_j (\phi\circ h)(s_i+s_j^*) \\ = \sum_{i,j=1}^N \xi_i\bar\xi_j \phi(h(s_i)+h(s_j^*))&=& {\left<\sum_{j=1}^N \xi_j K^{\phi} _{h(s_j)},\sum_{j=1}^N \xi_j K^{\phi} _{h(s_j)}\right>}_{\!\!\phi}, \end{aligned}$$ the condition $ V( K^{\phi\circ h}_{s}):= K^{\phi}_{h(s)} $ ($s\in S$) properly defines an isometry between ${\mathcal{H}}^{\phi\circ h}$ and ${\mathcal{H}}^{\phi}$. Since $h$ is onto, the range of $V$ is dense in ${\mathcal{H}}^\phi$, and so $V$ is a unitary operator. To finish the proof we need to show that $$\label{VuuV} h(u)_\phi Vf=V u_{\phi\circ h}f, \quad f\in{\mathcal{H}}^{\phi\circ h}.$$ This can be easily verified for $f\in{\mathcal{D}}^{\phi\circ h}$. Since all the operators appearing in (\[VuuV\]) are bounded, the proof is finished. Applying the above to the quotient semigroup $T={S/_\sim}$ and $h$ as the quotient mapping, together with the fact from [@biss-fact] that every $\phi\in{\frak{P}(S)}$ ($\phi\in{\frak{M}(S)}$) is of the form $\phi=\psi\circ h$ for some $\psi\in{\frak{P}({S/_\sim})}$ ($\psi\in{\frak{M}({S/_\sim})}$) gives the following. \[Nowyjeszcze\] Assume that $S$ is a $*$-semigroup with zero and $u\in S$ is such that $[2u]=0$, $[u]=[u^*]$. Then $$\begin{aligned} \dim\ker(u_\phi+I)&<&+\infty \textrm{ for every }\phi\in{\frak{P}(S)}\,\, (\phi\in{\frak{M}(S)})\\ \iff \dim\ker([u]_{\psi}+I)&<&+\infty\textrm{ for every }\psi\in{\frak{P}({S/_\sim})}\,\, (\psi\in{\frak{M}({S/_\sim})}) \end{aligned}$$ Examples ======== \[exm1\] Consider the semigroup $S={{\mathbb{Z}}}_2\times{\mathbb{N}}$ with standard addition and the identical involution. As usually (cf. [@BChR]) we identify $S^*$ with ${\left\{-1,1\right\}}\times{\mathbb{R}}$, note that $S$ is $*$-separative. The only nonzero element satisfying $u=u^*$, $2u=0$ is $u=(1,0)$. Let $\phi$ be a positive definite mapping, we will now compute the eigenspaces of $ u_\phi$. Since $S$ is semiperfect ([@biss-two]), there exists a Borel measure $\mu$ on $S^*$ such that $\phi={\mathcal{L}}(\mu)$. Having our interpretation of characters in mind we get $$\phi(x,n)=\int_{\mathbb{R}}(-1)^xt^nd\mu_-(t)+\int_{\mathbb{R}}t^nd\mu_+ (t),\quad x\in{{\mathbb{Z}}}_2,\,n\in{\mathbb{N}},$$ with $\mu_\pm:=\mu{\!\!\mid_{{{\left\{\pm1\right\}}}\times{\mathbb{R}}}}$. Let us define the functions $f_{x,n}:S^*\to S^*$ ($x\in{{\mathbb{Z}}}_2$, $n\in{\mathbb{N}}$) by $$f_{x,n}({\varepsilon},t):={\varepsilon}^xt^x ,\quad {\varepsilon}\in{\left\{-1,1\right\}},\,t\in{\mathbb{R}},\,x\in{{\mathbb{Z}}}_2,\,n\in{\mathbb{N}},$$ and note that they all are square integrable. By $\mathcal P^\mu$ we define the closure in $L^2(\mu)$ of the linear span of the functions $f_{x,n}$ ($x\in{{\mathbb{Z}}}_2$, $n\in{\mathbb{N}}$). The formula $$\label{Vdef} V(K^\phi_{x,n}):=f_{x,n},\qquad x\in{{\mathbb{Z}}}_2,\,n\in{\mathbb{N}},$$ constitutes a unitary isomorphism between ${\mathcal{H}}^\phi$ and $\mathcal P^\mu$. The shift operator $u_\phi$ is unitary equivalent (via $V$) to the following operator $M$ $$(M f)({\varepsilon},t):={\varepsilon}f ({\varepsilon},t),\qquad ({\varepsilon},t)\in{\left\{-1,1\right\}}\times{\mathbb{R}},\, f\in\mathcal{P}^\mu.$$ It is not hard to see that $$\ker(M\pm I)={\left\{f\in \mathcal{P}^\mu: f(\pm1,\cdot)=0\,\,\, \, \mu_\pm\textrm{--a.e}\right\}}.$$ Note that $\dim\ker(u_\phi\pm I)=\dim\ker(M\pm I)$ and the latter is finite dimensional if and only if the support of $\mu_{\mp}$ is a finite set. In particular there exists a mapping $\phi\in{\frak{P}(S)}={\frak{M}(S)} $ such that $\dim\ker( u_\phi+I)=\infty$. At this point we present a useful for our purposes construction. Let $S$ and $T$ be two disjoint $*$-semigroups (which only a formal restriction) and let $h:S\mapsto T$ be a $*$-homomorphism. We endow the set $S\cup T$ with the $*$-semigroup structure in the following way. The addition on $S\cup T$, denoted by the same symbol ‘$+$’, is defined by $$s+t:=\left\{\begin{array}{rcl} s+t & : & s,t\in S\textrm{ or } s,t\in T\\ s+h(t) & : & s\in S,\, t\in T\\ t+h(s) & : & t\in S,\, s\in T\\ \end{array}\right.$$ The involution on $S\cup T$ (still denoted by ‘$*$’) is such that its restriction to both $S$ and $T$ is the original involution on $S$ and $T$, respectively. We denote the above constructed semigroup by $\mathrm{U}(S,T,h)$. The reader can easily check a general fact, that if $S$ and $T$ are $*$-separative then $\mathrm{U}(S,T,h)$ is $*$-separative as well. \[exm2\] Consider a semigroup $S=\mathrm{U}({{\mathbb{Z}}}_2,{\mathbb{N}},h_0)$ where $h_0(x)=0$ ($x\in {{\mathbb{Z}}}_2$). The element $0_{{{\mathbb{Z}}}_2}$ is the neutral element of $S$. Take $u:=1_{{{\mathbb{Z}}}_2}$, clearly $u=u^*$ and $2u=0_{{{\mathbb{Z}}}_2}$. Let $\phi$ be any positive definite function on $S$ and suppose that $f\in\ker( u_\phi+I)$. This means that for $n\in{\mathbb{N}}$ $$f(n)=f(n+u)={\left<f,K^\phi_{n+u}\right>}={\left<f,u_\phi K^\phi_n\right>}={\left<u_\phi f, K^\phi_n\right>}=-f(n),$$ hence $f{\!\!\mid_{{\mathbb{N}}}}=0$. A similar calculation shows that $f(u)=-f(0_{{{\mathbb{Z}}}_2})$. Hence, the eigenspace $\ker( e_\phi+I)$ is spanned by the single function $$f(s)=\left\{\begin{array}{rcl} 0 &:& s\in{\mathbb{N}}\\ 1 & : & s=0_{{{\mathbb{Z}}}_2}\\ -1 & : & s=1_{{{\mathbb{Z}}}_2} \end{array} \right.$$ if $f\in{\mathcal{H}}^\phi$ or is trivial otherwise. Resuming, $\dim\ker(u_\phi+I)\leq 1$ for all positive definite $\phi$. Main result =========== \[main\] Let $S$ be a commutative $*$-semigroup with zero and let $u\in S$ be such that $[2u]=0$ and $[u]=[u^*]$. Then the following conditions are equivalent: - the set ${\left\{[s]\in S/_\sim\colon [u+s]\neq[s]\right\}}$ is finite[^2]; - [ $\sigma(u)=-1$ for only finitely many $\sigma\in S^*$;]{} - [$\dim\ker(u_\phi+I_{{\mathcal{H}}^\phi})<\infty$ for all $\phi\in{\frak{M}(S)} $;]{} - [$\dim\ker(u_\phi+I_{{\mathcal{H}}^\phi})<\infty$ for all $\phi\in{\frak{P}(S)}$.]{} Moreover, if [(i)]{} holds then $\dim\ker(u_\phi+I_{{\mathcal{H}}^\phi})$ is less or equal to the half of the number elements of the set mentioned in [(i)]{}. Let us stress here the solution of the main problem of the paper, which follows from Proposition \[krein\] and Theorem \[main\] : [*the operator $ u_\phi$ is a fundamental symmetry of a Pontryagin space for every $\phi\in{\frak{P}(S)}$ $(\!\!$ equivalently: $\phi\in{\frak{M}(S)}$$)$ if and only if $[2u]=0$, $[u]=[u^*]$ and the set ${\left\{[s]\in{S/_\sim}:[u+s]\neq[s]\right\}}$ is finite.*]{} \[stupidstyle\] Condition (ii) is equivalent to - [*$\sigma([u])=-1$ for only finitely many characters $\sigma$ on ${S/_\sim}$,*]{} since the characters on $S$ and ${S/_\sim}$ are in one-to-one correspondence. This, together with Corollary \[Nowyjeszcze\] (and the remarks above it) allows us to reduce the proof of Theorem \[main\] to the case when $S$ is $*$-separative. However, note that the conditions (i)–(iv) are equivalent to the following: - [*the set ${\left\{s\in S: u+s\neq s\right\}}$ is finite.*]{} Indeed, consider the following example. Let $S={{\mathbb{Z}}}_4\times{\mathbb{N}}$ with the natural operation $+$ and the identical involution. The greatest $*$-separative homomorphic image of $S$ is ($*$-isomorphic with) ${{\mathbb{Z}}}_2\times{\mathbb{N}}$ and the quotient homomorphism maps $u:=(2,0)$ to $(0,0)$. We have that $u+s\neq s$ for all $s\in S$ but the set mentioned in (i) is empty. It remains on open problem if the condition - [*the set ${\left\{s\in S: [u+s]\neq [s]\right\}}$ is finite.*]{} is equivalent to (i). Before the proof we introduce the notion of a $*$-archimedean component of a semigroup. We call a $*$-semigroup $H$ (not necessarily with 0) [*$*$-archimedean*]{} if for all $s,t\in H$ there exists $m\in{\mathbb{N}}\setminus{\left\{0\right\}}$ such that $m(s+s^*)\in t+ H$. An [*$*$-archimedean component*]{} of a $*$-semigroup $S$ is a maximal $*$-archimedean $*$-subsemigroup. Though $*$-archimedean component is $*$-semigroup for itself it is possible for it not to have the neutral element even if $S$ does. It can be shown that two elements $s,t$ belong to the same $*$-archimedean component of $S$ if and only if $m(s+s^*)\in t+S$ and $n(t+t^*)\in s+S$ for some $m,n\in{\mathbb{N}}\setminus{\left\{0\right\}}$. Furthermore, $S$ is the disjoint union of its $*$-archimedean components, see [@clipres Section 4.3] for the case of identical involution. The following Lemma was proven in [@biss-ext] (Lemma 2), the proof for an arbitrary involution requires minimal effort. \[extchar\] If $H$ is a $*$-archimedean component of a $*$-semigroup $S$ then every character on $H$ is everywhere nonzero and extends to a character on $S$. If $H$ and $K$ are two $*$-archimedean components of $S$ then $H+K$ is contained in one single $*$-archimedean component of $S$. If $(S_i)_{i\in I}$ is the family of all $*$-archimedean components of $S$ then we define the operation $+$ on $I$ by: $$i+j=k\textrm{ if and only if } S_i+S_j{\subseteq}S_k,\quad i,j,k\in I.$$ Since $S_i+S_i{\subseteq}S_i$ for all $i\in I$ we have that $i+i=i$. Therefore $I$ is a semilattice with the natural partial order given by the condition that $i\leq j$ if and only if $i+j=j$. The following easy lemma is left as an exercise for the reader. \[elat\] Let $S$ be a $*$-semigroup with zero and let $(S_i)_{i\in I}$ be the family of all $*$-archimedean components $S$. If $u\in S$ be such that $2u=0$, then $u$ belongs to the same $*$-archimedean component as 0. In particular, $u+S_i{\subseteq}S_i$ for all $i\in I$. As it was said in Remark \[stupidstyle\] we may assume that $S$ is $*$-separative. To prove (i)${ \Rightarrow}$(iv) let us put $$U:={\left\{ s\in S: u+ s= s\right\}}$$ and suppose that the set $ S\setminus U$ contains only a finite number $M$ of elements. Take $\phi\in{\frak{P}(S)}$. We show that $$\dim\ker( u_\phi+I)\leq M/2,$$ this will also prove the last statement of Theorem \[main\]. Let us fix an arbitrary $f\in\ker( u_\phi+I)$. We have $$f( s+ u)={\left<f,u_\phi K_{s}^\phi\right>}={\left<u_\phi f,K_s^\phi\right>}=-f( s),\quad s\in S.$$ This means that $f{\!\!\mid_{ U}}\equiv 0$. Observe that the relation $$aRb \Longleftrightarrow (a=u+b\textrm{ or }a=b)$$ is an equivalence relation on $ S\setminus U$ and that the each equivalence class contains exactly two elements. Take any representees $s_1{,\dots,}s_{M/2}$ of the equivalence classes of $R$. It is clear, that $$\ker( u_\phi+I){\subseteq}{\textrm{lin}}{\left\{\delta_{s_i}-\delta_{s_i+u}:i=1{,\dots,}M/2\right\}}.$$ Consequently $\dim\ker( u_\phi+I)\leq M/2$.\ (iv)${ \Rightarrow}$(iii) is obvious. (iii)${ \Rightarrow}$(ii) Suppose that ${\left\{\sigma_n:n\in{\mathbb{N}}\right\}}$ is an infinite set of characters satisfying $\sigma_n(u)=-1$ ($n\in{\mathbb{N}}$). We define a measure $\mu$ on $S^*$ by $\mu:=\sum_{n=0}^\infty 2^{-n} \delta_{\sigma_n}$ and we take a mapping $\phi={\mathcal{L}}(\mu)$. Note that for every $N\in{\mathbb{N}}$ and for every $s_0{,\dots,}s_N\in S$, $\xi_0{,\dots,}\xi_N\in{\mathbb{C}}$ we have $$\bigg|\sum_{j=0}^N\xi_j\sigma_n(s_j)\bigg|^2\leq 2^n\int_{S^*}\bigg|\sum_{j=0}^N\xi_j\sigma(s_j)\bigg|^2d\mu(\sigma)= 2^n\sum_{i,j=0}^N\xi_i\bar\xi_j\int_{S^*}\sigma(s_i+s_j^*)d\mu(\sigma)=$$ $$=2^n\sum_{i,j=0}^N\xi_i\bar\xi_j\phi(s_i+s_j^*),\quad n\in{\mathbb{N}}.$$ By the RKHS Test ([@test] and also [@test2]) we get that $\sigma_n\in{\mathcal{H}}^\phi$ for $n=1,2,\dots$. Now observe that $$u_\phi(\sigma_n)(s)=\sigma_n(u+s)=-\sigma_n(s),\quad s\in S,n=1,2\dots.$$ Therefore $\sigma_n\in\ker( u_\phi+I)$, $n=1,2,\dots$. It remains to show that the functions $\sigma_n$, $n\in{\mathbb{N}}$, are linearly independent. But this results from the well known fact that all characters are linearly independent[^3].\ (ii)${ \Rightarrow}$(i) Suppose that the number of elements $ s\in S$ satisfying $ u+ s\neq s$ is infinite. We show that there exists infinitely many characters $\sigma$ on $ S$ such that $\sigma( u)=-1$. Let $( S_i)_{i\in I}$ be the family of all $*$-archimedean components of $ S$. Set $$I_0:={\left\{i\in I: u+ s\neq s\textrm{ for some } s\in S_i\right\}}.$$ Our assumption implies that[^4] either $$\label{Case1} \textrm{ $ S_j$ is infinite for some $j\in I_0$}$$ or $$\label{Case2} I_0\textrm{ is infinite}.$$ Let us first assume (\[Case1\]). Take $s_0\in S_j$ such that $u+s_0\neq s_0$. By Lemma \[elat\] we have $ u+ s_0\in S_j$. The $*$-semigroup $S_j$ is $*$-separative as a subsemigroup of a $*$-separative semigroup $S$. Therefore, there exists a character $\sigma_0$ on $ S_j$ such that $\sigma_0( u+ s_0)\neq\sigma_0( s_0). $ By Lemma \[extchar\] $\sigma_0$ extends to some character $\tilde{\sigma_0}$ on $ S$. Since $ u= u^*$ and $2 u=0$ we have $\tilde{\sigma_0}( u)\in{\left\{-1,1\right\}}$. But $$\tilde{\sigma_0}( u)\sigma_0( s_0)=\sigma_0( u+ s_0)\neq\sigma_0( s_0).$$ Hence, $\tilde{\sigma_0}( u)=-1$, which means that $\sigma_0( u+ s_0)=-\sigma_0( s_0)$. Denote by $A$ the set of all those characters $\sigma$ on $ S_j$ satisfying $\sigma( u+ s_0)=-\sigma( s_0)$. Since $\sigma_0$ is everywhere nonzero on $ S_j$ (Lemma \[extchar\]), the mapping $$S_j^*\ni\sigma\longmapsto \sigma_0\sigma\in S_j^*$$ is bijective. Moreover, it maps $A$ onto $ S_j^*\setminus A$. Since $ S_j$ is infinite and $*$-separative, $ S_j^*$ is infinite as well. Hence, $A$ is infinite. By Lemma \[extchar\], there is an infinite number of characters $\sigma$ on $ S$ satisfying $\sigma( u+ s_0)=-\sigma( s_0)$ and consequently $\sigma( u)=-1$.\ Let us assume now (\[Case2\]). For each $i\in I_0$ we take a character $\sigma_i$ on $ S$ satisfying $\sigma_{i}( u)=-1$, $\sigma_{i}( s)\neq 0$ for $s\in S_{i}$ (such a character exist by repeating the proof from the previous case). We also define a family of characters $\chi_{i}\in S^*$ (${i\in I_0}$) by $$\chi_{i}( s)=\left\{\begin{array}{rl} 1 & \textrm{ if } s\in S_j\textrm{ and }j\leq i\\ 0 & \textrm{ otherwise} \end{array}\right.,\qquad s\in S,\,i\in I_0.$$ Finally, we put $\rho_i:=\sigma_i\chi_i$ ($i\in I_0$). By Lemma \[elat\] we have that $ u$ is in the same $*$-archimedean component as 0, denote this component by $ S_{j_0}$. It is clear that $j_0\leq i$ for all $i\in I$, therefore $\chi_i( u)=1$ and $\rho_i( u)=-1$ for all $i\in I_0$. The only thing that lasts is to show that $\rho_i\neq\rho_j$ for $i\neq j$. If $i\neq j$ then, by symmetry, we can assume that $j\nleq i$. Thus $\chi_i{\!\!\mid_{ S_j}}=0$ and $\rho_i{\!\!\mid_{ S_j}}=0$. But $\chi_j{\!\!\mid_{ S_j}}\equiv1$ and $\sigma_j$ is, by definition, everywhere nonzero on $ S_j$. Therefore $\rho_i{\!\!\mid_{ S_j}}\equiv0\neq\rho_j{\!\!\mid_{ S_j}}$.\ \[eitheror\] The alternative of (\[Case1\]) and (\[Case2\]) in the proof of (ii)${ \Rightarrow}$(i) becomes more clear if we observe that $ u+ s\neq s$ for $ s\in S_i$, provided that $i\leq j\in I_0$. Indeed, suppose that $ u+ s= s$ for some $ s\in S_i$, $ u+ s_0\neq s_0$ for some $ s_0\in S_j$ and $i\leq j$. This gives us $$( u+ s_0)+( s+ s_0)= u+ s+ s_0+ s_0= s_0+( s+ s_0).$$ By Lemma \[elat\] we have $ u+ s_0\in S_j$. Moreover, $ s+ s_0\in S_j$ because $i\leq j$. The semigroup $ S_j$ is cancellative as a $*$-archimedean component of a $*$-separable group (see [@biss-CC p.63]). This gives us $ u+ s_0= s_0$, contradiction. The example below shows that both (\[Case1\]) and (\[Case2\]) are possible. Let $S={{\mathbb{Z}}}_2\times{\mathbb{N}}$, with the natural addition on ${{\mathbb{Z}}}_2$ and ${\mathbb{N}}$ and the identical involution. The element $u=(1,0)$ is like in (\[Case1\]). Let us now consider the semigroup $T={{\mathbb{Z}}}_2\times{\mathbb{N}}$, with the natural addition on ${{\mathbb{Z}}}_2$ and maximum as the operation on ${\mathbb{N}}$, the involution is again set to identity. It is easy to see that $T$ is $*$–separable. The element $u=(1,0)$ is such that (\[Case2\]) is satisfied. This example shows one more thing. Namely, the condition - [ *$\dim\ker(u_\phi+I_{{\mathcal{H}}^\phi})<\infty$ for all $\phi\in{\frak{M}(S)} $ of compact support*]{} is equivalent to any of the conditions of Theorem \[main\]. Indeed, the characters on $T$ form a discrete, enumerable set. If the mapping $\phi\in{\frak{M}(S)} $ is compactly supported then it is finitely supported and consequently the space ${\mathcal{H}}^\phi$ is finite dimensional. Hence, $u=(1,0)$ satisfies the above condition, but does not satisfy (i). Note that in Proposition \[krein\] restricting to compactly supported moment functions is possible because the function $\phi$ constructed in the proof (ii)${ \Rightarrow}$(i) is supported by only one character. Functions with a finite number of negative squares ================================================== The condition $\dim\ker( u_\phi+I_{{\mathcal{H}}^\phi})<\infty$ can be written also in the language of negative squares. Precisely speaking, by [*the number of negative squares of a mapping* ]{} $\psi:S\to {\mathbb{C}}$ satisfying $$\label{symmetric} \psi(s)={\overline}{\psi(s^*)}, \qquad s\in S,$$ we understand the maximum, taken over all numbers $N\in{\mathbb{N}}$ and all sequences $s_0{,\dots,}s_N$, of the number of negative eigenvalues of the symmetric matrix $\left(\psi(s_i+s_j^*)\right)_{i,j=0}^N$. Note that if $[u]=[u^*]$ and $\phi\in{\frak{P}(S)}$ then the mapping $\psi:=\phi(\cdot+u)$ satisfies (\[symmetric\]). Indeed, take any character $\sigma$. Then $\sigma(s+u)=\sigma(s)\sigma(u)=\sigma(s)\sigma(u^*)=\sigma(s+u^*)$. By the result of [@biss-sep] we get $\phi(s+u)=\phi(s+u^*)$. Combining this with $\phi(t)={\overline}{\phi(t^*)}$ ($t\in S$) [@BChR 4.1.6] proofs the claim. \[(v)\] Let $S$ be a $*$–semigroup with zero and let $u\in S$ be such that $[2u]=0$ and $[u]=[u^*]$ and let $\phi\in{\frak{P}(S)}$. Then the number of negative squares of the mapping $\phi(\cdot+u)$ equals $\dim\ker(u_\phi+I)$. Consequently, conditions [(i)–(iv)]{} of Theorem \[main\] are equivalent to each of the following: - the mapping $\phi(\cdot+u)$ has a finite number of negative squares for every $\phi\in{\frak{P}(S)}$; - the mapping $\phi(\cdot+u)$ has a finite number of negative squares for every $\phi\in{\frak{M}(S)} $. (cf. [@berg; @langerio; @sasvari] for similar arguments) First let us assume that $\dim\ker(u_\phi+I)=m\in{\mathbb{N}}$. Consider the indefinite inner product space $({\mathcal{H}}^\phi, {\left<u_\phi\,\cdot\,,\,\cdot\,\right>})$. Since ${\mathcal{D}}^\phi$ is dense in ${\mathcal{H}}^\phi$ we can find elements $s_1{,\dots,}s_k\in S$ and vectors $\alpha^i=(\alpha^i_1{,\dots,}\alpha^i_k)\in{\mathbb{C}}^k$ ($i=1{,\dots,}m$) such that the elements $$\label{fi} f^i:=\sum_{j=1}^k\alpha^i_j K^\phi_{s_j}\qquad (i=1{,\dots,}m)$$ span an $m$-dimensional negative subspace ([@bognar Theorem IX.1.4]). Let $$\label{A} A:=\left(\phi(s_j+s_{j'}^*+u)\right)_{j,j'=1}^k\in{\mathbb{C}}^{k\times k}.$$ Note that for $i,l=1{,\dots,}m$ we have $$\label{fitransf} {\left<A \alpha^l,\alpha^i\right>}= \sum_{j,j'=1}^k\alpha^{l}_j{\overline}{\alpha^{i}_{j'}}K^\phi(s_j+u,s_{j'}) ={\left<u_\phi f^l,f^i\right>}.$$ Hence, the subspace ${\textrm{lin}}{\left\{\alpha^1{,\dots,}\alpha^k\right\}}$ is a negative subspace of the indefinite inner product space $({\mathbb{C}}^m,{\left<A\cdot,\cdot\right>})$. Since $f^1{,\dots,}f^m$ are linearly independent, the vectors $\alpha^1{,\dots,}\alpha^m$ are linearly independent as well. Therefore, the matrix $A$ has at least $m$ negative eigenvalues. Now let us assume that for some choice of $s_1{,\dots,}s_k\in S$ the matrix $A$ defined as in (\[A\]) has $m$ negative eigenvalues. Then there exists linearly independent vectors $\alpha^i=(\alpha^i_1{,\dots,}\alpha^i_k)\in{\mathbb{C}}^k$ ($i=1{,\dots,}m$) such that $${\left<A \alpha^l,\alpha^i\right>}=\delta_{il}\lambda_i\qquad i,l=1{,\dots,}m,$$ with some $\lambda_1{,\dots,}\lambda_m<0$. We define $f_1{,\dots,}f_m$ as in (\[fi\]) (with the new meaning of $s_1{,\dots,}s_m$). Using the calculation in (\[fitransf\]) we get that the space ${\textrm{lin}}{\left\{f_1{,\dots,}f_m\right\}}$ is a negative subspace of $({\mathcal{H}}^\phi, {\left<u_\phi\cdot,\cdot\right>})$. We show now that $f_1{,\dots,}f_m$ are linearly independent. If $\sum_{i=1}^m \beta_i f_i=0$ for some $\beta_1{,\dots,}\beta_m\in{\mathbb{C}}$ then, by (\[fitransf\]), $${\left<A \sum_{i=1}^m\beta_i\alpha^i,\sum_{i=1}^m\beta_i\alpha^i\right>}={\left<u_\phi \sum_{i=1}^m \beta_i f_i, \sum_{i=1}^m \beta_i f_i\right>}=0.$$ But $A$ is strictly negative on ${\textrm{lin}}{\left\{\alpha^1{,\dots,}\alpha^m\right\}}$ and $\alpha_1{,\dots,}\alpha_m$ are linearly independent. Hence, $\beta_1=\cdots=\beta_m=0$. More examples ============= The reader can easily check that Theorem  \[main\] can be applied to Examples  \[exm1\] and \[exm2\]. The next example concerns the estimation of the dimension of the eigenspace in Theorem  \[main\]. We will show that this dimension can be any number between $0$ and $M/2$, where $M$ is defined as in the proof of Theorem 2. Let $S={{\mathbb{Z}}}_2^m$ with identical involution and let $u=(1,0,0{,\dots,}0)$. We have $M=2^m$. The dual semigroup $S^*$ can be identified with ${\left\{-1,1\right\}}^m$. There are $2^{m-1}$ characters $\sigma$ on $S$ satisfying $\sigma(u)=-1$ and $2^{m-1}$ characters $\sigma$ satisfying $\sigma(u)=1$. We denote those characters by $\sigma_1{,\dots,}\sigma_{2^{m-1}}$ and $\rho_1{,\dots,}\rho_{2^{m-1}}$, respectively. For fixed $k,l\in{\left\{0{,\dots,}2^{m-1}\right\}}$ we put[^5] $\mu:=\sum_{i=1}^k \delta_{\sigma_i}+\sum_{j=1}^{l}\delta_{\rho_j}$ and $\phi:={\mathcal{L}}(\mu)$. Since the support of the measure is consists of $k+l$ points, the space ${\mathcal{H}}^\phi$ is $k+l$ dimensional. To see this one can use the interpretation of ${\mathcal{H}}^\phi$ as $\mathcal P^\mu$, as in Example \[exm1\]. Now let us observe that $$\label{sigmaker} \sigma_1{,\dots,}\sigma_k\in\ker( u_\phi+I),\quad \rho_1{,\dots,}\rho_{l}\in\ker( u_\phi-I),$$ by the same argument as in the proof of Theorem \[main\] (iii)${ \Rightarrow}$(ii). Since the characters are always linearly independent we get that $\dim\ker( u_\phi+I)\geq k$ and $\dim\ker( u_\phi-I)\geq l$. But the eigenspaces corresponding to $-1$ and $1$ are orthogonal, thus $\dim\ker( u_\phi+I)=k$ and $\dim\ker( u_\phi-I)=l$. Let us put $e=(0,1,0{,\dots,}0)$ and take two numbers $l_1\in{\left\{0{,\dots,}2^{m-1}\right\}}$ and $l_2\in{\left\{0{,\dots,}2^{m-2}\right\}}$. Using the same technique we can construct a mapping $\phi\in{\frak{P}(S)}$ such that $\dim\ker(u_\phi+I)=l_1$ and $\dim\ker(e_\phi+I)=l_2$. In the following example there are three elements satisfying $2u=0$ and $u=u^*$, with three different upper bounds for the dimensions of the kernel. Let us consider the semigroup $S=\mathrm{U}({{\mathbb{Z}}}_2^2,{{\mathbb{Z}}}_2,\pi)$ where $\pi(x,y)=x$ for $x,y\in{{\mathbb{Z}}}_2$. The involution on $S$ is identity. Note that $(1,0)+s\neq s$ for $s\in S$, but $(0,1)+s\neq s$ only for $s\in{{\mathbb{Z}}}_2^2$. Hence, the upper bounds for the dimensions of the kernels are three and two, respectively. The dimension of the kernel for $(0,0)$ is obviously zero. Let us take two $*$-separative semigroups $S$ and $T$, both having neutral elements ($0_S$ and $0_T$ respectively) and a $*$-homomorphism $h:S\to T$ satisfying $h(0_S)=0_T$. Take an element $u\in T$ satisfying $2u=0_T$ and $u=u^*$. The $*$-semigroup $\mathrm{U}(S,T,h)$ has a zero, namely $0_S$. However, the element $u$, understood as an element of $\mathrm{U}(S,T,h)$, does not satisfy the condition $2u=0_{\mathrm{U}(S,T,h)}$. Nevertheless, we have $3u=u$ and $u^*=u$, which by [@szaf-bdd] guaranties boundedness (and hence selfadjointness) of $u_\phi$ for any $\phi\in{\frak{P}(\mathrm{U}(S,T,h))}$. The indefinite inner product space $({\mathcal{H}}^\phi,{\left<u_\phi\cdot,-\right>})$ is then degenerate, i.e. $u_\phi$ has a non-trivial kernel. We could investigate, instead of positive definite mappings on $S$, the set of positive definite forms on $S$. Namely, we say that $\phi:S\times \mathcal{E}\times \mathcal{E}\to{\mathbb{C}}$ is a [ form over ($S,\mathcal{E}$)]{} if for every $s\in S$ the mapping $\phi(s,\,\cdot\,,-)$ is a hermitian bilinear form on the linear space $\mathcal{E}$. We say that a form is positive definite if for every finite sequences $(s_k)_k{\subseteq}S$, $(f_k)_k{\subseteq}\mathcal{E}$ we have $\sum_{i,j} \phi(s_j^*+s_i;f_i,f_j))\geq 0$. For a positive definite form $\phi$ we can construct a Hilbert space ${\mathcal{H}}^\phi$ which together with the functions $K_{s,f}^\phi$ ($s\in S$, $f\in\mathcal E$) constitute a RKHS. Like in the case of $\mathcal{E}={\mathbb{C}}$, cf. [@szaf-bdd] and also [@general], we can define the (closed) shift operator associated with an element $u\in S$ by $u_\phi(K_{s,f}^\phi)=K_{s+u,f}^\phi$. The following example shows, that in this setting the equivalence in Theorem \[main\] no longer holds. Let $S={{\mathbb{Z}}}_2$ (with the identical involution) and let $\mathcal{E}$ be an dimensional [Hilbert]{} space. Consider the following form $$\phi(x,f,g)=\left\{\begin{array}{rcl} {\left<f,g\right>}_\mathcal{E} & : & x=0\\ {\left<-f,g\right>}_\mathcal{E} & : & x=1 \end{array}\right. .$$ Note that $$\sum_{x,y=0,1} \phi(x+y,f_x,f_y)={\left<f_0,f_0\right>}+{\left<f_1,f_1\right>}-2\Re{\left<f_1,f_0\right>}={\left\Vertf_1-f_0\right\Vert}^2$$ which is greater or equal to zero for any choice of $f_0,f_1\in{\mathcal{H}}$. The space ${\mathcal{H}}^\phi$ can be realized as ${\mathcal{H}}^\phi=\mathcal{E}$ so as $K_{0,f}^\phi=f$ and $K_{1,g}=-g$, $f,g\in \mathcal{E}$. Take $u=1$. The condition (i) of Theorem \[main\] is satisfied because the semigroup is of finite cardinality. On the other hand $u_\phi K_{0,f}^\phi = K_{1,f}^\phi=-f$ for $f\in{\mathcal{H}}$. Hence, $\dim\ker(u_\phi+I)=\dim{\mathcal{H}}=\infty$. Final remarks ============= Our work is connected in a way with many other papers and books. Let us mention some of them. - The transformation $\phi\mapsto\phi(\cdot+u)$ has been investigated by Bisgaard. He showed in [@biss-fact] that it is always a sum of four positive definite mappings. - Functions with finite number of negative spaces on topological groups has been considered in the book of Sasvári [@sasvari]. - In [@berg] sequences on ${\mathbb{N}}$ with a finite number of negative squares are considered. - In [@BChR] the authors consider negative definite sequences, which is a subclass of mappings with a finite number of negative squares. - Finally, in [@biss-cor] definitizing ideals are investigated. [99]{} F.H. Szafraniec, Interpolation and domination by positive definite kernels, in [*Complex Analysis - Fifth Romanian-Finish Seminar*]{}, Part 2, Proceedings, Bucarest (Romania), 1981, eds. C. Andrean Cazacu, N. Boboc, M. Jurchescu and I. Suciu, Lecture Notes in Math., [**1014**]{}, 291-295, Springer, Berlin-Heidelberg, 1893. F.H. Szafraniec, The Sz.-Nagy “théorème principal” extended. Application to subnormality, [*Acta Sci. Math. [(]{}Szeged[)]{}* ]{}, [**57**]{}(1993), 249-262. F.H. Szafraniec, The reproducing kernel Hilbert space and its multiplication operators, [*Oper. Theory Adv. Appl.*]{}, [**114**]{}(2000), 253-263. B. Sz.-Nagy, [*Extensions of linear transformations in Hilbert space which extend beyond this space*]{}, Appendix to F. Riesz, B. Sz.–Nagy, [*Functional Analysis*]{}, Ungar, New York, [**1960**]{}. [^1]: The first author was supported by the MNiSzW grant N201 026 32/1350. He also would like to acknowledge an assistance of the EU Sixth Framework Programme for the Transfer of Knowledge “Operator theory methods for differential equations” (TODEQ) \# MTKD-CT-2005-030042. [^2]:  Cf. Remark \[stupidstyle\]. [^3]:  We can use the following argument: For every $s\in S$ the function $\hat s$ is a character on the dual semigroup $S^*$ and it is trivial that the family $(\hat s)_{s\in S}$ separates elements of $S^*$. Proposition 2 of [@lacunae] (see also [@BChR Proposition 6.1.8]) says that if $T$ is a semigroup and $C{\subseteq}T^*$ separates points, then the functions $\hat t{\!\!\mid_{C}}$, $t\in T$ are linearly independent in ${\mathbb{C}}^C$. We use this result for $T=S^*$ and $C={\left\{\hat s:s\in S\right\}}$, the functions $\hat\sigma{\!\!\mid_{C}}$ can be identified with characters on $S$. [^4]:  Remark \[eitheror\] shows that it is even equivalent to [^5]:  We use the convention $\sum_{i=1}^0 a_i:=0$

ArXiv

--- abstract: 'We report a beyond mean-field calculation of mass current and superfluidity fraction for a system of few bosons confined in a ring geometry in the presence of a rotating weak link induced by a potential barrier. We apply the Multiconfiguration Hartree Method for bosons to compute the ground state of the system and show the average superfluidity fraction for a wide range of interaction strength and barrier height, highlighting the behavior of density correlation functions. The decrease of superfluidity fraction due to the increase of barrier height is found whereas the condensation fraction depends exclusively on the interaction strength, showing the independence of both phenomena.' author: - 'Alex V. Andriati$^{1}$' - Arnaldo Gammal$^1$ bibliography: - 'ref.bib' title: Superfluidity fraction of few bosons in an annular geometry in the presence of a rotating weak link --- [^1] [^2] Introduction ============ The concepts of superfluidity [@Khalatnikov; @Leggett1998; @RevModPhys.Legget1999] and Bose-Einstein condensation [@Cornell; @Pethick] have dominated the research of cold bosonic systems. The presence of one do not necessarily imply the other, whereas superfluidity is related to dissipationless flow due to a minimum required energy to create excitations, Bose-Einstein condensation is characterized by a single macroscopically occupied state. Superfluid systems with very small condensation fraction around $10\%$, like liquid Helium, are widely known [@Penrose; @Sears], thereby characterizing independent effects. Nevertheless, many reports explore the superfluidity features of a Bose-Einstein condensate (BEC), as dilute cold bosonic gases are able to present both phenomena simultaneously [@Ketterle1999]. Specially, persistent flow, hallmark of superfluidity, has been reported for a BEC trapped in a ring shape format early in [@PhysRevLett.99.260401] and later in [@TunableWeakLink] in the presence of a tunable weak link. This boosted the interest to quantitatively study all the properties for the system due to a possible connection and quantum analogy with superconducting quantum interference devices (SQUID) [@SQUIDhandbook], that was experimentally implemented [@PhysRevLett.111.205301] generating Josephson junctions [@PhysRevLett.95.010402]. Many works studied several properties on imposing rotation for a BEC confined in a ring shape geometry, observing hysteresis (“swallow tail loops”) [@Mueller2002; @CampbellNature2014; @PhysRevA.87.013619; @Syafwan2016], excitation mechanisms [@PhysRevA.88.063633; @PhysRevA.99.043613; @PhysRevA.95.021602], spin superflow [@PhysRevLett.119.185302; @PhysRevLett.110.025301] and superfluid fraction [@Hadzibabic2010]. Although the theoretical studies rely mostly on the Gross-Pitaevskii (GP) equation that set a clear limitation on controlling the interactions to suppress the depletion from the condensate [@PhysRevLett.119.190404; @PhysRevLett.117.235303], this has changed in the past few years with the development of novel methods able to compute many-body observables and assure correctness for a wider range of interaction values. The employment of new methods pave a way to study independently the condensation phenomena and the superfluidity for a system of cold bosonic atoms and to sweep a wider range for interaction strength since depletion is included on the description. Moreover, they enable us a deep understanding of the physical system through new many-body quantities unseen in the mean field formalism, like correlations that has gained importance due to experimental measures in the past decade for cold atomic clouds [@Navon167; @Glauber; @10.1038/nphys2632; @PhysRevLett.118.240402]. Specifically, the Multi-Configuration Time Dependent Hartree method for Bosons (MCTDHB) [@MCTDHBderivation] has dragged attention for its quite straightforward generalization of the GP equation, since it is still based on variational principle but with more single-particle states (also known as orbitals) the atoms can occupy, therefore allowing the expansion of the many-body state in a configuration basis (Fock states) with each possible configuration expressed by a well-defined occupation number of the single-particle states. This procedure truncate the Hilbert space, whereas the coefficients of the basis expansion and the orbitals are determined by minimizing the action, enforcing an optimized basis. The MCTDHB has shown to be a powerful tool for many applications like bosons in optical lattices [@PhysRevA.97.043625], quench dynamics [@PhysRevA.92.033622] and other applications [@PhysRevX.9.011052; @PhysRevA.94.063648; @PhysRevA.93.063601] as well as a version for fermions [@PhysRevA.93.033635]. In the present report, we study beyond mean field the superfluidity fraction of a gas of bosons at zero temperature in the presence of a tunnable weak link moving in a periodic system (an effective ring), for a small number of atoms, using the MCTDHB to exploit strong interactions and to show the loss of the superfluidity fraction under a wide range of the physical parameters. Explanations for the superfluid properties are studied here through correlation of the bosonic field operator in the absence of rotation, which is directly related to the tunneling amplitude in the neighborhood of the barrier, therefore allowing a prediction if the atoms would be dragged when some rotation is applied. Model and methods {#secII} ================= The specific form of a barrier is generally unknown from an experimental perspective, though we must be able to define it through its thickness and height. As most of the experiments use lasers to physically implement a barrier [@TunableWeakLink], the height in the model plays the principal role as it is directly related to laser intensity, and the thickness can be fixed in a first moment. Indeed, an approach based on Dirac delta function for the barrier has been reported [@AnnaMinguzzi], which implies zero thickness. In any case, for a barrier rotating with velocity $v$, in the laboratory frame we thus have the one-body term of the Hamiltonian in the general form [ $$\hat{h}(t) = - \frac{\hbar^2}{2 m} \frac{\partial^2}{\partial \bar{x}^2} + U(\bar{x} - vt) \ ; \quad \bar{x} \in (-\pi R, \pi R] , \label{eq:labframe}$$ ]{} for a ring of radius $R$. The two-body part is assumed to be described by an effective contact interaction $V(\bar{x} - \bar{x}') = g_{1\mathrm{D}} \delta(\bar{x} - \bar{x}')$, where $g_{1\mathrm{D}}$ is related to the transverse harmonic trap frequency and the scattering length of the atoms. Using a unitary transformation to move to the rotating frame, the time dependence of Eq. is removed, resulting in the following many-body Hamiltonian in the second quantized formalism [ $$\begin{gathered} \mathcal{H} = \int_{-\pi R}^{\pi R} \!\!\!\!\! \mathrm{d}x \ \Psi^{\dagger}(x) \left[ \frac{\hbar^2}{2 m} \left( i \frac{\partial}{\partial x} + \frac{m v}{\hbar} \right)^2 + U(x) \right] \Psi(x) \ + \\ \frac{g_{1\mathrm{D}}}{2} \int \! \! \! \mathrm{d}x \ \Psi^{\dagger}(x) \Psi^{\dagger}(x) \Psi(x) \Psi(x) , \label{eq:secondquantized} \end{gathered}$$ ]{} where $x = \bar{x} - vt$. The MCTDHB is developed assuming a truncated Hilbert space where the many-body state is a superposition of all possible configurations $N_c$ of $N$ particles distributed over $M$ single-particle states, such that we can write $$| \Psi(t) \rangle \ \doteq \ \sum_{\alpha = 1}^{N_c} C_{\alpha}(t) | \vec{n}^{(\alpha)} \rangle , \quad N_c = \binom{N + M - 1}{ M - 1} , \label{eq:manybody-state}$$ where a valid configuration $| \vec{n}^{(\alpha)} \rangle$ is a Fock state where $\sum_{i}^{M} n_i^{(\alpha)} = N$, $\forall \alpha \in \mathbb{N} \ | \ 1 \leq \alpha \leq N_c$. Furthermore, the occupation number refers to a set of single-particle states $\{ \phi_k(x,t) \ | \ \int \!\! \mathrm{d}x \phi_l^{*}(x,t) \phi_k(x,t) = \delta_{lk}, \forall k,l = 1, ..., M \}$. Using this *Ansatz*, the time-dependent equations can be extracted from a minimization of the action with respect to the coefficients $C_{\alpha}$ in Eq. and the single-particle state, with the action defined by [$$\begin{gathered} \mathcal{S} \Big[ \mathbf{C} , \{ \phi_k, \phi_k^{*} \} \Big] = \int \!\!\! \mathrm{dt} \bigg[ \langle \Psi(t) | \dot{\Psi}(t) \rangle \ - \ \langle \Psi(t) | \mathcal{H} | \Psi(t) \rangle\ \\ - \sum_{k,l=1}^{M} \mu_{kl}(t) \langle \phi_k | \phi_l \rangle _t \bigg],\end{gathered}$$ ]{} where the $\mu_{kl}$ are introduced as Lagrangian multipliers to maintain orthonormality of the single-particle states. The variational principle conducts to $M$ nonlinear coupled partial differential equations for the set $\{ \phi_k(x,t)\}$ and a system of $N_c$ ordinary differential equation for the coefficients $C_{\alpha}$ [@Hans-Meyer; @MCTDHBderivation]. It is worth mentioning that the GP equation is a special case where we have just one possible configuration $| \vec{n} \rangle = | N, 0, ..., 0\rangle$, that yield for the macroscopically occupied orbital $\phi(x,t)$ the equation [$$i \hbar \frac{\partial \phi}{\partial t} = \Big[ \hat{h}' + g (N-1) |\phi(x,t)|^2 \Big] \phi (x,t) , \label{eq:gp}$$ ]{} with $\hat{h}' = \hbar^2 /2 m \left( i \partial / \partial x + m v /\hbar \right)^2 + U(x)$ the one-body Hamiltonian in the rotating frame. For numerical simulation purposes, we assume the following system of units: length measured in units of $(\pi R)$, probability/particle density in unis of $(\pi R)^{-1}$ and energy by $\hbar \zeta$ where $\zeta = (\hbar / 2 m \pi^2 R^2)$. Moreover, we introduce the dimensionless parameters $\Omega = m R v / \hbar$ and $\gamma = 2 m \pi R g_{1\mathrm{D}} / \hbar ^2$. All these transformations yield the following orthonormal condition for the set of orbitals: $\int_{-1}^{1} \!\! \mathrm{d} x \phi_l^{*}(x,t) \phi_k(x,t) = \delta_{lk}$. Here we developed our own codes to solve the MCTDHB equations with periodic boundary conditions. Our codes were extensively tested, matching results of the examples of the code available in Ref. [@MCTDHpackage], that has produced many results until now [@PhysRevA.93.033635; @PhysRevA.93.063601; @PhysRevA.94.063648; @PhysRevX.9.011052; @PhysRevA.92.033622]. Periodicity in energy spectrum and definition of superfluidity fraction {#secIII} ======================================================================= In the absence of a barrier, the single-particle energy levels as function of $\Omega$ are parabolas given by $E_j / (\hbar \zeta) = (j - \Omega)^2 \pi^2$, each one defined by the winding number of the phase ($j$), centered at $\Omega_j = j$, and crossing each other at $\tilde{\Omega}_j = (j + 1/2) ~$[@PhysRevA.87.013619]. ![(Color online) Energy per particle from GP equation as a function of $\Omega$ for different winding numbers ($j$) with (A) $\gamma (N-1) = 50$ and (B) $\gamma (N-1) = 10$. In both figures the soliton solutions energy are depicted in blue, whose connect the parabolas with $j = 0$ to $j = 1$, shown by dotted and full line respectively. Other values of winding numbers are shown in gray.[]{data-label="fig:nobarrier_energy"}](nobarrier_soliton_branches.pdf) ![Phase profile of soliton solution for some values of $\Omega$ corresponding to the red dots in \[fig:nobarrier\_energy\]. An abrupt transition occurs at $\Omega = 0.5$ that implies a transition in the winding numbers.[]{data-label="fig:nobarrier_soliton_phases"}](soliton_phases_gam10.pdf) As a first approach, we use the GP equation once the interaction is included in the description. In this case, still in the absence of the barrier, there are two kinds of analytical solutions, one with constant density, which results in the same energy of single-particle case, with the addition of an interaction contribution, yielding $E^{(GP)}_{j} / (\hbar \zeta) = (j - \Omega)^2 \pi^2 + \gamma (N-1) / 4$ as average energy per particle. The other is a soliton given in terms of Jacobi elliptic functions [@PhysRevA.62.063610; @Sato2016] that exists for a finite range of values of $\Omega$, whereas the extension of this range depends on the interaction strength. Fig. \[fig:nobarrier\_energy\] shows an energy landscape of the analytical solutions of the GP equation with the soliton solution energy connecting two parabolas from constant density solutions, where the dotted lines have winding number 1 and the filled line 0. From Fig. \[fig:nobarrier\_soliton\_phases\] we check that the soliton solution is responsible for the transition between different winding numbers and a discontinuity occurs in $\Omega = 0.5$ in its phase. Moreover, the soliton connects the values of angular momentum [@GuilleumasHysteresis] and this connection between the two lines is related to an hysteretic behavior by the presence of a “swallow-tail” loop [@GuilleumasHysteresis; @Mueller2002; @CampbellNature2014; @PhysRevA.87.013619]. The soliton branch in Fig. \[fig:nobarrier\_energy\] is an excited state [@PhysRevA.87.013619], however, it will not be discussed here, since the aim of the present work is to measure the superfluidity fraction for the ground state. In addition, Fig. \[fig:nobarrier\_energy\] reveals that the ground state energy has a periodic behavior with respect to the rotation $\Omega$, with kinks where the parabolas cross each other at $\tilde{\Omega}_j = (j + 1/2)$. This periodic structure remains even in the presence of a barrier as will be shown later. An important fact is that we can relate the mass current circulation with the energy, and use this periodicity to understand what happens with the current under action of fast rotating barriers. Here we start a derivation of mass current by looking at the time variation of the number of atoms within the range $[x_1 , x_2] \subseteq [-\pi R, \pi R]$, as [$$\frac{\mathrm{d}}{\mathrm{d} t} \int_{x_1}^{x_2} \!\!\!\!\!\! \mathrm{d} x \ \langle \Phi(t) | \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x) | \Phi(t) \rangle = \frac{i}{\hbar} \int_{x_1}^{x_2} \!\!\!\!\!\! \mathrm{d} x \ \langle \big[ \mathcal{H}, \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x) \big] \rangle_t , \label{eq:TimeVariationOfParticles}$$ ]{} where $\langle \cdot \rangle_t$ means the expectation value for an arbitrary many-body state $| \Phi (t) \rangle$. Using Eq. with the usual commutation relation for the boson field operator $[\hat{\Psi}(x),\hat{\Psi}^{\dagger}(x')] = \delta(x - x')$ to evaluate the commutator of the Hamiltonian with the density operator, the only terms that contribute are those carrying a derivative, and yields $$\begin{gathered} \big[ \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x), \mathcal{H} \big] = - \frac{\hbar^2}{2m} \left( \hat{\Psi}^{\dagger}(x) \frac{\partial^2 \hat{\Psi}(x)}{ \partial x^2} - \frac{\partial^2 \hat{\Psi}^{\dagger}(x)}{ \partial x^2} \hat{\Psi}(x) \right) \\ + i \hbar v \left( \hat{\Psi}^{\dagger}(x) \frac{\partial \hat{\Psi}(x)}{ \partial x} + \frac{\partial \hat{\Psi}^{\dagger}(x)}{ \partial x} \hat{\Psi}(x) \right) . \label{eq:density_H_commutator}\end{gathered}$$ It is straightforward to factor out the derivative with respect to $x$, and further using Eq. in Eq. yields [$$\frac{\mathrm{d}}{\mathrm{d}t} N([x_1,x_2];t) = - \Big[ \langle \hat{J} (x_2) \rangle_t - \langle \hat{J} (x_1) \rangle_t \Big]$$ ]{} where $ N([x_1,x_2];t) = \int_{x_1}^{x_2} \mathrm{d} x \langle \Phi(t) | \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x) | \Phi(t) \rangle$ is introduced and the particle number current operator $\hat{J}(x)$ is given by [$$\begin{gathered} \hat{J}(x) = - \frac{i \hbar}{2m} \left( \hat{\Psi}^{\dagger}(x) \frac{\partial \hat{\Psi}(x)}{ \partial x} - \frac{\partial \hat{\Psi}^{\dagger}(x)}{ \partial x} \hat{\Psi}(x) \right) \\ - v \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x) .\end{gathered}$$ ]{} The reduced single-particle density matrix (1-RDM) defined by $n^{(1)}(x,x';t) \doteq \langle \hat{\Psi}^{\dagger}(x') \hat{\Psi}(x) \rangle_t $ [@Glauber] has as a set of eigenvalues and eingenstates defined by the solution of $\int_{-\pi R}^{\pi R} \! \mathrm{d} x \ n^{(1)}(x,x';t) \psi(x',t) = \mathcal{N}(t) \psi(x,t)$, with $\mathcal{N}(t)$ the average occupation number in the eigenstate $\psi(x,t)$, here also called as *natural orbital*. Using these natural orbitals to express the reduced single-particle density matrix $n^{(1)}(x,x';t)$ allows us to express the current as a superposition, $\langle \hat{J}(x) \rangle_t = \sum_{k} j_k(x,t)$, where [$$j_k = - \left[ \frac{i \hbar}{2m} \left( \psi_k^{*} \frac{\partial \psi_k}{\partial x} - \psi_k \frac{\partial \psi_k^{*}}{\partial x} \right) + v|\psi_k|^2 \right] \mathcal{N}_k ,$$ ]{} with the position and time arguments omitted. For the ground state, the current $\langle \hat{J}(x) \rangle_t$ must be independent of position and time, because the density is not time-dependent. If we further average it over a period in the counter direction of the barrier velocity, yields [$$\langle \rho_s \rangle (v) = \tau \frac{1}{2\pi R} \int_{\pi R}^{-\pi R} \!\!\!\!\! \mathrm{d} x \ \left( \frac{\langle \hat{J} \rangle}{N} \right) , \label{eq:currentJ}$$ ]{} ![(Color online) Ground state energy and current fraction for 11 particles as a function of dimensionless rotation velocity $\Omega$ in the rotating frame. The ground state energy remains periodic as it was in Fig. \[fig:nobarrier\_energy\] but with a different landscape depending on the barrier height $\lambda$, and this periodicity implies a decrease on the current fraction for fast rotating barriers. Dimensionless interaction strength parameter used was $\gamma = 10$.[]{data-label="fig:CurrentXRotation"}](CurrentXrotation.pdf) ![Probability distribution for position(upper panel) where $n^{(1)}(x) \equiv n^{(1)}(x,x) = \langle \hat{\Psi}^{\dagger}(x) \hat{\Psi}(x) \rangle$ and angular momentum distribution(lower panel) for barrier height $\lambda = 1000$. From left to right $\Omega = 0, 0.5, 1.0$, corresponding to red crosses in Fig. \[fig:CurrentXRotation\], and as used before $\gamma = 10$ for 11 particles.[]{data-label="fig:OBprobabilities"}](OBprobabilities_H1000.pdf) where $\tau = 2 \pi R / v$ is the period of barrier rotation. This quantify the mean fraction of particles that go through the counter direction of the barrier in its period, that is from $\pi R$ to $-\pi R$ indicated by the limits of integration taken. Therefore, if $\langle \rho_s \rangle (v)$ takes the value 1, means a perfect superfluid since all the particles are flowing with velocity $-v$ in the rotating frame, that is, they remain at rest for a observer in the laboratory frame. Relations with other observables can be established, for instance, using the average momentum per particle [ $$\langle \rho_s \rangle(v) = \left(1 - \frac{\langle \hat{p} \rangle}{mv} \right), \quad \hat{p} = - \frac{i \hbar}{N} \int_{-\pi R}^{\pi R} \!\!\!\!\!\!\!\! \mathrm{d}x \ \Psi^{\dagger}(x) \frac{\partial}{\partial x} \Psi(x) , \label{eq:currentMomentum}$$ ]{} and a relation with the energy, by taking the derivative with respect to the barrier velocity [ $$\langle \rho_s \rangle (v) = \frac{1}{N m v} \frac{\partial E}{\partial v}, \quad E = \langle \mathcal{H} \rangle . \label{eq:currentEnergy}$$ ]{} The equation above can also be identified by the ratio between the moment of inertia of the atoms and the moment of inertia of a rigid body. Using $v = \omega R$, yields [$$\langle \rho_s \rangle(\omega) = \frac{1}{N m R^2} \left( \frac{1}{\omega} \frac{\partial E}{\partial \omega} \right) = \frac{I(\omega)}{I_{\mathrm{cl}}}. \label{eq:currentInertia}$$ ]{} Superfluidity fraction at rest (or simply superfluidity fraction), denoted here by $\langle \rho_s \rangle_0$ can be defined by taking the limit of $v \rightarrow 0$ in any of the forms ,, or and was studied in this way in previous works [@Hadzibabic2010; @RevModPhys.Legget1999; @Leggett1970; @Leggett1998]. With the dimensionless system of units and parameters introduced in the end of section \[secII\], we have a suitable expression for numerical calculations [$$\langle \rho_s \rangle (\Omega) = \left( \frac{1}{2 \pi^2 N \Omega} \frac{\partial E}{\partial \Omega} \right) , \ \langle \rho_s \rangle_0 = \lim_{\Omega \rightarrow 0} \langle \rho_s \rangle (\Omega) . \label{eq:rho_dimensionless}$$ ]{} Here we use the MCTDHB to find the ground state through imaginary time propagation for several parameters, and we first study the effect of rotation. Fig. \[fig:CurrentXRotation\] illustrate the behavior of the energy in panel (A) and the current fraction in panel (B) as function of dimensionless barrier frequency $\Omega$ for two different barrier heights, where the specific form used in Eq. was [ $$U(x) = \left\{ \begin{array}{lcl} \displaystyle (\hbar \zeta \lambda) \cos^2{\left( \frac{x}{2 R \sigma} \right)} & \mathrm{if} \ & | x | \leq \pi R \sigma \\ 0 & \mathrm{if} \ & \ \pi R \geq | x | > \pi R \sigma \end{array} \right. , \label{eq:barrierform}$$ ]{}  where $\lambda$ denotes the barrier height in dimensionless units and the width of the barrier was taken fixed $\sigma = 0.1$ . The energy of the ground state in Fig. \[fig:CurrentXRotation\](A) has a period 1 with respect to dimensionless rotation frequency for both cases of weak and strong barriers, while the difference relies on the maximum that occurs at $\Omega_j = j$, that is peaked or smooth. The current fraction shows a periodic behavior with a damped amplitude as function of $\Omega$ in Fig. \[fig:CurrentXRotation\](B), due to the periodicity of energy, where according to Eq. , the amplitude is damped by a factor of $1 / \Omega$. In the regions where $\langle \rho_s \rangle (\Omega) < 1$ the average momentum must increase together with the barrier velocity by Eq. . Indeed, that is what occurs in lower panel of Fig. \[fig:OBprobabilities\] that shows the angular momentum distribution fo some values of $\Omega$. Moreover, there is a critical dependence of the superfluid fraction on the barrier height, where Fig. \[fig:CurrentXRotation\] shows that, as $\Omega$ goes to zero, $\langle \rho_s \rangle (\Omega)$ becomes as smaller as higher is the barrier. This fact will be explored in the following. Decrease of superfluidity fraction due to increase of the barrier height {#secIV} ======================================================================== Numerical calculations of superfluidity fraction was carried out here using Eq. , finding the ground state by imaginary time propagation for $\Omega = 0$ and $\Omega = 0.02$, to approximate the derivative in $\Omega = 0.01$ and so get $\langle \rho_s \rangle(0.01)$. As showed by Fig. \[fig:CurrentXRotation\] the slope of current fraction goes to zero as $\Omega \rightarrow 0$, and therefore we use the value at $\Omega = 0.01$ as the proper superfluidity fraction, assuming the difference of $\langle \rho_s \rangle_0 - \langle \rho_s \rangle(0.01)$ to be close to zero. To assure this method is valid, we compare with the result using Eq. at $\Omega = 0.02$, to check if there is no appreciable(less than 1%) variation on the estimation of superfluidity fraction using a constant extrapolation of $\langle \rho_s \rangle(0.01)$. In Fig. \[fig:flow\_barrier\], we show the decrease of superfluid fraction for an increase in the barrier height for the form in Eq. , using different number of particles and interaction strength. Here the tunneling of particles through the barrier is as harder as higher is the barrier, thereby the system acquires momentum easily for stronger barriers because it drags almost every particle with it. This easy momentum gain for very strong barriers is responsible for the loss of superfluidity fraction $\langle \rho_s \rangle_0$. The superfluidity fraction decreases more rapidly for fewer particles and lower interaction strength, however, the number of particles and strenght of interactions have a small impact in the form of the curves of $\langle \rho_s \rangle_0$ as a function of $\lambda$. ![Probability distribution of position(upper panel) and angular momentum(lower panel) for 5 particles, $\gamma = 30$ and different barrier heights $\lambda = 10, 200, 10000$ in the left, center and right column respectively, in the absence of rotation $\Omega = 0$. The density distribution vanishes for $\lambda > 10^3$ at the peak of the barrier in $x = 0$, despite there is just a slight increase on the width of the angular momentum distribution.[]{data-label="fig:OBprobabilities_barrier"}](OBprobabilities_barrier.pdf) As can be seen in upper panel of Fig. \[fig:OBprobabilities\_barrier\], the barrier height $\lambda$ influence mostly the density at its peak, while the effect over the momentum distribution is a slight increase in its variance, but preserving $\langle \hat{L} \rangle = 0$, as can be checked by the lower panel. This effect can also be seen in Fig. \[fig:CurrentXRotation\], from which for $\lambda = 10$ the barrier has a critical value $\Omega = 0.5$ to start to move the particles, whereas for $\lambda = 1000$ it drags more easily the atoms since there $\langle \rho_s \rangle_0 \approx 0.46$. 1 10 30 ---- ----------------- ------------- ------------- 11 0.9936 / 0.9920 0.92 / 0.89 - 8 0.9946 / 0.9935 0.92 / 0.88 - 5 0.9962 / 0.9956 0.91 / 0.88 0.75 / 0.70 : maximum/minimum condensation fraction numbers given by the highest eingenvalue of $n^{(1)}(x,x')$, over the set of values of $\lambda$ in Fig. \[fig:flow\_barrier\]. The maximum and minimum values for each case have little influence from the barrier height whereas the superfluid fraction maximum and minimum values goes from 1 to 0 respectively. For $\gamma = 30$ we were able to perform the calculations only for 5 particles due to our code limitations.[]{data-label="tab:condensation_fraction"} ![(Color online) $| g^{(1)}(x,x') |^2$ mapped to colors in the ring, for 5 particles and $\gamma = 30$, where the horizontal and vertical axes represent $x$ and $x'$ values in units of $\pi R$, respectively. Values of barrier height used are $\lambda = 10$ in (A) and $\lambda = 10^4$ in (B) but sharing the same color scale.[]{data-label="fig:OBcorrelation"}](OBcorrelation_gam30.pdf) It is worth noting that this loss of superfluidity fraction due to increase in the barrier height is not related with the condensation fraction. As can be inferred from table \[tab:condensation\_fraction\], the condensation fraction depends mostly on the interaction strength and is minimally affected by the barrier height, particularly for small values of $\gamma$. Being able to investigate many-body quantities that are unreachable using the GP equation, we further investigate how the tunneling amplitude is affected by the barrier height, that is the transition amplitude for the system to move a particle from $x$ to $x'$. This can be achieved by $| \langle \hat{\Psi}^{\dagger}(x') \hat{\Psi}(x) \rangle |^2$ weighted by the probabilities to find the particles in respective positions given by $n^{(1)}(x)$ and $n^{(1)}(x')$. This is directly related to the first order normalized correlation function defined by [@Glauber; @PhysRevA.78.023615] $$g^{(1)}(x,x') = \frac{ \langle \hat{\Psi}^{\dagger}(x') \hat{\Psi}(x) \rangle }{\sqrt{n^{(1)}(x) n^{(1)}(x')}} . \label{eq:correlation1}$$ The values of $g^{(1)}$ shall be drastically affected by the barrier and must have an abrupt variation as $x x' > 0$ changes to $x x' < 0$, since the tunneling must be much harder if the shorter distance between two points has the barrier between them. Reminding that the system is periodic, this discussion applies just at the vicinity of either $x$ or $x'$ being zero, because if $x x' = -1$ they are actually the same point in the ring. The effect of barrier height mentioned above is in agreement with the images in Fig. \[fig:OBcorrelation\] that maps $|g^{(1)}(x,x')|^2$ values to colors. In panel (A), in the presence of weak barrier, it depends only on $|x - x'|$, that means the tunneling is smaller as higher is the distance between the two points, while in panel (B) this symmetry is lost, with an abrupt variation near at the barrier peak, $x$ or $x'$ approximately zero. Therefore, high barriers split the image in four square blocks, with the darker regions (small normalized tunneling probabilities) located on $x x' < 0$. This is consistent with previous studies in Ref. [@PhysRevA.78.013604], despite the different boundary conditions and interaction regimes. We further stress the relevance of applying a method that allows us to compute many-body quantities, since for example $| g^{(1)}(x,x') |^2$ would be identical to $1$ for all $x$ and $x'$ in case one uses the GP equation, that correspond to considering just one eigenstate of $n^{(1)}(x,x')$. Conclusions {#secV} =========== In this paper, from a general derivation of the number of atoms current in a ring, we studied the persistent flow under the rotation of a barrier, and explained its behavior under different conditions using MCTDHB, which allowed us to explore strong interaction regimes with few particles. This method enable us to check convergence for the observables presented here, enlarging the basis of the spanned space by using more single-particle states determined by action minimization and, therefore, achieve a beyond mean field theory and access to new observables as the one-body correlation function. Here we reported the periodicity of ground state energy under the MCTDHB and the effect of a barrier in the rotating frame, which changes the landscape of energy with respect to rotation velocity, producing narrow or smooth peaks periodically at dimensionless rotation frequency $\Omega_j = (j + 1/2)$ for weak or strong barriers respectively. The barrier also affects the particle number current fraction that either remains a perfect superfluid($\langle \rho_s \rangle (\Omega) = 1$) in the region $| \Omega | < 0.5$ for weak barriers, or has a fraction of the particles dragged by the system even for infinitesimal rotations, that is $\lim_{\Omega \rightarrow 0}\langle \rho_s \rangle (\Omega) < 1$. The superfluidity fraction decrease due to increase of the barrier height showed to be unrelated to condensation fraction since the value slightly changed under a wide range of values for the height of the barrier. However, the one body correlation function, that introduces a tunneling amplitude between two points weighted by the probability to find particles in the respective points is a key quantity to understand this observation. If the particles can pass through the obstacle without gaining momentum, in other words, tunnel through the barrier, this flow behaves as a perfect superfluid, that is, it stays at rest as the barrier starts to move. Indeed, this was quantitatively predicted by the one-body correlation function. Acknowledgements {#acknowledgements .unnumbered} ================ The authors thank the Brazilian agencies Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP) and Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq). We gratefully thanks to A. F. R. T. Piza, E. J. V. Passos and R. K. Kumar for the elucidating discussions. We are also grateful for conversations with A. U. J. Lode and M. C. Tsatsos about the implementation of codes to numerically solve the MCTDHB equations. [^1]: [email protected] [^2]: [email protected]

ArXiv

--- address: | Faculty of Mathematics, TU Dortmund University, 44221 Dortmund, Germany\ <[email protected]>\ author: - Frank Klinker title: 'An explicit description of $SL(2,{\mathbbm C})$ in terms of $SO^+(3,1)$ and vice versa ' --- Introduction ============ It is well known that for a pseudo-euclidean vector space $(V,g)$ the universal cover of the special orthogonal group $SO(V,g)$ is given by the so called spin group $Spin(V,g)$. For the case $V={\mathbbm R}^{p+q}$ and $g={\rm diag}(\mathbbm{1}_q,-\mathbbm{1}_p)$ we write $SO(p,q)$ and $Spin(p,q)$. The covering map is 2:1 for $\dim V>2$. The theoretic setting in which spin groups and related structures are best described is the Clifford algebra ${{\rm C}\ell}(V,g)$, see [@Che; @Har; @MichLaw] for example. Although spin groups in general refrain from being described by classical matrix groups for dimensional reason, there are accidental isomorphisms to such in dimension three to six, see Table \[tab:low\]. The isomorphisms are a consequence of the classification of Lie algebras and can for example be seen by recalling the connection to Dynkin diagrams. We use the notation from [@Helga] and recommend this book for details on the definition of the classical matrix groups. Due to the fact that the complexifications of the orthogonal groups are independent of the signature of the pseudo-Riemannian metric the groups in each column of Table \[tab:low\] are real forms of the same complex group for fixed dimension. $\displaystyle {\renewcommand{{1.5}}{1.5} \begin{array}{c|c||c|c||c|c||c|c} \multicolumn{2}{c||}{p+q=3} & \multicolumn{2}{c||}{p+q=4} &\multicolumn{2}{c||}{p+q=5} & \multicolumn{2}{c}{p+q=6} \\\hline (3,0)& SU(2) &(4,0)& SU(2)^2 &(5,0)& Sp(2) &(6,0)& SU(4)\\ (2,1)& SL(2,{\mathbbm R}) &(3,1)& SL(2,{\mathbbm C}) &(4,1)& Sp(1,1) &(5,1)& SL(2,{\mathbbm H})\\ & &(2,2)& SL(2,{\mathbbm R})^2 &(3,2)& Sp(4,{\mathbbm R}) &(4,2)& SU(2,2)\\ &&&&& &(3,3)& SL(4,{\mathbbm R}) \end{array} } $ \[tab:low\] Infinitesimally, i.e. on Lie algebra level, the 2:1 covering structure cannot be seen. Therefore, the description on this level is given by fixing bases in the respective Lie algebras. If we try to take over this to the groups we see that the exponential map enters in the construction. A useful and manageable description is not obtained in general due to the Baker-Campbell-Hausdorff formula. However, in dimension four such description is possible and we present explicit formulas for the maps that connect $SO(3,1)$ and $SL(2,{\mathbbm C})$. Some preliminaries ================== We will give some preliminaries on Lorentz transformations, Pauli matrices, $\mathfrak{gl}(2,{\mathbbm C})$, and $SL(2,{\mathbbm C})$, mainly to fix our notation. By $\mathfrak{gl}_n{\mathbbm{K}}$ we denote the set of all ($n\times n$)-matrices over the field ${\mathbbm{K}}$ and by $GL(n,{\mathbbm{K}})\subset\mathfrak{gl}_n{\mathbbm{K}}$ the group of all regular matrices. The set of Lorentz transformations $O(3,1)$ by definition contains all elements $T\in GL(4,{\mathbbm R})$ that obey $\|T\vec{x}\|^2=\|\vec{x}\|^2$ for $\vec{x}=(x^0,x^1,x^2,x^3)^t\in{\mathbbm R}^4$. Here we use $$\|\vec{x}\|^2=(x^0)^2-(x^1)^2-(x^2)^2-(x^3)^2=\sum\limits_{i,j=0}^3g_{ij}x^ix^j\,,$$ where $$\label{minkmetric} g=(g_{ij})_{i,j=0,\ldots,3}={\rm diag}(1,-1,-1,-1)$$ denotes the Minkowski metric and we write ${\mathbbm R}^{3,1}=({\mathbbm R}^4,g)$. We denote the matrix entries of an endomorphism $T$ by $T^i{}_j$ such that $(T\vec{x})^i=\sum\limits_{j=0}^3T^i{}_jx^j$. The Lorentz transformations form a subgroup of $GL(4,{\mathbbm R})$. As a submanifold of $GL(4,{\mathbbm R})$ the group structure is smooth such that $SO(3,1)$ is indeed a Lie group. This follows also from a more general fact stating that closed subgroups of Lie groups are Lie subgroups, see [@Helga Theorem II.2.3]. $SO(3,1)$ admits four connected components that are associated to orientability and time-orientability of ${\mathbbm R}^{3,1}$. The connected component of the identity is given by the Lorentz transformations that obey ${\rm det}(T)=1$ and $T^0{}_0>0$. The special Lorentz transformations form a subgroup denoted by $SO^+(3,1)$. We use the following Pauli matrices: $$\sigma_0 = \begin{pmatrix}1&0\\0&1 \end{pmatrix},\quad \sigma_1 = \begin{pmatrix}0&1\\1&0 \end{pmatrix},\quad \sigma_2 = \begin{pmatrix}0&-i\\i&0 \end{pmatrix},\quad \sigma_3 = \begin{pmatrix}1&0\\0&-1 \end{pmatrix}\,.$$ The matrices $\sigma_1,\sigma_2$, and $\sigma_3$ obey $$\begin{aligned} &\sigma_1^2=\sigma_2^2=\sigma^2_3=\sigma_0^2=\sigma_0\,, \\ &\sigma_1\sigma_2=-\sigma_2\sigma_1\,,\quad \sigma_1\sigma_3=-\sigma_3\sigma_1\,,\quad \sigma_2\sigma_3=-\sigma_3\sigma_2\,,\\ &\sigma_1\sigma_2=i\sigma_3\,,\quad \sigma_2\sigma_3=i\sigma_1\,,\quad \sigma_3\sigma_1=i\sigma_2\,. \end{aligned}\label{1}$$ This can be combined to $$\label{c} \sigma_i\sigma_j=\delta_{ij}\sigma_0+i\sum_{k=1}^3\epsilon_{ijk}\sigma_k$$ with $\epsilon_{ijk}$ totally skew symmetric and $\epsilon_{123}=1$. In particular, for $i=0,1,2,3$ we have the nice relation $$\label{orth} \delta_{ij}=\frac{1}{2}{\rm tr}(\sigma_i\sigma_j)\,,$$ and, moreover, from (\[c\]) we get for $i=1,2,3$ $$\label{d} \sum_{j=1}^3 \sigma_j\sigma_i\sigma_j =-\sigma_i\ \text{ and }\ \sum_{j=0}^3 \sigma_j\sigma_i\sigma_j =0\,.$$ We consider the natural ${\mathbbm R}$-linear map $$\begin{aligned} \Psi:{\mathbbm R}^4&\to\mathfrak{gl}(2,{\mathbbm C})\,,\\ \vec{x}=\begin{pmatrix}x^0\\x^1\\x^2\\x^3\end{pmatrix}&\mapsto \Psi(\vec{x})=\sum_{i=0}^3x^i\sigma_i = \begin{pmatrix}x^0+x^3 &x^1-ix^2 \\ x^1+ix^2 &x^0-x^3\end{pmatrix}\,. \end{aligned}\label{psi}$$ The image $\Psi(\vec{x})$ of $\vec x\in{\mathbbm R}^4$ is a Hermitian matrix, i.e. $$\Psi(\vec{x})\in \mathfrak{h}(2,{\mathbbm C}):=\{A\in\mathfrak{gl}(2,{\mathbbm C})\ |\ A=A^\dagger\}\,,$$ and, therefore, of type $\begin{pmatrix}a & \bar w\\\ w & b\end{pmatrix}$ with $a,b\in {\mathbbm R}$ and $w\in {\mathbbm C}$. The inverse map is given by $$\Psi^{-1}\left(\begin{pmatrix}a &\bar w\\ w & b\end{pmatrix}\right) =\begin{pmatrix}\frac{1}{2}(a+b) \\ Re(w) \\ Im(w)\\ \frac{1}{2}(a-b)\end{pmatrix}\,.$$ In particular, each Hermitian matrix $B\in \mathfrak{h}(2,{\mathbbm C})$ can be written as $B=x_0\sigma_0+x_1\sigma_1+x_2\sigma_2+x_3\sigma_3$ with $x_i\in{\mathbbm R}$ and we have $$\|\vec{x}\|^2=(x^0)^2-(x^1)^2-(x^2)^2-(x^3)^2=\det(\Psi(\vec{x}))\,.$$ As noticed in the title the special linear group $SL(2,{\mathbbm C})\subset\mathfrak{gl}(2,{\mathbbm C})$ will play an important role in the following and we will recall its definition: $$\label{sl} SL(2,{\mathbbm C})=\big\{ B\in\mathfrak{gl}(2,{\mathbbm C}) \,|\, \det(B)=1\big\}\,.$$ $SL(2,{\mathbbm C})$ is a Lie group, that has complex dimension three or real dimension six: in (\[sl\]) we have one complex equation for the four complex parameters. Each matrix in $SL(2,{\mathbbm C})$ can be written in the form $$\label{A} A=a^0\sigma_0+a^1\sigma_1+a^2\sigma_2+a^3\sigma_3$$ with $$\label{det} \det(A)=(a^0)^2-(a^1)^2-(a^2)^2-(a^3)^2=1\,.$$ Complex generators of $SL(2,{\mathbbm C})$ are, for example, $\sigma_1,\sigma_2$ and $\sigma_3$.[^1] Therefore, real generators are $\sigma_1$, $\sigma_2$, and $\sigma_3$, as well as $i\sigma_1$, $i\sigma_2$, and $i\sigma_3$. If we omit in (\[A\]) the condition on the determinant we get all of $\mathfrak{gl}(2,{\mathbbm C})$ by such linear combination. The product of two matrices $A=\sum\limits_{j=0}^3a^j\sigma_j$ and $B=\sum\limits_{j=0}^3b^j\sigma_j$ expands as $$\label{p} AB= \sum_{j=0}^3a_jb_j\sigma_0 +\sum_{j=1}^3(a_0b_j+b_0a_j)\sigma_j + i\sum_{j,k,\ell=1}^3\epsilon_{jk\ell}\,a_jb_k\,\sigma_\ell\,.$$ Given a matrix $A\in\mathfrak{gl}(2,{\mathbbm C})$ we define the conjugated matrix by $$\label{bar} A' = a^0\sigma_0- a^1\sigma_1- a^2\sigma_2- a^3\sigma_3 \,.$$ This conjugate obeys ${\rm det}(A')={\rm det}(A)$ and $A'B'=(BA)'$. In particular, the product of a matrix and its conjugated is given by $$A' A= A A' = \big((a^0)^2-(a^1)^2-(a^2)^2-(a^3)^2\big)\sigma_0\,,$$ such that its trace obeys $$\label{Atrace} \frac{1}{2}{\rm tr}(A'A)= \det(A)=(a^0)^2-(a^1)^2-(a^2)^2-(a^3)^2\,.$$ Moreover, for $A\in SL(2,{\mathbbm C})$ we have $A'\in SL(2,{\mathbbm C})$ and $A^{-1}= A'$ due to (\[det\]). We collect the symmetry properties (\[1\]) and the symmetry property (\[bar\]) as follows. We introduce signs $\varepsilon_i$ and $\varepsilon_{ij}$ defined by $\sigma_i'=\varepsilon_i\sigma_i$ and $\sigma_i\sigma_j=\varepsilon_{ij}\sigma_j\sigma_i$, i.e. $$\label{signs} \big(\varepsilon_i\big)_{i=0,\ldots,3}=\begin{pmatrix} 1\\-1\\-1\\-1 \end{pmatrix}\,,\quad \big(\varepsilon_{ij}\big)_{i,j=0,\ldots,3}=\begin{pmatrix} 1 & 1 & 1 & 1\\ 1 & 1 & -1 & -1\\ 1 & -1 & 1 & -1\\ 1 & -1 & -1 & 1 \end{pmatrix}\,.$$ In particular, in terms of $\varepsilon_i$ Minkowski metric (\[minkmetric\]) reads as $$g_{ij}=\varepsilon_i\delta_{ij}=\varepsilon_j\delta_{ij}\,.$$ $SO(3,1)$ in terms of $SL(2,{\mathbbm C})$ ========================================== We consider an action $\Phi$ of $SL(2,{\mathbbm C})$ on the set of Hermitian matrices that is defined by $$SL(2,{\mathbbm C})\times \mathfrak{h}(2,{\mathbbm C}) \ni (A,B)\mapsto ABA^\dagger \in \mathfrak{h}(2,{\mathbbm C})\,.$$ Writing $B=\Psi(\vec{x})$ the combination $\Psi^{-1}( A\Psi(\vec{x}) A^\dagger )$ yields an element in ${\mathbbm R}^{4}$. This defines an action $\Phi$ of $SL(2,{\mathbbm C})$ on ${\mathbbm R}^{4}$ via $$\label{X} \Phi(A) (\vec{x})= \Psi^{-1}(A\Psi(\vec{x})A^\dagger)$$ The map $\Phi(A):{\mathbbm R}^{4}\to{\mathbbm R}^{4}$ is ${\mathbbm R}$-linear. Furthermore, we have $$\begin{aligned} \|\Phi(A)(\vec{x})\|^2& = \|\Psi^{-1}( A\Psi(\vec{x}) A^\dagger )\|^2 = \det (A\Psi(\vec{x}) A^\dagger) \\ & =\det(A)\, \overline{\det(A)}\, \det(\Psi(\vec{x})) = \det (\Psi(\vec{x}))=\|\vec{x}\|^2\end{aligned}$$ such that $\Phi(A)$ is a Lorentz transformation. The Matrix entries of $T:=(\Phi(A)^i{}_j)_{i,j=0,\ldots,3}$ depend on the complex parameters $a_i$ from the decomposition of $A$ according to (\[A\]). They can explicitly be expressed by expanding and rearranging the right hand side of (\[X\]): $$\begin{aligned} \Psi\big(\Phi(A)(\vec{x})\big)=\ &\sum_{i,j=0}^4T^i{}_jx^j\sigma_i \nonumber\\ =\ &\quad \, \big( a^0\bar a^0 + a^1\bar a^1 + a^2\bar a^2 + a^3\bar a^3\big) x^0\sigma_0 \nonumber\\ &+\big( a^0\bar a^1 + a^1\bar a^0 - ia^2\bar a^3 + ia^3\bar a^2\big) x^1\sigma_0 \nonumber\\ &+\big( a^0\bar a^2 + a^2\bar a^0 + ia^1\bar a^3 - ia^3\bar a^1\big) x^2\sigma_0 \nonumber\\ &+\big( a^0\bar a^3 + a^3\bar a^0 - ia^1\bar a^2 + ia^2\bar a^1\big) x^3\sigma_0 \nonumber\\ &+\big( a^0\bar a^1 + a^1\bar a^0 + ia^2\bar a^3 - ia^3\bar a^2\big) x^0\sigma_1 \nonumber\\ &+\big( a^0\bar a^0 + a^1\bar a^1 - a^2\bar a^2 - a^3\bar a^3\big) x^1\sigma_1 \nonumber\\ &+\big( a^1\bar a^2 + a^2\bar a^1 + ia^0\bar a^3 - ia^3\bar a^0\big) x^2\sigma_1 \nonumber\\ &+\big( a^1\bar a^3 + a^3\bar a^1 - ia^0\bar a^2 + ia^2\bar a^0\big) x^3\sigma_1 \label{schwer}\\ &+\big( a^0\bar a^2 + a^2\bar a^0 - ia^1\bar a^3 + ia^3\bar a^1\big) x^0\sigma_2 \nonumber\\ &+\big( a^1\bar a^2 + a^2\bar a^1 - ia^0\bar a^3 + ia^3\bar a^0\big) x^1\sigma_2 \nonumber\\ &+\big( a^0\bar a^0 - a^1\bar a^1 + a^2\bar a^2 - a^3\bar a^3\big) x^2\sigma_2 \nonumber\\ &+\big( a^2\bar a^3 + a^3\bar a^2 + ia^0\bar a^1 - ia^1\bar a^0\big) x^3\sigma_2 \nonumber\\ &+\big( a^0\bar a^3 + a^3\bar a^0 + ia^1\bar a^2 - ia^2\bar a^1\big) x^0\sigma_3 \nonumber\\ &+\big( a^1\bar a^3 + a^3\bar a^1 + ia^0\bar a^2 - ia^2\bar a^0\big) x^1\sigma_3 \nonumber\\ &+\big( a^2\bar a^3 + a^3\bar a^2 - ia^0\bar a^1 + ia^1\bar a^0\big) x^2\sigma_3 \nonumber\\ &+\big( a^0\bar a^0 - a^1\bar a^1 - a^2\bar a^2 + a^3\bar a^3\big) x^3\sigma_3 \nonumber\,.\end{aligned}$$ By applying (\[orth\]) directly to (\[X\]) we see that (\[schwer\]) gets the following compact form. \[hat\] Consider $A\in SL(2,{\mathbbm C})$ and $\Phi:SL(2,{\mathbbm C})\to SO(3,1)$. Then the image of $T=\Phi(A)$ has the entries $$\label{leicht} \fbox{$\displaystyle T^i{}_j = \frac{1}{2}{\rm tr}(\sigma_i A\sigma_j A^\dagger)= \frac{1}{2}{\rm tr}( A\sigma_j A^\dagger\sigma_i)$}$$ As we saw, the map $\Phi(A)$ is a Lorentz transformation – but is it a special Lorentz transformation as well? From (\[schwer\]) we see directly that $\Phi(A)^0{}_0>0$ – but what about the determinant of $\Phi(A)$? Without calculating the determinant we can see the result as follows: The image of the map $\Phi: SL(2,{\mathbbm C})\to O(3,1)$ is connected because $\Phi$ is continuous and $SL(2,{\mathbbm C})$ is (simply) connected. Furthermore, the identity is in the image of $\Phi$ such that all of the image of $\Phi$ is contained in $SO^+(3,1)$. $SL(2,{\mathbbm C})$ in terms of $SO(3,1)$ ========================================== By explicitly inverting the system (\[schwer\])we show in this section that for any special Lorentz transformation $T\in SO^+(3,1)$ their exists a matrix $A\in SL(2,{\mathbbm C})$ with $T=\Phi(A)$. This matrix isn’t unique, because with $A$ its negative $-A\in SL(2,{\mathbbm C})$ obeys $\Phi(-A)=\Phi(A)$, too. This 2:1 behavior will be reflected in the existence of a square root during the process of solving the equation $T=\Phi(A)$, i.e. (\[schwer\]), for $a^i$. To solve (\[schwer\]) we write $\Phi(A)=T$ and consider (\[X\]) and (\[leicht\]), i.e. $\Psi(T\vec{x})= A\Psi(\vec{x}) A^\dagger $ and $T^i{}_j=\frac{1}{2}{\rm tr}(\sigma^iA\sigma_jA^\dagger)$. We define matrices $\tau_{(i)}(T)\in \mathfrak{gl}(2,{\mathbbm C})$ by $$\label{t} \tau_{(i)}(T):=\sum_{j,k=0}^3T^j{}_k\sigma_i'\sigma_j\sigma_i\sigma_k\,.$$ For example, $i=0$ yields $$\tau_{(0)}(T) = \sum_{i=0}^3 T^i{}_i\sigma_0+\sum_{i=1}^3(T^i{}_0 +T^0{}_i)\sigma_i+i\sum_{i,j,k=1}^3T^i{}_j\epsilon_{ijk}\sigma_k \,.$$ Moreover, due to (\[X\]), we have $$\begin{aligned} \tau_{(i)}(T) =\ & \sum_{k=0}^3\sigma_i'\big(\sum_{j=0}^3T^j{}_k\sigma_j\big)\sigma_i\sigma_k = \sum_{k=0}^3 \sigma_i'\Psi( T(\vec{e}_k) )\sigma_i\sigma_k \\ =\ & \sum_{k=0}^3 \sigma_i' A\sigma_k A^\dagger\sigma_i\sigma_k\,. \end{aligned}$$ For arbitrary $B=\sum\limits_{i=0}^3b^i\sigma_i$ we get $$\sum_{j=0}^3 \sigma_j B\sigma_j =b^0\sum_{j=0}^3(\sigma_j)^2\sigma_0 +\sum_{i=1}^3b^i\sum_{j=0}^3\sigma_j\sigma_i\sigma_j =4b^0 \sigma_0$$ by using (\[d\]). Doing the same calculations for $B\sigma_1$, $B\sigma_2$, and $B\sigma_3$ we get for $i=0,1,2,3$ $$\sum_{j=0}^3 \sigma_j B\sigma_i\sigma_j = 4b^i \sigma_0$$ such that $$\tau_{(i)}(T) = \sigma_i'A \sum_{j=0}^3\sigma_j A^\dagger\sigma_i\sigma_j =4\bar a^i \sigma_i'A\,.$$ We use (\[bar\]) and consider the following trace $$\begin{aligned} \frac{1}{2}{\rm tr}\big(\tau_{(i)}(T)' \tau_{(i)}(T)\big) & = \frac{1}{2} {\rm tr} \big((4\bar a^i\sigma_i'A)'(4\bar a^i \sigma_iA) \big)\\ & = 8(\bar a^i)^2{\rm tr}(A'\sigma_i\sigma_i A)\\ & = 16(\bar a^i)^2 \,. \end{aligned}$$ This special combination of $\tau_{(i)}(T)$ and $\tau_{(i)}(T)'$ yields the following statement. \[check\] Consider $T\in SO^+(3,1)$. Then there exist maps $$\hat\Phi^\pm_{(i)}: SO^+(3,1)\to SL(2,{\mathbbm C})$$ for $i=0,1,2,3$ such that the images $\hat\Phi^\pm_{(i)}(T)$ are the solutions of $\Phi(A)=T$ if ${\rm tr}\big(\tau_{(i)}(T)'\tau_{(i)}(T)\big)\neq 0$. The maps are given by $$\fbox{$\displaystyle \hat\Phi^\pm_{(i)}(T)= \pm\frac{1}{\sqrt{\frac{1}{2} {\rm tr}\big(\tau_{(i)}(T)'\tau_{(i)}(T)\big)}}\, \sigma_i'\tau_{(i)}(T) $} \label{sol}$$ We will formulate the result (\[sol\]) in terms of $T$ alone, i.e. without help of the map $\tau_{(i)}$. By using the signs (\[signs\]) a more explicit way to express (\[t\]) is $$\begin{aligned} \varepsilon_i\tau_{(i)}(T) & = \sum_{j,k=0}^3 T^j{}_k\sigma_i\sigma_j\sigma_i\sigma_k\\ &= \Big(\sum_{j=0}^3\varepsilon_{ij}T^j{}_j\Big)\sigma_0+\sum_{j=1}^3\big(\varepsilon_{ij}T^j{}_0+T^0{}_j\big)\sigma_j +i\sum_{j,k,\ell=1}^3\varepsilon_{ij}\epsilon_{jk\ell}T^j{}_k\sigma_\ell\,.\end{aligned}$$ We write $T_{(i)}$ for the matrix with entries $(T_{(i)}){}^j{}_k=\epsilon_{ij}T^j{}_k$, in particular $T_{(0)}=T$. For the expansion $\varepsilon_i\tau_{(i)}=t^0\sigma_0+t^1\sigma_1+t^2\sigma_2+t^3\sigma_3$ we have $$\begin{aligned} \frac{1}{2}{\rm tr}\big(\tau_{(i)}(T)'\tau_{(i)}(T)\big) =\ & (t^0)^2-(t^1)^2-(t^2)^2-(t^3)^2 \\ =\ & \big({\rm tr}(T_{(i)})\big)^2-\sum_{j=1}^3\Big( \varepsilon_{ij} T^j{}_0 +T^0{}_j +i\sum_{k,\ell=1}^3\varepsilon_{ik}\epsilon_{k\ell j}T^k{}_\ell\Big)^2\\ =\ & \big({\rm tr}(T_{(i)})\big)^2 -\sum_{j=1}^3\big( T^j{}_0T^j{}_0 + T^0{}_jT^0{}_j + 2\varepsilon_{ij} T^j{}_0T^0{}_j\big) \\ & -\!\!\!\sum_{j,k,\ell,m,n=1}^3\!\!\!\varepsilon_{ik}\varepsilon_{im}\epsilon_{k\ell j} \epsilon_{mnj}T^k{}_\ell T^m{}_n \\ & +2i \sum_{j,k,\ell=1}^3\epsilon_{k\ell j}(\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik} T^k{}_\ell\\ =\ & \big({\rm tr}(T_{(i)})\big)^2 -\sum_{i=1}^3\big( T^j{}_0T^j{}_0 + T^0{}_jT^0{}_j + 2\varepsilon_{ij}T^j{}_0T^0{}_j\big) \\ & + \sum_{j,k=1}^3 T^j{}_kT^j{}_k -\sum_{j,k=1}^3 \varepsilon_{ik}\varepsilon_{ij}T^j{}_kT^k{}_j \\ & -2i\!\! \sum_{j,k,\ell=1}^3\!\!\epsilon_{jk\ell}(\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik}T^k{}_\ell\,,\end{aligned}$$ where we used $\sum_{\ell=1}^3\epsilon_{jk\ell}\epsilon_{mn\ell}=2\delta^{jk}_{mn}$ in the last step. From the fact that $T$ is a Lorentz transformation we have $$\label{covariant} \sum_{i,k=0}^3T^i{}_jg_{ik}T^k{}_\ell=g_{j\ell}\,.$$ After considering $j=\ell$ and multiplying by the sign $\varepsilon_j$ we take the sum over $j$ and obtain $$\label{trace} \sum_{k,j=1}^3T^k{}_jT^k{}_j-\sum_{j=1}^3T^j{}_0T^j{}_0-\sum_{j=1}^3T^0{}_jT^0{}_j=4-(T^0{}_0)^2\,.$$ We use this and insert the positive signs $\varepsilon_{i0}$ to simplify $$\begin{aligned} \frac{1}{2}{\rm tr}(\tau_{(i)}'\tau_{(i)}) =\ & 4 + \big({\rm tr}(T_{(i)})\big)^2 -2i\!\! \sum_{j,k,\ell=1}^3\!\!\epsilon_{jk\ell}(\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik}T^k{}_\ell\\ & -\epsilon_{i0}\epsilon_{i0}T^0{}_0T^0{}_0 - 2\sum_{i=1}^3\varepsilon_{i0}\varepsilon_{ij}T^j{}_0T^0{}_j -\sum_{j,k=1}^3 \varepsilon_{ik}\varepsilon_{ij}T^j{}_kT^k{}_j \\ =\ & 4 + ({\rm tr}\big(T_{(i)})\big)^2 -\sum_{j,k=0}^3 \varepsilon_{ik}\varepsilon_{ij}T^j{}_kT^k{}_j \\ & -2i\!\! \sum_{j,k,\ell=1}^3\!\!\epsilon_{jk\ell}(\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik}T^k{}_\ell\\ =\ & 4 + \big({\rm tr}(T_{(i)})\big)^2 - {\rm tr}(T_{(i)}^2) -2i\!\! \sum_{j,k,\ell=1}^3\!\!\epsilon_{jk\ell}(\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik}T^k{}_\ell\,.\end{aligned}$$ This yields the following Corollary of Proposition \[check\] that contains the announced explicit description of $SL(2,{\mathbbm C})$ in terms of $SO(3,1)$. In terms of the entries of $T$ formula (\[sol\]) reads $$\tag{\ref{sol}${}_0$} \hat\Phi^\pm_{(0)}(T) =\pm \frac{\displaystyle {\rm tr}(T)\sigma_0 +\sum\limits_{j=1}^3\left( T^j{}_0 +T^0{}_j +i\sum\limits_{k,\ell=1}^3T^k{}_\ell\epsilon_{jk\ell}\right)\sigma_j }{ \sqrt{4+ ({\rm tr}(T))^2 - {\rm tr}(T^2) - 2i \sum\limits_{j,k,\ell=1}^3\epsilon_{jk\ell}(T^j{}_0+T^0{}_j)T^k{}_\ell} }\,$$ for $i=0$, as well as $$\tag{\ref{sol}${}_i$} \begin{aligned} \hat\Phi^\pm_{(i)}(T) = & \pm \frac{1}{ \sqrt{4+ \big({\rm tr}(T_{(i)})\big)^2 - {\rm tr}(T_{(i)}^2) - 2i \sum\limits_{j,k,\ell=1}^3\epsilon_{jk\ell} (\varepsilon_{ij}T^j{}_0+T^0{}_j)\varepsilon_{ik}T^k{}_\ell} }\times\\ &\qquad \times\left( \big( T^i{}_0 +T^0{}_i +i\!\!\sum\limits_{j,k=1}^3\!\! \varepsilon_{ij}\epsilon_{ijk}T^j{}_k\big) \sigma_0 +{\rm tr}(T_{(i)})\sigma_i \right.\\ &\qquad \qquad \left. +\sum\limits_{j=1}^3\Big( T^i{}_j-\varepsilon_{ij}T^j{}_i +i\sum\limits_{k=1}^3\epsilon_{ikj}(\varepsilon_{ik} T^k{}_0 +T^0{}_k ) \Big)\sigma_j\right) \end{aligned}$$ for $i=1,2,3$. All four combinations in (\[t\]) are needed to describe full $SL(2,{\mathbbm C})$ because formula (\[sol\]) only works for $\tau_{(i)}(T)\neq0$. For example, the choice $\hat\Phi_0$ only works for matrices $T$ such that $a^0\neq 0$. In particular, the matrices $\sigma_1$, $\sigma_2$, and $\sigma_3$ that correspond to $T={\rm diag}(1,1,-1,-1)$, $T={\rm diag}(1,-1,1,-1)$, and $T={\rm diag}(1,-1,-1,1)$, respectively, cannot be described. We consider $T\in SO^+(3,1)$ such that $$\begin{gathered} T_{(0)}=T={\begin{pmatrix} \cosh\alpha &\sinh\alpha & \\ \sinh\alpha &\cosh\alpha &\\&&\mathbbm{1}_2 \end{pmatrix}}\,,\quad T_{(1)}={\begin{pmatrix} \cosh\alpha &\sinh\alpha & \\ \sinh\alpha &\cosh\alpha &\\&&-\mathbbm{1}_2 \end{pmatrix}}\,,\\ T_{(2)}={\begin{pmatrix} \cosh\alpha &-\sinh\alpha &\\ \sinh\alpha &-\cosh\alpha &\\&&\sigma_3 \end{pmatrix}}\,,\quad T_{(3)}={\begin{pmatrix} \cosh\alpha &-\sinh\alpha &\\ \sinh\alpha &-\cosh\alpha &\\&&-\sigma_3 \end{pmatrix}}\,. \end{gathered}$$ Then $$\begin{gathered} {\rm tr}(T)=2(\cosh(\alpha)+1)\,,\quad {\rm tr}(T_{(1)})=2(\cosh(\alpha)-1)\,, \\ {\rm tr}(T_{(2)})={\rm tr}(T_{(3)})=0\,. \end{gathered}$$ Furthermore we have $$T^2=T_{(1)}^2=\begin{pmatrix} 2\cosh^2\alpha-1 &2\cosh\alpha\sinh\alpha & &\\2\cosh\alpha\sinh\alpha &2\cosh^2-1 & & \\& &\mathbbm{1}_2 \end{pmatrix},\quad T_{(2)}^2= T_{(3)}^2=\mathbbm{1}_4\,,$$ and $$\begin{aligned} {\rm tr}(\tau_{(0)}(T)'\tau_{(0)}(T))&=8(\cosh2\alpha+1)\,,\\ {\rm tr}(\tau_{(1)}(T)'\tau_{(1)}(T))&=8(\cosh2\alpha-1)\,,\\ {\rm tr}(\tau_{(2)}(T)'\tau_{(2)}(T))& ={\rm tr}(\tau_{(3)}(T)'\tau_{(3)}(T))=0\,.\end{aligned}$$ Therefore, we can consider $\hat\Phi^\pm_{(0)}(T)$ and $\hat \Phi^\pm_{(1)}(T)$ and get $$\begin{aligned} \hat\Phi^\pm_{(0)}(T)&=\frac{1}{\sqrt{2}} \left(\sqrt{\cosh\alpha+1}\,\sigma_0+\frac{\sinh\alpha}{\sqrt{\cosh\alpha+1}}\,\sigma_1\right)\,,\\ \hat\Phi_{(1)}^\pm(T)&=\frac{1}{\sqrt{2}} \left(\frac{\sinh\alpha}{\sqrt{\cosh\alpha-1}}\,\sigma_0+\sqrt{\cosh\alpha-1}\,\sigma_1\right).\end{aligned}$$ Both yield the same Matrix $A$ , namely $$A=\begin{pmatrix} \cosh\frac{\alpha}{2}&\sinh\frac{\alpha}{2} \\ \sinh\frac{\alpha}{2}& \cosh\frac{\alpha}{2} \end{pmatrix},$$ which follows from $2\sinh^2\frac{\alpha}{2}=\cosh\alpha-1$ and $2\cosh^2\frac{\alpha}{2}=\cosh\alpha+1$. Some concluding remarks ======================= - The results that we presented here in an elementary way have been discussed in parts in the literature. The particular choice $i=0$ in (\[sol\]) has been discussed in [@Joo p. 69] where the author states a variant of formula (\[sol\]${}_0$), in [@Hest p. 53] where the author emphasizes that the formula only holds in special cases, and in [@Lou p. 130] with reference to [@Hest] but without comment on the incompleteness. - On purpose we neglected the use of the theory of Clifford algebras and their representations although there is a strong relation. In fact, the Clifford algebra ${{\rm C}\ell}(4)$ is the framework in which the results above can be formulated and we will shortly recall how Pauli matrices enter into the discussion. Starting in dimension two we see that the set $\{\sigma_1,\sigma_2\}$ provides generators of the Clifford algebra ${{\rm C}\ell}(2)$ because $\sigma_i\sigma_j+\sigma_j\sigma_i=2\delta_{ij}$ for $i=1,2$. By adding the volume element $\sigma_3=-i\sigma_1\sigma_2$ we get generators $\{\sigma_1,\sigma_2,\sigma_3\}$ of ${{\rm C}\ell}(3)$ because the same relations as before hold but for $1\leq i\leq 3$. As we can check the set $\{\Sigma_0=\sigma_1\otimes\mathbbm{1},\Sigma_1=\sigma_2\otimes\sigma_1,\Sigma_2=\sigma_2\otimes\sigma_2,\Sigma_3=\sigma_2\otimes\sigma_3,\}$ obeys $\Sigma_i\Sigma_j+\Sigma_j\Sigma_i=2\delta_{ij}$ for $0\leq i\leq 3$ such that it yields generators of ${{\rm C}\ell}(4)$. Such doubling process can always be used when going from ${{\rm C}\ell}(2k-2)$ to ${{\rm C}\ell}(2k)$. This gives a iterative way to construct ${{\rm C}\ell}(2k)$ from ${{\rm C}\ell}(2)$, see for example [@BFGK]. The doubling process, of course, is not unique because for any set of generators $\{\Sigma_i\}$ and any unitary transformation $\Omega$ the set $\{\Omega\Sigma_i\Omega^\dagger\}$ yields generators, too. Although we restricted to the complex Clifford algebra above, generators of the real Clifford algebra according to a metric with signature can easily be obtained by adding some extra $i$ in front of some of the generators. For more details we again refer to the literature, for example [@Che; @Har; @MichLaw]. - Our choice for ${{\rm C}\ell}(4)$ above is the so called Weyl representation for which the subspace $\text{span}\big\{\Sigma_{ij}=\frac{1}{2}(\Sigma_i\Sigma_j-\Sigma_j\Sigma_i)\big\}\subset{{\rm C}\ell}(4)$ is block-diagonal. This subspace is isomorphic to the algebra of skew-symmetric ($4\times4$)-matrices and reflects the algebra isomorphism $\mathfrak{so}(4)\simeq\mathfrak{so}(2)\oplus\mathfrak{so}(2)$. In terms of Dynkin diagrams this is $D_2=A_1\oplus A_1$ and here $\mathfrak{sl}(2,{\mathbbm C})$ enters as the standard realization of $A_1$. A more geometric way to interpret the isomorphism is the notion of selfduality of two-forms in dimension four. In this particular dimension the Hodge operator provides an involution on the six-dimensional space of two-forms and, therefore, it splits into two three-dimensional eigenspaces, the so called self-dual and anti-self-dual two-forms. - The introduction of the sign $\epsilon_i$ into (\[covariant\]) to get (\[trace\]) is somewhat artificial. In a more geometric way this is due to the natural isomorphism ${\mathbbm R}^4\simeq({\mathbbm R}^4)^*$ defined by the Minkowski metric $g$. In terms of index-notation this is raising and lowering of indices. This isomorphism is needed when we want to calculate invariant traces of bilinear forms. In fact, (\[covariant\]) is a bilinear form rather than an endomorphism. - There is a last nice relation we like to mention. The isomorphism $\Psi$ from (\[psi\]) translates (\[p\]) to ${\mathbbm R}^4$. After writing ${\mathbbm R}^4={\mathbbm R}\times{\mathbbm R}^3$ this reflects the geometry of ${\mathbbm R}^3$, i.e. the Euclidean product $\langle\cdot,\cdot\rangle$ and the cross product $\times$. For this we write $\vec x=(x_0,\mathbf{x})$ with $\mathbf{x}\in{\mathbbm R}^3$. Then $$\Psi^{-1}\big(\Psi(\vec x)\Psi(\vec y)\big) =\begin{pmatrix} x_0y_0+\langle\mathbf{x},\mathbf{y}\rangle\\ x_0\mathbf{y}+y_0\mathbf{x}+\mathbf{x}\times\mathbf{y} \end{pmatrix},$$ and therefore $$\tfrac{1}{2}\Psi^{-1}\big(\Psi(\vec x)\Psi(\vec y)-\Psi(\vec y)\Psi(\vec x)\big) =\begin{pmatrix} 0\\ \mathbf{x}\times\mathbf{y} \end{pmatrix}.$$ [A]{} Helga Baum, Thomas Friedrich, Ralf Grunewald, Ines Kath: . B. G. Teubner Verlagsgesellschaft mbH, Stuttgart, 1991. Claude Chevalley: . In: Collected Works of Claude Chevalley, Vol. 2. Springer Verlag, 1996. F. Reese Harvey: (Perspectives in Mathematics). Academic Press, 1990) Sigurdur Helgason . American Mathematical Society, 2001. David Hestenes: (Documents on Modern Physics). Gordon and Breach Science Publishers, Inc. 1966. Hans Joos: . (1962), 65-146. Pertti Lounesto: . Cambridge University Press, 2nd Ed. 2001. Marie-Louise Michelson and H. Blaine Lawson: . Princeton University Press, 1989. [^1]: Because of (\[1\]) we have $\exp(b^1\sigma_1+b^2\sigma_2+b^3\sigma_3)=a^0+a^1\sigma_1+a^2\sigma_2+a^3\sigma_3$ with complex coefficients $a^i$ which depend on the $b^j$. This follows from a more general relation between Lie groups and their tangent space at the identity, i.e. their Lie algebra, see [@Helga Proposition II.1.6].

ArXiv

--- abstract: 'In the prequel of the paper (`arXiv:1803.02792`), we considered exact enumerations of the cored versions of a doubly-intruded hexagon. The result generalized Ciucu’s work about $F$-cored hexagons (*Adv. Math. 2017*). In this paper, we provide an extensive list of 30 tiling enumerations of hexagons with three collinear chains of triangular holes with alternating orientations. Besides two chains of holes attaching to the boundary of the hexagon, we remove one more chain of triangles that is slightly off the center of the hexagon. Two of our enumerations imply two conjectures posed by Ciucu, Eisenkölbl, Krattenthaler, and Zare (*J. Combin. Theory Ser. A 2001*) as two very special cases.' address: 'Department of Mathematics, University of Nebraska – Lincoln, Lincoln, NE 68588' author: - Tri Lai title: 'Tiling Enumeration of Hexagons with Off-central Holes' --- [^1] Introduction ============ Motivated by MacMahon’s elegant product formula for lozenge tilings a hexagon [@Mac], James Propp posed a problem asking for a tiling formula for an almost regular hexagon with the central unit triangle removed (see Problem 2 in [@Propp]). The problem has been solved and generalized by a number of authors (see e.g. [@Ciu2; @HG; @OK; @CEKZ; @CK; @Ciu1; @HoleDent]). A milestone on this line of work is when Ciucu, Eisenkölbl, Krattenthaler and Zare [@CEKZ] enumerated the *cored hexagon* (or *punctured hexagon*) $C_{x,y,z}(m)$ that are obtained by removing the central equilateral triangle of side-length $m$ from the hexagon $H$ of side-lengths $x, y+m, z,x+m,y,z+m$. We define this region in details in the next paragraph. ![(a) The cored hexagon $C_{2,6,4}(2)$. (b) The cored hexagon $C_{3,6,4}(2)$. (c) The cored hexagon $C_{2,5,4}(2)$. (d) The cored hexagon $C_{2,6,3}(2)$. []{data-label="fig:core"}](corehexagon.eps){width="12cm"} We start with an *auxiliary hexagon* $H_0$ with side-lengths[^2] $x,y,z,x,y,z$ (see the hexagons with the dashed boundary in Figure \[fig:core\]). Next, we push the north, the northeast, and the southeast sides of the hexagon $m$ units outward, and keep other sides staying put. This way, we get a larger hexagon $H$, called the *base hexagon*, of side-lengths $x, y+m, z,x+m,y,z+m$. Finally, we remove an up-pointing $m$-triangle such that its left vertex is located at the closest lattice point to the center of the auxiliary hexagon $H_0$. Precisely, there are four cases to distinguish based on the parities of $x,y,z$. When $x$, $y$ and $z$ have the same parity, the center of the hexagon is a lattice vertex and our removed triangle has the left vertex at the center. One readily sees that, in this case, the triangular hole stays evenly between each pair of parallel sides of the hexagon $H$. In particular, the distance between the north side of the hexagon and the top of the triangular hole and the distance between the base of the triangular hole and the south side of the hexagon are both $\frac{y+z}{2}$; the distances corresponding to the northeast and southwest sides of the hexagon are both $\frac{x+z}{2}$; the distances corresponding to the northwest and southeast sides of the hexagon are both $\frac{x+y}{2}$ (see Figure \[fig:core\](a); the hexagon with dashed boundary indicates the auxiliary hexagon). Next, we consider the case when $x$ has parity different from that of $y$ and $z$. In this case, the center of the auxiliary hexagon $H_0$ is *not* a lattice vertex anymore. It is the middle point of a horizontal unit lattice interval. We now place the triangular hole such that its leftmost is $1/2$ unit to the left of the center of the auxiliary hexagon (illustrated in Figure \[fig:core\](b); the larger shaded dot indicates the center of the auxiliary hexagon). Similarly, if $y$ has the opposite parity to $x$ and $z$, we place the triangular hole $1/2$ unit to the northwest of the center of the auxiliary hexagon $H_0$ (shown in Figure \[fig:core\](c)). Finally, if $z$ has parity different from that of $x$ and $y$, the hole is located $1/2$ unit to the southwest of the center of $H_0$ (see Figure \[fig:core\](d)). The work of Ciucu, Eisenkölbl, Krattenthaler and Zare has been generalized further by Ciucu and Krattenthaler when the central triangular hole was extended to a cluster of four triangles, called a ‘*shamrock*’ in [@CK], and later to a chain of alternating triangles, called ‘*fern*’ in [@Ciu1] (see e.g. [@CL; @Halfhex1; @Halfhex2; @Halfhex3] for more recent work about the fern structure). Recently, in the prequel of the paper [@HoleDent], the author extended even more the work of Ciucu to a triple of removed ferns, besides the central fern removed, we cut off two more ferns from the boundary of the hexagon. We note that in the above results, the hole must be located in the center of the region. In the general case, if the hole is moved away from the center, the tiling formula seems not to be simple anymore. However, Ciucu, Krattenthaler, Eisenkölbl and Zare [@CEKZ] observed that when the triangular hole in the cored hexagon $C_{x,y,z}(m)$ is moved $1$ unit away, the number of tilings seems to accept only several minor changes. In particular, they conjectured two explicit tiling formulas for the case when $x,y,z$ have the same parity and the triangular hole is 1-unit to the left of the center (see Figure \[fig:originoff\](a)), and the case when $x$ has parity different from that of $y$ and $z$ and the triangular hole is $3/2$-unit to the left of the center (see Figure \[fig:originoff\](b)). Next, we definition of the hyperfactorial function as follows: $$\label{hyper2} {\operatorname{H}}(n)=\begin{cases} \prod_{k=0}^{n-1}\Gamma(k+1) & \text{for $n$ a positive integer;}\\ \prod_{k=0}^{n-\frac{1}{2}}\Gamma(k+\frac{1}{2}) & \text{for $n$ a positive half-integer.} \end{cases}$$ where $\Gamma$ denotes the classical gamma function. Recall that $\Gamma(n+1)=n!$ and $\Gamma(n+\frac{1}{2})=\frac{(2n)!}{4^nn!}\sqrt{\pi}$, for a nonnegative integer $n$. \[con1\] Let $x,y,z,m$ be nonnegative integers, $x,y,z$ having the same parity. The number of the lozenge tilings of a hexagon with side-lengths $x,y+m,z,x+m,y,z+m$, with an equilateral triangle of side-length $m$ removed at $1$ unit to the left of the center, equals $$\begin{aligned} \label{phieq} \Phi_{x,y,z}(m):=&\frac{1}{4} P_1(x,y,z,m)\frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor ){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)}{{\operatorname{H}}(m+\frac{x+y}{2}+1){\operatorname{H}}(m+\frac{y+z}{2}){\operatorname{H}}(m+\frac{z+x}{2}-1)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor) {\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil) {\operatorname{H}}(\frac{x+y}{2}-1){\operatorname{H}}(\frac{y+z}{2}){\operatorname{H}}(\frac{z+x}{2}+1)}, \tag{T-1}\end{aligned}$$ where $P_1(x,y,z,m)$ is the polynomial given by $$P_1(x,y,z,m)= \begin{cases} (x+y)(x+z)+2xm & \text{ if $x$ is even,}\\ (x+y)(x+z)+2(x+y+z+m)m &\text{ if $x$ is odd.} \end{cases}$$ ![(a) Removal of the triangle which is off-center by one unit in the case $x,y,z$ have the same parity. (b) Removal of the triangle which is off-center by $3/2$ units in the case when $x$ has parity opposite to $y$ and $z$.[]{data-label="fig:originoff"}](corehexagonoff.eps){width="12cm"} \[con2\] Let $x,y,z,m$ be nonnegative integers, $x$ is of parity different from the parity of $y$ and $z$. The number of the lozenge tilings of a hexagon of side-lengths $x,y+m,z,x+m,y,z+m$, with an equilateral triangle of side-length $m$ removed at $3/2$ unit to the left of the center, equals $$\begin{aligned} \label{psieq1} \Psi_{x,z,y}(m):=&\frac{1}{16}P_2(x,y,z,m)\frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor) {\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil) {\operatorname{H}}(\left\lfloor\frac{x+y}{2}\right\rfloor-1){\operatorname{H}}(\frac{y+z}{2}){\operatorname{H}}(\left\lceil\frac{z+x}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor ){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)}{{\operatorname{H}}(m+\left\lceil\frac{x+y}{2}\right\rceil+1){\operatorname{H}}(m+\frac{y+z}{2}){\operatorname{H}}(m+\left\lfloor\frac{z+x}{2}\right\rfloor-1)}, \tag{T-2}\end{aligned}$$ where $P_2(x,y,z,m)$ is the polynomial given by $$P_2(x,y,z,m)= \begin{cases} ((x+y)^2-1)((x+z)^2-1)+4xm(x^2+2xy+y^2+2xz+3yz+z^2\\ \quad +2xm+3ym+3zm+2m^2-1) & \text{ if $x$ is even,}\\ ((x+y)^2-1)((x+z)^2-1)+4(x+y+z+m)m(x^2+xy-1) &\text{ if $x$ is odd.} \end{cases}$$ The formulas (\[phieq\]) and (\[psieq1\]) almost agree with the corresponding tiling formulas for cored hexagons. The only remarkable differences are the appearances of the quadratic factors $P_1(x,y,z,m)$ and $P_2(x,y,z,m)$. It is worth noticing that these conjectures were recently proved by Rosengren [@Rosen], using a different method from that in this paper. More precisely, he used lattice path combinatorics and Selberg integral to obtain a complicated determinantal formula for a weighted sum of lozenge tilings of a hexagon with a triangular hole at an arbitrary position. Then, by evaluating the determinant for the special cases when the triangular hole is close to the center, he was able to verify Ciucu–Eisenkölbl–Krattenthaler–Zare’s Conjectures \[con1\] and \[con2\]. Inspired by the above two conjectures and by the previous work in [@HoleDent], we would like to investigate similar situations for the case of hexagons with three ferns removed. Intuitively, we remove a fern from the interior of the hexagon so that its leftmost vertex (called the ‘*root*’ of the fern in this paper) is located at a lattice point next to the center of the hexagon. Then we remove two more ferns at the same level such that the new ferns are touching the boundary of the hexagons. Base on the parity of the side-lengths of our hexagons and the position of the inner fern, there are total *thirty* different regions to enumerate. One may be surprised that the number of situations needed to considered is significantly larger than that in the case of cored hexagons. The reason is that, unlike the cored hexagons, $60^{\circ}$-rotations do *not* preserve the structures of our regions. The rest of the paper is organized as follows. In Section 2, we define carefully all 30 off-central regions and state their exact tiling enumerations. In Section 3, we quote several fundamental results in enumeration of tilings, especially two versions of Kuo’s graphical condensation [@Kuo], that will be the key of our proof. Section 4 is devoted to the detailed proof of our main theorems. Precise statement of the main result {#Statement1} ==================================== Recall that in the prequel paper [@HoleDent], we enumerated $8$ different families of hexagons with three ferns removed (one fern is removed from the center and two more ferns are removed from the boundary). These families are divided into two groups based on the relative position of the lattice line $l$, that contains the three ferns and the west and east vertices of the hexagon. In particular, we have four ‘*$R$-families*’, $R^{\odot}$, $R^{\leftarrow}$, $R^{\nwarrow}$, and $R^{\swarrow}$, in the case when $l$ separates the east and west vertices of the hexagon, and four ‘*$Q$-families*’, $Q^{\odot}$, $Q^{\leftarrow}$, $Q^{\nwarrow}$, and $Q^{\nearrow}$, in the case when $l$ leaves both east and west vertices of the hexagon on a same side (see detailed definition in Section 2 of [@HoleDent]). Each of the above eight families of regions has several off-central counterparts, that will be defined carefully in the next eight subsections. In order to state our next tilling enumerations, we define five more functions, besides the functions $\Phi$ and $\Psi$ in Conjectures \[con1\] and \[con2\], as follows. We define the functions $\Theta_{x,y,z}(m)$ and $\Theta'_{x,y,z}(m)$ by: $$\begin{aligned} \label{thetaeq1} \Theta_{x,y,z}(m):=&Q_{1}(y,z,m) \cdot \frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor ){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)}{{\operatorname{H}}(m+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(m+\frac{y+z}{2}+1){\operatorname{H}}(m+\left\lfloor\frac{z+x}{2}\right\rfloor)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil) {\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)} {{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil){\operatorname{H}}(\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(\frac{y+z}{2}){\operatorname{H}}(\left\lceil\frac{z+x}{2}\right\rceil)} \tag{T-3}\end{aligned}$$ and $$\begin{aligned} \label{thetaeq2} \Theta'_{x,y,z}(m):=&Q_{1}(y,z,m) \cdot \frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor ){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)}{{\operatorname{H}}(m+\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(m+\frac{y+z}{2}-1){\operatorname{H}}(m+\left\lceil\frac{z+x}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil) {\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)} {{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{y+z}{2}+1){\operatorname{H}}(\left\lfloor\frac{z+x}{2}\right\rfloor)}, \tag{T'-3}\end{aligned}$$ where $Q_{1}(y,z,m)$ is $m+\frac{y+z}{2}$ if $x$ is even, and is $\frac{y+z}{2}$ if $x$ is odd. One readily sees that the $\Theta$- and $\Theta'$-functions are almost the same, except for certain differences at several hyperfactorial factors. We next define the function $\Lambda_{x,y,z}(m)$ and its variation $\Lambda'_{x,y,z}(m)$ by: $$\begin{aligned} \label{lambdaeq1} \Lambda_{x,y,z}(m):=&\frac{1}{8}Q_{2}(x,y,z,m) \cdot \frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)} {{\operatorname{H}}(m+\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(m+\frac{y+z}{2}+1){\operatorname{H}}(m+\left\lfloor\frac{z+x}{2}\right\rfloor-1)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil) {\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)} {{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{y+z}{2}-1){\operatorname{H}}(\left\lceil\frac{z+x}{2}\right\rceil+1)} \tag{T-4}\end{aligned}$$ and $$\begin{aligned} \label{lambdaeq2} \Lambda' _{x,y,z}(m):=&\frac{1}{8}Q_{2}(x,y,z,m) \cdot \frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)} {{\operatorname{H}}(m+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(m+\frac{y+z}{2}-1){\operatorname{H}}(m+\left\lceil\frac{z+x}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil) {\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)} {{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil){\operatorname{H}}(\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(\frac{y+z}{2}+1){\operatorname{H}}(\left\lfloor\frac{z+x}{2}\right\rfloor-1)}, \tag{T'-4}\end{aligned}$$ in which the polynomial $Q_{2}(x,y,z,m)$ is defined as $$Q_{2}(x,y,z,m)= \begin{cases} (x+z+2m-1)((z+1)(x+z-1)+2x(m-1))\\\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ +(y+1)((x+z)^2+4xm-1) &\text{ if $x$ is even,}\\ z((x+z+2m)^2-1)+y((x+z+2m)^2-4m(x+m)-1) &\text{if $x$ is odd.} \end{cases}$$ Finally, we define a counterpart of the function $\Psi$ in Conjecture \[con2\]: $$\begin{aligned} \label{psieq2} \Psi'_{x,y,z}(m):=&\frac{1}{16}\frac{{\operatorname{H}}(m+x){\operatorname{H}}(m+y){\operatorname{H}}(m+z){\operatorname{H}}(m+x+y+z)}{{\operatorname{H}}(m+x+y){\operatorname{H}}(m+y+z){\operatorname{H}}(m+z+x)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2})^2{\operatorname{H}}(\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\left\lceil\frac{z}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y}{2}\right\rceil){\operatorname{H}}(\frac{m}{2}+\frac{y+z}{2})^2{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{z+x}{2}\right\rfloor) {\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{z+x}{2}\right\rceil)}{{\operatorname{H}}(\frac{m}{2}+\left\lfloor\frac{x+y+z}{2}\right\rfloor){\operatorname{H}}(\frac{m}{2}+\left\lceil\frac{x+y+z}{2}\right\rceil) {\operatorname{H}}(\left\lceil\frac{x+y}{2}\right\rceil+1){\operatorname{H}}(\frac{y+z}{2}){\operatorname{H}}(\left\lfloor\frac{z+x}{2}\right\rfloor -1)}\notag\\ &\times \frac{{\operatorname{H}}(m+\left\lfloor\frac{x+y+z}{2}\right\rfloor ){\operatorname{H}}(m+\left\lceil\frac{x+y+z}{2}\right\rceil)}{{\operatorname{H}}(m+\left\lfloor\frac{x+y}{2}\right\rfloor-1){\operatorname{H}}(m+\frac{y+z}{2}){\operatorname{H}}(m+\left\lceil\frac{z+x}{2}\right\rceil +1)} P_2(x,y,z,m), \tag{T'-2}\end{aligned}$$ where $P_2(x,y,z,m)$ is the polynomial defined in the Conjecture \[con2\]. We also need the following well-known formula of Cohn, Larsen, and Propp for tiling number of dented semihexagon in the statements of our main theorems. For a sequence $\textbf{a}:=(a_i)_{i=1}^{m}$, we denote $o_a:=\sum_{\text{$i$ odd}} a_i$ and $e_a:=\sum_{\text{$i$ even}}a_i$. Let $S(a_1,a_2,\dotsc,a_m)$ denote the upper half of a hexagon of side-lengths $e_a,o_a,o_a,e_a,o_a,o_a$ in which $k:=\lfloor\frac{m}{2}\rfloor$ triangles of side-lengths $a_1,a_3,a_5,\dots,a_{2k+1}$ removed from the base, such that the distance between the $i$-th and the $(i+1)$-th removed triangles is $a_{2i}$ (see Figure \[fig:semihex\] for an example). We call the region $S(a_1,a_2,\dotsc,a_m)$ a *dented semihexagon*. Cohn, Larsen and Propp [@CLP] interpreted semi-strict Gelfand–Tsetlin patterns as lozenge tilings of the dented semihexagon $S(a_1,a_2,\dotsc,a_m)$, and obtained the following tiling formula $$\begin{aligned} \label{semieq} s(a_1,a_2,\dots,a_{2l-1})&=s(a_1,a_2,\dots,a_{2l})\\ &=\dfrac{1}{{\operatorname{H}}(a_1+a_{3}+a_{5}+\dotsc+a_{2l-1})}\dfrac{\prod_{\substack{1\leq i<j\leq 2l-1\\ \text{$j-i$ odd}}}{\operatorname{H}}(a_i+a_{i+1}+\dotsc+a_{j})}{\prod_{\substack{1\leq i<j\leq 2l-1\\ \text{$j-i$ even}}}{\operatorname{H}}(a_i+a_{i+1}+\dotsc+a_{j})},\end{aligned}$$ where $s(a_1,a_2,\dotsc,a_m)$ denotes the number of tilings of $S(a_1,a_2,\dotsc,a_m)$. \#1\#2\#3\#4\#5[ @font ]{} The $E^{(i)}$-type regions -------------------------- We will define our region similar to the cored hexagons. In particular, we start from an auxiliary hexagon $H_0$, and push our its sides in a certain way to obtain a larger hexagon, called the base hexagon. Then we will remove three collinear ferns from the base hexagon in a particular way to obtain our regions. Let us investigate first the case when the auxiliary hexagon $H_0$ of side-lengths $x,z,z,x,z,z$, for $x$ and $z$ of the same parity, and the middle fern is removed at $1$ unit from the center of the auxiliary hexagon. In this case, the center of $H_0$ is a lattice point, and there are six different lattice points, which have distance $1$ from the center (shown in Figure \[fig:offposition\](a); the darker shaded nodes indicates the center of the auxiliary hexagon $H_0$, and the lighter shaded nodes represent the possible off-central positions for the leftmost of the middle fern). We note that, in the case of cored hexagons, we only need to consider the regions corresponding to one of these six positions, since regions corresponding to other positions can be obtained from the latter region by $60^{\circ}$ rotations. However, this is *not* true anymore for our hexagons with three ferns removed. ![(a) Six off-central positions for $E^{(i)}$- and $\overline{E}^{(i)}$-type regions. (b) Eight off-central positions for $F^{(i)}$- and $\overline{F}^{(i)}$-type regions. (c) Eight off-central positions for $G^{(i)}$- and $\overline{G}^{(i)}$-type regions. (d) Eight off-central positions for $K^{(i)}$- an $\overline{K}^{(i)}$-type regions.[]{data-label="fig:offposition"}](offposition.eps){width="10cm"} Assume that $x,z$ are nonnegative integers with the same parity, that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}:=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three (may be empty) sequences of nonnegative integers, and that $y$ is an integer that may take negative values in certain condition. The three sequences $\textbf{a}, \textbf{b}, \textbf{c}$ determine the side-lengths of triangles in the left, the right, and the central ferns, respectively. Set $$\begin{aligned} e_a:=\sum_{i\ even}a_i, &\quad o_a:=\sum_{i \ odd} a_i,\notag\\ e_b:=\sum_{j\ even}b_j, &\quad o_b:=\sum_{j \ odd} b_j,\notag\\ e_c:=\sum_{t\ even}c_t, &\quad o_c:=\sum_{t \ odd} c_t,\end{aligned}$$ and $a:=a_1+a_2+\dotsc$, $b:=b_1+b_2+\dotsc$, $c:=c_1+c_2+\dotsc$. We first define the regions corresponding to the off-central positions $1$ and $4$ in Figure \[fig:offposition\](a). In this case, we assume further that $y\geq 0$ (the domain of the $y$-parameter changes in the other off-central cases). Let us start with the auxiliary hexagon $H_0$ of side-lengths $x,z,z,x,z,z$. We perform the following side-pushing process. We push outward all six sides of the hexagon (in the clockwise order, starting from the north side) a distance of $e_a+o_b+o_c, y+\max(b-a,0),\ b+c,\ b+c+y+\max(a-b,0),\ o_a+e_b+e_c+y+\max(a-b,0),\ a,\ a+y+\max(b-a,0)$ units, respectively. We get a larger hexagon $H$, called the *base hexagon*, has side-lengths $x+o_a+e_b+e_c,$ $2y+z+e_a+o_b+o_c+ |a-b|$, $z+o_a+e_b+e_c,$ $x+e_a+o_b+o_c$, $2y+z+o_a+e_b+e_c+ |a-b|,$ $z+e_a+o_b+o_c$. This is illustrated by Figure \[pushing1\] (for $j=0$); the auxiliary hexagon is the shaded one with the dashed boundary, and the base hexagon is the one restricted by the bold contour. \#1\#2\#3\#4\#5[ @font ]{} For $i=1,4$, we denote $E^{(i)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ the region obtained from the base hexagon $H$ by removing three ferns as follows. The middle fern consists of equilateral triangles of side-lengths $c_1,c_2,\dots,c_k$ as they appear from left to right. The first triangle (of side-length $c_1$) is up-pointing and the next triangles in the fern are oriented in alternating orientations. We remove the middle fern, such that its root (i.e. its leftmost point) is at the off-central position $i$ as shown in Figure \[fig:offposition\](a), for $i=1,4$. The left fern consists of triangles of side-lengths $a_1,a_2,\dots,a_m$ running from left to right and starts by a down-pointing triangle; while the right fern consists of triangles of side-lengths $b_1,b_2,\dots, b_n$ running from *right to left* and starts by an *up-pointing* triangles. These latter two ferns are removed such that they are at the same level as the middle fern, that the left fern is touching the southwest side of the hexagon, and that the right fern is touching the northeast side of the hexagon. See Figure \[fig:off1\](a) for an example. For $i=2,3,5,6$, unlike the case $i=1,4$ above, we start with an auxiliary hexagon of side-lengths $x,z+2,z,x,z+2,z$, we still perform the same side-pushing procedure to get the base hexagon of side-lengths $x+o_a+e_b+e_c,$ $2y+z+e_a+o_b+o_c+ |a-b|+2$, $z+o_a+e_b+e_c,$ $x+e_a+o_b+o_c$, $2y+z+o_a+e_b+e_c+ |a-b|+2,$ $z+e_a+o_b+o_c$. If $i=2,3$, we allow $y$ to take any integer values greater or equal to $\max(a-b,-2)$. In particular, $y\geq 0$ if $a\leq b$, however, $y$ may be negative when $a>b$. Our side-pushing process still works well in the same way as the definition of the $E^{(1)}$-type region, as even though $y$ may be negative, all the pushing distances are still nonnegative. Next, the three ferns are removed similarly, such that the root of the middle fern is at the off-central position $i$ in Figure \[fig:offposition\](a), for $i=2,3$. If $i=5,6$, we assume in addition that $y\geq \max (b-a,-2)$, and the $E^{(5)}$- and $E^{(6)}$-type regions are defined similarly. By, the symmetry it is enough to enumerate only three of the above regions, namely the $E^{(1)}$-, $E^{(2)}$- and $E^{(6)}$-type regions (see Figure \[fig:off1\] for examples[^3]), as the remaining regions can be obtained from these three regions by $180^{\circ}$-rotations. \[rmkE\] In the definition of the $E^{(1)}$-type regions, if we remove the three ferns such that the root of the middle one is at the center of the auxiliary hexagon $H_0$, then we obtain the region $R^{\odot}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.2 of [@HoleDent]. In some sense, the $E^{(i)}$-type regions can be viewed as counterparts of the $R^{\odot}$-type regions investigated in [@HoleDent]. We note that the positions of the two side ferns are determined uniquely by the level of the the middle fern. Thus, from now on, when define the three removed ferns, we only need to care about the position of the middle fern. \#1\#2\#3\#4\#5[ @font ]{} Before stating our enumerations, we note that one can always assume that each of our ferns has even number of triangles. Indeed, if a fern has an odd number of triangles, we can regard that the fern contains an even number of triangles by adding a triangle of side-length $0$ to its end. For the sake of simplicity, we assume this in the statements of our theorems throughout this paper. \[off1thm1\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of a even number of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three nonnegative integers, such that $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eq1} {\operatorname{M}}&(E^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{x,2y+z+2\max(a,b),z}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}-1+c_1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}-1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)}{{\operatorname{H}}(\max(a,b)+y)^2},\end{aligned}$$ where $\Phi_{x,y,z}(m)$ is defined as in (\[phieq\]) of Conjecture \[con1\]. \[off1thm2\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x,z\geq 0$, $y\geq\max(a-b,-2)$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eq2} {\operatorname{M}}&(E^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{2y+z+2\max(a,b)+2,x,z}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2},c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}+c_1,\dotsc,c_{k},\frac{x+z}{2},b_n,\dotsc,b_1,y+a-\min(a,b)+2\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2})}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2})}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+2){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)}.\end{aligned}$$ \[off1thm3\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers (for $m,n,k$ are even) and that $x,y,z$ are three integers, such that $x,z\geq 0$, $y\geq\max(b-a,-2)$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eq3} {\operatorname{M}}&(E^{(6)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{z,2y+z+2\max(a,b)+2,x}(c)\notag\\ &\times s\left(y+b-\min(a,b)+2,a_1,\dotsc, a_{m},\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}-1+c_1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)}.\end{aligned}$$ One can see that the factor involving function $\Phi$ in formula (\[off1eq2\]) (resp., formula (\[off1eq3\])) is obtained from that in the formula (\[off1eq1\]) by interchanging the $x$- and $y$-variables (resp., $x$- and $z$-variables). Next, we show that Theorem \[off1thm1\] implies Ciucu–Eisenkölbl–Krattenthaler–Zare’s Conjecture \[con1\]. Apply formula (\[off1eq1\]) in Theorem \[off1thm1\] to the region\ $E^{(1)}_{x,\frac{y-z}{2},z}(\emptyset;\ m;\ \emptyset)$, we prove the conjecture for the case when $y\geq z$ (all factors in (\[off1eq1\]), except for the first one, cancel out). The remaining case, when $y< z$, is obtained by applying Theorem \[off1thm1\] to a horizontal reflection of the region $E^{(1)}_{x,\frac{z-y}{2},y}(\emptyset;\ 0,m;\ \emptyset)$. One can rewrite the equation (\[off1eq1\]) in Theorem \[off1thm1\] as $$\begin{aligned} {\operatorname{M}}&(E^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))={\operatorname{M}}(C^{(1)}_{x,2y+z+2\max(a,b),z}(c))\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}-1+c_1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\frac{{\operatorname{PP}}(y+\max(a,b)-o_a+o_b+o_c,y+\max(a,b)+o_a-o_b+e_c,z)}{{\operatorname{PP}}(\frac{x+z}{2}-1,c,y+\max(a,b)){\operatorname{PP}}(y+\max(a,b)+c,y+\max(a,b),z)},\end{aligned}$$ where $C^{(1)}_{x,y,z}(m)$ is the (off-central) cored hexagon in Conjecture \[con1\] and $${\operatorname{PP}}(a,b,c):=\frac{{\operatorname{H}}(a){\operatorname{H}}(b){\operatorname{H}}(c){\operatorname{H}}(a+b+c)}{{\operatorname{H}}(a+b){\operatorname{H}}(b+c){\operatorname{H}}(c+a)}$$ is the tiling number of the centrally symmetric hexagon of side-lengths $a,b,c,a,b,c$ (it is also the number of plane partitions fitting in an $a \times b \times c$ box [@Mac]). We have similar interpretation for other theorems in this section. This would be interesting to have a combinatorial proof for this. The $F^{(i)}$-type regions -------------------------- \#1\#2\#3\#4\#5[ @font ]{} We now consider a similar situation to the case of $E^{(i)}$-type regions, however, $x$ and $z$ are now having opposite parities. Let $x,z$ be two nonnegative integers, $y$ an integer that may be negative, and $\textbf{a}, \textbf{b}, \textbf{c}$ three sequences of nonnegative integers that record the side-lengths of triangles in our ferns as usual. In this case, the center of our auxiliary hexagon $H_0$ is the middle point of a horizontal unit segment. There are total 8 off-central positions to remove the middle fern corresponding to the positions $1,2,3,\dots,8$ in Figure \[fig:offposition\](b). For $i=1,2,4,5,6,8$, we start with the auxiliary $H_0$ of side-lengths $x,z+2,z,x,z+2,z$, while $i=3,7$, we start with the auxiliary of side-lengths $x,z,z,x,z,z$. We assume in addition that $y\geq 0$ if $i=3,7$, $y\geq \max(a-b,-2)$ if $i=4,5,6$, and $y\geq \max(b-a,-2)$ if $i=2,1,8$. Next, we apply the same side-pushing procedure as in the definition of the $E^{(i)}$-type regions to obtain the base hexagon $H$. Finally, we remove similarly three collinear ferns from $H$ at the same level, such that the root of the middle fern is at the off-central position $i$ as in Figure \[fig:offposition\](b) (and the two side ferns are defined uniquely such that left fern is touching the southwest side and the right fern is touching the northeast side of the base hexagon). Denote by $F^{(i)}_{x,y,z}(\textbf{a}; \textbf{c}; \textbf{b})$ the resulting region, $i=1,2,\dotsc,8$. However, by symmetry, we only need to enumerate four of them: the ones corresponding to the first, the second, the third, and the fourth off-central positions (see Figure \[fig:offF\] for examples of these regions). \[rmkF\] In the definition of the $F^{(3)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the left of the center of the auxiliary hexagon $H_0$, then we obtain the region $R^{\leftarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.3 of [@HoleDent]. In some sense, the $F^{(i)}$-type regions can be viewed as counterparts of the $R^{\leftarrow}$-type regions. \[off32thm1\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-2)$, $z\geq 0$, and $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq1} {\operatorname{M}}&(F^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta_{x,2y+z+2\max(a,b)+2,z}(c)\notag\\ &\times s\left(y+b-\min(a,b)+2,a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}\max(a,b)-o_a+o_b+o_c+y+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)},\end{aligned}$$ where $\Theta_{x,y,z}(m)$ is given by (\[thetaeq1\]). Unlike the tiling formulas for the $E^{(i)}$-type regions above, which has the non-linear term $P_1(x,y,z,m)$, this formula is a ‘*simple*’ product formula, in the sense the largest factor is linear in the parameters of the region. Moreover, the corresponding cored hexagons of these regions have *not* appeared in the previous work of Ciucu–Eisenkölbl–Krattenthaler–Zare [@CEKZ]. Theorem \[off32thm1\] gives an exact enumeration for a new type of cored hexagons with the triangular hole down-pointing. \[cor1\] Let $x,y,z,m$ be nonnegative integers. The number of tilings of the cored hexagon, in which the down-pointing triangular hole of side $m$ has its leftmost at position $1$ as in Figure \[fig:offposition\](b), equals $\Theta'_{x,y,z}(m)$. The case when $y\geq z$ is obtained directly from Theorem \[off32thm1\] for the region $F^{(1)}_{x,\frac{y-z}{2},z}(\emptyset;\ 0,m;\ \emptyset)$. The case when $y<z$ is obtained by applying Theorem \[off32thm1\] to a vertical reflection of the region $F^{(1)}_{x,\frac{z-y}{2},z}(\emptyset;\ 0, m;\ \emptyset)$. \[off32thm2\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-2)$, $z\geq 0$, and $x$ and $z$ have opposite parities. We have $$\begin{aligned} \label{off32eq2} {\operatorname{M}}&(F^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda_{x,2y+z+2\max(a,b)+2,z}(c)\notag\\ &\times s\left(y+b-\min(a,b)+2,a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor-1,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor-1+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)} \frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+2){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)}.\end{aligned}$$ where $\Lambda_{x,y,z}(m)$ is defined as in (\[lambdaeq1\]). \[off32thm3\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three nonnegative integers, such that $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq3} {\operatorname{M}}&(F^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi_{x,2y+z+2\max(a,b),z}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor-1,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor-1+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)} \frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)}{{\operatorname{H}}(\max(a,b)+y)^2},\end{aligned}$$ where $\Psi_{x,y,z}(m)$ is given by the formula (\[psieq1\]) in Conjecture \[con2\]. Next, we show that Theorem \[off32thm3\] has Conjecture \[con2\] as a special case. Apply Theorem \[off32thm3\] to the region $F^{(1)}_{x,\frac{y-z}{2},z}(\emptyset;\ m \; \emptyset)$, we verify the conjecture for $y\geq z$, as all factors in (\[off32eq1\]), except for the first one, cancel out. The case $y<z$ is obtained by applying the theorem to a horizontal reflection of the region $F^{(1)}_{x,\frac{z-y}{2},z}(\emptyset;\ 0, m \; \emptyset)$. We have an exact enumeration of a new type of cored hexagons whose triangular hole is down-pointing: \[cor3\] Let $x,y,z,m$ be nonnegative integers. The number of tilings of the cored hexagon, in which the down-pointing triangular hole has its leftmost vertex at the third position as in Figure \[fig:offposition\](b), equals $\Psi'_{x,y,z}(m)$ in (\[psieq2\]). We now have a tiling enumeration for the $F^{(4)}$-type region: \[off32thm4\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq 0$, and $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq4} {\operatorname{M}}&(F^{(4)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{x,z,2y+z+2\max(a,b)+2}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil,b_n,\dotsc,b_1,y+a-\min(a,b)+2\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor)} \frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+2){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)},\end{aligned}$$ where $\Lambda' _{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \[cor2\] Let $x,y,z,m$ be nonnegative integers. The number of tilings of the cored hexagon, in which the down-pointing triangular hole has its leftmost vertex at the second position as in Figure \[fig:offposition\](b), equals $\Lambda' _{x,y,z}(m)$ in (\[lambdaeq2\]). The case when $y\geq z$ is obtained directly from Theorem \[off32thm2\] for the region $F^{(2)}_{x,\frac{y-z}{2},z}(\emptyset;\ 0,m;\ \emptyset)$. The case when $y<z$ is obtained by applying Theorem \[off32thm4\] to a horizontal reflection of the region $F^{(1)}_{x,\frac{z-y}{2},z}(\emptyset;\ m;\ \emptyset)$. The $G^{(i)}$-type regions -------------------------- We now consider the case when the center of the auxiliary hexagon is the middle point of a southeast-to-northwest unit lattice interval. There are also $8$ off-central positions labeled by $1,2,\dots, 8$ around the center of the auxiliary hexagon $H_0$ as shown in Figure \[fig:offposition\](c). Let $x,z$ be two nonnegative integers with the same parity. In particular, for $i=1,2,5,6$, we consider the auxiliary hexagon $H_0$ of side-lengths $x,z+1,z,x,z+1,z$, while for $i=3,4,7,8$, the auxiliary hexagon has side-lengths $x,z+3,z,x,z+3,z$. The domain of the $y$-parameter can be defined in general as follows. If the position $i$ is $d$ units above the center of the auxiliary hexagon ($d=1/2$ or $3/2$ here), then $y\geq \max(a-b,-2d)$; if $i$ is $d$ units below the center, then $y\geq \max (b-a,-2d)$. Next, we apply the same side-pushing procedure as in the definition of the $E^{(i)}$-type region to the new auxiliary hexagon to get the base hexagon $H$. We then remove the three ferns at the same level similarly, such that the middle fern has the root at the off-central position $i$ ($i=1,2,\dots,8$). Let us denote by $G^{(i)}_{x,y,z}(\textbf{a}; \textbf{c}; \textbf{b})$ the newly defined regions (see Figure \[fig:offG\] for examples). Again, by the symmetry, we only need to enumerate four of the eights regions above, namely $G^{(1)}$-, $G^{(2)}$-, $G^{(3)}$-, and $G^{(4)}$-type regions. \[rmkG\] In the definition of the $G^{(1)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the northwest of the center of the auxiliary hexagon $H_0$, then we obtain the region $R^{\nwarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.4 of [@HoleDent]. It means that the $G^{(i)}$-type regions can be viewed as counterparts of the $R^{\nwarrow}$-type regions. \#1\#2\#3\#4\#5[ @font ]{} \[off32thm5\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-1)$, $z\geq 0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off32eq5} {\operatorname{M}}&(G^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta'_{2y+z+2\max(a,b)+1,z,x}(c)\notag\\ &\times s\left(y+b-\min(a,b)+1,a_1,\dotsc, a_{m},\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}-1+c_1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)} {{\operatorname{H}}(\max(a,b)+y+1){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where $\Theta'_{x,y,z}(m)$ is defined as in (\[thetaeq2\]). \[off32thm6\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-1)$, $z\geq 0$, and $x$ and $z$ have the same parity. We have $$\begin{aligned} \label{off32eq6} {\operatorname{M}}&(G^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{2y+z+2\max(a,b)+1,z,x}(c))\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}-1+c_1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)+1\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}-1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+1){\operatorname{H}}(\max(a,b)+c+y+z)} {{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+1)} {{\operatorname{H}}(\max(a,b)+y+1){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \[off32thm7\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq 0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off32eq7} {\operatorname{M}}&(G^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi'_{2y+z+2\max(a,b)+3,z,x}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2},c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}+c_1,\dotsc,c_{k},\frac{x+z}{2},b_n,\dotsc,b_1,y+a-\min(a,b)+3\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2})}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2})}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+3){\operatorname{H}}(\max(a,b)+c+y+z)} {{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+3)} {{\operatorname{H}}(\max(a,b)+y+3){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where $\Psi'_{x,y,z}(m)$ is defined as in (\[psieq2\]). \[off32thm8\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq 0$, and $x$ and $z$ have the same parity. We have $$\begin{aligned} \label{off32eq8} {\operatorname{M}}&(G^{(4)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda_{z,2y+z+2\max(a,b)+3,x}(c)\notag\\ &\times s\left(y+b-\min(a,b),a_1,\dotsc, a_{m},\frac{x+z}{2}+1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n-1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\frac{x+z}{2}+c_1+1,\dotsc,c_{k},\frac{x+z}{2}-1,b_n,\dotsc,b_1,y+a-\min(a,b)+3\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}+1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}+1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+3){\operatorname{H}}(\max(a,b)+c+y+z)} {{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+3)} {{\operatorname{H}}(\max(a,b)+y+3){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where $\Lambda_{x,y,z}(m)$ is defined as in (\[lambdaeq1\]). Off-central counterparts of the $R^{\swarrow}$-type regions ----------------------------------------------------------- \#1\#2\#3\#4\#5[ @font ]{} We consider the situation when the center of the auxiliary hexagon $H_0$ is the middle point of a southwest-to-northeast unit lattice interval. There are also $8$ off-central positions labeled by $1,2,\dots, 8$ around the center of the auxiliary hexagon $H_0$ as illustrated in Figure \[fig:offposition\](d). Let $x$ and $z$ be two nonnegative integers with opposite parities. For $i=1,4,5,8$, we consider the auxiliary hexagon $H_0$ of side-lengths $x,z+1,z,x,z+1,z$, while for $i=2,3,6,7$, the auxiliary hexagon has side-lengths $x,z+3,z,x,z+3,z$. The domain of the $y$-parameter is determined by the level $d$ of the position $i$ above or below the center of $H_0$ as in the definition of the $G^{(i)}$-type region ($d=1/2$ or $3/2$). Next, we apply the same side-pushing procedure as in the definition of the $E^{(i)}$-type regions and remove the three ferns at the same level, such that the middle fern has its root at the off-central position $i$ ($i=1,2,\dots,8$). Let us denote by $K^{(i)}_{x,y,z}(\textbf{a}; \textbf{c}; \textbf{b})$ the newly defined regions (see Figure \[fig:offK\] for examples). Again, by the symmetry, we only need to enumerate four of the eights regions, namely $K^{(1)}$-, $K^{(2)}$-, $K^{(3)}$-, and $K^{(4)}$-type regions. \[rmkK\] In the definition of the $K^{(1)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the southwest of the center of the auxiliary hexagon $H_0$, then we obtain the region $R^{\swarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.5 of [@HoleDent]. Therefore, we can view the $K^{(i)}$-type regions as counterparts of the $R^{\swarrow}$-type regions. \[off32thm9\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three nonnegative integers, such that $x\geq 0$, $y\geq \max(b-a,-1)$, $z\geq 0$, and $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq9} {\operatorname{M}}&(K^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta'_{z,x,2y+z+2\max(a,b)+1}(c)\notag\\ &\times s\left(y+b-\min(a,b)+1,a_1,\dotsc, a_{m},\left\lceil\frac{x+z}{2}\right\rceil,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lceil\frac{x+z}{2}\right\rceil+c_1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lceil\frac{x+z}{2}\right\rceil)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lceil\frac{x+z}{2}\right\rceil)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)} {{\operatorname{H}}(\max(a,b)+y+1){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where the region $\Theta'_{x,y,z}(m)$ is defined as in (\[thetaeq2\]). \[off32thm10\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-3)$, $z\geq 0$, and $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq10} {\operatorname{M}}&(K^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{z,x,2y+z+2\max(a,b)+3}(c)\notag\\ &\times s\left(y+b-\min(a,b)+3,a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+2)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+2)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+3)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+3) {\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+3){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)} {{\operatorname{H}}(\max(a,b)+y+3){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where the region $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \[off32thm11\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-3)$, $z\geq 0$, and $x$ has parity opposite to $x$. We get $$\begin{aligned} \label{off32eq11} {\operatorname{M}}&(K^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi'_{z,x,2y+z+2\max(a,b)+3}(c)\notag\\ &\times s\left(y+b-\min(a,b)+3,a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor-1,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+1+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor-1+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+3)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+3) {\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+3){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)} {{\operatorname{H}}(\max(a,b)+y+3){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where the region $\Psi'_{x,y,z}(m)$ is defined as in (\[psieq2\]). \[off32thm12\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(b-a,-1)$, $z\geq 0$, and $x$ has parity opposite to $z$. Then $$\begin{aligned} \label{off32eq12} {\operatorname{M}}&(K^{(4)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda_{2y+z+2\max(a,b)+1,x,z}(c)\notag\\ &\times s\left(y+b-\min(a,b)+1,a_1,\dotsc, a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor-1,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(a_1,\dotsc, a_{m-1},a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor+c_1-1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+z+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y+z)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)-o_a+o_b+o_c+y+1){\operatorname{H}}(\max(a,b)+o_a-o_b+e_c+y)} {{\operatorname{H}}(\max(a,b)+y+1){\operatorname{H}}(\max(a,b)+y)},\end{aligned}$$ where the region $\Lambda_{x,y,z}(m)$ is defined as in (\[lambdaeq1\]). The $\overline{E}^{(i)}$-type regions ------------------------------------- In the previous four families of off-central regions, the horizontal line $l$, that contains the three removed ferns, separates the east and the west vertices of base hexagon $H$. We next consider the other situation when the line $l$ leaves the west and the east vertices of the base hexagon on the same side. Without loss of generality, we assume that these vertices are both above the line $l$. Similar to the cases of the previous four off-central families we define our next region as follows. We (1) start with a certain auxiliary hexagon $H_0$ of certain side-lengths, (2) perform the an edge-pushing procedure on the sides of $H_0$ to get the base hexagon $H$, and (3) remove three ferns from the resulting base hexagon at the same level, such that the leftmost point of the middle fern is at one of the off-central positions as shown in Figure \[fig:offposition\] and that the left and the right ferns touch the *northwest* and the *northeast* sides of the base hexagon. The only major difference is that, in the step (2), we will perform a different pushing procedure. In particular we now push respectively the north, northeast, southeast, south, southwest, northwest sides of $H_0$ a distance of $o_a+o_b+o_c,b+c, b+c+y+\max(0,a-b), e_a+e_b+e_c+2y+|a-b|,a+y+\max(0,b-a),a$ units. This procedure is illustrated in Figure \[pushing2\]. \#1\#2\#3\#4\#5[ @font ]{} Let $x,z$ be two nonnegative integers of the same parity, and let $y$ be an integer that may be negative as indicated particularly in the statements of the theorems below. For $i=1$ or $4$, we perform the above pushing procedure on the auxiliary hexagon $H_0$ of side-lengths $x,z,z,x,z,z$. This way we get the base hexagon $H$ of side-lengths $x+e_a+e_b+e_c, y+z+o_a+o_b+o_c+\max(a-b,0),y+z+e_a+e_b+e_c+\max(b-a,0),x+o_a+o_b+o_c,y+z+e_a+e_b+e_c+\max(a-b,0),y+z+o_a+o_b+o_c+\max(b-a,0)$. Denote by $\overline{E}^{(i)}$, the region corresponding to the choice of the middle fern whose leftmost is at the off-central position $i$ in Figure \[fig:offposition\](a), for $i=1,4$. We do the same in the case $i=2,3$, the only difference is that we start with an auxiliary hexagon of side-lengths $x,z+2,z,x,z+2,z$ (as opposed to the one of side-lengths $x,z,z,x,z,z$). See Figure \[fig:offE2\] for examples. We note that in this family (and in the next three families) we will only allow the root of the middle fern to be above or on the same level as the center of the auxiliary hexagon $H_0$. In particularly, our region here is not defined for the case $i=5,6$. Again, by symmetries, it is enough for us to enumerate only three of the four new regions, namely $\overline{E}^{(1)}$-, $\overline{E}^{(2)}$-,$\overline{E}^{(3)}$-type regions, as any $\overline{E}^{(4)}$-type region is simply a reflection of some $\overline{E}^{(1)}$-type one over a vertical line. \[rmkQE\] In the definition of the $\overline{E}^{(1)}$-type regions, if we remove the three ferns such that the root of the middle one is at the center of the auxiliary hexagon $H_0$, then we obtain the region $Q^{\odot}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.5 of [@HoleDent]. In some sense, the $\overline{E}^{(i)}$-type regions can be viewed as counterparts of the $Q^{\odot}$-type regions. \#1\#2\#3\#4\#5[ @font ]{} \[off1thmQ1\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three nonnegative integers, such that $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQ1} {\operatorname{M}}&(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{x,2y+z+2\max(a,b),z}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1-1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}-1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y)}{{\operatorname{H}}(\max(a,b)+y)^2},\end{aligned}$$ where $\Phi_{x,y,z}(m)$ denotes the formula in (\[phieq\]) in Conjecture \[con1\]. \[off1thmQ2\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQ2} {\operatorname{M}}&(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{2y+z+2\max(a,b)+2,x,z}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2},c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1,\dotsc,c_{k},\frac{x+z}{2},b_n,\dotsc,b_1,y+a-\min(a,b)+2\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2})}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2})}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+2){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)}.\end{aligned}$$ \[off1thmQ3\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQ3} {\operatorname{M}}&(\overline{E}^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Phi_{z,x,2y+z+2\max(a,b)+2}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2}+1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n-1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1+1,\dotsc,c_{k},\frac{x+z}{2}-1,b_n,\dotsc,b_1,y+a-\min(a,b)+2\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)}.\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} The $\overline{F}^{(i)}$-type regions ------------------------------------- We now consider the case when $x$ and $z$ have opposite parities. For $i=4,5,6$, we start with an auxiliary hexagon $H_0$ of side-lengths $x,z+2,z,x,z+2,z$, and for $i=3,7$, we start with the auxiliary hexagon of side-lengths $x,z,z,x,z,z$. We still push all six sides of the auxiliary hexagon as in the definition of the $\overline{E}^{(i)}$-type regions to get the base hexagon $H$, and remove the three ferns from the resulting hexagon $H$, such that the root of the middle fern is now at the off-central position $i$ as shown in Figure \[fig:offposition\](b) (the left and right ferns are also determined by the position of the middle one). Similar to the case of the $\overline{E}^{(i)}$-type regions, there are no counterparts for the case $i=1,2,8$. Denote by $\overline{F}^{(i)}_{x,y,z}(\textbf{a}; \ \textbf{c};\ \textbf{b})$ the resulting regions (illustrated in Figure \[fig:offF2\]). By the symmetry, we only need to enumerate four of the five regions (as any $\overline{F}^{(7)}$-type region is a vertical refection of an $\overline{F}^{(3)}$-type one). \[rmkQF\] In the definition of the $\overline{F}^{(3)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the left of the center of the auxiliary hexagon $H_0$, then we obtain the region $Q^{\leftarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.6 of [@HoleDent]. \[off1thmQF3\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three nonnegative integers, such that $x$ and $z$ have the different parities. Then $$\begin{aligned} \label{off1eqQF3} {\operatorname{M}}&(\overline{F}^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi_{z,2y+z+2\max(a,b),x}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor-1,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor+c_1-1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil+1,b_n,\dotsc,b_1,y+a-\min(a,b)\right)\notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y)}{{\operatorname{H}}(\max(a,b)+y)^2},\end{aligned}$$ where $\Psi_{x,y,z}(m)$ denotes the formula in (\[psieq1\]). \[off1thmQF4\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQF4} {\operatorname{M}}&(\overline{F}^{(4)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{x,z,2y+z+2\max(a,b)+2}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil,b_n,\dotsc,b_1,y+a-\min(a,b)+2 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lfloor\frac{x+z}{2}\right\rfloor)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+2){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)},\end{aligned}$$ where $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \[off1thmQF5\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQF5} {\operatorname{M}}&(\overline{F}^{(5)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta'_{x,2y+z+2\max(a,b)+2,z}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lceil\frac{x+z}{2}\right\rceil,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lceil\frac{x+z}{2}\right\rceil+c_1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor,b_n,\dotsc,b_1,y+a-\min(a,b)+2 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2})}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2})}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(a-b|+e_a+e_b+e_c+2y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)},\end{aligned}$$ where $\Theta'_{x,y,z}(m)$ is defined as in (\[thetaeq2\]). \[off1thmQF6\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-2)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQF6} {\operatorname{M}}&(\overline{F}^{(6)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{x,2y+z+2\max(a,b)+2,z}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lceil\frac{x+z}{2}\right\rceil+1,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor+b_n-1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lceil\frac{x+z}{2}\right\rceil+c_1+1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor-1,b_n,\dotsc,b_1,y+a-\min(a,b)+2 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+2)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+2)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+2)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+2)},\end{aligned}$$ where $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \#1\#2\#3\#4\#5[ @font ]{} The $\overline{G}^{(i)}$-type regions ------------------------------------- Our regions in this section is defined by applying the pushing procedure in the definition of the $\overline{E}^{(i)}$-type region to the auxiliary hexagon as in the case of the $G^{(i)}$-type region. In particular, we are assuming that $x$ and $z$ are two nonnegative integers having the same parity, and that $y$ is an integer with domain defined particularly in the theorems below. For $i=2,5$, we start with an auxiliary hexagon of side-lengths $x,z+1,z,x,z+1,z$, for $i=3,4$, we have the auxiliary hexagon of side-lengths $x,z+3,z,x,z+3,z$. Next, we perform the 2-stage pushing to all six sides of the auxiliary hexagons above in the same way as in the definition of the $\overline{E}^{(i)}$-type regions to obtain the base hexagon $H$. Finally, the three ferns are removed from the base hexagon $H$, such that the leftmost of the middle fern is at the off-central positions $i$ as indicated in Figure \[fig:offposition\](c). There are no counterparts for $i=1,6,7,8$. Denote by $\overline{G}^{(i)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ the resulting regions, for $i=2,3,4,5$. \[rmkQG\] In the definition of the $\overline{G}^{(2)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the northwest of the center of the auxiliary hexagon $H_0$, then we obtain the region $Q^{\nwarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.7 of [@HoleDent]. This way, the $\overline{G}^{(i)}$-type regions can be viewed as counterparts of the $Q^{\nwarrow}$-type regions. \[off1thmQG2\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-1)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQG2} {\operatorname{M}}&(\overline{G}^{(2)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{2y+z+2\max(a,b)+1,z,x}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2}-1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n+1,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1-1,\dotsc,c_{k},\frac{x+z}{2}+1,b_n,\dotsc,b_1,y+a-\min(a,b)+1 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}-1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}-1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+1){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+1)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+1)},\end{aligned}$$ where $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). \[off1thmQG3\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQG3} {\operatorname{M}}&(\overline{G}^{(3)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi'_{2y+z+2\max(a,b)+3,z,x}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2},c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1,\dotsc,c_{k},\frac{x+z}{2},b_n,\dotsc,b_1,y+a-\min(a,b)+3 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2})}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2})}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+3){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+3)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+3)},\end{aligned}$$ where $\Psi'_{x,y,z}(m)$ is the formula in (\[psieq2\]). \[off1thmQG4\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQG4} {\operatorname{M}}&(\overline{G}^{(4)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda_{z,2y+z+2\max(a,b)+3,x}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2}+1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1}-1,\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1+1,\dotsc,c_{k},\frac{x+z}{2}-1,b_n,\dotsc,b_1,y+a-\min(a,b)+3 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}+1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}+1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2}+1)}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2}+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z+3){\operatorname{H}}(\max(a,b)+c+y+z)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+3)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+3)},\end{aligned}$$ where $\Lambda_{x,y,z}(m)$ is defined as in (\[lambdaeq1\]). \[off1thmQG5\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-1)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQG5} {\operatorname{M}}&(\overline{G}^{(5)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta_{2y+z+2\max(a,b)+1,z,x}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\frac{x+z}{2}+1,c_1,\dotsc,c_{k}+\frac{x+z}{2}+b_n,b_{n-1}-1,\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\frac{x+z}{2}+c_1+1,\dotsc,c_{k},\frac{x+z}{2}-1,b_n,\dotsc,b_1,y+a-\min(a,b)+1 \right) \notag\\ &\times\frac{{\operatorname{H}}(c+\frac{x+z}{2}-1)}{{\operatorname{H}}(c){\operatorname{H}}(\frac{x+z}{2}-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\frac{x+z}{2})}{{\operatorname{H}}(\max(a,b)+c+y+\frac{x+z}{2})}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+1)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+1)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+1)},\end{aligned}$$ where $\Theta_{x,y,z}(m)$ is defined as in (\[thetaeq1\]). \#1\#2\#3\#4\#5[ @font ]{} The $\overline{K}^{(i)}$-type regions ------------------------------------- We now define the final group of the off-central regions considered in this paper. We also start with an auxiliary hexagon $H_0$ whose side-lengths are $x,z+1,z,x,z+1,z$ for $i=5,8$, and are $x,z+3,z,x,z+3,z$, for $i=6,7$ ($x$ and $z$ have different parities in this case). After pushing the sides of the auxiliary hexagon $H_0$ in the same way as in the definition of the $\overline{E}^{(i)}$-type regions, we remove three ferns such that the root of the middle fern is at the off-central positions $i$ as shown in Figure \[fig:offposition\](d). There are no counterparts for the case $i=1,2,3,4$ here. Let $\overline{K}^{(i)}$ denote the newly defined region, $i=5,6,7,8$. \[rmkQK\] In the definition of the $\overline{K}^{(5)}$-type regions, if we remove the three ferns such that the root of the middle one is $1/2$-unit to the southwest of the center of the auxiliary hexagon $H_0$, then we obtain the region $Q^{\swarrow}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b})$ in Theorem 2.8 of [@HoleDent]. One can vew the $\overline{K}^{(i)}$-type regions as counterparts of the $Q^{\swarrow}$-type regions. \[off1thmQK5\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-1)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQK5} {\operatorname{M}}&(\overline{K}^{(5)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Theta_{z,x,2y+z+2\max(a,b)+1}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lfloor\frac{x+z}{2}\right\rfloor,c_1,\dotsc,c_{k}+\left\lceil\frac{x+z}{2}\right\rceil+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lfloor\frac{x+z}{2}\right\rfloor+c_1,\dotsc,c_{k},\left\lceil\frac{x+z}{2}\right\rceil,b_n,\dotsc,b_1,y+a-\min(a,b)+1\right) \notag\\ &\times\frac{{\operatorname{H}}(c+\left\lceil\frac{x+z}{2}\right\rceil)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lceil\frac{x+z}{2}\right\rceil)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+1}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+1)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+1)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+1)},\end{aligned}$$ where $\Theta_{x,y,z}(m)$ is defined as in (\[thetaeq1\]). \[off1thmQK6\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQK6} {\operatorname{M}}&(\overline{K}^{(6)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda_{z,x,2y+z+2\max(a,b)+3}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lceil\frac{x+z}{2}\right\rceil,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor,+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lceil\frac{x+z}{2}\right\rceil+c_1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor,b_n,\dotsc,b_1,y+a-\min(a,b)+3\right) \notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+2)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+2)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+3)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+3)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+3)},\end{aligned}$$ where $\Lambda_{x,y,z}(m)$ is defined as in (\[lambdaeq1\]). \[off1thmQK7\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even) and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-3)$, $z\geq0$, and $x$ and $z$ have different parities. Then $$\begin{aligned} \label{off1eqQK7} {\operatorname{M}}&(\overline{K}^{(7)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Psi_{z,x,2y+z+2\max(a,b)+3}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lceil\frac{x+z}{2}\right\rceil+1,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor-1,+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lceil\frac{x+z}{2}\right\rceil+c_1+1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor-1,b_n,\dotsc,b_1,y+a-\min(a,b)+3\right) \notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil+1)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+3)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+3)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+3)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+3)},\end{aligned}$$ where $\Psi_{x,y,z}(m)$ is defined as in (\[psieq1\]). \[off1thmQK8\] Assume that $\textbf{a}=(a_1,a_2,\dotsc,a_m)$, $\textbf{b}=(b_1,b_2,\dotsc,b_n)$, $\textbf{c}=(c_1,c_2,\dotsc,c_k)$ are three sequences of nonnegative integers ($m,n,k$ are all even)and that $x,y,z$ are three integers, such that $x\geq 0$, $y\geq \max(a-b,-1)$, $z\geq0$, and $x$ and $z$ have the same parity. Then $$\begin{aligned} \label{off1eqQK8} {\operatorname{M}}&(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\textbf{c};\textbf{b}))=\Lambda'_{2y+z+2\max(a,b)+1,x,z}(c)\notag\\ &\times s\left(a_1,\dotsc, a_{m}+\left\lceil\frac{x+z}{2}\right\rceil+1,c_1,\dotsc,c_{k}+\left\lfloor\frac{x+z}{2}\right\rfloor-1,+b_n,b_{n-1},\dotsc,b_1\right)\notag\\ &\times s\left(y+b-\min(a,b), a_1,\dotsc,a_{m},\left\lceil\frac{x+z}{2}\right\rceil+c_1+1,\dotsc,c_{k},\left\lfloor\frac{x+z}{2}\right\rfloor-1,b_n,\dotsc,b_1,y+a-\min(a,b)+1\right) \notag\\ &\times\frac{{\operatorname{H}}(c+\left\lfloor\frac{x+z}{2}\right\rfloor-1)}{{\operatorname{H}}(c){\operatorname{H}}(\left\lfloor\frac{x+z}{2}\right\rfloor-1)}\frac{{\operatorname{H}}(\max(a,b)+y+\left\lceil\frac{x+z}{2}\right\rceil)}{{\operatorname{H}}(\max(a,b)+c+y+\left\lceil\frac{x+z}{2}\right\rceil)}\notag\\ &\times \frac{{\operatorname{H}}(\max(a,b)+y+z){\operatorname{H}}(\max(a,b)+c+y+z+1)}{{\operatorname{H}}(o_a+o_b+o_c+z)){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+z+1)}\notag\\ &\times \frac{{\operatorname{H}}(o_a+o_b+o_c){\operatorname{H}}(|a-b|+e_a+e_b+e_c+2y+1)}{{\operatorname{H}}(\max(a,b)+y){\operatorname{H}}(\max(a,b)+y+1)},\end{aligned}$$ where $\Lambda'_{x,y,z}(m)$ is defined as in (\[lambdaeq2\]). Proofs of the main results for off-central regions {#sec:proofoff} ================================================== Organization of the proof {#subsec:organize} ------------------------- Recall that our 8 families of regions, $E^{(i)}$, $F^{(i)}$, $G^{(i)}$, $K^{(i)}$, $\overline{E}^{(i)}$, $\overline{F}^{(i)}$, $\overline{G}^{(i)}$ and $\overline{K}^{(i)}$, are all obtained from a certain base hexagon $H$ by removing three ferns along a common lattice line $\ell$. We call the perimeter of the base hexagon the *quasi-perimeter* of the regions, denoted by $p$ in the rest of the proof. One readily sees that \[claimp\] We always have for all regions in the eight families $$p \geq 2x+4z.$$ If we start with an auxiliary hexagon $H_0$ of side-lengths $x,z+j,z,x,z+j,z$ (for $j=0,1,2,3$) in the definition of our region, then the quasi perimeter is always $$p=2x+4y+4z+3a+3b+2|a-b|+2j.$$ In this case, we always have $y\geq\max(-|a-b|,-j)$. It means that $4y+2|a-b|+2j\geq 0$, so $p\geq 2x+4z+3a+3b\geq2x+4z$. We aim to prove *all* 30 tiling formulas in Theorems \[off1thm1\]–\[off1thmQK8\] at once by induction on $h:=p+x+z$, where $p$ is the quasi-perimeter of the region. Our proof is organized as follows. In Section 3.2, we quote the particular versions the Kuo condensation that will be employed in our proofs. Next, in Sections 3.3–3.10, we will present carefully 66 recurrences for our 8 families of regions obtained by applying Kuo condensation. Each family of regions will have more than one recurrences, depending on whether $a> b$, $a=b$, or $a< b$. We would like to emphasize that, due to the difference in the structures of our regions, the universal recurrence seems *not* to exist. In Section 3.11, we investigate the two extremal cases of our proof. Section 3.12 is devoted to the main arguments of the inductive proof. Kuo condensation and other preliminary results {#sec:kuo} ---------------------------------------------- A *forced lozenge* in a region $R$ on the triangular lattice is a lozenge contained in any tilings of $R$. Removal of forced lozenges does not change the tiling number of a region. A region on the triangular lattice is said to be *balanced* if it has the same number of up- and down-pointing unit triangles (this is a necessary condition that a region admits a tiling). The following useful lemma allows us to decompose a large region into several smaller ones. \[RS\] Let $R$ be a balanced region on the triangular lattice. Assume that a blanched sub-region $Q$ of $R$ satisfies the condition that there is only one type of unit triangles running along each side of the border between $Q$ and $R-Q$. Then ${\operatorname{M}}(R)={\operatorname{M}}(Q)\, {\operatorname{M}}(R-Q).$ Let $G$ be a finite graph without loops, however the multiple-edges are allowed. A *perfect matching* of $G$ is a collection of disjoint edges covering all vertices of $G$. The *(planar) dual graph* of a region $R$ on the triangular lattice is the graph whose vertices are unit triangles in $R$ and whose edges connects precisely two unit triangles sharing an edge. We can identify the tilings of a region and perfect matchings of its dual graph. In this point of view, we use the notation ${\operatorname{M}}(G)$ for the number of perfect matchings of the graph $G$. The following two theorems of Kuo are the key of our proofs in this paper. \[kuothm1\] Let $G=(V_1,V_2,E)$ be a bipartite planar graph in which $|V_1|=|V_2|$. Assume that $u, v, w, s$ are four vertices appearing in a cyclic order on a face of $G$ so that $u,w \in V_1$ and $v,s \in V_2$. Then $$\label{kuoeq1} {\operatorname{M}}(G){\operatorname{M}}(G-\{u, v, w, s\})={\operatorname{M}}(G-\{u, v\}){\operatorname{M}}(G-\{ w, s\})+{\operatorname{M}}(G-\{u, s\}){\operatorname{M}}(G-\{v, w\}).$$ \[kuothm2\] Let $G=(V_1,V_2,E)$ be a bipartite planar graph in which $|V_1|=|V_2|$. Assume that $u, v, w, s$ are four vertices appearing in a cyclic order on a face of $G$ so that $u,v \in V_1$ and $w,s \in V_2$. Then $$\label{kuoeq2} {\operatorname{M}}(G-\{u, s\}){\operatorname{M}}(G-\{v, w\})={\operatorname{M}}(G){\operatorname{M}}(G-\{u, v, w, s\})+{\operatorname{M}}(G-\{u,w\}){\operatorname{M}}(G-\{v, s\}).$$ Theorems \[kuothm1\] and \[kuothm2\] are usually mentioned two variants of *Kuo condensation*. Kuo condensation (or *graphical condensation* as called in [@Kuo]) can be considered as a combinatorial interpretation of the well-known *Dodgson condensation* in linear algebra (which is based on the Jacobi–Desnanot identity, see e.g. [@Abeles], [@Dod] and [@Mui], pp. 136–148, and [@Zeil] for a bijective proof). The Dodgson condensation was named after Charles Lutwidge Dodgson (1832–1898), better known by his pen name Lewis Carroll, an English writer, mathematician, and photographer. The preliminary version of Kuo condensation (when the for vertices $u,v,w,s$ in Theorem \[kuothm1\] form a $4$-cycle in the graph $G$) was originally conjectured by Alexandru Ionescu in context of Aztec diamond graphs, and was proved by Propp in 1993 (see e.g. [@Propp2]). Eric H. Kuo introduced Kuo condensation in his 2004 paper [@Kuo] with four different versions, two of them are Theorems \[kuothm1\] and \[kuothm2\] stated above. Kuo condensation has become a powerful tool in the enumeration of tilings with a number of applications. We refer the reader to [@Ciucu; @Ful; @Knuth; @Kuo06; @speyer; @YYZ; @YZ] for various aspects and generalizations of Kuo condensation, and e.g. [@CF; @CK; @CL; @KW; @LMNT; @Lai15a; @Tri1; @Tri2; @Halfhex1; @Halfhex2; @Halfhex3; @Minor; @LM; @LR; @Ranjan1; @Ranjan2] for recent applications of the method. Recurrences for $E^{(i)}$-type regions {#subsec:offrecurE} -------------------------------------- We apply Kuo condensation in Theorem \[kuothm1\] to the dual graph $G$ of the region $E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ with the four vertices $u,v,w,s$ chosen as in Figure \[fig:off1\]. In particular, the unit triangle corresponding to $u$ is the shaded one on the northeast corner of the region, the $v$-triangle is appended to the left of the right fern, the $w$-triangle is at the east corner of the region, and the $s$-triangle is at the southwest corner. Let us consider the region corresponding to the graph $G-\{u,v,w,s\}$ (see Figure \[fig:kuooff1\](b)). The removal of the $u$-, $w$-, $s$-triangles yields several forced lozenges on the boundary of the region. The removal of the $v$-triangle yields forced lozenges on the side of the last down-pointing triangle of the $b$-fern. After removing these forced lozenges, we get a $R^{\leftarrow}$-type region considered in Theorem 2.3 of [@HoleDent]. We note that Figure \[fig:kuooff1\](c) illustrates the case when the $b$-fern has an even number of triangles, i.e. it ends with a down-pointing triangle. In the case $b$-fern has an odd number of triangles, we can regard that the $b$-fern ends with a down-pointing triangle of side-length $0$. In particular, we obtain the region $R^{\leftarrow}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})$, where $\textbf{b}^{+1}$ denotes the sequence obtained from $\textbf{b}$ by adding 1 to the last term if $\textbf{b}$ has an even number of terms, in the case $\textbf{b}$ has an odd number of terms, we include a new term $1$ to the end of the sequence. We note that, besides changing the side-length of the bases hexagon and the $b$-fern, the removal of forced lozenges may change the center of the region. As a consequence, it may change the type of our region, in particular, from $E^{(1)}$ to $R^{\leftarrow}$ in this case. Since the removal of the forced lozenges does not change the tiling number, we get $$\label{kuothm1eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(R^{\leftarrow}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})).$$ By considering forced lozenges yields by the removal of $u$-, $v$-, $w$-, $s$-triangles as in Figures \[fig:kuooff1\](d), (e) and (f), we have $$\label{kuothm1eq2} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(G^{(1)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b})),$$ $$\label{kuothm1eq3} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(E^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ and $$\label{kuothm1eq4} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(R^{\leftarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).$$ Finally, we remove forced lozenges from the region corresponding to $G-\{u,v\}$, $180^{\circ}$-rotate the resulting region, and obtain $$\label{kuothm1eq5} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(K^{(1)}_{x,y-1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})),$$ where $\overline{\textbf{c}}$ is the sequence obtained from $\textbf{c}$ by reverting the order of terms its if $\textbf{c}$ has an even number of terms, otherwise, we revert $\textbf{c}$ and include a new 0 term in the beginning (see Figure \[fig:kuooff1\](c)). Plugging equations (\[kuothm1eq1\])–(\[kuothm1eq5\]) into the equation in Kuo’s Theorem \[kuothm1\], we have the following $E^{(1)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurE1a} {\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(R^{\leftarrow}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(1)}_{x,y-1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})) {\operatorname{M}}(G^{(1)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(E^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(R^{\leftarrow}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ Working similarly for the case $a>b$, we get a slightly different $E^{(1)}$-recurrence: $$\begin{aligned} \label{offcenterrecurE1b} {\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(R^{\leftarrow}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(1)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})) {\operatorname{M}}(G^{(1)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(E^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(R^{\leftarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} Apply Kuo condensation to the dual graph $G$ of the region $E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ with the same choice of the four vertices $u,v,w,s$ as above, we get the $E^{(2)}$-recurrence for $a\leq b$ as $$\begin{aligned} \label{offcenterrecurE2a} {\operatorname{M}}(E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(F^{(1)}_{x,y-1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(K^{(2)}_{x,y-1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})) {\operatorname{M}}(R^{\nwarrow}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(E^{(2)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(F^{(1)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})),\end{aligned}$$ and the $E^{(2)}$-recurrence for $a> b$: $$\begin{aligned} \label{offcenterrecurE2b} {\operatorname{M}}(E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(F^{(1)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(K^{(2)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})) {\operatorname{M}}(R^{\nwarrow}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(E^{(2)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(F^{(1)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ We note that the removal of forced lozenges from the region corresponding with $G-\{w,s\}$ yield a $R^{\nwarrow}$-type region in Theorem 2.4 in [@HoleDent]. Finally, we have the recurrence for $E^{(6)}$-type region by the same application of Kuo condensation. For $a\leq b$, we have $$\begin{aligned} \label{offcenterrecurE6a} {\operatorname{M}}(E^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(F^{(1)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(R^{\swarrow}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(G^{(4)}_{x,y-1,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(E^{(6)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(F^{(1)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})),\end{aligned}$$ and for $a> b$ $$\begin{aligned} \label{offcenterrecurE6b} {\operatorname{M}}(E^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(F^{(1)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(R^{\swarrow}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(G^{(4)}_{x,y-1,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(E^{(6)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(F^{(1)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})).\end{aligned}$$ The region corresponding with $G-\{u,v\}$ is congruent with, after removing forced lozenges, a region of type $R^{\swarrow}$ in Theorem 2.5 of [@HoleDent]. Recurrences for $F^{(i)}$-type regions {#subsec:offrecurF} -------------------------------------- Applying Kuo Theorem \[kuothm1\] to the dual graph $G$ of region $F^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$, for $a>b$, with the same choice of the four vertices $u,v,w,s$ as in the case for $E^{(i)}$-type regions in the previous subsection (see Figure \[fig:offF\]). By considering forced lozenges yielded by the removal of the $u$-, $v$-, $w$-,$s$-triangles as in Figures \[fig:kuooff2\](b)–(f), we get respectively $$\label{kuothm2eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(E^{(2)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a}))$$ $$\label{kuothm2eq2} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(R^{\nwarrow}_{x,y+1,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a}))$$ $$\label{kuothm2eq3} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(K^{(2)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm2eq4} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(F^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm2eq5} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(E^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a})).$$ We note that the regions on the right-hand sides of (\[kuothm2eq1\]), (\[kuothm2eq2\]) and (\[kuothm2eq5\]) obtained by $180^{\circ}$-rotating the leftover regions (i.e. the remaining ones after removing forced lozenges; these regions are illustrated as the ones with bold contour). Plugging the above 5 equations into the equation in Kuo’s Theorem \[kuothm1\], we get the $F^{(1)}$-recurrence for $a>b$ $$\begin{aligned} \label{offcenterrecurF1b} {\operatorname{M}}(F^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(R^{\nwarrow}_{x,y+1,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(2)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a})).\end{aligned}$$ Working similarly for the case $a\leq b$, we have the ‘sibling’ recurrence for $a\leq b$ $$\begin{aligned} \label{offcenterrecurF1a} {\operatorname{M}}(F^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x,y-1,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(R^{\nwarrow}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(2)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(2)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{ \textbf{c}}; \ \textbf{a})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} Similar application of Kuo condensation we have the following recurrences for other $F^{(i)}$-type regions. The $F^{(2)}$-recurrence for $a\leq b$ is $$\begin{aligned} \label{offcenterrecurF2a} {\operatorname{M}}(F^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(1)}_{x,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(3)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(2)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(6)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})),\end{aligned}$$ and its $a>b$ counterpart is $$\begin{aligned} \label{offcenterrecurF2b} {\operatorname{M}}(F^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(1)}_{x,y+1,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(3)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(2)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(6)}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ Next, we have the $F^{(3)}$-recurrences: $$\begin{aligned} \label{offcenterrecurF3a} {\operatorname{M}}(F^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(2)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(4)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(1)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1}))\end{aligned}$$ if $a\leq b$, and $$\begin{aligned} \label{offcenterrecurF3b} {\operatorname{M}}(F^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(2)}_{x,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(4)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(1)}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1}))\end{aligned}$$ if $a>b$. The last recurrence in this subsection is the $F^{(4)}$-recurrences below. For $a\leq b$, we have $$\begin{aligned} \label{offcenterrecurF4a} {\operatorname{M}}(F^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(3)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(1)}_{x,y,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(F^{(4)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(2)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})),\end{aligned}$$ and for $a>b$ $$\begin{aligned} \label{offcenterrecurF4b} {\operatorname{M}}(F^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(G^{(3)}_{x,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(K^{(1)}_{x,y,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(F^{(4)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(E^{(2)}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ Recurrences for $G^{(i)}$-type regions {#subsec:offrecurG} -------------------------------------- For the $G^{(i)}$-type regions, we still apply Kuo’s Theorem \[kuothm1\], however the four vertices $u,v,w,s$ are chosen differently as in Figure \[fig:offG\]. In particular, the $u$- and $s$-triangles are the shaded one appended to the ends of the right and left ferns, respectively. The $v$ and $w$-triangles are located at the east and southwest corners of the region, respectively. Let us consider the dual graph $G$ of the region $G^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ in the case $a>b$. Working on forced lozenges formed from the removal of the $u$-, $v$-, $w$-,$s$-triangles as in Figures \[fig:kuooff3\](b)–(f), we have respectively $$\label{kuothm3eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(E^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))$$ $$\label{kuothm3eq2} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(K^{(1)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))$$ $$\label{kuothm3eq3} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(F^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm3eq4} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(G^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1}))$$ $$\label{kuothm3eq5} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).$$ We note that the region on the right-hand side of (\[kuothm3eq2\]) is obtained from the leftover region (the one restricted by the bold contour in Figure \[fig:kuooff3\](c)) by a $180^{\circ}$-rotation. Plugging the above 5 equalities into the equation in Kuo’s Theorem \[kuothm1\], we get the $G^{(2)}$-recurrences for $a>b$: $$\begin{aligned} \label{offcenterrecurG2b} {\operatorname{M}}(G^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(1)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ Similarly, we get the $G^{(2)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurG2a} {\operatorname{M}}(G^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(1)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(3)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})),\end{aligned}$$ and for for $a=b$ $$\begin{aligned} \label{offcenterrecurG2c} {\operatorname{M}}(G^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(1)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The same application of Kuo condensation gives us the recurrence of other $G^{(i)}$-type regions. The $G^{(1)}$-recurrence for $a<b$ $$\begin{aligned} \label{offcenterrecurG1a} {\operatorname{M}}(G^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y-1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(R^{\swarrow}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(F^{(2)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))\end{aligned}$$, for $a> b$ $$\begin{aligned} \label{offcenterrecurG1b} {\operatorname{M}}(G^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y-1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(R^{\swarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(F^{(2)}_{x,y-1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b})),\end{aligned}$$ and for $a=b$ $$\begin{aligned} \label{offcenterrecurG1c} {\operatorname{M}}(G^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y-1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(R^{\swarrow}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(F^{(2)}_{x,y-1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} The $G^{(3)}$-recurrence for $a> b$ is $$\begin{aligned} \label{offcenterrecurG3a} {\operatorname{M}}(G^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(2)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(4)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ When $a> b$, this recurrence becomes $$\begin{aligned} \label{offcenterrecurG3b} {\operatorname{M}}(G^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(4)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})),\end{aligned}$$ and when $a=b$ we have $$\begin{aligned} \label{offcenterrecurG3c} {\operatorname{M}}(G^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(K^{(2)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(4)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(G^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ Finally, the application of Kuo condensation gives the $G^{(4)}$-recurrence as $$\begin{aligned} \label{offcenterrecurG4a} {\operatorname{M}}(G^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}^{+1}))&={\operatorname{M}}(K^{(3)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(1)}_{x,y+1,z-1}(\textbf{b};\ \textbf{c};\ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(G^{(4)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\end{aligned}$$ for $a< b$, $$\begin{aligned} \label{offcenterrecurG4b} {\operatorname{M}}(G^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}^{+1}))&={\operatorname{M}}(K^{(3)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c};\ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(G^{(4)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\end{aligned}$$ for $a>b$, and $$\begin{aligned} \label{offcenterrecurG4c} {\operatorname{M}}(G^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}^{+1}))&={\operatorname{M}}(K^{(3)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(F^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c};\ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(G^{(4)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(E^{(6)}_{x,y,z}(\textbf{b};\ \overline{\textbf{c}};\ \textbf{a}))\end{aligned}$$ for $a=b$. Recurrences for $K^{(i)}$-type regions {#subsec:offrecurK} -------------------------------------- For the case of $K^{(i)}$-regions, we still apply Kuo’s Theorem \[kuothm1\] with the three vertices $u,v,w$ chosen similarly to that in the cases of $E^{(i)}$- and $F^{(i)}$-type regions, the only difference is that the $s$-triangle is now located at the west corner (see Figure \[fig:offK\]). Let us work in detail for the case of $K^{(1)}$-type regions with $a>b$. Let $G$ be the dual graph of the region $K^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ (for $a<b$). Consider forced lozenges as in Figures \[fig:kuooff4\](b)–(f), we get $$\label{kuothm4eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(E^{(1)}_{x-1,y,z-1}(E^{(1)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))$$ $$\label{kuothm4eq2} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(E^{(1)}_{x,y+1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))$$ $$\label{kuothm4eq3} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(K^{(1)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm4eq4} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(R^{\leftarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm4eq5} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(G^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).$$ We need to $180^{\circ}$-rotate the leftover regions in the case of (\[kuothm4eq1\]), (\[kuothm4eq2\]), (\[kuothm4eq5\]). By Eqs. (\[kuothm4eq1\])–(\[kuothm4eq5\]) and Theorem \[kuothm1\], we have the $K^{(1)}$-recurrence for $a>b$: $$\begin{aligned} \label{offcenterrecurK1b} {\operatorname{M}}(K^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(E^{(1)}_{x,y+1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(1)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(R^{\leftarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ By applying Kuo condensation similarly, we get other recurrences: The $K^{(1)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurK1a} {\operatorname{M}}(K^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(E^{(1)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(1)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(R^{\leftarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(2)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} The $K^{(2)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurK2a} {\operatorname{M}}(K^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(E^{(2)}_{x,y,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(2)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(3)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ The $K^{(2)}$-recurrence for $a> b$: $$\begin{aligned} \label{offcenterrecurK2b} {\operatorname{M}}(K^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(2)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a}))&={\operatorname{M}}(E^{(2)}_{x,y+1,z-1}(\textbf{b}^{+1};\ \overline{\textbf{c}}; \ \textbf{a})) {\operatorname{M}}(K^{(2)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(3)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ The $K^{(3)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurK3a} {\operatorname{M}}(K^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(E^{(6)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(K^{(3)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(4)}_{x-1,y-1,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ The $K^{(3)}$-recurrence for $a> b$: $$\begin{aligned} \label{offcenterrecurK3b} {\operatorname{M}}(K^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(6)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(E^{(6)}_{x,y+1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(K^{(3)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(4)}_{x-1,y,z}(\textbf{b}^{+1};\ \overline{\textbf{c}};\ \textbf{a})).\end{aligned}$$ The $K^{(4)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurK4a} {\operatorname{M}}(K^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(E^{(1)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(K^{(4)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(1)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ The $K^{(4)}$-recurrence for $a> b$: $$\begin{aligned} \label{offcenterrecurK4b} {\operatorname{M}}(K^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(E^{(1)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(E^{(1)}_{x,y+1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(K^{(4)}_{x-1,y-1,z+1}(\textbf{a};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(F^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(G^{(1)}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ Recurrences for $\overline{E}^{(i)}$-type regions {#subsec:offrecurE2} ------------------------------------------------- We apply Theorem \[kuothm1\] to the $\overline{E}^{(1)}$- and $\overline{E}^{(2)}$-type regions with the particular choice of the four vertices $u,v,w,s$ as shown in Figures \[fig:offE2\](a) and (b), respectively. For the $\overline{E}^{(3)}$-type region, we apply Theorem \[kuothm2\] as in Figure \[fig:offE2\](c). Let us work out in detail for case of $\overline{E}^{(1)}$-type regions when $a\leq b$. We also apply Kuo’s Theorem \[kuothm1\] to the dual graph $G$ of the region $\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$. Consider forced lozenges as in Figures \[fig:kuooff5\](b)–(f), we have $$\label{kuothm5eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(Q^{\leftarrow}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))$$ $$\label{kuothm5eq2} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(\overline{K}^{(5)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))$$ $$\label{kuothm5eq3} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(\overline{G}^{(5)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}))$$ $$\label{kuothm5eq4} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(\overline{E}^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm5eq5} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(Q^{\leftarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).$$ The region of type $Q^{\leftarrow}$ was enumerated in Theorem 2.7 of [@HoleDent]. The sequence $\textbf{c}^{\leftrightarrow}$ obtained by reverting the sequence $\textbf{c}$ when it has an odd number of terms, otherwise, we revert and add a new $0$ term in the beginning. More precisely, the leftover region after the removal of forced lozenges in Figure (\[kuothm5eq3\]) is not of our interests. We need to reflect the this region to get $\overline{G}^{(5)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a})$. Plugging the above five equations into that of Theorem \[kuothm1\], we get the $\overline{E}^{(1)}$-recurrence for $a\leq b$: $$\begin{aligned} \label{offcenterrecurE1qb} {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(5)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{G}^{(5)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\leftarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ Working out similarly, we get other recurrences: The $\overline{E}^{(1)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurE1qa} {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(5)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{G}^{(5)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(1)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\leftarrow}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} The $\overline{E}^{(2)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurE2qa} {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nwarrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{F}^{(5)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(5)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nwarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{E}^{(2)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurE2qb} {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nwarrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{F}^{(5)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(5)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nwarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{E}^{(3)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurE3qa} {\operatorname{M}}(Q^{\nearrow}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{G}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{E}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x,y+1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\leftarrow}_{x-1,y+1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})).\end{aligned}$$ The $\overline{E}^{(3)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurE3qb} {\operatorname{M}}(Q^{\nearrow}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{G}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{E}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{E}^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\leftarrow}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})).\end{aligned}$$ Recurrences for $\overline{F}^{(i)}$-type regions {#subsec:offrecurF2} ------------------------------------------------- We apply Kuo’s Theorem \[kuothm1\] to the $\overline{F}^{(3)}$-, $\overline{F}^{(4)}$- and $\overline{F}^{(5)}$-type regions as shown in Figures \[fig:offF2\](a), (b) and (c). The second Kuo’s theorem, Theorem \[kuothm2\], is used for the $\overline{F}^{(6)}$-type regions as in Figure \[fig:offF2\](d). Let us demonstrate in detail for the case of $\overline{F}^{(6)}$-type region when $a\geq b$. We apply Kuo’s Theorem \[kuothm2\] to the dual graph $G$ of the region $\overline{F}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$ with the choice of the four vertices $u,v,w,s$ shown in Figure \[fig:kuooff6\](d). By considering forced lozenges as in Figures \[fig:kuooff6\](a),(b),(d),(e),(f), we get $$\label{kuothm6eq1} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(\overline{G}^{(5)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}))$$ $$\label{kuothm6eq2} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm6eq3} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))$$ $$\label{kuothm6eq4} {\operatorname{M}}(G-\{u,w\})={\operatorname{M}}(\overline{F}^{(6)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm6eq5} {\operatorname{M}}(G-\{v,s\})={\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})).$$ We note that we reflected the leftover regions by a vertical line in (\[kuothm6eq3\]) and (\[kuothm6eq5\]) to get respectively $\overline{E}^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})$ and $\overline{E}^{(1)}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})$. Plugging the above 5 equalities into that of Kuo’s Theorem \[kuothm2\], we have the $\overline{F}^{(6)}$-recurrence for $a\geq b$: $$\begin{aligned} \label{offcenterrecurF6qb} {\operatorname{M}}(\overline{G}^{(5)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{F}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(6)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})).\end{aligned}$$ Similar application of Kuo condensation give the other 6 recurrences: The $\overline{F}^{(3)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurF3qa} {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{G}^{(2)}_{x,y-1,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(8)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{E}^{(1)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ The $\overline{F}^{(3)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurF3qb} {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{G}^{(2)}_{x,y,z-1}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(8)}_{x,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(3)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}^{+1})).\end{aligned}$$ The $\overline{F}^{(4)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurF4qa} {\operatorname{M}}(\overline{F}^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{E}^{(2)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{G}^{(2)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(4)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{F}^{(4)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurF4qb} {\operatorname{M}}(\overline{F}^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{G}^{(2)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(4)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{K}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{F}^{(5)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurF5qa} {\operatorname{M}}(\overline{F}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nearrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{E}^{(3)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nwarrow}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nearrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{F}^{(5)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurF5qb} {\operatorname{M}}(\overline{F}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nearrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{E}^{(3)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nwarrow}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\nearrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{F}^{(6)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurF6qa} {\operatorname{M}}(\overline{G}^{(5)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{F}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x,y+1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{F}^{(6)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{E}^{(1)}_{x-1,y+1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow};\ \textbf{a}^{+1})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} Recurrences for $\overline{G}^{(i)}$-type regions {#subsec:offrecurG2} ------------------------------------------------- We apply Theorem \[kuothm1\] to the $\overline{G}^{(2)}$- and $\overline{G}^{(3)}$-type regions as in Figures \[fig:offG2\](a) and (b). For the case of $\overline{G}^{(4)}$- and $\overline{G}^{(5)}$-type regions, we employ Theorem \[kuothm2\] with the choice of the vertices $u,v,w,s$ shown in Figures \[fig:offG2\](c) and (d). Let us work out in detail for the case of $\overline{G}^{(3)}$-type region when $a< b$. We apply Kuo’s Theorem \[kuothm1\] to the dual graph $G$ of the region $\overline{G}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$. By considering forced lozenges as in Figures \[fig:kuooff7\](b)–(f), we get $$\label{kuothm7eq1} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))$$ $$\label{kuothm7eq2} {\operatorname{M}}(G-\{u,v\})={\operatorname{M}}(\overline{K}^{(6)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1}))$$ $$\label{kuothm7eq3} {\operatorname{M}}(G-\{w,s\})={\operatorname{M}}(\overline{F}^{(4)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm7eq4} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(\overline{G}^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1}))$$ $$\label{kuothm7eq5} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).$$ Plugging the above 5 recurrences into the equation in Kuo’s Theorem \[kuothm1\], we obtain the $\overline{G}^{(3)}$-recurrence for $a< b$: $$\begin{aligned} \label{offcenterrecurG3qa} {\operatorname{M}}(\overline{G}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(6)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(4)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ Similarly, we have the other 7 recurrences: The $\overline{G}^{(2)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurG2qa} {\operatorname{M}}(\overline{G}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(5)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(3)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(2)}$-recurrence for $a> b$ $$\begin{aligned} \label{offcenterrecurG2qb} {\operatorname{M}}(\overline{G}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(2)}$-recurrence for $a= b$ $$\begin{aligned} \label{offcenterrecurG2qc} {\operatorname{M}}(\overline{G}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(1)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(5)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} The $\overline{G}^{(3)}$-recurrence for $a>b$ $$\begin{aligned} \label{offcenterrecurG3qb} {\operatorname{M}}(\overline{G}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(6)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(4)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(3)}$-recurrence for $a=b$ $$\begin{aligned} \label{offcenterrecurG3qc} {\operatorname{M}}(\overline{G}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{E}^{(2)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{K}^{(6)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(4)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(3)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(4)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurG4qa} {\operatorname{M}}(\overline{F}^{(5)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{E}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{G}^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nearrow}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(4)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\nearrow}_{x-1,y+1,z}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(4)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurG4qb} {\operatorname{M}}(\overline{F}^{(5)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{E}^{(3)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{G}^{(4)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\nearrow}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(4)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(Q^{\nearrow}_{x-1,y,z}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{G}^{(5)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurG5qa} {\operatorname{M}}(Q^{\leftarrow}_{x,y+1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}))&={\operatorname{M}}(\overline{G}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{K}^{(5)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(5)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})).\end{aligned}$$ The $\overline{G}^{(5)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurG5qb} {\operatorname{M}}(Q^{\leftarrow}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}))&={\operatorname{M}}(\overline{G}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{K}^{(5)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{G}^{(5)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{K}^{(5)}_{x-1,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})).\end{aligned}$$ Recurrences for $\overline{K}^{(i)}$-type regions {#subsec:offrecurK2} ------------------------------------------------- For the final groups of regions, $\overline{K}^{(5)}$–$\overline{K}^{(8)}$, we apply Kuo condensation in the same way we did for the $\overline{G}^{(i)}$-type regions above. The placements of the $u$-, $v$-, $w$- and $s$- triangles are shown particularly in Figure \[fig:offK2\]. Let us demonstrate in detail for the case of $\overline{K}^{(8)}$-type region when $a\geq b$. We apply Kuo’s Theorem \[kuothm2\] to the dual graph $G$ of the region $\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})$. By considering forced lozenges as in Figures \[fig:kuooff8\](a),(b),(d),(e),(f), we get $$\label{kuothm8eq1} {\operatorname{M}}(G-\{u,s\})={\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))$$ $$\label{kuothm8eq2} {\operatorname{M}}(G-\{v,w\})={\operatorname{M}}(\overline{F}^{(3)}_{x,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}))$$ $$\label{kuothm8eq3} {\operatorname{M}}(G-\{u,v,w,s\})={\operatorname{M}}(\overline{G}^{(2)}_{x,y-1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))$$ $$\label{kuothm8eq4} {\operatorname{M}}(G-\{u,w\})={\operatorname{M}}(\overline{K}^{(8)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b}))$$ $$\label{kuothm8eq5} {\operatorname{M}}(G-\{v,s\})={\operatorname{M}}(\overline{G}^{(2)}_{x-1,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}).$$ One can realize that we reflected the leftover regions by a vertical line in (\[kuothm8eq1\]), (\[kuothm8eq2\]), (\[kuothm8eq3\]) and (\[kuothm8eq5\]). Plugging the above 5 equalities into that of Kuo’s Theorem \[kuothm2\], we have the $\overline{K}^{(8)}$-recurrence for $a\geq b$ (illustrated in Figure \[fig:kuooff8\]) $$\begin{aligned} \label{offcenterrecurK8qb} {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})) {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}))&={\operatorname{M}}(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{G}^{(2)}_{x,y-1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(8)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{G}^{(2)}_{x-1,y-1,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})).\end{aligned}$$ We also have 7 more recurrences by applying Kuo condensation: The $\overline{K}^{(5)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurK5qa} {\operatorname{M}}(\overline{K}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(Q^{\nwarrow}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\leftarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(5)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurK5qb} {\operatorname{M}}(\overline{K}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(Q^{\leftarrow}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(Q^{\nwarrow}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(1)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(Q^{\leftarrow}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(6)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurK6qa} {\operatorname{M}}(\overline{K}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{F}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{G}^{(4)}_{x-1,y-1,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(2)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(6)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(6)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurK6qb} {\operatorname{M}}(\overline{K}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{F}^{(5)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}^{+1}))&={\operatorname{M}}(\overline{G}^{(4)}_{x-1,y,z}(\textbf{a};\ \textbf{c}; \ \textbf{b}^{+1})) {\operatorname{M}}(\overline{E}^{(2)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(6)}_{x-1,y,z-1}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b}^{+1})) {\operatorname{M}}(\overline{F}^{(5)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(7)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurK7qa} {\operatorname{M}}(\overline{E}^{(3)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{F}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{K}^{(7)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{G}^{(5)}_{x,y+1,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(7)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{G}^{(5)}_{x-1,y+1,z}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(7)}$-recurrence for $a\geq b$ $$\begin{aligned} \label{offcenterrecurK7qb} {\operatorname{M}}(\overline{E}^{(3)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b})) {\operatorname{M}}(\overline{F}^{(6)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b}))&={\operatorname{M}}(\overline{K}^{(7)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{G}^{(5)}_{x,y,z-1}(\textbf{a}^{+1};\ \textbf{c}; \ \textbf{b}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(7)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{G}^{(5)}_{x-1,y,z}(\textbf{a}^{+1};\ \textbf{c};\ \textbf{b})).\end{aligned}$$ The $\overline{K}^{(8)}$-recurrence for $a< b$ $$\begin{aligned} \label{offcenterrecurK8qa} {\operatorname{M}}(\overline{E}^{(1)}_{x,y+1,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})) {\operatorname{M}}(\overline{F}^{(3)}_{x,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}))&={\operatorname{M}}(\overline{K}^{(8)}_{x,y,z}(\textbf{a};\ \textbf{c};\ \textbf{b})){\operatorname{M}}(\overline{G}^{(2)}_{x,y,z-1}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1}))\notag\\ &+ {\operatorname{M}}(\overline{K}^{(8)}_{x+1,y,z-1}(\textbf{a};\ \textbf{c};\ \textbf{b})) {\operatorname{M}}(\overline{G}^{(2)}_{x-1,y,z}(\textbf{b};\ \textbf{c}^{\leftrightarrow}; \ \textbf{a}^{+1})).\end{aligned}$$ \#1\#2\#3\#4\#5[ @font ]{} . Two extremal cases for off-central regions {#sec:extremcase} ------------------------------------------ This subsection is devoted to two special cases when certain parameters in the off-central regions achieve their minimal values. Our first special case is when some of triangles in our ferns have side-length $0$. \#1\#2\#3\#4\#5[ @font ]{} \[lem3\] For any off-central regions, we can find a new region of the same type (1) whose number of tilings is the same, (2) whose $h$-parameter is at most that of the original region, (3) whose left and right ferns consist of triangles with positive side-lengths, and (4) whose middle fern contain at most one possible triangle of side length $0$ at the beginning. The proof is essentially the same as that of Lemma 3.5 in [@HoleDent], however in order to make our paper more self-contained, we still present the full proof here. We only consider the $R:=E^{(2)}_{x,y,z}(\textbf{a}; \textbf{c}; \textbf{b})$ region, the other regions can be treated similarly. We will show how to eliminate $0$-triangles from the three ferns without changing the tiling number or increasing the $h$-parameter. We consider the following three $0$-eliminating procedures for the left fern: \(1) If $a_1=a_2=\dotsc=a_{2i}=0$, for some $i\geq 1$, we can simply truncate the first $2i$ zero terms in the sequence $\textbf{a}$. The new region is ‘exactly’ the old one, however, strictly speaking, it has less $0$-triangles in the left fern. \(2) If $a_1=0$ and $a_2>0$, then we can remove forced lozenges along the southwest side of the region $R$ and obtain the region $E^{(2)}_{x,y,z}(a_3,\dots,a_m;\ \textbf{c};\ \textbf{b})$ (see Figure \[Specialoff\](a)). The new region has the same number of tilings as the original one, the $h$-parameter $a_1$-unit less than $h$, and less $0$-triangles in the left fern. \(3) If $a_i=0$, for some $i>1$, then we can eliminate this $0$-triangle by combining the $(i-1)$-th and the $(i+1)$-th triangles in the fern (as shown in Figure \[Specialoff\](b)). Repeating these three procedures if needed, one can eliminate all $0$-triangles from the left fern. Working similarly for the right fern, we obtain a region with no $0$-triangle in the left and right ferns. For the middle fern, we apply the procedure (3) to eliminate all $0$-triangles, except for a possible $0$-triangle at the beginning. This finishes our proof. In the next special case, we deal with off-central regions whose $y$-parameter achieves it minimal value. \[lem4\] For any off-central region $R$ with the $y$-parameter minimal, we can find another off-central region whose number of tilings is the same as that of $R$ and whose $h$-parameter is strictly smaller than that of $R$. By Lemma \[lem3\], we can assume that $a_i$’s, $b_j$’s, $c_t$’s are all positive, for $i\geq 1$, $j\geq 1$, $t\geq 2$. Recall that if $R$ has the leftmost of its middle fern $d$ unit above (resp., below) the center of the auxiliary hexagon $H_0$, for $d=0,\frac{1}{2}, 1, \frac{3}{2}$, then $y \geq \max(a-b,-2d)$ (resp., $y \geq \max(b-a,-2d)$). We call these regions *upper regions* (resp., *lower regions*). We note that all $\overline{E}^{(i)}$’s, $\overline{F}^{(i)}$’s, $\overline{G}^{(i)}$’s, and $\overline{K}^{(i)}$’s regions are upper ones. There are six cases to distinguish: *Case 1a: $R$ is an upper region, and $a\geq b$.* We consider the case $R$ has the leftmost of the middle fern $d$-unit above the center of the auxiliary hexagon, for $d=1/2,1,3/2$. If $a\geq b$, then $y\geq 0$, we can find a ‘bar’ or ‘un-bar’ counterpart of the region that has the same number of tilings and has $h$-parameter strictly smaller. In particular, by considering forced lozenges, we get $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(i)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(F^{(j)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(j)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(G^{(k)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(k)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(K^{(l)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(l)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b})) \end{aligned}$$ for $i=1,2,3,4$, $j=3,4,5,6$, $k=2,3,4,5$, and $l=5,6,7,8$ (see Figure \[fig:Specialoff1\] for examples). Similarly, we get $$\begin{aligned} {\operatorname{M}}(\overline{E}^{(i)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(E^{(i)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(\overline{F}^{(j)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(F^{(j)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(\overline{G}^{(k)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(G^{(k)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b}))\\ {\operatorname{M}}(\overline{K}^{(l)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(K^{(l)}_{x,\min(a_1, a-b),z}(a_2,\dotsc,a_m;\textbf{c}; \textbf{b})) \end{aligned}$$ for $i=1,2,3,4$, $j=3,4,5,6$, $k=2,3,4,5$, and $l=5,6,7,8$ (see Figure \[fig:Specialoff2\] for examples). \#1\#2\#3\#4\#5[ @font ]{} \#1\#2\#3\#4\#5[ @font ]{} \#1\#2\#3\#4\#5[ @font ]{} *Case 1b: $R$ is an upper region, and $0<2d\leq b-a$.* We have in this case $y\geq -2d$. If $y=-2d$, then $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(i)}_{x,\min(b_1, b-a)-2d,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(F^{(j)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(j)}_{x,\min(b_1, b-a)-2d,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(G^{(k)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(k)}_{x,\min(b_1, b-a)-2d,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(K^{(l)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(l)}_{x,\min(b_1, b-a)-2d,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n)), \end{aligned}$$ for $i=2,3$, $j=4,5,6$, $k=2,3,4,5$, and $l=5,6,7,8$ (illustrated in Figure \[fig:Specialoff4\](a)). Recall that we use the notation ${}^0\textbf{c}$ for the sequence obtained from $\textbf{c}$ by including a new $0$ term in the front. It is easy to see that the regions on the right-hand sides of the above equalities satisfy the conditions of the lemma. *Case 1c: $R$ is an upper region, and $0< b-a\leq 2d$.* We have here $y\geq a-b$. If $y=a-b$, then $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,a-b,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(i)}_{x,a-b,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(F^{(j)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(j)}_{x,a-b,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(G^{(k)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(k)}_{x,a-b,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(K^{(l)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(l)}_{x,a-b,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b})), \end{aligned}$$ for $i=2,3$, $j=4,5,6$, $k=2,3,4,5$, and $l=5,6,7,8$ (see Figure \[fig:Specialoff4\](b) for an example). It is easy to see that the regions on the right-hand sides of the above equalities satisfy the conditions of the lemma. *Case 2a: $R$ is a lower region and $a\leq b$.* Then $y\geq 0$, and we get by removing forced lozenges: $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(3+i \mod 6)}_{x,\min(b_1, b-a),z}(b_2,\dotsc,b_n;\overline{\textbf{c}}; \textbf{a}))\\ {\operatorname{M}}(F^{(j)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(4+j \mod 8)}_{x,\min(b_1, b-a),z}(b_2,\dotsc,b_n;\overline{\textbf{c}}; \textbf{a}))\\ {\operatorname{M}}(G^{(k)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(4+k \mod 8)}_{x,\min(b_1, b-a),z}(b_2,\dotsc,b_n;\overline{\textbf{c}}; \textbf{a}))\\ {\operatorname{M}}(K^{(l)}_{x,0,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(4+l \mod 8)}_{x,\min(b_1, b-a),z}(b_2,\dotsc,b_n;\overline{\textbf{c}}; \textbf{a})), \end{aligned}$$ for $i=1,4,5,6$, $j=1,2,3,7,8$, $k=1,6,7,8$, and $l=1,2,3,4$. Recall that $\overline{\textbf{c}}$ is the sequence obtained from $\textbf{c}$ by reverting the order of the terms if we have an even number of terms, otherwise, we revert the sequence and include a new $0$ in front of the resulting one. One readily sees that the regions on the right-hand sides of the above equalities satisfy the conditions of the lemma. \#1\#2\#3\#4\#5[ @font ]{} \#1\#2\#3\#4\#5[ @font ]{} *Case 2b: $R$ is a lower region and $ 0<2d\leq a-b$.* We have $y\geq -2d$. If $y=-2d$, then by investigating forced lozenges $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,a-b,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(3+i \mod 6)}_{x,\min(a_1,a-b)-2d,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(F^{(j)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(4+j \mod 8)}_{x,\min(a_1,a-b)-2d,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(G^{(k)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(4+k \mod 8)}_{x,\min(a_1,a-b)-2d,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b}))\\ {\operatorname{M}}(K^{(l)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(4+l \mod 8)}_{x,\min(a_1,a-b)-2d,z}(a_2,\dotsc,a_m;\textbf{c};\textbf{b})), \end{aligned}$$ for $i=5,6$, $j=1,2,8$, $k=1,6,7,8$, and $l=1,2,3,4$ (see Figure \[fig:Specialoff5\](a) for an example). Again, we can check that the new regions (the one on the right-hand sides of the above identities) have $h$-parameter strictly less than that of the corresponding region on the left-hand side. *Case 2c: $R$ is a lower region and $0<a-b\leq 2d$.* We have $y\geq b-a$ (in this case $b-a$ is negative). If $y=b-a$, then $$\begin{aligned} {\operatorname{M}}(E^{(i)}_{x,a-b,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{E}^{(3+i \mod 6)}_{x,b-a,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(F^{(j)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{F}^{(4+j \mod 8)}_{x,b-a,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(G^{(k)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{G}^{(4+k \mod 8)}_{x,b-a,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n))\\ {\operatorname{M}}(K^{(l)}_{x,-2d,z}(\textbf{a}; \textbf{c}; \textbf{b}))&={\operatorname{M}}(\overline{K}^{(4+l \mod 8)}_{x,b-a,z}(\textbf{a};\ {}^0\textbf{c};\ b_2,\dotsc,b_n)), \end{aligned}$$ for $i=5,6$, $j=1,2,8$, $k=1,6,7,8$, and $l=1,2,3,4$ (illustrated in Figure \[fig:Specialoff5\](b)). One readily sees that the regions on the right-hand sides of the above equalities satisfy the conditions of the lemma. This finishes our proof. Combined proof for all off-central regions ------------------------------------------ We prove all $30$ tiling formulas in Section 2 at the same time by induction on $h=p+x+z$. Recall that $p$ denotes the quasi-perimeter of the region, i.e. the perimeter of the base hexagon that our region is obtained by removing three ferns from. First by Lemma \[lem3\], we can assume that our three ferns consists of all triangles of positive side lengths, except for a possible $0$-triangle in front of the middle fern. The base cases are the cases $x=0$, $z=0$, and $p<6$. \#1\#2\#3\#4\#5[ @font ]{} \#1\#2\#3\#4\#5[ @font ]{} If $z=0$, we divide our region into two dented semihexagons with dents by cutting along the horizontal lattice line containing our three ferns. This semihexagons corresponds to the two $s$-terms in our tiling formulas (shown in Figure \[Basecaseoff1\]). By the Region-splitting Lemma \[RS\], the tiling number of our region is equal to the product of tiling numbers of the two dented semihexagons. Then our theorems follow from Cohn–Larsen–Propp’s formula (\[semieq\]). If $x=0$, we also partition the region into two parts along the same lattice line above, the only difference is that we now add two ‘bump’ of the same size as the gaps between the three ferns. The upper part is a dented semihexagon, and the lower part, after removing several vertical forced lozenges at the two bumps, is congruent with another dented semihexagon (illustrated in Figure \[Basecaseoff2\]). Again, the theorems follows from Cohn–Larsen–Propp formula (\[semieq\]) and the Region-splitting Lemma \[RS\]. If $p< 6$, by Lemma \[claimp\], we always have $p\geq 2x+4z$. This implies that $2x+4z<6$, and at least one of $x$ and $z$ is $0$. This base case is then reduced to the above base cases. For the induction step, we assume that $x$ and $z$ are positive, that $p\geq 6$, and that all our 30 tiling formulas hold for off-central regions whose $h$-parameter strictly less than $h=p+x+z$. If $y$ does not achieve its minimal value, by the $66$ recurrences in Sections 3.3–3.10, the tiling number of each of our off-central regions can be written in terms of tiling numbers of other regions, that are either the regions in the central case treated in [@HoleDent Theorems 2.2–2.9] or other off-central regions with $h$-parameter strictly less than $h$. This means that by Theorem 2.2–2.9 in [@HoleDent] and by the induction hypothesis, we get an explicit formula for the tiling number of each of our regions. Then, by performing a straightforward simplification, one can verify that the latter explicit tiling formula agrees with that in our theorem. If $y$ achieves its minimal value, then our recurrences dot not work well anymore. However, by Lemma \[lem4\], each of our regions has the same tiling number as that of a off-central region whose $h$-parameter strictly less than. Thus, our theorems follow from the induction hypothesis. This finishes our proof. Acknowledgement {#acknowledgement .unnumbered} =============== The author would like to thank Dennis Stanton and Hjalmar Rosengren for pointing out to him paper [@Rosen]. [50]{} F. Abeles, Dodgson condensation: The historical and mathematical development of an experimental method, *Linear Algebra Appl.* **429** (2–3) (2008), 429–438 G. E. Andrews, Plane partitions (III): The weak Macdonald conjecture, *Invent. Math.*, **53** (1979), 193–225. M. Ciucu, Enumeration of lozenge tilings of punctured hexagons, *J. Combin. Theory Ser. A*, **83** (1998), 268–272. M. Ciucu, Plane partition I: A generalization of MacMahon’s formula, *Memoirs of Amer. Math. Soc.*, **178** (2005), no. 839, 107–144. M. Ciucu, A generalization of Kuo condensation, *J. Combin. Theory Ser. A*, **134** (2015), 221–241. M. Ciucu, Another dual of MacMahon’s theorem on plane partitions, *Adv. Math.*, **306** (2017), 427–450. M. Ciucu, T. Eisenkölbl, C. Krattenthaler, and D. Zare, Enumeration of lozenge tilings of hexagons with a central triangular hole, *J. Combin. Theory Ser. A*, **95** (2001), 251–334. M. Ciucu and I. Fischer, Lozenge tilings of hexagons with arbitrary dents, *Adv. Appl. Math.*, **73** (2016), 1–22. M. Ciucu and C. Kattenthaler, The number of centered lozenge tilings of a symmetric hexagon, *J. Combin. Theory Ser. A*, **86** (1999), 103–126. M. Ciucu and C. Krattenthaler, A dual of MacMahon’s theorem on plane partitions, *Proc. Natl. Acad. Sci. USA*, **110** (2013), 4518–4523. M. Ciucu and T. Lai, Lozenge tilings doubly-intruded hexagons, Preprint (2017): `http://arxiv.org/abs/1712.08024`. H. Cohn, M. Larsen, and J. Propp, The shape of a typical boxed plane partition, *New York J. Math.*, **4** (1998), 137–165. C.L. Dodgson, Condensation of determinants, *Proc. Roy. Soc. London*, **15** (1866), 150– 155. M. Fulmek, Graphical Condensation, Overlapping Pfaffians and Superpositions of Matchings, *Electron. J. Combin.*, **17**(1) (2010), \#R83. T. Eisenkölbl, Rhombus tilings of a hexagon with three fixed border tiles, *J. Combin. Theory Ser. A*, **88** (2) (1999), 368–378. H. Helfgott and I. Gessel, Tilings of diamonds and hexagons with defects, *Electron. J. Combin.*, **6** (1999), \#R16. R. Kenyon and D. Wilson, The space of circular planar electrical networks, *SIAM J. Discrete Math.*, **31**(1), 1–28. D.E. Knuth, Overlapping pfaffians, *Electron. J. Combin.*, **3** (1996), R5. C. Koutschan, M. Kauer, and Zeilberger, A proof of George Andrews’ and David Robbins’ $a$ TSPP-conjecture. *Proc. Natl. Acad. Sci. USA*, **108** (2011), 2196–2199. C. Krattenthaler, Plane partitions in the work of Richard Stanley and his school, “*The Mathematical Legacy of Richard P. Stanley*" P. Hersh, T. Lam, P. Pylyavskyy and V. Reiner (eds.), Amer. Math. Soc., R.I., 2016, pp. 246-277. E. H. Kuo, Applications of graphical condensation for enumerating matchings and tilings, *Theor. Comput. Sci.*, **319** (2004), 29–57. E. H. [Kuo]{}, Graphical Condensation Generalizations Involving Pfaffians and Determinants, , May 2006. `arXiv:math/0605154`. G. Kuperberg, Symmetries of plane partitions and the permanent-determinant method, *J. Comnin. Theory Ser. A*, **68** (1994), 115–151. M. Leoni, G. Musiker, S. Neel, and P. Turner, Aztec Castles and the dP3 Quiver, *J. Phys. A: Math. Theor.*, **47** (2014), \#474011. T. Lai, A generalization of Aztec dragons, *Graph Combin.*, **2**(5) (2016), 1979–1999. T. Lai, A $q$-enumeration of lozenge tilings of a hexagon with three dents, *Adv. Applied Math.*, **82** (2017), 23–57. T. Lai, A $q$-enumeration of a hexagon with four adjacent triangles removed from the boundary, *European J. Combin.*, **64** (2017), 66–87. T. Lai, Lozenge Tilings of a Halved Hexagon with an Array of Triangles Removed from the Boundary, *SIAM J. Discrete Math.*, **32**(1) (2018), 783–814. T. Lai, Lozenge Tilings of a Halved Hexagon with an Array of Triangles Removed from the Boundary, Part II, *Electron. J. Combin.*, **25**(4) (2018), \#P4.58. T. Lai, Tiling Enumeration of Doubly-intruded Halved Hexagons , *Preprint `arXiv:1801.00249`*. T. Lai, Lozenge Tilings of Hexagons with Central Holes and Dents, *Preprint `arXiv:1803.02792`*. T. Lai, Proof of a Conjecture of Kenyon and Wilson on Semicontiguous Minors, *J. Combin. Theory Ser. A*, **116** (2019), 134–163. T. Lai and G. Musiker, Beyond Aztec castles: Toric cascades in the dP3 quiver, *Comm. Math. Phys.* , **356**(3) (2017), 823–88. T. Lai and R. Rohatgi, Cyclically Symmetric Tilings of a Hexagon with Four Holes, *Adv. Appl. Math.*, **96** (2018), 249–285. P. A. MacMahon, Memoir on the theory of the partition of numbers—Part V. Partition in two-dimensional space, *Phil. Trans. R. S.*, 1911, A. T. Muir, The Theory of Determinants in the Historical Order of Development, vol. I, Macmillan, London, 1906. S. Okada and C. Krattenthaler, The number of rhombus tilings of a ‘punctured’ hexagon and the minor summation formula, *Adv. in Appl. Math.* ,**21** (1998), 381–404. J. Propp, Enumeration of matchings: Problems and progress, *New Perspectives in Geometric Combinatorics*, Cambridge Univ. Press, Cambridge, 1999, 255–291. J. Propp, Generalized domino-shuffling, *Theoret. Comput. Sci.*, **303** (2003), 267–301. R. Proctor, Odd symplectic groups, *Invent. Math*, **92**(2) (1988), 307–332. R. Rohatgi, Enumeration of lozenge tilings of halved hexagons with a boundary defect, *Electron. J. Combin.*, **22**(4) (2015), P4.22. R. Rohatgi, Enumeration of tilings of a hexagon with a maximal staircase and a unit triangle removed, *Australas. J. Combin.*, **65**(3) (2016), 220–231. H. Rosengren, Selberg integrals, Askey–Wilson polynomials and lozenge tilings of a hexagon with a triangular hole, *J. Combin. Theory Ser. A* **138** (2016), 29–59. D. E [Speyer]{}, Perfect Matchings and the Octahedron Recurrence, *J. Algebraic Combin.*, **25**(6) (2007), 309–348. R. Stanley, Symmetries of plane partitions, *J. Combin. Theory Ser. A*, **43** (1986), 103–243. J. R. Stembridge, The enumeration of totally symmetric plane partitions, *Adv. in Math.*, **111** (1995), 227–243. W. Yan, Y. Yeh, and F. Zhang, Graphical condensation of plane graphs: A combinatorial approach, *Theoret. Comput. Sci.*, **349**(3) (2005), 452–461. W. Yan and F. Zhang, Graphical condensation for enumerating perfect matchings, *J. Combin. Theory Ser. A*, **110** (2005), 113–125. D. Zeilberger, Dodgson’s Determinant-Evaluation Rule proved by Two-Timing Men and Women, *Electron. J. Combin.*, **4**(2)(1997),\#R22. [^1]: This research was supported in part by Simons Foundation Collaboration Grant (\# 585923). [^2]: From now on, we always list the side-lengths of a hexagon in the clockwise order from the north side. [^3]: In this section, we ignore the four shaded triangles of labels $u,v,w,s$ in the illustrative figures; these triangles will be used later in Section \[sec:proofoff\].

ArXiv

--- abstract: | Let $A \in \{0,1\}^{n \times n}$ be a matrix with $z$ zeroes and $u$ ones and $x$ be an $n$-dimensional vector of formal variables over a semigroup $(S, \circ)$. How many semigroup operations are required to compute the linear operator $Ax$? As we observe in this paper, this problem contains as a special case the well-known range queries problem and has a rich variety of applications in such areas as graph algorithms, functional programming, circuit complexity, and others. It is easy to compute $Ax$ using $O(u)$ semigroup operations. The main question studied in this paper is: can $Ax$ be computed using $O(z)$ semigroup operations? We prove that in general this is not possible: there exists a matrix $A \in \{0,1\}^{n \times n}$ with exactly two zeroes in every row (hence $z=2n$) whose complexity is $\Theta(n\alpha(n))$ where $\alpha(n)$ is the inverse Ackermann function. However, for the case when the semigroup is commutative, we give a constructive proof of an $O(z)$ upper bound. This implies that in commutative settings, complements of sparse matrices can be processed as efficiently as sparse matrices (though the corresponding algorithms are more involved). Note that this covers the cases of Boolean and tropical semirings that have numerous applications, e.g., in graph theory. As a simple application of the presented linear-size construction, we show how to multiply two $n\times n$ matrices over an arbitrary semiring in $O(n^2)$ time if one of these matrices is a 0/1-matrix with $O(n)$ zeroes (i.e., a complement of a sparse matrix). author: - 'Alexander Kulikov[^1]' - 'Ivan Mikhailin[^2]' - 'Andrey Mokhov[^3]' - 'Vladimir Podolskii[^4]' bibliography: - 'text.bib' title: Complexity of Linear Operators --- Introduction ============ Problem Statement and New Results --------------------------------- Let $A \in \{0,1\}^{n \times n}$ be a matrix with $z$ zeroes and $u$ ones, and $x=(x_1, \dotsc, x_n)$ be an $n$-dimensional vector of formal variables over a semigroup $(S, \circ)$. In this paper, we study the complexity of the *linear operator* $Ax$, i.e., how many semigroup operations are required to compute a vector whose $i$-th element is $$\sum_{1 \le j \le n\,\bigwedge\,A_{ij}=1}x_j$$ where the summation is over the semigroup operation $\circ$.[^5] More specifically, we are interested in lower and upper bounds involving $z$ and $u$. Matrices with $u=O(n)$ are usually called *sparse*, whereas matrices with $z=O(n)$ are called *complements of sparse matrices*. Computing all $n$ outputs of $Ax$ directly (i.e. using the above definition) takes $O(u)$ semigroup operations. The main question studied in this paper is: can $Ax$ be computed using $O(z)$ semigroup operations? Note that it is easy to achieve $O(z)$ complexity if $\circ$ has an inverse. Indeed, in this case $Ax$ can be computed via subtraction: $Ax = (U-\overline{A})x = Ux - \overline{A}x$, where $U$ is the all-ones matrix whose linear operator can be computed trivially using $O(n)$ semigroup operations, and $\overline{A}$ is the complement of $A$ and therefore has only $z = O(n)$ ones. ### Commutative Case Our first main result shows that in the commutative case, complements of sparse matrices can be processed as efficiently as sparse matrices. Specifically, we prove that if the semigroup is commutative, $Ax$ can be computed in $O(z)$ semigroup operations; or, more formally, there exists a circuit of size $O(z)$ that uses $x=(x_1, \dotsc, x_n)$ as an input and computes $Ax$ by only applying the semigroup operation $\circ$ (we provide the formal definition of the computational model in Section \[subsec:circuits\]). Moreover, the constructed circuits are *uniform* in the sense that they can be generated by an efficient algorithm. Hence, our circuits correspond to an elementary algorithm that uses no tricks like examining the values $x_j$, i.e., the semigroup operation $\circ$ is applied in a (carefully chosen) order that is independent of the specific input $x$. [theorem]{}[upperthm]{} \[thm:upperbound\] Let $(S, \circ)$ be a commutative semigroup, and $A \in \{0,1\}^{n \times n}$ be a matrix with $z=\Omega(n)$ zeroes. There exists a circuit of size $O(z)$ that uses a vector $x = (x_1,\ldots, x_n)$ of formal variables as an input, uses only the semigroup operation $\circ$ at internal gates, and outputs $Ax$. Moreover, there exists a randomized algorithm that takes the positions of $z$ zeroes of $A$ as an input and outputs such a circuit in time $O(z)$ with probability at least $1-\frac{O(\log^5n)}{n}$. There also exists a deterministic algorithm with running time $O(z+n\log^4n)$. We state the result for square matrices to simplify the presentation. Theorem \[thm:upperbound\] generalizes easily to show that $Ax$ for a matrix $A \in \{0,1\}^{m \times n}$ with $z=\Omega(n)$ zeroes can be computed using $O(m+z)$ semigroup operations. Also, we assume that $z=\Omega(n)$ to be able to state an upper bound $O(z)$ instead of $O(z+n)$. Note that when $z<n$, the matrix $A$ is forced to contain all-one rows that can be computed trivially. The following corollary generalizes Theorem \[thm:upperbound\] from vectors to matrices. [corollary]{}[matrixmultcor]{} \[cor:matrixmultiplication\] Let $(S, \circ)$ be a commutative semigroup. There exists a deterministic algorithm that takes a matrix $A \in \{0,1\}^{n \times n}$ with $z=O(n)$ zeroes and a matrix $B \in S^{n \times n}$ and computes the product $AB$ in time $O(n^2)$. ### Non-commutative Case As our second main result, we show that *commutativity is essential*: for a faithful non-commutative semigroup $S$ (the notion of faithful non-commutative semigroup is made formal later in the text), the minimum number of semigroup operations required to compute $Ax$ for a matrix $A \in \{0,1\}^{n \times n}$ with $z=O(n)$ zeroes is $\Theta(n\alpha(n))$, where $\alpha(n)$ is the inverse Ackermann function. [theorem]{}[lowerthm]{} \[thm:lowerbound\] Let $(S, \circ)$ be a faithful non-commutative semigroup, $x = (x_1,\ldots, x_n)$ be a vector of formal variables, and $A \in \{0,1\}^{n \times n}$ be a matrix with $O(n)$ zeroes. Then $Ax$ is computable using $O(n\alpha(n))$ semigroup operations, where $\alpha(n)$ is the inverse Ackermann function. Moreover, there exists a matrix $A \in \{0,1\}^{n \times n}$ with exactly two zeroes in every row such that the minimum number of semigroup operations required to compute $Ax$ is $\Omega(n\alpha(n))$. Motivation ---------- The complexity of linear operators is interesting for many reasons. Range queries. : In the *range queries* problem, one is given a vector $x=(x_1, \dotsc, x_n)$ over a semigroup $(S, \circ)$ and multiple queries of the form $(l,r)$, and is required to output the result $x_l \circ x_{l+1} \circ \dotsb \circ x_r$ for each query. It is a classical problem in data structures and algorithms with applications in many fields, such as bioinformatics and string algorithms, computational geometry, image analysis, real-time systems, and others. We review some of the less straightforward applications in Section \[subseq:rmqapp\], as well as a rich variety of algorithmic techniques for the problem in Section \[subsec:approaches\]. The linear operator problem is a natural generalization of the range queries problem: each row of the matrix $A$ defines a subset of the elements of $x$ that need to be summed up and this subset is not required to be a contiguous range. The algorithms (Theorem \[thm:upperbound\] and Corollary \[cor:matrixmultiplication\]) and hardness results (Theorem \[thm:lowerbound\]) for the linear operator problem presented in this paper are indeed inspired by some of the known results for the range queries problem. Graph algorithms. : Various graph path/reachability problems can be reduced naturally to matrix multiplication. Two classic examples are: (i) the all-pairs shortest path problem (APSP) is reducible to min-plus matrix multiplication, and (ii) the number of triangles in an undirected graph can be found by computing the third power of its adjacency matrix. It is natural to ask what happens if a graph has $O(n)$ edges or $O(n)$ anti-edges (as usual, by $n$ we denote the number of nodes). In many cases, an efficient algorithm for sparse graphs ($O(n)$ edges) is straightforward whereas an algorithm with the same efficiency for complements of sparse graphs ($O(n)$ anti-edges) is not. For example, it is easy to solve APSP and triangle counting on sparse graphs in time $O(n^2)$, but achieving the same time complexity for complements of sparse graphs is more complicated. Theorem \[thm:upperbound\] and Corollary \[cor:matrixmultiplication\] can be used to solve various problems on complements of sparse graphs in time $O(n^2)$. Matrix multiplication over semirings. : Fast matrix multiplication methods rely essentially on the ring structure of the underlying set of elements. The first such algorithm was given by Strassen, the current record upper bound is $O(n^{2.373})$ [@DBLP:conf/stoc/Williams12; @DBLP:conf/issac/Gall14a]. The removal of the inverse operation often drastically increases the complexity of algorithmic problems over algebraic structures, and even the complexity of standard computational tasks are not well understood over tropical and Boolean semirings (see, e.g. [@Williams14; @GrigorievP15]). For various important semirings, we still do not know an $n^{3-\varepsilon}$ (for a constant $\varepsilon>0$) upper bound for matrix multiplication, e.g., the strongest known upper bound for min-plus matrix multiplication is $n^3/\exp(\sqrt{\log n})$ [@Williams14]. The interest in computations over such algebraic structures has recently grew substantially throughout the Computer Science community with the cases of Boolean and tropical semirings being of the main interest (see, for example, [@Jukna16; @Williams14; @butkovic10systems]). From this perspective, the computation complexity over sparse and complements of sparse matrices is one of the most basic questions. Theorem \[thm:upperbound\] and Corollary \[cor:matrixmultiplication\] therefore characterise natural special cases when efficient computations are possible. Functional programming. : Algebraic data structures for graphs developed in the functional programming community [@mokhov2017algebraic] can be used for representing and processing densely-connected graphs in linear (in the number of vertices) time and memory. As we discuss in Section \[sec-dense-graph\], Theorem \[thm:upperbound\] yields an algorithm for deriving a linear-size algebraic graph representation for complements of sparse graphs. Circuit complexity. : Computing linear operators over a Boolean semiring $(\{0,1\}, \lor)$ is a well-studied problem in circuit complexity. The corresponding computational model is known as *rectifier networks*. An overview of known lower and upper bounds for such circuits is given by Jukna [@DBLP:books/daglib/0028687 Section 13.6]. Theorem \[thm:upperbound\] states that very dense linear operators have linear rectifier network complexity. Background ========== Semigroups and Semirings ------------------------ A *semigroup* $(S, \circ)$ is an algebraic structure, where the operation $\circ$ is *closed*, i.e., $\circ : S\times S \rightarrow S$, and *associative*, i.e., $x \circ (y \circ z) = (x \circ y) \circ z$ for all $x$, $y$, and $z$ in $S$. *Commutative* (or *abelian*) semigroups introduce one extra requirement: $x \circ y = y \circ x$ for all $x$ and $y$ in $S$. A commutative semigroup $(S, \circ)$ can often be extended to a *semiring* $(S, \circ, \bullet)$ by introducing another associative (but not necessarily commutative) operation $\bullet$ that *distributes* over $\circ$, that is $$x \bullet (y \circ z) = (x \bullet y) \circ (x \bullet z)\\$$ $$(x \circ y) \bullet z = (x \bullet z) \circ (y \bullet z)$$ hold for all $x$, $y$, and $z$ in $S$. Since $\circ$ and $\bullet$ behave similarly to numeric addition and multiplication, it is common to give $\bullet$ a higher precedence to avoid unnecessary parentheses, and even omit $\bullet$ from formulas altogether, replacing it by juxtaposition. This gives a terser and more convenient notation, e.g., the left distributivity law becomes: $x (y \circ z) = x y \circ x z$. We will use this notation, insofar as this does not lead to ambiguity. See Subsection  \[subsec:algstr\] for an overview of commonly used semigroups and semirings. Range Queries Problem and Linear Operator Problem ------------------------------------------------- In the [*range queries problem*]{}, one is given a sequence $x_1, x_2, \dotsc, x_n$ of elements of a fixed semigroup $(S, \circ)$. Then, a *range query* is specified by a pair of indices $(l,r)$, such that $1 \le l \le r \le n$. The answer to such a query is the result of applying the semigroup operation to the corresponding range, i.e., $x_l \circ x_{l+1} \circ \dotsb \circ x_r$. The range queries problem is then to simply answer all given range queries. There are two regimes: online and offline. In the [*online regime*]{}, one is given a sequence of [*values*]{} $x_1=v_1, x_2=v_2, \dotsc, x_n=v_n$ and is asked to preprocess it so that to answer efficiently any subsequent query. By “efficiently” one usually means in time independent of the length of the range (i.e., $r-l+1$, the time of a naive algorithm), say, in time $O(\log n)$ or $O(1)$. In this paper, we focus on the [*offline*]{} version, where one is given a sequence together with all the queries, and are interested in the minimum number of semigroup operations needed to answer all the queries. Moreover, we study a more general problem: we assume that $x_1, \dotsc, x_n$ are formal variables rather than actual semigroup values. That is, we study the [*circuit size*]{} of the corresponding computational problem. The [*linear operator*]{} problem generalizes the range queries problem: now, instead of contiguous ranges one wants to compute sums over arbitrary subsets. These subsets are given as rows of a 0/1-matrix $A$. Circuits {#subsec:circuits} -------- We assume that the input consists of $n$ formal variables $\{x_1, \dotsc, x_n\}$. We are interested in the minimum number of semigroup operations needed to compute all given words $\{w_1, \dotsc, w_m\}$ (e.g., for the range queries problem, each word has a form $x_l\circ x_{l+1}\circ \dotsb \circ x_r$). We use the following natural [*circuit*]{} model. A circuit computing all these queries is a directed acyclic graph. There are exactly $n$ nodes of zero in-degree. They are labelled with $\{1, \dotsc, n\}$ and are called [*input gates*]{}. All other nodes have positive in-degree and are called [*gates*]{}. Finally, some $m$ gates have out-degree 0 and are labelled with $\{1, \dotsc, m\}$; they are called [*output gates*]{}. The [*size*]{} of a circuit is its number of edges (also called [*wires*]{}). Each gate of a circuit computes a word defined in a natural way: input gates compute just $\{x_1, \dotsc, x_n\}$; any other gate of in-degree $r$ computes a word $f_1 \circ f_2 \circ \dotsb \circ f_r$ where $\{f_1, \dotsc, f_r\}$ are words computed at its predecessors (therefore, we assume that there is an underlying order on the incoming wires for each gate). We say that the circuit computes the words $\{w_1, \dotsc, w_m\}$ if the words computed at the output gates are equivalent to $\{w_1, \dotsc, w_m\}$ over the considered semigroup. For example, the following circuit computes range queries $(l_1,r_1)=(1,4)$, $(l_2,r_2)=(2,5)$, and $(l_3,r_3)=(4,5)$ over inputs $\{x_1, \dotsc, x_5\}$ or, equivalently, the linear operator $Ax$ where the matrix $A$ is given below. ///in [0/4/x1/1, 1/4/x2/2, 2/4/x3/3, 3/4/x4/4, 4/4/x5/5, 2/3/a/ , 1/2/b/1, 3/2/c/2, 4/2/d/3]{} () at (,) [$\t$]{}; /in [x2/a, x3/a, x4/a, x1/b, a/b, x5/c, a/c, x4/d, x5/d]{} () – (); at (8,3) [$A=\begin{pmatrix}1&1&1&1&0\\0&1&1&1&1\\0&0&0&1&1\end{pmatrix}$]{}; For a 0/1-matrix $A$, by $C(A)$ we denote the minimum size of a circuit computing the linear operator $Ax$. A [*binary circuit*]{} is a circuit having no gates of fan-in more than two. It is not difficult to see that any circuit can be converted into a binary circuit of size at most twice the size of the original circuit. For this, one just replaces every gate of fan-in $k$, for $k>2$, by a binary tree with $2k-2$ wires (such a tree contains $k$ leaves hence $k-1$ inner nodes and $2k-2$ edges). In the binary circuit the number of gates does not exceed its size (i.e., the number of wires). And the number of gates in a binary circuit is exactly the minimum number of semigroup operations needed to compute the corresponding function. We call a circuit $C$ computing $A$ *regular* if for every pair $(i,j)$ such that $A_{ij}=1$, there exists exactly one path from the input $j$ to the output $i$. A convenient property of regular circuits is the following observation. \[obs:transpose\] Let $C$ be a regular circuit computing a 0/1-matrix $A$ over a commutative semigroup. Then, by reversing all the wires in $C$ one gets a circuit computing $A^T$. Instead of giving a formal proof, we provide an example of a reversed circuit from the example given above. It is because of this observation that we require circuit outputs to be gates of out-degree zero (so that when reversing all the wires the inputs and the outputs exchange places). ///in [0/4/x1/1, 1/4/x2/2, 2/4/x3/3, 3/4/x4/4, 4/4/x5/5, 2/3/a/ , 1/2/b/1, 3/2/c/2, 4/2/d/3]{} () at (,) [$\t$]{}; /in [x2/a, x3/a, x4/a, x1/b, a/b, x5/c, a/c, x4/d, x5/d]{} () – (); at (8,3) [$A^T=\begin{pmatrix}1&0&0\\1&1&0\\1&1&0\\1&1&1\\0&1&1\end{pmatrix}$]{}; Commutative Case {#sec-commutative} ================ This section is devoted to the proofs of Theorem \[thm:upperbound\] and Corollary \[cor:matrixmultiplication\] which we remind below. We start by proving two simpler statements to show how commutativity is important. \[lemma:easy\] Let $S$ be a semigroup (not necessarily commutative) and let $A \in \{0,1\}^{n \times n}$ contain at most one zero in every row. Then $C(A) = O(n)$. To compute the linear operator $Ax$, we first precompute all prefixes and suffixes of $x=(x_1, \dotsc, x_n)$. Concretely, let $p_i=x_1 \circ x_2 \circ \dotsb \circ x_i$. All $p_i$’s can be computed using $(n-1)$ binary gates as follows: $$p_1=x_1, p_2=p_1 \circ x_2, p_3=p_2 \circ x_3, \dotsc, p_i=p_{i-1} \circ x_i, \dotsc, p_n=p_{n-1}\circ x_n.$$ Similarly, we compute all suffixes $s_j=x_j \circ x_{j+1} \dotsb \circ x_n$ using $(n-1)$ binary gates. From these prefixes and suffixes all outputs can be computed as follows: if a row of $A$ contains no zeroes, the corresponding output is $p_n$; otherwise if a row contains a zero at position $i$, the output is $p_{i-1} \circ s_{i+1}$ (for $i=1$ and $i=n$, we omit the redundant term). In the rest of the section, we assume that the underlying semigroup is commutative. Allowing at most two zeroes per row already leads to a non-trivial problem. We give only a sketch of the solution below, since we will further prove a more general result. It is interesting to compare the following lemma with Theorem \[thm:lowerbound\] that states that in the non-commutative setting matrices with two zeroes per row are already hard. \[lem:at\_most\_2\] Let $A \in \{0,1\}^{n \times n}$ contain at most two zeroes in every row. Then $C(A) = O(n)$. Consider the following undirected graph: the set of nodes is $\{1,2,\dotsc,n\}$; two nodes $i$ and $j$ are joined by an edge if there is a row having zeroes in columns $i$ and $j$. In the worst case (all rows are different and contain exactly two zeroes), the graph has exactly $n$ edges and hence it contains a cut $(L,R)$ of size at least $n/2$. This cut splits the columns of the matrix into two parts ($L$ and $R$). Now let us also split the rows into two parts: the top part $T$ contains all columns that have exactly one zero in each $L$ and $R$; the bottom part $B$ contains all the remaining rows. What is nice about the top part of the matrix ($T \times (L \cup R)$) is that it can be computed by $O(n)$ gates (using Lemma \[lemma:easy\]). For the bottom part, let us cut all-1 columns out of it and make a recursive call (note that this requires the commutativity). The corresponding recurrence relation is $T(n) \le cn + T(n/2)$ for a fixed constant $c$, implying $T(n)=O(n)$, and hence $C(A) = O(n)$. We now state a few auxiliary lemmas that will be used as building blocks in the proof of Theorem \[thm:upperbound\]. \[lemma:decompose\] There exists a binary regular circuit of size $O(n\log n)$ such that any range can be computed in a single additional binary gate using two gates of the circuit. It can be generated in time $O(n\log n)$. \[lemma:blocks\] There exists a binary regular circuit of size $O(n)$ such that any range of length at least $\log n$ can be computed in two binary additional gates from the gates of the circuit. It can be generated by an algorithm in time $O(n)$. \[lemma:permute\] Let $m \le n$ and $A \in \{0,1\}^{m \times n}$ be a matrix with $z=\Omega(n)$ zeroes and at most $\log n$ zeroes in every row. There exists a circuit of size $O(z)$ computing $Ax$. Moreover, there exists a randomized $O(z)$ time algorithm that takes as input the positions of $z$ zeros and outputs a circuit computing $Ax$ with probability at least $1-\frac{O(\log^5n)}{n}$. There also exists a deterministic algorithm with running time $O(n\log^4n)$. Denote the set of rows and the set of columns of $A$ by $R$ and $C$, respectively. Let $R_0 \subseteq R$ be all the rows having at least $\log n$ zeroes and $R_1=R \setminus R_0$. We will compute all the ranges of $A$. From these ranges, it takes $O(z)$ additional binary gates to compute all the outputs. We compute the matrices $R_0 \times C$ and $R_1 \times C$ separately. The main idea is that $R_0 \times C$ is easy to compute because it has a small number of rows (at most $z/\log n$), while $R_1 \times C$ is easy to compute because it has a small number of zeroes in every row (at most $\log n$). The matrix $R_1 \times C$ can be computed using Lemma \[lemma:permute\]. To compute $R_0 \times C$, it suffices to compute $C \times R_0$ by a regular circuit, thanks to the Observation \[obs:transpose\]. Let $|R_0|=t$. Clearly, $t \le z/\log n$. Using Lemma \[lemma:decompose\], one can compute all ranges of $C \times R_0$ by a circuit of size $$O(t\log t+z)=O\left(\frac{z}{\log n} \cdot \log z+z\right)=O(z+n)=O(z)\, ,$$ since $z =O(n^2)$. The algorithm for generating the circuit is just a combination of the algorithms from Lemmas \[lemma:decompose\] and \[lemma:permute\]. We adopt the divide-and-conquer construction by Alon and Schieber [@Alon87optimalpreprocessing]. Split the input range $(1,n)$ into two half-ranges of length $n/2$: $(1,n/2)$ and $(n/2+1,n)$. Compute all suffixes of the left half and all prefixes of the right half. Using these precomputed suffixes and prefixes one can answer any query $(l,r)$ such that $l \le n/2 \le r$ in a single additional gate. It remains to be able to answer queries that lie entirely in one of the halves. We do this by constructing recursively circuits for both halves. The resulting recurrence relation $T(n) \le 2T(n/2)+O(n)$ implies that the resulting circuit has size at most $O(n\log n)$. We use the block decomposition technique for constructing the required circuit. Partition the input range $(1,n)$ into $n/\log n$ ranges of length $\log n$ and call them blocks. Compute the range corresponding to each block (in total size $O(n)$). Build a circuit from Lemma \[lemma:decompose\] on top of these blocks. The size of this circuit is $O(n)$ since the number of blocks is $n/\log n$. Compute all prefixes and all suffixes of every block. Since the blocks partition the input range $(1,n)$, this also can be done with an $O(n)$ size circuit. Consider any range of length at least $\log n$. Note that it cannot lie entirely inside the block. Hence, any such range can be decomposed into three subranges: a suffix of a block, a range of blocks, and a prefix of a block (where any of the three components may be empty). For example, for $n=16$, a range $(3,13)$ is decomposed into a suffix $(3,4)$ of the first block, a range $(2,3)$ of blocks $(B_1, B_2, B_3, B_4)$, and a prefix $(13,13)$ of the last block: in [1,...,16]{} at (,2) ; (2.5,0.5) rectangle (13.5,1.5); in [1,...,15]{} (+0.5,0.5) – (+0.5,1.5); (0.5,0.5) rectangle (16.5,1.5); in [4,8,12]{} (+0.5,0.4) – (+0.5,1.6); /iin [2/1, 6/2, 10/3, 14/4]{} at (+0.5,0) [$B_{\i}$]{}; It remains to note that all these three components are already precomputed. The $z$ zeroes of $A$ breaks its rows into ranges. Let us call a range [*short*]{} is its length is at most $\log n$. We will show that it is possible to permute the columns of $A$ so that the total length of all short ranges is at most $O(\frac{n}{\log n})$. Then, all such short ranges can be computed by a circuit of size $O(\frac{\log n}{n} \cdot n)=O(n)=O(z)$. All the remaining ranges can be computed by a circuit of size $O(n)$ using Lemma \[lemma:blocks\]. Randomized algorithm. : Permute the columns randomly. A uniform random permutation of $n$ objects can be generated in time $O(n)$ [@DBLP:books/lib/Knuth98 Algorithm P (Shuffling)]. Let us compute the expectation of the total length of short ranges. Let us focus on a single row and a particular cell in it. Denote the number of zeroes in the row by $t$. What is the probability that the cell belongs to a short segment? There are two cases to consider. 1. The cell lies close to the border, i.e., it belongs to the first $\log n$ cells or to the last $\log n$ cells (the number of such cells is $2\log n$). Then, this cell belongs to a short range iff there is at least one zero in $\log n$ cells close to it (on the side opposite to the border). Hence, one zero must belong to a set of $\log n$ cells while the remaining $t-1$ zeroes may be anywhere. The probability is then at most $$\log n \cdot \frac{\binom{n}{t-1}}{\binom{n}{t}}=\log n \cdot \frac{t}{n-t+1}=O\left(\frac{\log^2n}{n}\right) \, .$$ 2. It is not close to the border (the number of such cells is $n-2\log n$). Then, there must be a zero on both sides of the cell. The probability is then at most $$\log^2 n \cdot \frac{\binom{n}{t-2}}{\binom{n}{t}}=\log^2n \cdot \frac{t(t-1)}{(n-t+1)(n-t+2)}=O\left(\frac{\log^4 n}{n^2}\right) \, .$$ Hence, the expected total length of short ranges in one row is $$O\left( 2\log n \cdot \frac{\log^2 n}{n} + (n-2\log n) \cdot \frac{\log^4 n}{n^2}\right)=O\left(\frac{\log^4 n}{n}\right) \, .$$ Thus, the expected length of short ranges in the whole matrix $A$ is $O(\log^4n)$. By Markov inequality, the probability that the length of all short ranges is larger than $\frac{n}{\log n}$ is at most $O(\frac{\log^5 n}{n})$. Deterministic algorithm. : It will prove convenient to assume that $A$ is a $t \times t$ matrix with exactly $t$ zeros with at most $\log t$ zeroes in every row. For this, we let $t=\max\{n, z\}$ and add a number of all-ones rows and columns if needed. This enlargement of the matrix does not make the computation simpler: additional rows mean additional outputs that can be ignored and additional columns correspond to redundant variables that can be removed (substituted by 0) once the circuit is constructed. Below, we show how to deterministically construct a circuit of size $O(t)$ for $A$. For this, we present a greedy algorithm for permuting the columns of $A$ in such a way that the total length of all short segments is $O(\log^4n)$. This will follow from the fact that all short ranges in the resulting matrix $A$ will lie within the last $O(\log^2 t)$ columns. We construct the required permutation of columns step by step by a greedy algorithm. After step $r$ we will have a sequence of the first $r$ columns chosen and we will maintain the following properties: - For each $i \leq r$, the first $i$ columns contain at least $i$ zeros. - There are no short ranges within the first $r$ rows (apart from those, that can be extended by adding columns on the right). The process will work for at least $t - \log^2 t$ steps, so short ranges are only possible within the last $\log^2 t + \log t = O(\log^2 t)$ columns. On the first step, we pick any column that has a zero in it. Suppose we have reached step $r$. We explain how to add a column on step $r+1$. Consider the last $\log t$ columns in the currently constructed sequence. Consider the set $R$ of rows that have zeros in them. These are exactly the rows that constrain our choice for the next column. There are two cases. 1. There are at most $\log t$ rows in $R$. Then, for each row in $R$, there are at most $\log t$ columns that have zeros in this row. In total, there are at most $\log^2 t$ columns that have zeros in some row of $R$. Denote the set of this columns by $F$. If there is an unpicked column outside of $F$ that has at least one zero in it, we add this column to our sequence. Clearly, both properties are satisfied and the step is over. Otherwise, all other columns contain only ones, so we add all of them to our sequence, place the columns from $F$ to the end of the sequence, and the whole permutation is constructed. 2. There are more that $\log t$ rows in $R$. This means that the last $\log t$ columns of the current sequence contain more than $\log t$ zeros. By the first property, the first $r - \log t$ columns contain at least $r - \log t$ zeros. So overall, in the current sequence of $r$ columns there are more than $r$ zeros. Thus, in the remaining $t-r$ columns there are less then $t-r$ zeros and there is a column without zeros. We add this column to the sequence. To implement this algorithm in time $O(t\log^{4}t)$, we store, for each column $j$ of $A$, a sorted array of rows $i$ such that $A_{ij}=0$. Since the total number of zeros $z$ is at most $t\log t$, these arrays can be computed in time $O(t\log^2t)$: if $c_1, \dotsc, c_t$ are the numbers of zeros in the columns, then sorting the corresponding arrays takes time $$\sum_{i=1}^{t}c_i \log c_i \le \log(t \log t) \cdot \sum_{i=1}^{t}c_i \le \log(t \log t) \cdot t\log t \, .$$ At every iteration, we need to update the set $R$. For this, we need to remove some rows from it (from the column that no longer belongs to the stripe of columns of width $\log t$) and to add the rows of the newly added column. Since the size of $|R|$ is always at most $t$ and the total number of zeros is $z \le t\log t$, the total running time for all such updates is $O(t\log^2t)$ (if one uses, e.g., a balanced binary search tree for representing $R$). If $|R| > \log t$, one just takes an all-one column (all such columns can be stored in a list). If $|R| \le \log t$, we need to find a column outside of the set $F$. For this, we just scan the list of the yet unpicked columns. For each column, we first check whether it belongs to the set $F$. This can be checked in time $O(\log^2t)$: for every row in $|R|$, one checks whether this row belongs to the sorted array of the considered column using binary search in time $O(\log t)$. Since $|F| \le \log^2t$, we will find a column outside of $F$ in time $O(\log^4 t)$. One deterministically generates a circuit for $A$ of size $O(n)$ in time $O(n\log^4n)=O(n^2)$ by Theorem \[thm:upperbound\]. This circuit can be used to multiply $A$ by any column of $B$ in time $O(n)$. For this, one constructs a topological ordering of the gates of the circuits and computes the values of all gates in this order. Hence, $AB$ can be computed in time $O(n^2)$. Non-commutative Case {#sec-non-commutative} ==================== In the previous section, we have shown that for commutative semigroups dense linear operators can be computed by linear size circuits. A closer look at the circuit constructions reveals that we use commutativity crucially: it is important that we may reorder the columns of the matrix. In this section, we show that this trick is unavoidable: for non-commutative semigroups, it is not possible to construct linear size circuits for dense linear operators. Namely, we prove Theorem \[thm:lowerbound\]. Faithful semigroups ------------------- We consider computations over general semigroups that are not necessarily commutative. In particular, we will establish lower bounds for a large class of semigroups and our lower bound does not hold for commutative semigroups. This requires a formal definition that captures semigroups with rich enough structure and in particular requires that a semigroup is substantially non-commutative. Previously lower bounds in the circuit model for a large class of semigroups were known for the Range Queries problem [@DBLP:conf/stoc/Yao82; @DBLP:journals/ijcga/ChazelleR91]. These result are proven for a large class of commutative semigroups that are called *faithful* (we provide a formal definition below). Since we are dealing with non-commutative case we need to generalize the notion of faithfulness to non-commutative semigroups. To provide formal definition of faithfulness it is convenient to introduce the following notation. Suppose $(S, \circ)$ is a semigroup. Let $X_{S,n}$ be a semigroup with generators $\{x_1,\ldots, x_n\}$ and with the equivalence relation consisting of identities in variables $\{x_1,\ldots, x_n\}$ over $(S,\circ)$. That is, for two words $W$ and $W'$ in the alphabet $\{x_1,\ldots,x_n\}$ we have $W\sim W'$ in $X_{S,n}$ iff no matter which elements of the semigroup $S$ we substitute for $\{x_1,\ldots, x_n\}$ we obtain a correct equation over $S$. In particular, note that if $S$ is commutative (respectively, idempotent), then $X_{S,n}$ is also commutative (respectively, idempotent). The semigroup $X_{S,n}$ is studied in algebra under the name of relatively free semigroup of rank $n$ of a variety generated by semigroup $S$ [@neumann2012varieties]. We will often omit the subscript $n$ and write simply $X_S$ since the number of generators will be clear from the context. Below we will use the following notation. Let $W$ be a word in the alphabet $\{x_1,\ldots, x_n\}$. Denote by ${\texttt{Var}}(W)$ the set of letters that are present in $W$. We are now ready to introduce the definition of a commutative faithful semigroup. A commutative semigroup $(S, \circ)$ is *faithful commutative* if for any equivalence $W\sim W'$ in $X_S$ we have ${\texttt{Var}}(W)={\texttt{Var}}(W')$. Note that this definition does not pose any restrictions on the cardinality of each letter in $W$ and $W'$. This allows to capture in this definition important cases of idempotent semigroups. For example, semigroups $(\{0,1\}, \vee)$ and $(\mathbb{Z},\min)$ are commutative faithful. We need to study the non-commutative case, and moreover, our results establish the difference between commutative and non-commutative cases. Thus, we need to extend the notion of faithfulness to non-commutative semigroups to capture their non-commutativity in the whole power. At the same time we would like to keep the case of idempotency. We introduce the notion of faithfulness for the non-commutative case inspired by the properties of free idempotent semigroups [@GreenR52]. To introduce this notion we need several definitions. The *initial mark* of $W$ is the letter that is present in $W$ such that its first appearance is farthest to the right. Let $U$ be the prefix of $W$ consisting of letters preceding the initial mark. That is, $U$ is the maximal prefix of $W$ with a smaller number of generators. We call $U$ the *initial* of $W$. Analogously we define the *terminal mark* of $W$ and the *terminal* of $W$. \[def:strong\_non\_commutativity\] We say that a semigroup $X$ with generators $\{x_1,\ldots, x_n\}$ is *strongly non-commutative* if for any words $W$ and $W'$ in the alphabet $\{x_1,\ldots, x_n\}$ the equivalence $W\sim W'$ holds in $X$ only if the initial marks of $W$ and $W'$ are the same, terminal marks are the same, the equivalence $U \sim U'$ holds in $X$, where $U$ and $U'$ are the initials of $W$ and $W'$, respectively, and the equivalence $V \sim V'$ holds in $X$, where $V$ and $V'$ are the terminals of $W$ and $W'$, respectively. In other words, this definition states that the first and the last occurrences of generators in the equivalence separates the parts of the equivalence that cannot be affected by the rest of the generators and must therefore be equivalent themselves. We also note that this definition exactly captures the idempotent case: for a free idempotent semigroup the condition in this definition is “if and only if”[@GreenR52]. \[def:faithful\] A semigroup $(S, \circ)$ is *faithful non-commutative* if $X_S$ is strongly non-commutative. We note that this notion of faithfulness is relatively general and is true for semigroups $(S,\circ)$ with considerable degree of non-commutativity in their structure. It clearly captures free semigroups with at least two generators. It is also easy to see that the requirements in Definition \[def:faithful\] are satisfied for the free idempotent semigroup with $n$ generators (if $S$ is idempotent, then $X_{S,n}$ is also clearly idempotent and no other relations are holding in $X_{S,n}$ since we can substitute generators of $S$ for $x_1, \ldots, x_n$). Next we observe some properties of strongly non-commutative semigroups that we need in our constructions. \[lem:prefix\_equivalence\] Suppose $X$ is strongly non-commutative. Suppose the equivalence $W \sim W'$ holds in $X$ and $|{\texttt{Var}}(W)|=|{\texttt{Var}}(W')|=k$. Suppose $U$ and $U'$ are minimal (maximal) prefixes of $W$ and $W'$ such that $|{\texttt{Var}}(U)| = |{\texttt{Var}}(U')| = l\leq k$. Then the equivalence $U \sim U'$ holds in $X$. The same is true for suffixes. The proof is by induction on the decreasing $l$. Consider the maximal prefixes first. For $l=k$ and maximal prefixes we just have $U=W$ and $U'=W'$. Suppose the statement is true for some $l$, and denote the corresponding prefixes by $U$ and $U'$, respectively. Then note that the maximal prefixes with $l-1$ variables are initials of $U$ and $U'$. And the statement follows by Definition \[def:strong\_non\_commutativity\]. The proof of the statement for minimal prefixes is completely analogous. Note that on the step of induction the prefixes differ from the previous case by one letter that are initial marks of the corresponding prefixes. So these additional letters are also equal by the Definition \[def:strong\_non\_commutativity\]. The case of suffixes is completely analogous. The next lemma is a simple corollary of Lemma \[lem:prefix\_equivalence\]. \[lem:variables\_order\] Suppose $X$ is strongly non-commutative. Suppose $W \sim W'$ holds in $X$. Let us write down the letters of $W$ in the order in which they appear first time in $W$ when we read it from left to right. Let’s do the same for $W'$. Then we obtain exactly the same sequences of letters. The same is true if we read the words from right to left. Proof Strategy -------------- We now proceed to the proof of Theorem \[thm:lowerbound\]. The upper bound follows easily by a naive algorithm: split all rows of $A$ into ranges, compute all ranges by a circuit of size $O(n\alpha(n))$ using Yao’s construction [@DBLP:conf/stoc/Yao82], then combine ranges into rows of $A$ using $O(n)$ gates. Thus we will concentrate on lower bounds. We will view the computation of the circuit as a computation in a strongly non-commutative semigroup $X=X_S$. We will use the following proof strategy. First we observe that it is enough to prove the lower bound for the case of idempotent strongly non-commutative semigroups $X$. Indeed, if $X$ is not idempotent, we can factorize it by idempotency relations and obtain a strongly non-commutative idempotent semigroup $X_{id}$. A lower bound for the case of $X_{id}$ implies lower bound for the case of $X$. We provide a detailed explanation in Section \[sec:noncommutative\_extension\]. Hence, [ from this point we can assume that $X$ is idempotent and strongly non-commutative]{}. Next for idempotent case we show that our problem is equivalent to the commutative version of the range query problem. For a semigroup $X$ with generators $\{x_1,\ldots, x_n\}$ denote by $X_{sym}$ its factorization under commutativity relations $x_i x_j \sim x_j x_i$ for all $i,j$. Note that if $X$ is idempotent and strongly non-commutative, then $X_{sym}$ is just the semigroup in which $W \sim W'$ iff ${\texttt{Var}}(W)={\texttt{Var}}(W')$ (this is free idempotent commutative semigroup). \[thm:equivalence\] For an idempotent strongly non-commutative $X$ and for any $s=\Omega(n)$ we have that (commutative) range queries problem over $X_{sym}$ has size $O(s)$ circuits iff (non-commutative) dense linear operator problem over $X$ has size $O(s)$ circuits. Using this theorem, it is straightforward to finish the proof of Theorem \[thm:lowerbound\]. Indeed, by Theorem \[thm:equivalence\] if non-commutative dense linear operator problem has size $s$ circuit, then the commutative range queries problem also does. However, for the latter problem it is proved by Chazelle and Rosenberg [@DBLP:journals/ijcga/ChazelleR91] that $s=\Omega(n \alpha(n))$. Moreover, in our construction for the proof of Theorem \[thm:equivalence\] it is enough to consider dense linear operators with exactly two zeroes in every row. From this the second part of Theroem \[thm:lowerbound\] follows. Note that for the proof of Theorem \[thm:lowerbound\] only one direction of Theorem \[thm:equivalence\] is needed. However, we think that the equivalence in Theorem \[thm:equivalence\] might be of independent interest, so we provide the proof for both directions. Thus, it remains to prove Theorem \[thm:equivalence\]. We do this by showing the following equivalences for any $s = \Omega(n)$. =\[rectangle,draw,inner sep=1mm,text width=36mm,above right,minimum height=20mm\] \(a) at (0,0) [(commutative) range queries problem over $X_{sym}$ has $O(s)$ size circuits]{}; \(b) at (6.5,0) [(non-commutative) range queries problem over $X$ has $O(s)$ size circuits]{}; \(c) at (13,0) [(non-commutative) dense linear operator problem over $X$ has $O(s)$ size circuits]{}; (a.10) edge\[-&gt;\] node\[above\] [Lemma \[lem:intervals\]]{} (b.170); (b.190) edge\[-&gt;\] node\[below\] [special case]{} (a.-10); (b.10) edge\[-&gt;\] node\[above\] [straightforward]{} (c.170); (c.190) edge\[-&gt;\] node\[below\] [Lemma \[lem:dense\_matrices\]]{} (b.-10); Note that two of the reductions on this diagram are trivial. The other two are formulated in the following lemmas. \[lem:dense\_matrices\] If the (non-commutative) dense linear operator problem over $X$ has size $s$ circuit then the (non-commutative) range queries problem over $X$ has size $O(s)$ circuit. \[lem:intervals\] If the (commutative) version of the range queries problem over $X_{sym}$ has size $s$ circuits then the (non-commutative) version over $X$ also does. The proofs of these lemmas are presented in Sections \[sec:operators\_to\_queries\] and \[sec:non-commutative\_to\_commutative\] respectively. From Idempotent Semigroups to General Semigroups {#sec:noncommutative_extension} ------------------------------------------------ In this section we provide a detailed explanation of the reduction in Theorem \[thm:lowerbound\] from general semigroups to idempotent semigroups. Consider an arbitrary strongly non-commutative semigroup $X$. Consider a new semigroup $X_{id}$ over the same set of generators that is a factorization of $X$ by idempotency relations $W^2\sim W$ for all words $W$ in the alphabet $\{x_1,\ldots, x_n\}$. \[lem:idempotisation\] If $X$ is strongly non-commutative, then $X_{id}$ is also strongly non-commutative. Suppose $W$ and $W'$ are words in the alphabet $\{x_1,\ldots, x_n\}$ and $W \sim W'$ in $X_{id}$. This means that there is a sequence $W_0,\ldots, W_k$ of words in the same alphabet such that $W=W_0$, $W'=W_k$ and for each $i$ either $W_i \sim W_{i+1}$ in $X$, or $W_{i+1}$ is obtained from $W_i$ by one application of the idempotency equivalence to some subword of $W_i$. Clearly, it is enough to check that the conditions of Definition \[def:strong\_non\_commutativity\] are satisfied in $X_{id}$ for each consecutive pair $W_i$ and $W_{i+1}$. If $W_i \sim W_{i+1}$ in $X$, then the conditions of Definition \[def:strong\_non\_commutativity\] follows from the strong non-commutativity of $X$. Suppose now that $W_{i+1}$ is obtained from $W_{i}$ by substituting some subword $A$ by $A^2$ (the symmetrical case is analyzed in the same way). We will show that initial marks of $W_i$ and $W_{i+1}$ are the same and $U_{i} \sim U_{i+1}$ in $X_{id}$, where $U_{i}$ and $U_{i+1}$ are initials of $W_i$ and $W_{i+1}$ respectively. For the terminals and terminal marks the proof is completely analogous. Suppose $A$ lies to the left of initial mark in $W_i$ and we substitute $A$ by $A^2$. Then the initial mark is unaltered and in the initial $U_i$ we also substitute $A$ by $A^2$. Thus in this case $U_{i+1}$ is obtained from $U_i$ by idempotency relation. Suppose $A$ contains initial mark of $W_i$ or lies to the right of it. Then after the substitution of $A$ by $A^2$ the initial mark is still the same and the initial $U_i$ also does not change. Now we outline the reduction of the lower bound in Theorem \[thm:lowerbound\] from idempotent semigroup to the general case. Suppose $X$ is strongly non-commutative and suppose that for $X$ all dense operators can be computed by circuits of size at most $s$. Consider a semigroup $X_{id}$ as introduced above. By Lemma \[lem:idempotisation\] $X_{id}$ is also strongly non-commutative. On the other hand, since $X_{id}$ is a factorization of $X$ any circuit computing dense operator over $X$ also computes the same dense operator over $X_{id}$. Thus, by our assumption there are circuits of size at most $s$ for all dense operators over $X_{id}$. Finally, $X_{id}$ is idempotent, so by the special case of our theorem we have $s = \Omega(n \alpha(n))$ and we are done. Reducing Dense Linear Operator to Range Queries {#sec:operators_to_queries} ----------------------------------------------- In this subsection, we prove Lemma \[lem:dense\_matrices\]. Intuitively, the lemma holds as the best way to compute rows of a dense matrix is to combine input variables in the natural order. This intuition is formalized in Lemma \[lemma:correctorder\] below. Given this, it is easy to reduce dense linear operator problem to the range queries problem: we just “pack” each range query into a separate row, i.e., for a query $(l,r)$ we introduce a $0/1$-row having two zeroes in positions $l-1$ and $r+1$ (hence, this row consists of three ranges: $(1,l-1)$, $(l,r)$, $(r+1,n)$). Then, if a circuit computing the corresponding linear operator has a nice property of always using the natural order of variables (guaranteed by Lemma \[lemma:correctorder\]), one may extract the answer to the query $(l,r)$ from it. It should be mentioned, at the same time, that the semigroup $X$ might be complicated. In particular, the idempotency is tricky. For example, it can be used to simulate commutativity: one can turn $xy$ into $yx$, by first multiplying $xy$ by $y$ from the left and then multiplying the result by $x$ from the right (obtaining $(y(xy))x=(yx)(yx)=yx$). Using similar ideas, one can place new variables inside of already computed products. To get $xyz$ from $xz$, one multiplies it by $xyz$ first from the left and then from the right: $(xyz)xz(xyz)=xy(zxzx)yz=xy(zx)yz=xyz$. This is not extremely impressive, since to get $xyz$ we multiply by $xyz$, but the point is that this is possible in principle. We proceed to the formal proofs. Let’s call the word $W$ in the alphabet $\{x_1,\ldots,x_n\}$ *increasing* if it is a product of variables in the increasing order. A binary circuit is called an [*increasing circuit*]{} if each of its gates computes a word equivalent in $X$ to increasing word. Note that if a gate in an increasing circuit is fed by two gates $G$ and $H$, then the increasing words computed by $G$ and $H$ are matching in a sense that some suffix of $G$ (possibly an empty suffix) is equal to some prefix of $H$. Otherwise, the result is not equal to a product of variables in the increasing order, due to Lemma \[lem:variables\_order\]. Analogously, a binary circuit is called a [*range circuit*]{} if each of its gates computes a word that is equivalent to a range. The proof of Lemma \[lem:dense\_matrices\] follows from the following two lemmas. \[lemma:correctorder\] Given a binary circuit computing $Ax$, one may transform it into an increasing circuit of the same size computing the same function. \[lemma:matrixranges\] Given an increasing circuit computing $Ax$, one may transform it into a range circuit of the same size computing all ranges of $A$. Given $n$ ranges, pack them into a matrix $A \in \{0,1\}^{n \times n}$ with at most $2n$ zeroes. Take a size-$s$ circuit computing $Ax$ and convert it into a binary circuit. Then, transform it into an increasing circuit using Lemma \[lemma:correctorder\]. Finally, extract the answers to all the ranges from this circuit using Lemma \[lemma:matrixranges\]. Note that the second statement of Theorem \[thm:lowerbound\] follows since the proof of Lemma \[lem:dense\_matrices\] deals with matrices with exactly two zeroes in every row. Take an increasing circuit ${\cal C}$ computing $Ax$ and process all its gates in some topological ordering. Each gate $G$ of ${\cal C}$ computes a (word that is equivalent to an) increasing word. We split this increasing word into ranges and we put into correspondence to $G$ an ordered sequence $G_1,\ldots, G_k$ of gates of the new circuit. Each of this gates compute one of the ranges of the word computed by $G$ and $G \sim G_1\circ\ldots \circ G_k$. Consider a gate $G$ of ${\cal C}$ and suppose we have already computed all gates of the new circuit corresponding to previous gates of ${\cal C}$. $G$ is the product $F \circ H$ of previous gates of ${\cal C}$, for which new range gates are already computed. Since ${\cal C}$ is increasing we have that $F$ and $H$ are matching, that is some suffix (maybe empty) of the increasing word computed in $F$ is equal to some prefix (maybe empty) of the increasing word computed in $H$ and there are no other common variables in these increasing words. It is easy to see that ranges for the sequence corresponding to $G$ are just the ranges for the sequences for $F$ and $H$ with possibly two of them united. If needed, we compute the product of gates of the new circuit corresponding to the united ranges and the sequence of new gates for $G$ is ready. Thus, to process each gate of ${\cal C}$ we need at most one operation in the new circuit and the size of the new circuit is at most the size of ${\cal C}$. For output gates of ${\cal C}$ we have gates in the new circuit that compute exactly ranges of output gates. Thus, in the new circuit all ranges of $A$ are computed. Consider a binary circuit ${\cal C}$ computing $Ax$ and its gate $G$ together with a variable $x_i$ it depends on. We say that $x_i$ is *good* in $G$ if there is a path in ${\cal C}$ from $G$ to an output gate, on which the word is never multiplied from the left by words containing variables greater than or equal to $x_i$. Note that if $x_i$ and $x_{i'}$ are both contained in $G$, $i<i'$, and $x_i$ is good in $G$, then $x_{i'}$ is good in $G$, too. That is, the set of all good variables in $G$ is closed upwards. Consider the largest good variable in $G$ (if there is one), denote it by $x_k$ ($x_k$ is actually just the largest variable in $G$, unless of course there are no good variables in $G$). Let us focus on the first occurrence of $x_k$ in $G$. All first occurrences of other good variables in $G$ must be to the left of the first occurrence of $x_k$. Suppose that a good variable $x_i$ has the first occurrence to the right of (the first occurrence of) $x_k$. Consider an output gate $H$ such that there is a path from $G$ to $H$ and along this path there are no multiplications of $G$ from the left by words containing variables greater than $x_i$. Then we have $H \sim LGR$, where all variables of $L$ are smaller then $x_i$. Then in $H$ the variable $x_i$ appears before $x_k$ when we read from left to right, but at the same time we have that $x_k$ appears before $x_i$ in $LGR$. This contradicts Lemma \[lem:variables\_order\]. Now, for a gate $G$, define two words ${\texttt{MIN}}_G$ and ${\texttt{MAX}}_G$. Both these words are products of variables in the increasing order: ${\texttt{MIN}}_G$ is the product of good variables of $G$ in the increasing order, ${\texttt{MAX}}_G$ is the product (in the increasing order) of all variables that has first occurrences before (the first occurrence of) $x_k$. Note that ${\texttt{MIN}}_G$ is a suffix of ${\texttt{MAX}}_G$. If there are no good variables in $G$ we just let ${\texttt{MIN}}_g={\texttt{MAX}}_g=\lambda$ (the empty word). For the word $W$ that has the form of the product of variables in the increasing order, we call $x_j$ a *gap variable* if it is not contained in $W$ while $W$ contains variables $x_i$ and $x_k$ with $i < j < k$. Below we show how for a given circuit ${\cal C}$ to construct an increasing circuit ${\cal C}'$ that for each gate $G$ of ${\cal C}$ computes some intermediate product $P_G$ between ${\texttt{MIN}}_G$ and ${\texttt{MAX}}_G$: ${\texttt{MIN}}_g$ is a suffix of $P_G$ and $P_G$ is a suffix of ${\texttt{MAX}}_g$. The size of ${\cal C}'$ is at most the size of ${\cal C}$. For an output gate $G$, ${\texttt{MIN}}_g={\texttt{MAX}}_g=g$ hence the circuit ${\cal C}'$ computes the correct outputs. To construct ${\cal C}'$, we process the gates of ${\cal C}$ in a topological ordering. If $G$ is an input gate, everything is straightforward: in this case ${\texttt{MAX}}_G={\texttt{MIN}}_G$ is either $\lambda$ or $x_j$. Assume now that $G$ is an internal gate with predecessors $F$ and $H$. Consider the set of good variables in $G$. If there are none, we let $P_G=\lambda$. If all first occurrences of good variables of $G$ are lying in one of the predecessors ($F$ and $H$), then they are good in the corresponding input gate. We then set $P_G$ to $P_F$ or $P_H$. The only remaining case is that some good variables have their first occurrence in $F$ while some others have their first occurrence in $H$. Then the largest variable $x_k$ of $G$ has the first occurrence in $H$ and all variables of $F$ are smaller than $x_k$. \[cl: h is good\] There are no gap variables for ${\texttt{MAX}}_H$ in $F$. Suppose that some variable $x_i$ in $F$ is a gap variable for ${\texttt{MAX}}_H$. Consider an output $C$ such that there is a path from $G$ to $C$ and along this path there are no multiplications of $G$ from the left by words containing variables greater than $x_k$. Then we have $C \sim LGR$ where all variables of $L$ are smaller then $x_k$. Consider the prefix $P$ of $C$ preceding the variable $x_k$ and the prefix $Q$ of $LG$ preceding the variable $x_k$. Then by Lemma \[lem:prefix\_equivalence\] we have $P \sim Q$. But then the variables of $P$ and $Q$ appear in the same order if we read the words from right to left. But this is not true (the variable in $P$ are in the decreasing order and in $Q$ the variable $x_i$ is not on its place), a contradiction. \[cl: f is good\] There are no gap variables for ${\texttt{MAX}}_F$ in $H$. Suppose that a variable $x_i$ in $H$ is a gap variable for ${\texttt{MAX}}_F$. Consider an output $C$ such that there is a path from $G$ to $C$ and along this path there are no multiplications of $G$ from the left by words containing variables greater than $x_l$, the largest variable of $F$. Then we have $C \sim LGR$, where all variables of $L$ are smaller then $x_l$. Consider the prefix $P$ of $C$ preceding $x_l$ and the prefix $Q$ of $LG$ preceding $x_l$. Then by Lemma \[lem:prefix\_equivalence\] we have $P \sim Q$. But then the variables of $P$ and $Q$ appear in the same order if we read the words from right to left. But this is not true (the variables in $P$ are in the decreasing order and in $Q$ the variable $x_i$ is not on its place), a contradiction. We are now ready to complete the proof of Lemma \[lemma:correctorder\]. Consider $P_F$ and $P_H$. By Claims \[cl: h is good\] and \[cl: f is good\], we know that they are ranges in the same sequence of variables ${\texttt{Var}}(P_F)\cup {\texttt{Var}}(P_H)$. We know that the largest variables of $P_H$ is greater than all variables of $P_f$. Then either $P_F$ is contained in $P_H$, and then we can let $P_G=P_H$ (it contains all good variables of $G$), or we have $P_F =PQ$ and $P_H=QR$ for some words $P, Q, R$. In this case we let $P_G = P_F \circ P_H = PQQR=PQR$. Clearly, ${\texttt{MIN}}_G$ is the suffix of $P_G$ and $P_G$ itself is the suffix of ${\texttt{MAX}}_G$. Reducing Non-commutative Range Queries to Commutative Range Queries {#sec:non-commutative_to_commutative} ------------------------------------------------------------------- In this subsection we prove Lemma \[lem:intervals\]. We will show that any computation of ranges over $X_{sym}$ can be reconstructed without increase in the number of gates in such a way that each gate computes a range (recall, that we call this a range circuit). It is easy to see that then this circuit can be reconstructed as a circuit over $X$ each gate of which computes the same range with the variables in the increasing order. Indeed, we need to make sure that each gate computes a range in such a way that all variables are in the increasing order and this is easy to do by induction. Each gate computes a product of two ranges $a$ and $b$. If one of them is contained in the other, we simplify the circuit, since the gate just computes the same range as one of its inputs (due to idempotency and commutativity). It is impossible that $a$ and $b$ are non-intersecting and have a gap between them, since then our gate does not compute a range (in a range circuit). So, if $a$ and $b$ are non-intersecting, then they are consecutive and we just need to multiply them in the right order. If the ranges are intersecting, we just multiply then in the right order and apply idempotency. Thus it remains to show that each circuit for range query problem over $X_{sym}$ can be reconstructed into a range circuit. For this we will need some notation. Suppose we have some circuit ${\cal C}$. For each gate $G$ denote by ${\texttt{left}}(G)$ the smallest index of the variable in $G$ (the leftmost variable). Analogously denote by ${\texttt{right}}(G)$ the largest index of the variable in $G$. Denote by ${\texttt{gap}}(G)$ the smallest $i$ such that $x_i$ is not in $G$, but there are some $j,k$ such that $j<i<k$ and $x_j$ and $x_k$ (the smallest index of the variable that is in the gap in $G$). Next, fix some topological ordering of gates in ${\cal C}$ (the ordering should be proper, that is inputs to any gate should have smaller numbers). Denote by ${\texttt{num}}(G)$ the number of a gate in this ordering. Finally, by ${\texttt{out}}(G)$ denote the out-degree of $G$. For each gate that computes a non-range consider the tuple $${\texttt{tup}}(G)=({\texttt{left}}(G),{\texttt{gap}}(G),{\texttt{num}}(G),-{\texttt{out}}(G)).$$ For the circuit ${\cal C}$ consider ${\texttt{tup}}({\cal C}) = \min_G {\texttt{tup}}(G)$, where the minimum is considered in the lexicographic order and is taken over all non-range gates. If there are no non-range gates we let ${\texttt{tup}}({\cal C})=\infty$. This is our semi-invariant, we will show that if we have a circuits that is not a range circuit, we can reconstruct it to increase its ${\texttt{tup}}$ (in the lexicographic order) without increasing its size. Since ${\texttt{tup}}$ ranges over a finite set, we can reconstruct the circuit repeatedly and end up with a range circuit. Now we are ready to describe a reconstruction of a circuit. Consider a circuit ${\cal C}$ that is not a range circuit. And consider a gate $G$ such that ${\texttt{tup}}(G)={\texttt{tup}}({\cal C})$ (it is clearly unique). Denote by $A$ and $B$ two inputs of $G$. Let $i={\texttt{left}}(G)$ and $j={\texttt{gap}}(G)$, that is $x_i$ is the variable with the smallest index in $G$ and $x_j$ is the first gap variable of $G$ (it is not contained in $G$). The variable $x_i$ is contained in at least one of $A$ and $B$. Consider the gate among $A$ and $B$ that contains $x_i$. This gate cannot have $x_j$ or earlier variable as a gap variable: it would contradict minimality of $G$ (by the second or the third coordinate of ${\texttt{tup}}$). Thus this gate is a range $[x_i,x_{j'})$ for some $j'\leq j$ (by this we denote the product of variables from $x_i$ to $x_{j'}$ excluding $x_{j'}$). In particular, only one of $A$ and $B$ contains $x_i$: otherwise they are both ranges and $x_j$ is not a gap variable for $G$. From now on we assume that $A$ contains $x_i$, that is $A=[x_i,x_{j'})$. Now we consider all gates $H_1,\ldots, H_k$ that have edges leading from $G$. Denote by $F_1,\ldots, F_k$ their other inputs. If $k$ is equal to $0$, we can remove $G$ and reduce the circuit. Now we consider cases. ///in [0/4/f1/F\_1, 2/4/fl/F\_l, 4/4/fk/F\_k, 6/4/g/G, 1/2/h1/H\_1, 3/2/hl/H\_l, 5/2/hk/H\_k, 5/6/a/A, 7/6/b/B]{} () at (,) [$\t$]{}; ///in [1/4/fdots1/…, 3/4/fdots2/…, 2/2/hdots1/…, 4/2/hdots2/…]{} () at (,) [$\t$]{}; /in [f1/h1, fl/hl, fk/hk, a/g, b/g, g/h1, g/hl, g/hk]{} () – (); *Case 1.* Suppose that there is $l$ such that ${\texttt{left}}(F_l) \leq {\texttt{left}}(G)$. If ${\texttt{left}}(F_l) < {\texttt{left}}(G)$, then $F_l$ must contain all variables $x_i, \ldots, x_j$, since otherwise either $F_l$ or $H_l$ will have smaller ${\texttt{tup}}$ then $G$. Thus $F_l$ contains $A$. Then, we can restructure the circuit by feeding $B$ to $H_l$ instead of $G$. This does not change the value of the gate computed by $H_l$ and reduces ${\texttt{out}}(G)$. Thus ${\texttt{tup}}({\cal C})$ increases and we are done. If ${\texttt{left}}(F_l) = {\texttt{left}}(G)$, then $F_l$ still cannot have gap variables among $x_i, \ldots, x_{j-1}$ as it would contradict the minimality of $G$. Thus, $F_l$ is either a range, or it is not a range, but contain all variables $x_i, \ldots, x_{j-1}$. In the latter case again $F_l$ contains $A$. In the former case $F_l$ either contains $A$, or is contain in $G$. If $F_l$ contains $A$, we can again simplify the circuit as above. If $F_l$ is contained in $G$, we have $G=H_l$, so we can remove $H_l$ from the circuit and reduce the size of the circuit. *Case 2.* Suppose that for all $l$ we have ${\texttt{left}}(F_l)>{\texttt{left}}(G)$. Consider $l$ such that $F_l$ has the minimal ${\texttt{right}}(F_l)$ (if there are several such $l$ pick among them the one with the minimal ${\texttt{num}}(F_l)$). For convenience of notation let $l=k$. Now we restructure the circuit in the following way. We feed $F_k$ to $G$ instead of $A$. We feed $A$ to $H_k$ instead of $F_k$. We feed $H_k$ to all other $H_p$’s instead of $G$. ///in [-2/4/f1/F\_1, 2/4/fk1/F\_[k-1]{}, 3/6/fk/F\_k, 6/4/g/G, 1/1/h1/H\_1, 5/1/hk1/H\_[k-1]{}, 4/3.5/hk/H\_k, 5/6/a/A, 7/6/b/B]{} () at (,) [$\t$]{}; ///in [0/4/fdots1/…, 3/1/hdots1/…]{} () at (,) [$\t$]{}; /in [f1/h1, fk1/hk1, b/g, g/hk]{} () – (); /in [fk/hk, a/g, g/h1, g/hk1]{} () – (); /in [fk/g, a/hk, hk/h1, hk/hk1]{} () – (); Observe that all these reconstructions are valid, that is, they do not create directed cycles in the circuit. To verify this we need to check that there are no cycles using new edges. Indeed, there cannot be a cycle going through one of the edges $(H_k,H_p)$ since this would mean that there was a directed path from $H_p$ to one of the vertices $F_k$, $A$ and $G$ on the original circuit. Such a path to $A$ or $G$ would mean a cycle in the original circuit. Such a path to $F_k$ violates the minimality property of $F_k$ (minimal ${\texttt{right}}(F_k)$). Next, there cannot be a cycle going through both edges $(F_k,G)$ and $(A,H_k)$, since substituting these edges by $(F_k,H_k)$ and $(A,G)$ we obtain one or two cycles in the original circuit. Next, there cannot be a cycle going through the edge $(A,H_k)$ only, since $H_k$ is reachable from $A$ in the original circuit and this would mean a cycle in the original circuit. Finally, there cannot be a cycle going only through the edge $(F_k,G)$ since this would mean a directed path from $G$ to $F_k$ in the original circuit and this contradicts ${\texttt{left}}(F_k)>{\texttt{left}}(G)$. Note that our reconstruction might require reordering of the circuit gates, since we create edges between previously incomparable $H$-gates and between $F_k$ and $G$. But the reordering affect only the gates with ${\texttt{num}}$ greater than ${\texttt{num}}(G)$ and may only reduce ${\texttt{num}}(F_k)$ to be smaller than ${\texttt{num}}(G)$. But this can only increase ${\texttt{tup}}(G)$ and since ${\texttt{left}}(F_k)>{\texttt{left}}(G)$ this can only increase ${\texttt{tup}}({\cal C})$. Observe, that the circuit still computes the outputs correctly. The changes are in the gates $H_1\ldots, H_k$ (and also in $G$, but $H_1,\ldots, H_k$ are all of its outputs). $H_k$ does not change. Other $H_p$’s might have changed, they now additionally include variables of $F_k$. But note that all of these variables are in between of ${\texttt{left}}(H_p)$ and ${\texttt{right}}(H_p)$, so they must be presented in the output gates connected to $H_p$ anyway (recall that at the output gates we compute ranges). Now, observe that ${\texttt{tup}}(G)$ has increased (by the first coordinate). There are no new gates with smaller ${\texttt{left}}$. Among gates with the minimal ${\texttt{left}}$ there are no new gates with smaller ${\texttt{gap}}$. Among gates with minimal $({\texttt{left}},{\texttt{gap}})$ all gates have larger ${\texttt{num}}$ then $G$. Thus ${\texttt{tup}}({\cal C})$ increased and we are done. Open Problems ============= There are two natural problems left open. 1. Design a deterministic $O(z)$ time algorithm for generating a circuit in the commutative case. For this, it suffices to design an $O(n)$ deterministic algorithm for the following problem: given a list of positions of $n$ zeroes of an $n \times n$ 0/1-matrix with at most $\log n$ zeroes in every row, permute its columns so that the total length of all segments of length at most $O(\log n)$ is $O(\frac{n}{\log n})$. 2. Determine the asymptotic complexity of the linear operator in terms of the number of zeroes in the non-commutative case. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Paweł Gawrychowski for pointing us out to the paper [@DBLP:journals/ijcga/ChazelleR91]. We thank Alexey Talambutsa for fruitful discussions on the theory of semigroups. Review ====== algebraic\_structures range\_queries\_applications approaches dense\_graph\_repr [^1]: Steklov Mathematical Institute at St. Petersburg, Russian Academy of Sciences, email: <[email protected]> [^2]: Department of Computer Science and Engineering, University of California, San Diego, email: <[email protected]> [^3]: School of Engineering, Newcastle University, UK, email: <[email protected]> [^4]: Steklov Mathematical Institute, Russian Academy of Sciences and National Research University Higher School of Economics, Moscow, email: <[email protected]> [^5]: Note that the result of summation is undefined in case of an all-zero row, because semigroups have no neutral element in general. One can trivially sidestep this technical issue by adding an all-one column $n+1$ to the matrix $A$, as well as the neutral element $x_{n+1}$ into the vector. Alternatively, we could switch from semigroups to *monoids*, but we choose not to do that, since we have no use for the neutral element and associated laws in the rest of the paper.

ArXiv

--- abstract: 'It is well-known that the pancake graphs are widely used as models for interconnection networks [@Akers]. In this paper, some properties of the pancake graphs are investigated. We first prove that the pancake graph, denoted by $P_n~(n\geq 4),$ is super-connected and hyper-connected. Further, we study the symmetry of $P_n$ and completely determine its full automorphism group, which shows that $P_n~(n\geq 5)$ is a graphical regular representation of $S_n.$' author: - | Yun-Ping Deng, Xiao-Dong Zhang$^{\dagger}$\ [Department of Mathematics, Shanghai Jiao Tong University]{}\ [800 Dongchuan road, Shanghai, 200240, P.R. China]{}\ [Emails: [email protected], [email protected]]{}\ title: ' Automorphism Groups of the Pancake Graphs [^1]' --- \[section\] \[theorem\][Corollary]{} \[theorem\][Definition]{} \[theorem\][Conjecture]{} \[theorem\][Question]{} \[theorem\][Lemma]{} \[theorem\][Proposition]{} \[theorem\][Example]{} \[theorem\][Problem]{} c Ø ø $H$  ¶[$P$ ]{} ł [[**Key words:**]{} Interconnection networks; pancake graph; super-connected; hyper-connected; efficient dominating sets; automorphism group. ]{}\ [[**AMS Classifications:**]{} 05C25, 05C69]{} 0.5cm Introduction ============ For a simple graph $\T,$ we denote its vertex set, edge set and full automorphism group respectively by $V(\T), E(\T)$ and $\Aut(\T)$. $\T$ is said to be [*vertex-transitive*]{} or [*edge-transitive*]{} if $\Aut(\T)$ acts transitively on $V(\T)$ or $E(\T),$ respectively. Let be a finite group and $S$ a subset of not containing the identity element $1$ with $S=S^{-1}.$ The [*Cayley graph*]{} $\T:=\Cay(G,S)$ on with respect to $S$ is defined by $$V(\T){=}G,~E(\T){=}\{(g,gs): g{\in} G,\ s{\in} S\}.$$ Clearly, $\T$ is a $|S|$-regular and vertex-transitive graph, since $\Aut(\T)$ contains the left regular representation $L(G)$ of . Moreover, $\T$ is connected if and only if is generated by $S.$ A [*permutation*]{} $\sigma$ on the set $X=\{1,2,\cdots,n\}$ is a bijective mapping from $X$ to $X.$ As usual, we denote by $S_n$ the group of all permutations on $X,$ which is called the [*symmetric group*]{}. The [*pancake graph*]{} $P_n,$ also called the [*prefix-reversal graph*]{} is the Cayley graph $\Cay(S_n,PR_n),$ where $PR_n= \{{r_{1j}:2\leq j\leq n}\}$ and $r_{1j}=(\begin{array}{ll} 1~~~~2~~\,\cdots~j~j+1~\cdots~n\\ j~j-1\cdots~1~j+1~\cdots~n \end{array}).$ The pancake graph is well-known because of the famous unsolved combinatorial problem about computing its diameter, which has been introduced by [@Dweighter], and has been studied in several papers[@Gates; @Hyedari1; @Hyedari2]. The pancake graph was often used as a model for interconnection networks of parallel computers [@Akers] due to its attractive properties regarding degree, diameter, symmetry, embeddings and self similarity. The pancake graph $P_n$ corresponds to the $n$-dimensional pancake network in computer science such that this network has processors labeled by permutations on $X$ and two processors are connected when the label of one is obtained from the other by some $r_{1j}, 2\leq j\leq n.$ The diameter of this network corresponds to the worst communication delay for transmitting information in a system. Morover, many researchers (see [@Kanevsky; @Lin; @Sheu]) have investigated some other properties of $P_n,$ such as the hamilton-connectedness, cycle-embedding problem, super-connectivity. A graph is said to be [*super-connected*]{} [@Boesch] if each minimum vertex cut is the neighbor set of a single vertex in . A graph is said to be [*hyper-connected*]{} [@Hamidoune] if for every minimum vertex cut $D$ of , $X-D$ has exactly two components, one of which is an isolated vertex. In [@Li], Li investigated the super-connectedness and hyper-connectedness of the reversal Cayley graph and pointed out that it is unknown for the pancake graph. Here we solve this problem and prove that the pancake graph $P_n~(n\geq 4)$ is super-connected and hyper-connected. An independent set $D$ of vertices in a graph is called an [*efficient dominating set*]{} [@Bange1; @Bange2] if each vertex not in $D$ is adjacent to exactly one vertex in $D.$ In [@Dejter], Dejter investigated the efficient dominating sets of Cayley graphs on the symmetric groups, which implied that there exists the efficient dominating sets in the pancake graph. In addtion, the efficient dominating sets are used in optimal broadcasting algorithms for multiple messages on the pancake graphs [@Qiu]. Motivated by these results, we completely characterize all the efficient dominating sets in $P_n~(n\geq 3).$ A graph $\T=(V,E)$ is a [*graphical regular representation (GRR)*]{}[@Nowitz] of the finite group $G$ if $\Aut(\T)=G$ and $\Aut(\T)$ acts regularly on $V.$ It is well-known that for the interconnection networks modeled by Cayley graphs, the symmetry is one of the problems focused by many researchers. In [@Lakshmivarahan], Lakshmivarahan investigated the symmetry of the pancake graph and showed that $P_n$ is not edge-transitive and hence not distance-transitive. In this paper, we further study the symmetry of $P_n$ and completely determine the automorphism group of $P_n,$ which shows that $P_n~(n\geq 5)$ is a graphical regular representation of $S_n$ and hence not edge-transitive and distance-transitive. The rest part of this paper is organized as follows. In Section 2, we first prove that the pancake graph $P_n~(n\geq 4)$ is super-connected and hyper-connected, then we show that there are exactly $n$ efficient dominating sets $B^{(i)}~(i=1,2,\cdots,n)$ in $P_n~(n\geq 3),$ where $B^{(i)}=\{\pi\in S_n:\,\pi(1)=i\}.$ In section 3, we prove that the full automorphism groups of $P_n~(n\geq5)$ is the left regular representation of $S_n,$ i.e. $\Aut(P_n)=L(S_n).$ Some properties of $P_n$ ======================== In table 1 of [@Li], it has been pointed out that the super-connectedness and hyper-connectedness of $P_n$ are unknown. In this section, we first prove that the pancake graph $P_n~(n\geq 4)$ is super-connected and hyper-connected. Following [@Li], we introduce some notations and terminologies. Let $X$ be a graph and $F$ a subset of $V(X).$ Set $N(F)=\{x\in V(X)\setminus F:\exists\,y\in F,\,s.t.\,xy\in E(X)\},\,C(F)=F\cup N(F),\,R(F)=V(X)\setminus C(F).$ A subset $F\subseteq V(X)$ is a [*fragment*]{} if $|N(F)|=\kappa(X)$ and $R(F)\neq\emptyset,$ where $\kappa(X)$ is the vertex-connectivity of $X.$ A fragment $F$ with $2\leq|F|\leq|V(X)|-\kappa(X)-2$ is called a [*strict fragment*]{}. A strict fragment with minimum cardinality is called a [*superatom*]{}. The following result is due to Mader [@Mader]: ${{\fs\cite{Mader}}}$ \[yl-2.1\] If $X$ is a connected undirected graph which is a vertex-transitive and $K_4$-free, then $\kappa(X)=\delta(X),$ where $\delta(X)$ denotes the minimum degree of $X.$ \[yl-2.2\] $\kappa(P_n)=\delta(P_n)=n-1$ for any $n\geq 3.$ By [@Sheu], we obtain that $g(P_n)=6,$ where $g(P_n)$ is the girth of $P_n.$ So $P_n$ is $K_4$-free, by Lemma \[yl-2.1\], the assertion holds. In the following Lemma, we shall state some facts without proof. Some of these facts may be found in [@Sheu], and others follow immediately from the definition of the pancake graph. \[yl-2.3\] Let $B^{(i)}=\{\pi\in S_n:\,\pi(1)=i\},\,B_{(j)}=\{\pi\in S_n:\,\pi(n)=j\},\,B_{(j)}^{(i)}=B^{(i)}\cap B_{(j)}.$ Then the following (i)-(iii) hold: \(i) For any $i\neq j,$ each vertex in $B^{(i)}$ is adjacent to exactly one vertex in $B^{(j)};$ \(ii) For any $i\neq j,$ each vertex in $B_{(j)}^{(i)}$ is adjacent to exactly one vertex in $B_{(i)}^{(j)}$ and exactly one vertex in $B_{(j)}^{(k)}$ for each $k\neq i,j.$ \(iii) The mapping $\varphi:\,S_{n-1}\rightarrow B_{(j)}$ defined as $\varphi(\pi)=(j,n)\pi$ is an isomorphism from $P_{n-1}$ to $P_n[B_{(j)}],$ where $P_n[B_{(j)}]$ is the subgraph of $P_n$ induced by $B_{(j)}.$ \[dl-2.4\] If $n\geq 4,$ then $P_n$ is super-connected. It is enough to show that $P_n$ contains no superatom. Suppose on the contrary that $A$ is a superatom of $P_n$ and consider the following possible cases: [**Case 1.**]{} $A\subseteq B_{(i)}$ for some $i\in\{1,2,\cdots,n\}.$ By Lemmas \[yl-2.2\] and \[yl-2.3\], we have $\kappa(P_{n}[B_{(i)}])=\kappa(P_{n-1})=n-2$ for $n-1\geq 3,$ so $|N(A)\cap B_{(i)}|\geq n-2$ for $n\geq 4.$ Hence $|N(A)|=|N(A)\cap B_{(i)}|+|N(A)\cap(\bigcup_{j\neq i} B_{(j)})|\geq (n-2)+|A|\geq(n-2)+2=n>n-1=\kappa(P_n),$ which is a contradiction. [**Case 2.**]{} $A\nsubseteq B_{(i)}$ for any $i\in\{1,2,\cdots,n\}.$ Then there exist $i,j~(i\neq j)$ such that $A\cap B_{(i)}\neq \emptyset$ and $A\cap B_{(j)}\neq \emptyset.$ Hence $|N(A)|\geq|N(A)\cap B_{(i)}|+|N(A)\cap B_{(j)}|\geq 2(n-2)\geq n-1=\kappa(P_n),$ which is a contradiction. [**Remark.**]{} If $n=3,$ then $P_3=C_6,$ clearly it is not super-connected. \[dl-2.5\] If $n\geq 4,$ then $P_n$ is hyper-connected. By Theorem \[dl-2.4\], $P_n$ is super-connected for $n\geq 4.$ Consider the vertex-transitivity of $P_n,$ it suffices to show that $P_n-N[I]$ is connected, where $N[I]$ is the closed neighbourhood of the identity element $I.$ We proceed by the induction on $n.$ If $n=4,$ one can easily check that $P_4-N[I]$ is connected. If $n>4,$ then $P_{n}[B_{(i)}]-N[I]$ is connected for any $i<n$ since $|N[I]\cap B_{(1)}|=1$ and $|N[I]\cap B_{(i)}|=0$ for any $1<i<n.$ By induction, $P_{n}[B_{(n)}]-N[I]=P_{n-1}-N[I]$ is connected. By Lemma \[yl-2.3\], each vertex in $B_{(i)}^{(n)}$ is adjacent to exactly one vertex in $B_{(n)}^{(i)}$ for any $i<n.$ So for each $i<n$ there exists a vertex in $P_{n}[B_{(i)}]-N[I]$ which is adjacent to some vertex in $P_{n}[B_{(n)}]-N[I].$ Thus $P_n-N[I]=\bigcup_{i=1}^{n}P_{n}[B_{(i)}]-N[I]$ is connected. Next we turn to consider the efficient dominating sets of $P_n.$ By the definition of efficient dominating set, it is easy to see that any efficient dominating set $D$ in $P_n$ has $(n-1)!$ elements and $d(u,v)\geq 3$ for any $u,v\in D,$ where $d(u,v)$ is the distance between two vertice $u$ and $v$ in $P_n.$ Konstantinova in the abstract [@Elena] obtained the following result on the efficient dominating set. For the completeness of this paper, here we present a proof of the result. ${{\fs\cite{Elena}}}$ \[dl-2.6\] There are exactly $n$ efficient dominating sets $B^{(i)}~(1\leq i\leq n)$ in $P_n~(n\geq 3).$ Clearly each $B^{(i)}~(1\leq i\leq n)$ is an efficient dominating set in $P_n.$ So it suffices to prove that for any efficient dominating set $D$ in $P_n,$ if $D\cap B^{(i)}\neq\emptyset,$ then $D=B^{(i)}.$ Set $D^{(i)}=D\cap B^{(i)},\,D_{(j)}=D\cap B_{(j)},\,D_{(j)}^{(i)}=D\cap B^{(i)}_{(j)},\,R_{(j)}^{(i)}=B^{(i)}_{(j)}\setminus D_{(j)}^{(i)}.$ We consider the following cases: [**Case 1.**]{} There exists a $D_{(j)}^{(i)}$ such that $D_{(j)}^{(i)}=B^{(i)}_{(j)}.$ By Lemma \[yl-2.3\], $N(B^{(i)}_{(j)})\cap B_{(i)}=B^{(j)}_{(i)},\,N(N(B^{(i)}_{(j)}))\cap B_{(i)}=B_{(i)}\setminus B^{(j)}_{(i)}.$ Since $B^{(i)}_{(j)}\subseteq D$ and $d(u,v)\geq 3$ for any $u,v\in D,$ so we have $D\cap B^{(j)}_{(i)} =\emptyset,\,D\cap (B_{(i)}\setminus B^{(j)}_{(i)})=\emptyset,$ i.e. $D\cap B_{(i)}=\emptyset.$ By Lemma \[yl-2.3\] again, $N(B_{(i)})=B^{(i)}$ and each vertex in $B_{(i)}$ is adjacent to exactly one vertex in $B^{(i)}.$ Hence $B^{(i)}\subseteq D.$ Since $|B^{(i)}|=|D|=(n-1)!,$ we have $D=B^{(i)}.$ [**Case 2.**]{} There exists a $D_{(j)}^{(i)}$ such that $\emptyset\neq D_{(j)}^{(i)}\varsubsetneq B^{(i)}_{(j)}.$ By Lemma \[yl-2.3\], we have $X_{(i)}:=N(D_{(j)}^{(i)})\cap B_{(i)}\subseteq B_{(i)}^{(j)},\,Y_{(i)}:=N(R_{(j)}^{(i)})\cap B_{(i)}\cap D\subseteq B_{(i)}^{(j)},\,Z_{(i)}:=B_{(i)}^{(j)}\setminus(X_{(i)}\cup Y_{(i)}),\,W_{(i)}:=N(Z_{(i)})\cap D\subseteq B_{(i)}\setminus B_{(i)}^{(j)},\,Y_{(j)}:=N(R_{(j)}^{(i)})\cap B_{(j)}\cap D\subseteq B_{(j)}\setminus B_{(j)}^{(i)}.$ Now we claim that $D_{(i)}=Y_{(i)}\cup W_{(i)},\,D_{(j)}=D_{(j)}^{(i)}\cup Y_{(j)}.$ Clearly $D_{(i)}\supseteq Y_{(i)}\cup W_{(i)},\,D_{(j)}\supseteq D_{(j)}^{(i)}\cup Y_{(j)}.$ For any $x\in D_{(i)}\setminus Y_{(i)},$ then $x\in B_{(i)}\setminus (X_{(i)}\cup N(X_{(i)})\cup Y_{(i)}\cup N(Y_{(i)})\cup Z_{(i)})=N(Z_{(i)})\cap B_{(i)}$ and so $x\in N(Z_{(i)})\cap B_{(i)}\cap D=W_{(i)}.$ Hence $D_{(i)}\subseteq Y_{(i)}\cup W_{(i)}.$ For any $y\in D_{(j)}\setminus D_{(j)}^{(i)},$ then $y\in B_{(j)}\setminus (D_{(j)}^{(i)}\cup N(D_{(j)}^{(i)})\cup R_{(j)}^{(i)})=N(R_{(j)}^{(i)})\cap B_{(j)}$ and so $y\in N(R_{(j)}^{(i)})\cap B_{(j)}\cap D=Y_{(j)}.$ Hence $D_{(j)}\subseteq D_{(j)}^{(i)}\cup Y_{(j)}.$ Clearly $|X_{(i)}|=|D_{(j)}^{(i)}|$ and $|Y_{(i)}|+|Y_{(j)}|=|R_{(j)}^{(i)}|,$ so $|X_{(i)}|+|Y_{(i)}|+|Y_{(j)}|=|B_{(j)}^{(i)}|=(n-2)!.$ Since $|X_{(i)}|+|Y_{(i)}|+|Z_{(i)}|=|B_{(i)}^{(j)}|=(n-2)!,$ we have $|W_{(i)}|=|Z_{(i)}|=|Y_{(j)}|.$ By the definition of efficient dominating set and Lemma \[yl-2.3\], for $k=i,j,$ each vertex in $B_{(k)}\setminus (D_{(k)}\cup N(D_{(k)})$ is adjacent to exactly one vertex in $D^{(k)},$ each vertex in $D^{(k)}$ is adjacent to exactly one vertex in $B_{(k)}\setminus (D_{(k)}\cup N(D_{(k)}).$ So $|D^{(i)}|=|B_{(i)}\setminus (D_{(i)}\cup N(D_{(i)})| =(n-1)!-(n-1)|D_{(i)}| =(n-1)!-(n-1)(|Y_{(i)}|+|W_{(i)}|) =(n-1)!-(n-1)((n-2)!-|X_{(i)}|)=(n-1)|X_{(i)}|,\, |D^{(j)}|=|B_{(j)}\setminus (D_{(j)}\cup N(D_{(j)})|=(n-1)!-(n-1)|D_{(j)}|=(n-1)!-(n-1)(|D_{(j)}^{(i)}|+ |Y_{(j)}|)=(n-1)!-(n-1)((n-2)!-|Y_{(i)}|)=(n-1)|Y_{(i)}|.$ Hence $|\bigcup_{k\neq i,j}D_{(k)}^{(i)}|=|D^{(i)}|- |D_{(j)}^{(i)}|=(n-1)|X_{(i)}|-|X_{(i)}|=(n-2)|X_{(i)}|,\,|\bigcup_{k\neq i,j}D_{(k)}^{(j)}|=|D^{(j)}|-|D_{(i)}^{(j)}|=(n-1)|Y_{(i)}|-|Y_{(i)}|=(n-2)|Y_{(i)}|$ and $|\bigcup_{k,l\neq i,j}D_{(k)}^{(l)}|=|D|-|D^{(i)}|-|D^{(j)}|-|W_{(i)}|-|Y_{(j)}|= (n-1)!-(n-1)|X_{(i)}|-(n-1)|Y_{(i)}|-|W_{(i)}|-|Y_{(j)}|= (n-1)((n-2)!-|X_{(i)}|-|Y_{(i)}|)-2|Z_{(i)}|=(n-3)|Z_{(i)}|.$ By the definition of efficient dominating set and Lemma \[yl-2.3\] , for any a fixed $l_0\neq i,j,$ each vertex in $\bigcup_{k\neq i,j}B_{(k)}^{(l_0)}$ either belongs to $\bigcup_{k\neq i,j}D_{(k)}$ or is adjacent to exactly one vertex in $\bigcup_{k\neq i,j}D_{(k)},$ each vertex in $\bigcup_{k\neq i,j}D_{(k)}$ either belongs to $\bigcup_{k\neq i,j}B_{(k)}^{(l_0)}$ or is adjacent to exactly one vertex in $\bigcup_{k\neq i,j}B_{(k)}^{(l_0)},$so $(n-3)(n-2)!=|\bigcup_{k\neq i,j}B_{(k)}^{(l_0)}|=|\bigcup_{k\neq i,j}D_{(k)}|=|\bigcup_{k\neq i,j}D_{(k)}^{(i)}|+|\bigcup_{k\neq i,j}D_{(k)}^{(j)}|+|\bigcup_{k,l\neq i,j}D_{(k)}^{(l)}|=(n-2)|X_{(i)}|+(n-2)|Y_{(i)}|+(n-3)|Z_{(i)}|= (n-3)(|X_{(i)}|+|Y_{(i)}|+|Z_{(i)}|)+|X_{(i)}|+|Y_{(i)}|=(n-3)(n-2)!+|X_{(i)}|+|Y_{(i)}|,$ hence $|D_{(j)}^{(i)}|=|X_{(i)}|=0,$ which is a contradiction. The automorphism group of $P_n$ =============================== In this section, we completely determine the full automorphism group of $P_n.$ First we introduce some definitions. Let $\Sym(\Omega)$ denote the set of all permutations of a set $\Omega.$ A [*permutation representation*]{} of a group is a homomorphism from into $\Sym(\Omega)$ for some set $\Omega.$ A permutation representation is also referred to as an action of on the set $\Omega,$ in which case we say that acts on $\Omega.$ Furthermore, if $\{g\in G:x^g=x,\,\forall x\in \Omega\}=1,$ we say the action of $G$ on $\Omega$ is [*faithful*]{}, or $G$ acts [*faithfully*]{} on $\Omega.$ \[dl-3.1\] For $n\geq 5,$ if $N(X)=B^{(i)}$ and $|X|=|B^{(i)}|,$ where $X\subseteq V(P_n)=S_n$ and $i\in \{1,2,\cdots,n\},$ then $X=B_{(i)}.$ For $n=5,$ one can easily check that the assertion holds. We proceed by induction on $n.$ First since $N(X)=\{y\in V(P_n)\setminus X:\exists\,x\in X,\,s.t.\,xy\in E(P_n)\},$ we have $X\cap N(X)=\emptyset,$ i.e. $X\cap B^{(i)}=\emptyset.$ Next we shall show that $X=B_{(i)}$ by the following three Claims: [**Claim 1.**]{} Either $X=B_{(i)}$ or $X\cap B_{(i)}=\emptyset.$ Set $X_i:=X\cap B_{(i)},\,\overline{X_i}:=X\setminus X_i.$ Suppose on the contrary that $\emptyset\neq X_i\subsetneq B_{(i)}.$ By Lemma \[yl-2.3\] (iii), $P_n[B_{(i)}]\cong P_{n-1},$ so $P_n[B_{(i)}]$ is connected, which implies that $N(X_i)\cap B_{(i)}\neq\emptyset.$ Since $N(X_i)\subseteq B_{(i)}\cup B^{(i)}$ and $\overline{X_i}\cap (B_{(i)}\cup B^{(i)})=\emptyset,$ we have $N(X_i)\cap \overline{X_i}=\emptyset,$ i.e. $N(X_i)\subseteq N(X).$ So $N(X)\cap B_{(i)}\neq\emptyset,$ which contradicts $N(X)=B^{(i)},$ hence Claim 1 holds. [**Claim 2.**]{} Set $X_k=X\cap B_{(k)},\,B_{(n-1)\rightarrow i,\,n\rightarrow k}=\{\pi\in S_n:\,\pi(n-1)=i,\,\pi(n)=k\}.$ If $X\neq B_{(i)},$ then $X_k=B_{(n-1)\rightarrow i,\,n\rightarrow k}$ for any $k\neq i.$ By $X\neq B_{(i)}$ and Claim 1, $X\cap (B^{(i)}\cup B_{(i)})=\emptyset.$ By Lemma \[yl-2.3\] (ii), $B^{(i)}_{(k)}\cap N(X_l)=\emptyset$ for any $k\neq l.$ So we have $B^{(i)}_{(k)}\subseteq B^{(i)}=N(X)=N(\bigcup_{k\neq i}X_k)\subseteq\bigcup_{k\neq i}N(X_k)\Rightarrow B^{(i)}_{(k)}\subseteq N(X_k)\Rightarrow B^{(i)}_{(k)}\subseteq B_{(k)}\cap N(X_k).$ On the other hand, $B_{(k)}\cap N(X_k)\subseteq B_{(k)}\cap N(X)=B_{(k)}\cap B^{(i)}=B^{(i)}_{(k)}.$ Thus $B_{(k)}\cap N(X_k)=B^{(i)}_{(k)}.$ By Theorem \[dl-2.6\], $B^{(i)}$ is an efficient dominating set of $P_n,$ so $|X_k|\geq |B^{(i)}_{(k)}|\Rightarrow|X|=\sum_{k\neq i}|X_k|\geq\sum_{k\neq i}|B^{(i)}_{(k)}|=|B^{(i)}|,$ note that $|X|=|B^{(i)}|,$ and so $|X_k|=|B^{(i)}_{(k)}|.$ By Lemma \[yl-2.3\] (iii), $P_n[B_{(k)}]\cong P_{n-1}$ and $B^{(i)}_{(k)}$ is an efficient dominating set of $P_n[B_{(k)}].$ Since $B_{(k)}\cap N(X_k)=B^{(i)}_{(k)}$ and $|X_k|=|B^{(i)}_{(k)}|,$ by induction, we have $X_k=B_{(n-1)\rightarrow i,\,n\rightarrow k},$ hence Claim 2 holds. [**Claim 3.**]{} If $X\neq B_{(i)},$ then $n=3.$ By $X\neq B_{(i)}$ and Claim 2, $X_k=B_{(n-1)\rightarrow i,\,n\rightarrow k}$ for any $k\neq i.$ Since $X_k\subseteq B_{(k)}$ and $P_n[B_{(k)}]\cong P_{n-1},$ which is a $(n-2)$-regular graph, we have $|N(x_k)\cap B_{(k)}|=n-2$ for any $x_k\in X_k,$ note that $|N(x_k)|=n-1,$ and so $|N(x_k)\cap(\bigcup_{l\neq k} B_{(l)})|=1.$ Set $N(x_k)\cap(\bigcup_{l\neq k} B_{(l)})=\{x_l\},$ where $x_l\in B_{(l)}$ for some $l\neq k,i.$ Since $x_l\in N(x_k)\cap B_{(l)}$ and $N(x_k)\cap B_{(l)}\cap B^{(i)}=\emptyset$ (by Lemma \[yl-2.3\]), we have $x_l\not\in B^{(i)}.$ Note that $x_l\in N(x_k)$ and $N(X)=B^{(i)},$ then $x_l\in X_l=B_{(n-1)\rightarrow i,\,n\rightarrow l}$ (by Claim 2) and there exists a $r_{1j}\in PR_n$ such that $x_k=x_lr_{1j}.$ Now we show that $j=n.$ otherwise, we have $j\neq n\Rightarrow r_{1j}(n)=n\Rightarrow k=x_k(n)=x_lr_{1j}(n)=x_l(n)=l,$ which contradicts $k\neq l.$ So $i=x_k(n-1)=x_lr_{1n}(n-1)=x_l(2)\Rightarrow n-1=x_l^{-1}(i)=2\Rightarrow n=3,$ hence Claim 3 holds. By Claim 3, if $X\neq B_{(i)},$ then $n=3,$ which contradicts $n\geq 5.$ Hence $X=B_{(i)},$ the assertion holds. [**Remark.**]{} For $n=3,4,$ one can easily check that the result of Theorem \[dl-3.1\] is not true. For example, in $P_3,~ N(\{(1\,2),(1\,3\,2)\})=B^{(1)}$ and $|\{(1\,2),(1\,3\,2)\}|=|B_{(1)}|=2,$ however, $\{(1\,2),(1\,3\,2)\}\neq B_{(1)};$ In $P_4,~ N(\{(1\,2),(1\,2)(3\,4),(1\,3\,2),(1\,3\,4\,2),(1\,4\,2),(1\,4\,3\,2)\})=B^{(1)}$ and $|\{(1\,2),(1\,2)(3\,4),(1\,3\,2),(1\,3\,4\,2),(1\,4\,2),(1\,4\,3\,2)\}|=|B_{(1)}|=6,$ however, $\{(1\,2),(1\,2)(3\,4),(1\,3\,2),\\ (1\,3\,4\,2),(1\,4\,2),(1\,4\,3\,2)\}\neq B_{(1)}.$ \[dl-3.2\] If $n\geq 5,$ then $\Aut(P_n)=L(S_n),$ where $L(S_n)$ is the left regular representation. For $n=5,$ a Nauty [@Mckay] computation shows that $|\Aut(P_5)|=120.$ Since $|\Aut(P_5)|\geq |L(S_5)|=120,$ we have $\Aut(P_5)=L(S_5).$ We proceed by induction on $n.$ Clearly any automorphism of $P_n$ must permute the efficient dominating sets of $P_n.$ Let ${\mathcal B}=\{B^{(i)}: i=1,2,\cdots,n\}.$ By Theorem \[dl-2.6\], $\Aut(P_n)$ naturally acts on ${\mathcal B}.$ Next we shall show that the action of $\Aut(P_n)$ on ${\mathcal B}$ is faithful. Assume that $\phi\in \Aut(P_n)$ such that $\phi(B^{(i)})=B^{(i)}$ for each $i\in\{1,2,\cdots,n\}.$ By Lemma \[yl-2.3\], $N(B_{(i)})=B^{(i)},$ so we have $N(\phi(B_{(i)}))=\phi(B^{(i)})=B^{(i)},\, |\phi(B_{(i)})|=|B_{(i)}|=|B^{(i)}|.$ By Theorem \[dl-3.1\], $\phi(B_{(i)})=B_{(i)}$ for each $i\in\{1,2,\cdots,n\}.$ Hence $\phi$ can be treated as an automorphism of $P_n[B_{(n)}]=P_{n-1},$ that is, the restriction $\phi{\upharpoonright}{B_{(n)}}\in\Aut(P_{n-1})=L(S_{n-1})$ by induction. For the identity element $I\in S_n,$ set $y=\phi(I),$ then $y,I\in B_{(n)}^{(1)}\subseteq B_{(n)}$ and $\phi{\upharpoonright}{B_{(n)}}=L(y).$ Hence $$\begin{aligned} \phi(I)=y&\Rightarrow&\phi(N(I)\cap B_{(n)}^{(i)})=N(y)\cap B_{(n)}^{(i)}\\ &\Rightarrow&L(y)(r_{1,i})=yr_{1,y^{-1}(i)}\\ &\Rightarrow&yr_{1,i}=yr_{1,y^{-1}(i)}\\ &\Rightarrow&y(i)=i,\end{aligned}$$ where $i=2,3,\cdots,n.$ So we have $\phi(I)=y=I,$ that is, $\phi$ fixes $I.$ Since $\phi(B^{(i)}_{(j)})=B^{(i)}_{(j)}$ for each $i,j\in\{1,2,\cdots,n\},$ by Lemma \[yl-2.3\] (ii) and the connectedness of $P_n,$ $\phi$ fixes all vertice of $P_n,$ so $\phi=1,$ which implies that the action of $\Aut(P_n)$ on ${\mathcal B}$ is faithful. Thus $Aut(P_n)\lesssim \Sym({\mathcal B})\Rightarrow |Aut(P_n)|\leq n!.$ On the other hand, $|Aut(P_n)|\geq |L(S_n)|=n!.$ Hence $Aut(P_n)=L(S_n).$ The assertion holds. [**Remark.**]{} If $n=3,$ then $P_3=C_6,$ so $\Aut(P_3)=D_{12},$ where $D_{12}$ is the dihedral group of order $12.$ If $n=4,$ a Nauty computation shows that $|\Aut(P_4)|=48,$ so $L(S_4)$ is a normal subgroup of $\Aut(P_4).$ By Godsil [@Godsil], $\Aut(P_4)$ is the semiproduct $L(S_4)\rtimes\Aut(S_4,PR_4),$ where $\Aut(S_4,PR_4)=\{\phi\in \Aut(S_4):\,\phi(PR_4)=PR_4\}=\{1,c((2\,3))\},$ here we denote by $1$ the identity automorphism and by $c((2\,3))$ the automorphism induced by the conjugacy of $(2\,3)$ on $S_4.$ S.B. Akers, B. Krishnamurthy, A group-theoretic model for symmetric interconnection networks, IEEE Trans. Comput. (4)38(1989) 555-566. D.W. Bange, A.E. Barkauskas, P.J. Slater, Efficient near-domination of grid graphs, Congr. Numer. 58(1986) 83-92. D.W. Bange, A.E. Barkauskas, L.H. Host, P.J. Slater, Generalized domination and efficient domination in graphs, Discrete Math. 159(1996) 1-11. F. Boesch, R. Tindell, Circulants and their connectivities, J. Graph Theory 8(1984) 487-499. I.J. Dejter, O. Serra, Efficient dominating sets in Cayley graphs, Discrete Applied Math. 129(2003) 319-328. H. Dweighter, E 2569 in: Elementary problems and solutions, Amer. Math. Monthly (1)82(1975) 1010. Elena Konstantinova, Perfect codes in the pancake networks, available at http://www.math.uniri.hr/NATO-ASI/abstracts/Konstantinova\_abstract.pdf. W.H. Gates, C.H. Papadimitriou, Bounds for sorting by prefix-reversal, Discrete Math. 27(1979) 47-57. C.D. Godsil, On the full automorphism group of a graph, Combinatorica 1(1981) 243-256. Y.O. Hamidoune, Subsets with small sums in Abelian group’s. I: The Vosper property, European J. Combin. 18(1997) 541-556. M.H. Hyedari, I.H. Sudborough, On the diameter of the pancake network, J. Algorithms (1)25(1997) 67-94. M.H. Hyedari, I.H. Sudborough, A Quadratic Lower Bound for Reverse Card Shuffle. In Proc. 26th S.E. Conf. Combinatorics, Graph Theory, and Computing, 1995. A. Kanevsky, C. Feng, On the embedding of cycles in pancake graphs, Parallel Comput. 21(1995) 923-936. S. Lakshmivarahan, J.S. Jwo, S.K. Dhall, Symmetry in interconnection networks based on Cayley graphs of permutation groups: A survey, Parallel Comput. (4)19(1993) 361-407. R. Li, J.X. Meng, Reversals Cayley graph of symmetric groups, Information Processing Letters 109(2008) 130-132. C.K. Lin, H.M. Huang, L.H. Hsu, The super connectivityof the pancake graphs and the super laceability of the star graphs, Theoretical Computer Science 339(2005) 257-271. W. Mader, $\ddot{u}$ber den zusammen symmetricher graphen, Arch. Math. 21(1970) 331-336. Brender D. Mckay, Practical graph isomorphism, Congressus Numerantium 30(1981) 45-87, Nauty available from http://cs.anu.edu.au/people/bdm/nauty/. L.A. Nowitz, M.E. Watkins, Graphical Regular Representations of Non-abelian Groups, Canad. J. Math. 24(1972) 993-1008. K. Qiu, Optimal broadcasting algorithms for multiple messages on the star and pancake graphs using minimum dominating sets, Congressus Numerantium 181(2006) 33-39. J.J. Sheu, J.M. Tan, L.H. Hsu, M.Y. Lin, On the cycle embedding of pancake graphs, available at http://dspace.lib.fcu.edu.tw/bitstream/2377/3179/1/ce07ncs001999000212.pdf. [^1]: This work is supported by National Natural Science Foundation of China (No:10971137), the National Basic Research Program (973) of China (No.2006CB805900), and a grant of Science and Technology Commission of Shanghai Municipality (STCSM, No: 09XD1402500) . $^{\dagger}$Correspondent author: Xiao-Dong Zhang (Email: [email protected])

ArXiv

--- abstract: 'We review the basic techniques for extracting information about quasar structure and kinematics from the broad emission lines in quasars. We consider which lines can most effectively serve as virial estimators of black hole mass. At low redshift the Balmer lines,particularly broad H$\beta$, are the lines of choice. For redshifts greater than 0.7 – 0.8 one can follow H$\beta$ into the IR windows or find an H$\beta$ surrogate. We explain why UV C[iv]{}$\lambda$1549 is not a safe virial estimator and how Mg[ii]{}$\lambda$2800 serves as the best virial surrogate for H$\beta$ up to the highest redshift quasar known at $z \approx 7$. We show how spectral binning in a parameter space context (4DE1) makes possible a more effective comparison of H$\beta$ and Mg[ii]{}. It also helps to derive more accurate mass estimates from appropriately binned spectra and, finally, to map the dispersion in $M_\mathrm{BH}$ and Eddington ratio across the quasar population. FWHM MgII is about 20$\%$ smaller than FWHM H$\beta$ in the majority of type 1 AGN requiring correction when comparing $M_\mathrm{BH}$ estimates from these two lines. The 20$\%$ of sources showing narrowest FWHM H$\beta$ ($< 4000$ km s$^{-1}$) and strongest FeII ($R_\mathrm{Fe} \gtrsim 1.0$) emission (we call them bin A3-4 sources) do not show this FWHM difference and a blueshift detected in MgII for these sources suggests that FWHM H$\beta$ is the safer virial estimator for these extreme Eddington emitters.' address: - 'Instituto de Astrof[í]{}sica de Andaluc[í]{}a (CSIC), Granada, 18008, España' - 'INAF, Osservatorio Astronomico di Padova, Padova,35122, Italia' - 'Instituto de Astrof[í]{}sica de Andaluc[í]{}a (CSIC), Granada, 18008, España' author: - 'Jack W. Sulentic' - Paola Marziani - 'Ascensión del Olmo and Ilse Plauchu-Frayn' title: 'Techniques for Profile Binning and Analysis of Eigenvector Composite Spectra: Comparing H$\beta$ and Mg[ii]{}$\lambda$2800 as Virial Estimators ' --- line:formation; line: profile; quasars: emission lines; quasars: general =0.5 cm Introduction ============ Quasars were discovered in 1963 as radio-loud blue-stellar sources showing very broad emission lines in their high redshifted spectra. Fifty years later we know that most quasars are radio-quiet and show a wide diversity of spectroscopic properties united under the umbrella of Active Galactic Nuclei (AGN). The large population of AGN showing broad emission lines can be viewed as a luminous high-accreting parent population extending from local Seyfert galaxies like NGC1068 to the quasar with highest known redshift at $z \approx$7.03. Spectroscopy is the fundamental method we employ to “resolve” the central structure of line emitting gas (broad line region: BLR; see the review by @gaskell09) surrounding the supermassive black hole thought to lie at the center of each AGN. We study the broad emission lines to infer BLR structure and kinematics. More recently we have begun using the width of select broad lines to estimate the mass of the black hole and which also gives us an estimate of the source Eddington ratio. Broad emission lines characterise an important and dominant spectroscopic class of high accreting active galactic nuclei (AGN). Most are accreting at a high enough rate to fuel a reasonably stable broad line emitting region (BLR). The (likely) lowest accreting sources among them e.g. NGC 1097 [@storchi-bergmannetal93] and NGC 4151 [@bonetal12] are less stable and show broad lines only part of the time. Some are obscured (by a torus?) and show broad lines only in polarized light e.g. NGC 1068 [@antonuccimiller85]. At low redshift, where a host galaxy can be seen, they are called Seyfert 1 galaxies and at higher redshift, where they are largely unresolved point sources, we call them Type 1 quasars. Seyferts 1’s were long ago identified by Carl Seyfert [@seyfert43] while discovery of the quasars occured fifty years ago [@schmidt63]. Broad lines alone cannot define the AGN phenomenon because many subtypes lacking them are now known (e.g. BL Lacs, Type 2 AGN and LINERs). Our focus in this paper will be exclusively on Type 1 AGN. Whether hosted in a detectable galaxy or not, the region producing the broad emission lines is spatially unresolved. Such source compactness, and associated short variability timescales, are consistent with the hypothesis that the broad lines are produced in a region no larger than a fraction of a light year in diameter and connected with gas accreting onto a supermassive black hole [e.g. @gaskellsparke86; @padovanirafanelli88]. Spectroscopy therefore is, and will likely remain, the means by which we “resolve” this broad line region (BLR). Broad line widths and shapes then provide the most direct clues about BLR geometry and kinematics.All broad lines in an individual source do not necessarily show the same properties and any specific line can as well show striking diversity in different quasars. The latter fact offers an immediate motivation for attempts to systematize measures of specific broad lines. In addition to providing clues about BLR structure and kinematics, broad lines are now used increasingly to estimate the mass of the central black hole which measure is of great importance both for models of BLR structure and (potentially) for cosmology (see contribution by Marziani & Sulentic in this proceeding). Which lines can we use for black hole mass estimation and over what redshift range is a particular line useful? We seek lines that arise in a virialized medium which leads us to the Balmer lines at low redshift. H$\beta$ is the most useful Balmer line for such studies because it usually shows only small shifts and asymmetries – it thus appears to be a “safe” virial estimator for most Type 1 AGN. If it arises from a flattened distribution of emitting clouds or, even from a Keplerian accretion disk, as is often assumed, then the virial assumption is not unreasonable. H$\beta$  suffers from contamination by nearby lines of Fe[ii]{}, He[ii]{}$\lambda$4686 and (narrow) \[O[iii]{}\]$\lambda\lambda$4959,5007 but these can be reasonably modeled and subtracted to facilitate evaluation of line properties. H$\beta$ is lost to optical spectroscopy in the redshift range $z \sim 0.7 - 0.9$. It can be followed through the JHK infrared windows out to $z \approx 3.7$ using suitably large telescopes although this approach is still rather costly in terms of telescope time. Estimates of black hole masses for large high redshift samples require an H$\beta$ surrogate. Two lines, C[iv]{}$\lambda$1549 and Mg[ii]{}$\lambda$2800 [e.g., @mclurejarvis02], are the best candidates in terms of line strength and low levels of contamination. High ionization C[iv]{}$\lambda$1549 is so dangerous that many prefer to avoid it entirely [@sulenticetal07; @netzeretal07]. Yes ironically, various studies suggest that it is virialized [@gaskell88; @petersonwandel99]. CIV often shows profile blueshifts and asymmetries interpreted as signatures of winds or outflows [@gaskell82; @gaskellgoosmann13]. These C[iv]{}$\lambda$1549 distortions are often seen in quasars where H$\beta$ is most symmetric and well behaved. The problem with CIV is therefore not the issue of virialization but the likelyhood that the line emission arises highly anisotropically in a structure with unknown geometry. FWHM C[iv]{}$\lambda$1549 does not show a clear correlation with FWHM H$\beta$ the low $z$ virial estimator of choice and this precludes a simple geometric relation between C[iv]{}$\lambda$1549 and H$\beta$ as well as Mg[ii]{}$\lambda$2800 emissions [@shenetal08]. Thus C[iv]{}$\lambda$1549 is falling from (actually it was never in) favor, and Mg[ii]{}$\lambda$2800 is replacing it, as the preferred high redshift virial estimator. This paper compares H$\beta$ with Mg[ii]{}$\lambda$2800 and presents evidence in support of the latter line as a virial estimator for 80-90% of high redshift quasars–all except the highest accretors. This paper, and the emergence of Mg[ii]{}$\lambda$2800, are largely due to the advent of the SDSS which provides Mg[ii]{}$\lambda$2800 spectra for thousands of quasars. Most importantly for 500+ sources where both H$\beta$ and Mg[ii]{}$\lambda$2800 can be measured in the same spectra. Low ionization Mg[ii]{}$\lambda$2800 is reasonably symmetric and unshifted like H$\beta$ making it the best candidate as a surrogate. Mg[ii]{}$\lambda$2800 has even been measured (K band) in the quasar with highest current redshift ($z \approx$ 7.035; @mortlocketal11). Estimating Black Hole Mass ========================== When estimating black hole mass one requires both a velocity dispersion and a radius of the BLR. If the virial assumption is valid then $M_\mathrm{BH} = f (\delta v)^2r/G$. $f$ is a parameter that accounts for line-of-sight effects on the line profile due to geometry and kinematics. It may impose the ultimate limitation on the accuracy of $M_\mathrm{BH}$ estimates. We do not know the correct value of $f$ and, more importantly, how much it changes across the quasar population. The value of $r$ (the effective BLR radius) can be directly estimated using reverberation techniques as it simply involves the delay time in the response of an emission line to continuum changes. Reverberation mapping is prohibitively expensive in terms of required telescope time meaning that only about 60 of the brightest nearby (and hence predominantly low luminosity) sources have been reverberation mapped. The reverberation radii tend to correlate with measures of source luminosity leading to a radius – luminosity relation [@dibai77; @koratkargaskell91; @kaspietal05] $r \propto L^{\alpha}$ ($\alpha \approx $0.5) which can provide, via extrapolation, $r$ estimates for all sources where $L$ can be reliably measured. Note that we have only a vague notion of what $ r$ means since the BLR is unlikely to be a shell surrounding the central continuum source [@negreteetal13 and Alenka Negrete’s contribution in this proceedings]. Measures of $\delta v$ (the virial velocity dispersion) come from direct measurement of a line that is considered a reliable virial estimator. If a given broad line is reasonably symmetric and unshifted we adopt FWHM H$\beta$ (or FWHM Mg[ii]{}$\lambda$2800?) as the virial estimator. We do this because a symmetric line without inflections implies a reasonably coherent single source of line emission involving motions centered around the quasar rest frame (usually inferred from the redshift of narrow lines like \[O[iii]{}\]$\lambda\lambda$4959,5007). This is the simplest expectation for a bound virialized emitting region. If the line profile shows inflections we do not know what to do because inflections imply spectroscopic resolution of the BLR. The spatially unresolved source could involve multiple kinematically distinct emitting components –a binary black hole comes to mind – if each black hole is surrounded by a BLR. This is perhaps less likely than such profile complexities indicating stratification of the emitting region and/or radial motions of all or part of the line emitting gas (winds, outflows or infall). Line profile inflections are a valuable clue telling us that multiple emitting components are present and that all or part of the line cannot be safely regarded as a virial estimator. The study of profile inflections is still in its infancy because they require high resolution (like SDSS) and high s/n ( like only a few hundred SDSS) spectra. Three broad line components have been identified in H$\beta$ [@marzianietal10] and presumably only one of them is/might be a virial estimator. The situation is not so bad because many sources show only a single component ($\sim$50% of quasars in the case of H$\beta$). Naturally for the other 50% we adopt the least shifted broad component for black hole mass estimation. The larger blueshifts and asymmetries observed in C[iv]{}$\lambda$1549 represent a more serious problem–in this case we are not even sure if any part of the line can be trusted. Another problem with C[iv]{}$\lambda$1549 involves the difficulty of subtracting a narrow line component which if uncorrected will cause underestimation of the black hole mass [@sulenticmarziani99; @sulenticetal07]. C[iv]{}$\lambda$1549 is not the road to more accurate black hole masses. Note that we do not want to use many different lines for virial mass estimation because this simply adds uncertainty when we compare them and try to tie them together over different redshift ranges. That is why it is likely that H$\beta$ and Mg[ii]{} represent the best and safest set of virial estimators that can cover the full redshift range. In other words we do not need or want C[iv]{}$\lambda$1549 especially as a (more uncertain) third estimator. FWHM H$\beta$ vs. $\sigma$ as the Virial Estimator? =================================================== The second moment of the line profile $\sigma$ (also called the line dispersion) has sometimes been favored over profile width as a more physical and/or reliable virial measure. The relative merits of the two measures as virial estimators were discussed in @collinetal06. Both FWHM and $\sigma$ are simply numbers (albeit one in units of line-of-sight velocity) and 1) if the line profile of the adopted virial estimator were similar from source-to-source and 2) if a uniform range of profile width was observed, then the two numbers should be fully equivalent (allowing only for the possibility that one might be measured more accurately than the other, @petersonetal04) Unfortunately broad lines sometimes show inflections indicating 2 or 3 emitting components. Since they cannot all be virial estimators neither a FWHM nor $\sigma$ measure for the full profile should be used. The line profile can at least be used to model the line components [@marzianietal10] while $\sigma$  is of no use and has no meaning. If one models the individual line components and isolates a FWHM measure for the component most likely to be a virial estimator then there seems little additional value in computing $\sigma$ and using it instead of FWHM as the virial measure. Profile modeling is the road to more accurate black hole mass estimates in the future but it can only be applied if spectral S/N is high enough. A S/N$\geq$20computed in the continuum near, but not including, the adopted virial estimator is a useful rule of thumb. In addition to broad line inflections there is often the well known inflection between broad and narrow line components. This is especially important when using H$\beta$ as the virial estimator. When the inflection is clear–especially for sources where broad H$\beta$ shows FWHM $>$ 3000 km s$^{-1}$ – it can be used to guide subtraction of the narrow component. Black hole mass estimates from a broad line without correction for the narrow component will be underestimates. Again reasonable spectra S/N is desirable if one wishes to obtain an estimate with less than 1dex uncertainty. Contextualization of Broad Line Properties: 4DE1 ================================================ Fortunately narrow line and broad line inflections do not occur randomly but show trends. What we need is a context in which to interpret broad line spectra. The PCA analysis of high S/N PG spectra [@borosongreen92] opened the door to a new era of source contextualization. These results provided the first hints that all quasar spectra are [*not*]{} self similar and that indiscriminate averaging of quasar spectra [@vandenberketal01] – no matter how tempting – is not the path to progress. We have built upon the PG results using larger samples of quasars and have proposed a 4D Eigenvector 1 (4DE1) parameter space [@sulenticetal00a] involving correlations between: 1) FWHM H$\beta$ (the virial estimator of choice), 2) $R_\mathrm{Fe{\sc ii}} = I($Fe[ii]{}$\lambda$4570/I(H$\beta$), 3) soft X-ray photon index $\Gamma_\mathrm{soft}$ see also [@wangetal96] and 4) velocity shift of the C[iv]{}$\lambda$1549 centroid at FWHM. Results so far are consistent with the assumption that source occupation in 4DE1 has a physical basis with source Eddington ratio as the principal driver [@marzianietal01]. How can a context involving diagnostic measures help us derive more accurate black holes masses? Can it also help us to compare the relative merits of H$\beta$ and Mg[ii]{}$\lambda$2800 as virial estimators? Figure \[fig:e1\] shows the optical plane of 4DE1 which plots source FWHM H$\beta$ vs. $R_\mathrm{Fe{\sc ii}} $ measures. It is shown here for a large magnitude limited sample of bright SDSS quasars [@zamfiretal10] where a clear sequence of source occupation is observed. Sources with FWHM H$\beta <$4000 km s$^{-1}$ (population A) usually show a symmetric unshifted H$\beta$ profile while sources with FWHM H$\beta >$4000 km s$^{-1}$ (population B) require a double Gaussian model to describe the line. Pop A sources show Lorentz-like profiles so FWHM H$\beta$ derived from Gaussian fits to the line will result in overestimation of $M_\mathrm{BH}$. Pop B sources fit with a single function and FWHM (or $\sigma$) measured from the fit, will result in serious overestimation of $M_\mathrm{BH}$. Pop. B includes the majority of radio-loud quasars so this overestimation often affects comparisons of $M_\mathrm{BH}$ for radio-quiet and radio-loud quasars. The double Gaussian required to fit Pop. B sources involve a broad relatively unshifted BLR component (BC) plus a very broad (FWHM$\sim$10000km s$^{-1}$) and redshifted ($\Delta v_\mathrm{r}$=1000 – 2000 km s$^{-1}$) VBC component. We have no choice but to adopt the unshifted component as the virial estimator – but FWHM of this component is often broadened by the VBC if the profile is not modelled. Most spectra lack high enough S/N to allow modelling or even recognition of the composite profiles for the almost half of quasars that show pop B characteristics. This is where the 4DE1 formalism can help. We know where the sources with simple and complex spectra are located in 4DE1 space. We can bin the optical plane of 4DE1 as shown in Figure \[fig:e1\] allowing us to generate median composite spectra for quasars in each bin. We thus avoid binning together dissimilar spectra which might be reflecting different BLR physics. The resultant composite spectra show much higher S/N facilitating modeling of line profiles and more accurate measures of black hole mass and Eddington ratio. We can map median $M_\mathrm{BH}$ and $L/L_\mathrm{Edd}$ for sources across the 4DE1 optical plane. Binned median spectra also open the door to a more refined comparison between H$\beta$ and Mg[ii]{}$\lambda$2800 as virial estimators. ![The optical plane of the 4D eigenvector 1 parameter space, using data from [@zamfiretal10]. The plane has been subdivided in two regions occupied by Pop. A and B. A finer subdivision in spectral types (bins in FWHM and $R_\mathrm{Fe{\sc ii}}$) is also shown. The halftone arrow symbolizes several trends along the sequence of sources in the plane: among them, the increase in high ionization line blueshift and in Eddington ratio toward extreme Pop. A sources (A3 and A4). Circled sources are radio-loud according to the criteria defined by @zamfiretal08. \[fig:e1\]](e1fserbia.pdf){width="13.25cm"} Comparing H$\beta$ and Mg[ii]{}$\lambda$2800 as Virial Estimators ================================================================= H$\beta$ is well established as the virial estimator of choice at $z<$ 0.7 – 0.9. There is no suitable low redshift alternative except for H$\alpha$ which can only be used for sources below $z \approx 0.2$. Of course both can be followed into the IR, H$\beta$ can be used up to z=3.7. Suitably high S/N single spectra and luminosity binned composite spectra suggest that the effect of nonvirial components becomes more serious at high redshift [@sulenticetal06]. $M_\mathrm{BH}$ has likely been overestimated for a large fraction of high $z$ quasars. Such IR spectra are sufficiently costly in telescope time that they are not likely to provide measures for more than modest handfuls of sources [@marzianietal09]. Mg[ii]{}$\lambda$2800 is a low ionization resonance doublet that might reasonably be expected to show similar propertied to H$\beta$ [@grandiphillips79; @netzer80] but with the added difficulty of narrow and broad-line absorption in an unknown fraction of sources – possibly more common in Pop. A quasars. SDSS has opened the door to detailed statistical studies of Mg[ii]{}$\lambda$2800 and to comparison with H$\beta$ [e.g., @trakhtenbrotnetzer12]. We find 680 sources in SDSS DR8 with spectra that provided high resolution line profiles for both H$\beta$ and Mg[ii]{}$\lambda$2800. They are observed in the redshift range from $z$ = 0.4 – 0.75. There are more sources in this redshift range but their spectra are too noisy to permit reliable 4DE1 spectral bin assignments. Bright sources in this redshift range cover a limited range of source luminosity. This means that median composite spectra that we compute for this bright quasar sample will reflect properties of sources within 1 dex of $\log L_\mathrm {BOL} \approx$ 46. This is not bad because it represents typical intermediate luminosity sources. The composites trace a supposedly stable profile that is shared by most sources. They are conceptually different from rms spectra computed for single sources in reverberation studies. Our assumption is that a large numbers of sources scatter around a well defined median/average in each 4DE1 bin. Using this procedure we are able to generate composite spectra for the 8 most occupied bins in the 4DE1 optical plane. ![Continuum subtracted spectral regions of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800 (right) for spectral types A1 (top) and A2 (bottom). Abscissa is rest frame wavelength in Å and ordinate is continuum-normalized intensity. Orange lines trace narrow lines and Fe[i]{} emission at $\approx$ 2900 Å. The green line represents Fe[ii]{} emission. Thick solid lines show the broad components of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800. \[fig:a1a2\]](a1a2serbia.pdf){width="13.25cm"} Comparison for Spectral Bins A1 and A2 ====================================== Figure \[fig:a1a2\] shows median composite spectra for H$\beta$ and Mg[ii]{}$\lambda$2800 in bins A1 and A2. A2 is the most populated bin in Population A (48% of the sample) with 50% of the Pop. A sample. The A bins are largely a sequence of increasing $R_\mathrm{Fe{\sc ii}}$ that subsumes the region formally defined for NLSy1 sources. Given the 1 Å resolution of SDSS spectra we are forced to model Mg[ii]{}$\lambda$2800 as a 2793/2806$\AA$ doublet of fixed ratio 1.25. The composites in Figure 2 show [iraf specfit]{} results superimposed including: 1) Fe[ii]{} modelling, 2) fits to narrow \[O[iii]{}\]4959,5007 and H$\beta$, 3) He4686 and 4) broad H$\beta$. When we use [iraf specfit]{} we must specify what the program should look for or it will never find a reasonable solution with so many free parameter lines and components present. In this case we assume that both components of Mg[ii]{}$\lambda$2800 are unshifted Lorentz profiles. This provides a good fit judging from the residuals shown at the bottom of the fits in Fig. 2. Gaussian models provide an inferior fit consistent with what we always find for H$\beta$. The main result of this comparison indicates that FWHM Mg[ii]{}$\lambda$2800 is 20% smaller than FWHM H$\beta$ meaning that Mg[ii]{}$\lambda$2800 used as the virial estimator will yield systematically smaller black holes masses than H$\beta$ [in agreement with the BLR self-shielding model of @gaskelletal07]. In general if one employs Mg[ii]{}$\lambda$2800 one must pay attention to the resolution of the spectra being measured relative to the separation of the Mg[ii]{}$\lambda$2800 doublet. Otherwise one will overestimate $M_\mathrm{BH}$ and one will find less than a 20% FWHM difference relative to H$\beta$ (e.g. in bin A1 FWHM H$\beta$=3180 while for Mg[ii]{}$\lambda$2800 the full profile FWHM $\approx$ 3040 km s$^{-1}$ while individual multiplets show FWHM Mg[ii]{}$\lambda$2800 $\approx$ 2710 km s$^{-1}$). Note that in this comparison a black hole mass estimation also requires a value of $r$ which is obtained from extrapolation of the Kaspi relation. The luminosities used to derive $r$ will not be the same for the two lines. $L$ is derived from the continuum near each line: 3100 Å and 5100 Å for Mg[ii]{}$\lambda$2800 and H$\beta$ respectively. Best estimate median log $M_\mathrm{BH}$ values for bins A1/A2 are 8.67/8.57 and 8.62/8.45 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively (solar units). Best log Eddington ratio estimates for A1/A2 are -0.60/-0.53 and -0.56/-0.40 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively. It is assumed that the smaller of the two FWHM values (e.g. FWHM Mg[ii]{}$\lambda$2800 BC which is a single VBC corrected term of the doublet) is the best estimate with 0.5 dex uncertainty. ![Continuum subtracted spectral regions of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800 (right) for spectral types A3 (top) and A4 (bottom). Abscissa is rest frame wavelength in Å and ordinate is continuum-normalized intensity. Orange lines trace narrow line and FeI emission, the latter at $\approx$ 2900 Å. The green line represents Fe[ii]{} emission. Thick solid lines show the broad components of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800. \[fig:a3a4\]](a3a4serbia.pdf){width="13.25cm"} Comparison for spectral bins A3 and A4 ====================================== Figure \[fig:a3a4\] shows composite spectra for bins A3 and A4 which involve 43 and 15 sources respectively. Results of best [specfit]{} models are superimposed with residuals displayed below. In many ways these bins represent the most extreme quasars in a low redshift sample. They show the strongest Fe[ii]{} emission ($R_\mathrm{Fe{\sc ii}} >$1.0), a C[iv]{}$\lambda$1549 blueshift/asymmetry with some shift amplitudes exceeding 1000km s$^{-1}$ and a soft X-ray excess ($\Gamma_\mathrm{soft}>$ 2.7). The well known NLSy1 quasar I Zw 1 is found in bin A3. Computed $ \log M_\mathrm{BH}$ values for these sources A3/A4 (8.33/8.27 and 8.41/8.56 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively) coupled with source luminosities lead us to infer that these sources are radiating closest to the Eddington limit (A3/A4 –0.28/–0.18 and –0.37/–0.47 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively). See @marzianietal03d [@marzianietal06] and communication in this proceeding for the cosmological potential of such sources. These sources show surprisingly large scatter in a FWHM Mg[ii]{}$\lambda$2800 vs FWHM H$\beta$ plot and gave the impression that the difference between FWHM H$\beta$ and FWHM Mg[ii]{}$\lambda$2800 was converging towards zero at low FWHM values (see e.g. @wangetal09). In fact they show anomalous properties and are sources where Mg[ii]{}$\lambda$2800 cannot be trusted as a virial estimator. We find that indeed FWHM Mg[ii]{}$\lambda$2800 is equal too or even slightly greater than FWHM H$\beta$ – the only bins where this is observed. In additional the median Mg[ii]{}$\lambda$2800 profiles for these two bins show a systematic blueshift that is largest in bin A4 (–265km s$^{-1}$). The blueshift is attributed to a wind or outflow associated with the high $L/L_\mathrm{Edd}$ values for these sources. The profile of H$\beta$ for these two bins also shows a small blueshifted component perhaps related to the same outflow process. Since it can be well modeled and does not affect H$\beta$ at the FWHM level we conclude that H$\beta$can be trusted as a virial estimator. Comparison for spectral B bins ============================== Figure 4 shows composite spectra for bins B1 and B1+ with associated specfits and residuals superposed and displayed below, respectively. Population B represents a series of bins (52% of our sample) with increasing FWHM H$\beta$. They share a constant range of $R_{FE}<$0.5 except for bin B2 which shows a mean $R_\mathrm{Fe{\sc ii}} \approx 0.5 - 1.0$. Only results for Bins B1 and B1+ are shown. With 218 and 115 sources, respectively, they represent 90% of the population B sample. These sources require a double Gaussian (BC+VBC) fit to both H$\beta$ and Mg[ii]{}$\lambda$2800. Given the large uncertainies about: 1) the widths of both components, 2) the BC/VBC intensity ratio and 3) the amplitude of the VBC redshift, these fits are the least well constrained. We know that the BC/VBC intensity ratio can show a wide range [@marzianietal09]. Some sources in fact show a ratio near to zero (e.g. PG1416-129, @sulenticetal00b and 3C110 @marzianietal10). While compiling our SDSS sample for the H$\beta$ – Mg[ii]{}$\lambda$2800 comparison we found additional examples of what we call super VBC quasars. Figure 5 shows the H$\beta$ and Mg[ii]{}$\lambda$2800 profiles for PG1201+436 which is clearly a case where the BC/VBC ratios for H$\beta$ and Mg[ii]{}$\lambda$2800 profiles are very different. Bins B1 and B1+ show $M_\mathrm{BH}$ values B1/B1+ of 9.15/9.29 and 8.98/9.06 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively. The estimated Eddington ratios are -1.13/-1.44 and -0.96/-1.21 for H$\beta$ and Mg[ii]{}$\lambda$2800 respectively. Since the Mg[ii]{}$\lambda$2800 VBC is weaker in the Mg[ii]{}$\lambda$2800 composites we consider the $M_\mathrm{BH}$ and $L/L_\mathrm{Edd}$ values for that line to be more reliable. ![Continuum subtracted spectral regions of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800 (right) for spectral types B1 (top) and B1$^+$ (bottom). Abscissa is rest frame wavelength in Å and ordinate is continuum-normalized intensity. Orange lines trace narrow line emission, the latter at $\approx$ 2900 Å. The green line represents Fe[ii]{} emission. Thick solid lines show the broad components of H$\beta$ (left) and of Mg[ii]{}$\lambda$2800. \[fig:b1b1p\]](b1b1pserbia.pdf){width="13.25cm"} ![Continuum subtracted spectral regions of H$\beta$ (top) and of Mg[ii]{}$\lambda$2800 (bottom) for source PG 1201+436. Abscissa is radial velocity difference in km s$^{-1}$ with respect to line rest frame wavelength and ordinate is continuum-normalized intensity. \[fig:pg1201\]](pg1201serbia.pdf){width="13.25cm"} Conclusions =========== We are still struggling with $M_\mathrm{BH}$ (and consequently $L/L_\mathrm{Edd}$) estimates largely with $\pm$1dex uncertainties. However we are laying the foundation for estimations with uncertainties of a few 0.1dex–especially bin-to-bin uncertainties which are almost as valuable as absolute ones. The path towards more accurate estimates lies within a context like 4DE1 where spectral binning can greatly increase S/N of the line measures. The latest bin results (for log $L_\mathrm{bol} = 46 \pm $0.5 –a typical quasar luminosity) ) show a trend in the 4DE1 optical plane of decreasing $M_\mathrm{BH}$ from bins B1+/B1++ (9.1) to bin A4 (8.3). Resultant $L/L_\mathrm{Edd}$ values increase from –1.5 to –0.2. The full $M_\mathrm{BH}$ range is likely 7.0 – 9.5 (with $L/L_\mathrm{Edd}$ range from –2.0 to 0.0 in logarithm) but smaller bins and a larger range in source luminosity is needed to explore it. If $M_\mathrm{BH}$ grows with time via merging and accretion then the 4DE1 trend may also represent an evolutionary sequence with quasars in this sample middle-aged and the youngest sources in bins A3/A4 radiating at or near the Eddington limit. #### Acknowledgements 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 listed on the webpage http://www.sdss.org/collaboration/credits.html. [36]{} natexlab\#1[\#1]{}\[2\][\#2]{} , , , () . , , , , , , , , , , () . , , , () . , , , , , () . , , () . , , () . , , () . Gaskell, C. M., Close supermassive binary black holes. Nature 463 (2010), E1. , , , () . , , , , (). , , , () . , S. A. & [Phillips]{}, M. M., , 232 (1979), 659–669. , , , , , , , () . , , , () . , , , , , in New Developments in Black Hole Research, , , p. . , , , , , , , () . , , , , , , () . , , , , , , , , () . , , , , , , () . , R. J. & [Jarvis]{}, M. J.,  337 (2002), 109–116. , , , , , , , , , , , , , , , , , , () . , , , , , () . , H.,  236 (1980), 406–418. , , , , , , () . , , , () . , , , , , , , , , , , , , () . , , , () . , , () . , , () . , Y., [Greene]{}, J. E., [Strauss]{}, M. A., [Richards]{}, G. T., & [Schneider]{}, D. P.,  680 (2008), 169–190. , , , , () . , , , , , , () . , , , () . , , , , () . , , , , , , () . , , , , , , , () . Trakhtenbrot, B., Netzer, H. 2012. Black hole growth to z = 2 - I. Improved virial methods for measuring M$_{BH}$ and L/L$_{Edd}$. Monthly Notices of the Royal Astronomical Society 427, 3081-3102. Vanden Berk, D. E., and 61 colleagues 2001. Composite Quasar Spectra from the Sloan Digital Sky Survey. Astrono. J. 122, 549-564. , , , , , , , , , , () . , , , , () . , , , , () . , , , , , () .

ArXiv

--- author: - | $^a$, Chris Monahan$^b$, Christine Davies$^c$, Eduardo Follana$^d$, Ron Horgan$^e$, Peter Lepage$^f$, Junko Shigemitsu$^g$\ ALCF, Argonne National Laboratory, Argonne, IL 60439, USA\ Department of Physics, College of William and Mary, Williamsburg, VA 23187, USA\ SUPA, School of Physics & Astronomy, University of Glasgow, Glasgow, G12 8QQ, UK\ Departamento de Fisica Teorica, Universidad de Zaragoza, E-50009 Zaragoza, Spain\ DAMTP, Cambridge University, Cambridge, CB3 0WA, UK\ LEPP, Cornell University, Ithaca, NY 14853, USA\ Department of Physcis, The Ohio State University, Columbus, OH 43210, USA\ E-mail: title: Precise Determinations of the Decay Constants of $B$ and $D$ mesons --- Introduction ============ Investigating the flavor structure of the Standard Model (SM) is important for its own sake; however, it is even more interesting since it can lead to physics beyond the SM. Furthermore, the data accumulations and new analysis emerging from the LHC suggest more and more that the Higgs is very close to the SM Higgs. They even do not find any hint of new particles yet. In this situation, precise understanding of the SM in the flavor sector becomes more critical. Decay constants of heavy-light pseudoscalar mesons have been studied from lattice QCD for quite some time. We can determine the corresponding CKM matrix elements by combining decay constants from theory and decay rates from experiments. Moreover, decay constants are basic quantities related to many hadronic quantities. For instance, $f_B$ is an important input parameter for inclusive determinations of the CKM matrix elements. The decay constants can also be used for testing the lattice formalism, since the calculations have typically smaller errors and the procedures are relatively straightforward. This paper presents new calculations for $B$, $B_s$, $D$ and $D_s$ meson decay constants from HPQCD. We have completed the projects, and published the results in two papers [@fb][@fd]. So, essentially this proceeding consists of a brief summary of the two papers and a discussion of their impact on the CKM matrix elements $|V_{cd}|$ and $|V_{cs}|$. $B$ and $B_s$ meson decay constants =================================== We used the MILC AsqTad $N_f=2+1$ gauge configurations with NRQCD (Nonrelativistic QCD) $b$ quarks for this project. The previous HPQCD calculation [@fb2] used AsqTad light and strange valence quarks; however, in this work we used the highly improved staggered quark (HISQ) action for the valence quarks. The HISQ action has much smaller discretization effects, so one can expect improvements in the continuum extrapolation errors. We include one more ensemble (F0); $40^3\times96$ with $m_l/m_s=0.0031/0.031$, which is a more-chiral fine ensemble. Details of the lattice configurations are in Tab. \[enb\]. Moreover, this calculation is done with the new scale parameter $r_1$ values. HPQCD was using $r_1=0.321(5)$ fm extracted from $\Upsilon$ splittings [@oldr1]. In 2010 HPQCD published a much more accurate $r_1$ determination, $r_1=0.3133(23)$ fm, based on several physical quantities and an improved continuum extrapolation with 5 lattice spacings [@newr1]. We needed to re-tune valence quark masses due to the scale changes, and updating for such different new settings is one of the main purposes of this analysis. We used the spin averaged $\Upsilon$ mass to tune the bare bottom quark mass, and (fictitious) $\eta_s$ for the strange quark mass. Fig. \[tune\] shows the tuning for $b$ (left) and $s$ (right) quarks. As one can see, our physical target meson mass has a large error compared to the deviations of the alignment of the tuning measurements. We found that it is important to tune quark masses precisely up to the statistical errors or $r_1/a$ errors of the tuning measurements. This precise tuning ensures correct estimation for $\chi^2$ of chiral and continuum extrapolation. If one has large deviations between the tuning measurements more than its statistical errors, then the chiral and continuum extrapolations may suffer from additional systematic errors. Set $r_1/a$ $m_l/m_s$ (sea) $N_{conf}$ $N_{tsrc}$ $L^3 \times N_t$ ----- --------- ----------------- ------------ ------------ ------------------ C1 2.647 0.005/0.050 1200 2 $24^3 \times 64$ C2 2.618 0.010/0.050 1200 2 $20^3 \times 64$ C3 2.644 0.020/0.050 600 2 $20^3 \times 64$ F0 3.695 0.0031/0.031 600 4 $40^3 \times 96$ F1 3.699 0.0062/0.031 1200 4 $28^3 \times 96$ F2 3.712 0.0124/0.031 600 4 $28^3 \times 96$ : Simulation details on three “coarse” and three “fine” MILC AsqTad ensembles. $N_{conf}$ is the number of the configurations that were used in the simulation, and $N_{tsrc}$ is the number of time sources per configuration.[]{data-label="enb"} ![Tuning of the $b$ (left) and $s$ (right) quark masses. []{data-label="tune"}](mkin.ps "fig:"){width=".45\textwidth"} ![Tuning of the $b$ (left) and $s$ (right) quark masses. []{data-label="tune"}](etas.final.ps "fig:"){width=".45\textwidth"} We calculated operator matching factors in full QCD at one-loop through order $\alpha_s$, $\frac{\Lambda_{QCD}}{M}$, $\frac{\alpha_s}{aM}$, $a \alpha_s$, and $\alpha_s \frac{\Lambda_{QCD}}{M}$. These matching calculations were presented separately at this conference [@mat]. We also employed random-wall sources for the HISQ propagators and Gaussian smearing sources for the NRQCD propagators. Including all statistical and systematic errors, we obtained $$f_B = 0.191(9) {\rm GeV}, \;\;\;\;\;\; f_{B_s}=0.228(10) {\rm GeV},$$ and $$\frac{f_{B_s}}{f_B}=1.188(18).$$ These calculations are a definite improvement on our previous calculations [@fb2], $f_B=0.190(13) {\rm GeV}$, $f_{B_s}=0.231(15) {\rm GeV}$, and $f_{B_s}/f_B=1.226(26)$. The largest source of errors is the operator matching error for the decay constants. The ratio $f_{B_s}/f_B$ has very small errors, since most of the matching factors are canceled. If one calculates the decay constants without matching factors, one could reduce around 5 % errors down to 1 $\sim$ 2 % errors. Recently, HPQCD has calculated $f_{B_s}$ without matching factors [@fbs], and obtained $f_{B_s}=0.225(4) {\rm GeV}$ with only 1.8 % errors. Essentially, this very precise calculation utilizes the HISQ action for the $b$ quark. The HISQ action can be used to simulate the charm quark, but for the bottom quark it is very difficult with current technology. The clever idea was that in fact we can simulate a heavy quark heavier than the charm quark but lighter than the bottom quark. Once one gets heavy meson correlators depending on multiple heavy quark masses, then one can extrapolate to the physical bottom quark mass. In this way, one can determine the decay constants without matching factors. Of course, we can apply this heavy HISQ method for $f_B$, but it would be difficult. First of all, the light quark is much more expensive than the strange quark, and for $f_B$ we need to perform additional chiral extrapolation. Thus, for now we fix $f_B$ by combining ${f_{B_s}}/{f_B}$ from the NRQCD analysis and $f_{B_s}$ from the heavy HISQ analysis; $$f_B \equiv \bigg [ \frac{f_{B_s}}{f_B} \bigg]^{-1}_{NRQCD} \times f_{B_s}^{HISQ} = 0.189(4) {\rm GeV}.$$ This $f_B$ result with 2 % total error is the most accurate $f_B$ available today. Comparisons of results for $f_B$ (left) and $f_{B_s}$ (right) are shown in Fig.  \[bmeson\] ![Comparisons of results for $f_B$ (left) and $f_{B_s}$ (right) from this analysis, previous HPQCD, Fermilab/MILC and ETM collaborations.[]{data-label="bmeson"}](fb.ps "fig:"){width=".45\textwidth"} ![Comparisons of results for $f_B$ (left) and $f_{B_s}$ (right) from this analysis, previous HPQCD, Fermilab/MILC and ETM collaborations.[]{data-label="bmeson"}](fbs.ps "fig:"){width=".45\textwidth"} $D$ and $D_s$ meson decay constants =================================== With the same simulation setting shown in Tab. \[enb\], we determined $D$ and $D_s$ meson decay constants. The discretization error of the HISQ action starts at $\mathcal{O}(\alpha_s (am_h)^2 v^2/c^2)$ and $\mathcal{O}((am_h)^4 v^2/c^2)$, and this provides enough accuracy to simulate relativistic charm quarks on current typical lattices. So, we apply the HISQ action for all valence quarks including the charm quark. Thus, we can evaluate the decay constants without matching factors, since the HISQ action exhibits the chiral symmetry in the continuum limit. The decay constant can be written with the heavy-light axial vector current $A_\mu = \overline{\Psi}_q \gamma_\mu \gamma_5 \Psi_c$, with $q=s$ or $d$, as $$<0|A_\mu|D> = p_\mu f_{D_q}.$$ We can express the decay constant in terms of the pseudoscalar density $PS=\overline{\Psi}_q \gamma_5 \Psi_c$, as we used for light meson decay constants, $f_\pi$ and $f_K$, $$f_{D_q} = \frac{m_c+m_q}{M_{D_q}^2} <0|PS|D_q>.$$ As we did for the $B$ decay constants, we also re-tuned the charm quark mass for the new $r_1=0.3133(23)$ fm. We used $\eta_c$ mass to fix the charm quark mass. HPQCD updated $f_{D_s}$ with the new $r_1$ in 2010 [@fds] already, so this work is mainly to update $f_D$ with the new $r_1$. Our final results are $$f_D = 208.3 (1.0)_{stat.}(3.3)_{sys.} {\rm MeV}, \;\;\;\;\;\; f_{D_s}= 246.0(0.7)_{stat.} (3.5)_{sys.} {\rm MeV},$$ and, $$\frac{f_{D_s}}{f_D}=1.187(4)_{stat.}(12)_{sys.},$$ which show good agreement with our previous determinations. One interesting question would be what the impact of the scale change is. Tab. \[impact\] summarizes HPQCD’s determinations of the decay constants with the old and new scale factor $r_1$. We found no significant effect due to the scale change. (One exception is for $f_{D_s}$ with the old $r_1$ and the most accurate result with the new $r_1$.) It appears that for the decay constants re-tuning of quark masses largely compensates the shift from overall scale change. Thus, predicting the impact of the scale change before the actual calculations would be risky. We will investigate the impact of the scale change further in the future. This study would lead to better estimation of systematic errors for the scale setting. Old $r_1$ New $r_1$ --------------- ----------- ----------------- $f_{D_s}$ 241(3) 246(4), 248(3) $f_D$ 207(4) 208(3) $f_{D_s}/f_D$ 1.164(11) 1.187(12) $f_{B_s}$ 231(15) 228(10), 225(4) $f_B$ 190(13) 191(9), 189(4) $f_{B_s}/f_B$ 1.226(26) 1.188(18) $f_K$ 157(2) 159(2) $f_\pi$ 132(2) 132(2) : Decay constants from HPQCD with the old $r_1=0.321(5) $fm and the new $r_1=0.3133(23) $fm. The unit of the decay constants is MeV, and the ratios are dimensionless. []{data-label="impact"} In Fig. \[dmeson\], we compare $f_D$ and $f_{D_s}$ results from FNAL/MILC [@milc], HPQCD [@fds][@fds2], ETMC [@etmc], and PACS-CS [@pacs]. The comparisons include FNAL/MILC’s preliminary results with $N_f=2+1+1$ including simulations at the physical pion mass, and ETMC’s preliminary results with $N_f=2+1+1$. In their preliminary results, they achieve a good precision that is comparable to our best results, and they show very good agreement with HPQCD. ![Comparisons of results for $f_D$ (left) and $f_{D_s}$ (right). The results of this proceeding are shown under HPQCD 2012. []{data-label="dmeson"}](fD.ps "fig:"){width=".45\textwidth"} ![Comparisons of results for $f_D$ (left) and $f_{D_s}$ (right). The results of this proceeding are shown under HPQCD 2012. []{data-label="dmeson"}](fDs.ps "fig:"){width=".45\textwidth"} Combining our decay constant results with branching fractions from experiments, we can obtain corresponding CKM matrix elements. See Fig. \[ckm\] for the comparisons. For $|V_{cd}|$, one can immediately notice that the leptonic determination and the semileptonic determination are in good agreement. Those two results were obtained with completely different systematics for both lattice and experiment analysis, since the decay channels are quite different. This is a highly non-trivial check for lattice formulation and experiments. Those two lattice determinations of $|V_{cd}|$ demonstrate good agreement with the unitarity point as well. Thus, we do not see any signature of new physics here yet. We note that now the precision of the lattice determination of $|V_{cd}|$, especially with the preliminary branching fraction result from BES III is actually better than the accuracy of the determination from neutrino experiments. So far, the PDG quotes the neutrino experiment result for $|V_{cd}|$. This is simply because lattice determinations had much larger errors in the past. ![Comparisons of results for $|V_{cd}|$ (left) and $|V_{cs}|$ (right) from leptonic decays and semileptonic decays. []{data-label="ckm"}](Vcd.ps "fig:"){width=".45\textwidth"} ![Comparisons of results for $|V_{cd}|$ (left) and $|V_{cs}|$ (right) from leptonic decays and semileptonic decays. []{data-label="ckm"}](Vcs.ps "fig:"){width=".45\textwidth"} For $|V_{cs}|$, the situation is more interesting. The right plot of Fig. \[ckm\] shows that the semileptonic determination and the unitarity point are in good agreement. However, for the leptonic determinations, it shows some discrepancies depending on the decay channels and experiments. If we only consider average leptonic determinations with averages of HFAG, then the $|V_{cs}|$ represents more than 1 $\sigma$ discrepancy between the unitarity point, which is the identical to observation of the $f_{D_s}$ puzzle. The $f_{D_s}$ puzzle [@puzzle] was a 4 $\sigma$ tension between experiments and lattice determinations of $f_{D_s}$ in around 2010. The puzzle is no longer a puzzle, since new experiments and lattice analysis results have moved closer to each other. Now the difference is about 1.6 $\sigma$. When experiments determine $f_{D_s}$, they use the unitarity $|V_{cs}|$. Thus, the $f_{D_s}$ puzzle indicates a difference between $|V_{cs}|$ from the unitarity point and our determination from the leptonic decay of $D_s$ meson. In Fig. \[ckm\], we show the results with two Belle’s 2012 preliminary results, the first is for $D_s \rightarrow \mu \nu$ and the second is for $D_s \rightarrow \tau \nu$; these results suggest that we need to wait to see the experiments attain more accuracy. It may still be possible to see some discrepancy in determinations of $|V_{cs}|$, and this could be a hint for new physics. This work was supported by the DOE and the NSF in the U.S., by the STFC in the U.K., by MICINN and DGIID-DGA in Spain, and by ITN-STRONGnet in the EU. Numerical simulations were carried out on facilities of the USQCD collaboration funded by the Office of Science of the DOE and at the Ohio Supercomputer Center. We thank the MILC collaboration for use of their gauge configurations. [99]{} H. Na [*et al.*]{} \[HPQCD\], Phys. Rev. D [**86**]{}, 034506 (2012). H. Na [*et al.*]{} \[HPQCD\], Phys. Rev. D [**86**]{}, 054510 (2012). E. Gamiz [*et al.*]{} \[HPQCD\], Phys. Rev. D [**80**]{}, 014503 (2009). A. Gray [*et al.*]{} \[HPQCD\], Phys. Rev. D [**72**]{}, 094507 (2005). C. Davies [*et al.*]{} \[HPQCD\], Phys. Rev. D [**81**]{}, 034506 (2010). C. Monahan [*et al.*]{} \[HPQCD\], \[[arXiv:1210.7016]{}\], ,\ C. Monahan [*et al.*]{} \[HPQCD\], \[[arXiv:1211.6966]{}\]. C. McNeile [*et al.*]{} \[HPQCD\], Phys. Rev. D [**85**]{}, 031503 (2012). C. Davies [*et al.*]{} \[HPQCD\], Phys. Rev. D [**82**]{}, 114504 (2010). C. Aubin [*et al.*]{} \[MILC\], Phys. Rev. Lett [**95**]{}, 122002 (2005),\ A. Bazavov [*et al.*]{} \[MILC\], Phys. Rev. D [**85**]{}, 114506 (2012),\ A. Bazavov [*et al.*]{} \[MILC\], \[[arXiv:1210.8431]{}\], . E. Follana [*et al.*]{} \[HPQCD\], Phys. Rev. Lett [**100**]{}, 062002 (2008). B. Blossier [*et al.*]{} \[ETMC\], JHEP 0907, 043 (2009),\ P. Dimopoulos [*et al.*]{} \[ETMC\], JHEP 01, 046 (2012),\ K. Jansen \[ETMC\], BNL lattice workshop (2012). Y. Namekawa [*et al.*]{} \[PACS-CS\], Phys. Rev. D [**84**]{}, 074505 (2011) A. Kronfeld, \[[arXiv:0912.0543]{}\].

ArXiv

--- abstract: 'We have developed a novel technique for detection of spin polarization with a quantum dot weakly coupled to the objective device. The disturbance to the object in this technique is very small since the detection is performed through sampling of single electrons in the object with very slow rate. We have applied the method to a quantum point contact (QPC) under a spin-orbit interaction. A high degree of spin polarization in the vicinity of the QPC was detected when the conductance stayed on a plateau at a half of the unit conductance quantum ($G_{\rm q}/2\equiv e^2/h$), and also on another plateau at $2e^2/h$. On the half-quantum plateau, the degree of polarization $P$ decreased with the bias source-drain voltage of the QPC while $P$ increased on the single-quantum plateau, manifesting that different mechanisms of polarization were working on these plateaus. Very long spin relaxation times in the detector quantum dot probably due to dynamical nuclear spin polarization were observed. Anomalous decrease of $P$ around zero-bias was observed at a Kondo-like resonance peak.' author: - Sunwoo Kim - Yoshiaki Hashimoto - Taketomo Nakamura - Shingo Katsumoto title: 'Spin-polarization in the vicinity of quantum point contact with spin-orbit interaction' --- Introduction ============ Creation and detection of spin polarization in quantum structures of non-magnetic semiconductors are key-techniques in semiconductor spintronics.[@awschalom2007challenges] The use of spin-orbit interaction (SOI) is a candidate for the creation of spin current and a number of combinations with quantum structures such as a bent quantum wire,[@PhysRevB.72.115321] a quantum point contact (QPC),[@doi:10.1143/JPSJ.74.1934] or an Aharonov-Bohm (AB) interferometer[@PhysRevB.84.035323] have been proposed though in the authors’ view, there has been no report of sound experimental evidence of these devices. A conductance plateau at a half of quantum conductance $G_{\rm q}$ ($G_{\rm q}\equiv 2e^2/h$, $e$ being the elementary charge, $h$ the Planck constant) in a quantum point contact (QPC) with an SOI (here we call it the “0.5 anomaly”) was found in experiments [@debray2009all] and considered as the sign of perfectly spin-polarized current. The possible spin polarization is now attracting the attention of researchers for the use as a spin source. However there is no evidence that the current flowing at the 0.5 anomaly is spin-polarized, other than the transport throughout the samples including the shot noise measurement.[@kohda2012spin] A theoretical explanation of the 0.5 anomaly was given by Wan [*et al.*]{}.[@PhysRevB.80.155440] They showed the possibility that the Coulomb interaction and the Rashba-type SOI (RSOI) form a spin-dependent potential, which gives different threshold gate voltages to the spin subbands. The spin-dependence is reversed for the electrons with the opposite momentum, hence this results in a persistent spin-current through the QPCs. Their calculation also shows a kind of spin-standing wave in the vicinity of QPC, which means the appearance of some non-uniform spin accumulation. Another explanation was given by Kohda [*et al.* ]{}, who considered non-uniformity in the effective magnetic field of RSOI due to some gradient in the confinement potential of the QPC.[@kohda2012spin] The field gradient gives a direction-dependent force on electron spins, just like the Stern-Gerlach experiment. Spin polarization due to the RSOI was also predicted on the ordinary conductance plateau at $G_{\rm q}$.[@doi:10.1143/JPSJ.74.1934] The polarization in this case is caused by non-equilibrium occupation of quantum levels split by the RSOI. The prediction, however, has not been examined in experiment, mainly because the polarization in this mechanism occurs after electrons pass through the narrowest part of the QPC and no anomaly appears in the conductance. In addition, the small number of polarized electrons are rapidly diluted when they flow out into broader electrodes though the degree of polarization in the vicinity of the QPC should be very high. According to the fundamental law of symmetry, such a spontaneous spin polarization does not occur in infinite uniform systems, in other words, systems with the spatial translational symmetry, without any mechanism to break the time-reversal symmetry such as external magnetic field or ferromagnetic interaction between magnetic moments. While the former introduces the symmetry breaking directly into single-electron Hamiltonian, the latter brings in spontaneous symmetry breaking through a many-body effect. (Even with the ferromagnetic interaction, it is proved that there is no long range ferromagnetic order in one-dimensional systems, though this issue is out of the present scope.) In finite systems, on the other hand, local spin polarization naturally appears. A quantum dot with an odd number of electrons or in a high-spin state is a representative example. Spontaneous polarization due to the breaking of uniformity, or translational symmetry is also predicted and has been claimed to be the origin of so called the 0.7 anomaly in QPCs with very small SOI. In this paper we propose a novel method to detect spin polarization with very small disturbance to the objective systems. It utilizes a quantum dot coupled to a side-edge of the system. The dot samples the electrons from the edge with very slow rate, the lower bound of which is determined by the spin-relaxation time in the quantum dot. We have applied the method to QPCs with a moderate strength of spin-orbit coupling and found that high spin polarization rate is obtained not only on the plateau with the conductance $G_{\rm q}/2$ (0.5 plateau) but also on the one with $G_{\rm q}$ (1.0 plateau). The bias dependencies of the polarization on these plateaus were opposite to each other, manifesting the different mechanisms for the polarization. The spin-relaxation time estimated from the duration-controlled experiment is very long probably due to dynamic nuclear polarization when the polarization rate of electrons is high. We also observe Kondo-like enhancement of spin scattering rate around the zero-bias. Experiment ========== Principle of polarization detection ----------------------------------- The principle of the polarization detection is schematically illustrated in Fig.\[fig\_exp\](a1) and (a2). The method is based on the measurement of two-electron tunneling rate from an objective into a QD, which is prepared in known electronic states. For rough sketch of the principle, we assume that the total electron concentration in the objective $n_{\rm t}\equiv n_\uparrow+n_\downarrow$ ($n_\sigma,\;\sigma=\uparrow,\;\downarrow$, is the electron concentration with spin $\sigma$) is fixed and that the tunnel coupling between the objective and the QD is common for all the spin and the orbital states. Here we write the averaged time for a tunneling event of a single electron as $\tau_{\rm p}$. ![\[fig\_exp\] (a1) and (a2) illustrate the principle for detection of spin-polarization by the use of two-electron tunneling processes. (a1) is for a spin-unpolarized target (objective). The Pauli principle does not affect the tunneling rate. (a2) is for 100 % ($P=1$) spin polarization. In 0$\rightarrow$2 transition, the Pauli principle blocks the second electron to tunnel. In 1$\rightarrow$3 transition, the second electron is allowed to tunnel when the initially occupying electron has down-spin. “s” and “p” denote the first orbital state and the second one respectively. (b) Cross sectional view of the layered structure. (c) Scanning electron micrograph of a sample showing the gate configuration. Whiter regions are Au/Ti Schottky gates. ](exp.eps){width="\linewidth"} We first consider the case in which the initial state of the QD has no electron (the process 0$\rightarrow$2). When the spin polarization $$P\equiv (n_\uparrow-n_\downarrow)/(n_\uparrow+n_\downarrow) \label{eq_polarization_definition}$$ in the target device (objective) is equal to 1, [*i.e.*]{}, the objective has only spin-up electrons, the lowest energy level in the QD denoted as “s” in Fig.\[fig\_exp\](a1) and (a2) can accommodate only one spin-up electron due to the Pauli exclusion principle until the spin flip occurs in the QD. Hence if we take the averaged spin flip time $\tau_{\rm f}$ as infinite, the time-averaged charge $q_{\rm av}$ in the QD over a period $\tau_0$ should be $e(\tau_0-\tau_{\rm p})/\tau_0$ for $\tau_0\gg \tau_{\rm p}$. In the other extreme of $P=0$, $n_\uparrow=n_\downarrow=n_{\rm t}/2$ and $q_{\rm av}$ should be $2e(\tau_0-2\tau_{\rm p})/\tau_0$. When $\tau_0\gg \tau_{\rm p}$, naturally $q_{\rm av}$ is twice of that in the case $P=1$. The two extrema show that $q_{\rm av}$ can work as a measure of $P$. In experiments, the number of electrons and $q_{\rm av}$ in QDs can be measured [*e.g.*]{}, with remote sensing. However, single measurement of $q_{\rm av}$ does not give the value of $P$ directly and some supporting data such as the value of $\tau_{\rm p}$ are required. There are various ways to solve it and here we consider to utilize the tunneling process of 1$\rightarrow$3. For $P=1$ and an initial state of the QD with an up-spin electron, the Pauli principle also blocks the tunneling to the s-state while it does not block the one to the second state (p-state). When the initially occupying electron has down-spin, the tunneling to s-state is allowed. We assume the initial up/down spin states appear with even probability and obtain $q_{\rm av}$ in this case as $3e(\tau_0-\tau_{\rm p})/2\tau_0$. Hence taking the ratio of $q_{\rm av}$ for these two processes, one can obtain the value of $P$. In more realistic models, the analysis should be more complicated. Within the relaxation time approximation, however, simple analytic formulas can be obtained with solving some rate equations as shown in Appendix. Sample configuration -------------------- A two-dimensional electron system (2DES) in a pseudomorphic ${\rm In_{0.1}Ga_{0.9}As}$ quantum well was grown on a GaAs (001) substrate with ordinary molecular beam epitaxy. Figure \[fig\_exp\](b) shows the layered structure, in which two GaAs-(In,Ga)As interfaces are placed in the electric field produced by the modulation-doped layer. The RSOI in such 2DES structures is known to be strong due to the asymmetry in the two interfaces.[@raey; @PhysRevB.68.035315] The carrier concentration and the Hall mobility are $9.8\times 10^{11}~{\rm cm^{-2}}$ and $7.4\times 10^{4}~{\rm cm^2/Vs}$ respectively at 4.2 K. QPCs and a QD were defined by Au/Ti split gates fabricated with electron-beam lithography. Figure \[fig\_exp\](c) shows a scanning electron micrograph of the metallic gates (whiter regions) with captions being overlapped. The constriction shown in the right side worked as the QPC for remote charge detection (d-QPC) while that in the left side worked as the QPC which showed the 0.5 anomaly (t-QPC). Between these two one-dimensional channels, a QD, which worked as the detector of polarization, was placed and the connection to the left channel existed just next to the narrowest point (slightly upper side in the figure). An advantage in such side-coupled QD structure is the access to $N_{\rm D}=0$ state without killing tunneling conductance to the electrode (in the present case, t-QPC).[@doi:10.1143/JPSJ.76.084706] We fabricated three samples (sample A, B, C) with the same gate configuration. Low-temperature measurements were preformed on the three and essentially the same results were obtained for the spin-polarization on the 0.5 plateaus. Measurements of the spin-polarization for various t-QPC conductances and bias conditions were carried out on one of the three (sample C). The Kondo-like conductance peak was also observed in this sample. Measurement ----------- The specimens were cooled down to 100 mK in a dilution fridge. A care was taken for the cooling time to be longer than 12 hours from room temperature to 4.2 K. During the cooling process, the split gates were biased at $+0.2$ V for getting leakage-free Schottky gate characteristics with smaller hysteresis in the response of conductance.[@PhysRevB.72.115331] In order to avoid the cross-coupling between the plunger gate and the other terminals, and also to go up to microwave frequencies, the lines to the gates of QD and d-QPC were through coaxial cables. The electrostatic potential of the detector QD was driven by a square wave superposed on the gate voltage. Tunneling of electrons from t-QPC to the QD is detected as the modulation of conductance in d-QPC (remote charge detection). To obtain the charge sensing signal, we adopted time averaging synchronized to a square wave on the QD gate voltage with lock-in technique.[@/content/aip/journal/apl/84/23/10.1063/1.1757023] The remote charge detection is based on the electrostatic sensitivity of d-QPC to the electric field formed by the electrons in the QD. However the gate voltage of QD also modulates the G-V characteristics in t-QPC effectively as shifts in the gate voltage of t-QPC. Thus the modulation was compensated with a software control of the gate voltage in t-QPC for keeping charge sensitivity constant against the gate voltage in QD as far as possible and the residual modulation was normalized with the sensitivity measurement prior to the charge sensing measurement. The signal can thus be transformed to the time-averaged charge $q_{\rm av}$ (with reference to the initial state) with the knowledge of system parameters. Results and Discussion ====================== ![\[fig\_tqpc\] Typical gate voltage ($V_{\rm tQPC}$) dependence of conductance through t-QPC ($G_{\rm tQPC}$). Conductance quantization in units of half conductance quantum ($G_{\rm q}/2=e^2/h$) is observed up to $2G_{\rm q}$. []{data-label="fig_gate_voltage_response"}](tqpc.eps){width="0.9\linewidth"} Figure \[fig\_gate\_voltage\_response\] shows a typical gate-voltage dependence of the conductance (G-V characteristics) in t-QPC $G_{\rm tQPC}$. Clear plateau structures appeared around $G_{\rm q}$ and also around $G_{\rm q}/2$ reproducing preceding studies.[@debray2009all; @kohda2012spin] ![\[fig\_signal1\] (a) Typical charge detection signal as a function of the dot plunger gate voltage. The insets are chemical potential diagrams for t-QPC and detector QD. Dip signal appears when $E_{\rm F}$ in t-QPC is within the square wave swing of QD discrete levels hence the dip width is the same as the swing width $V_{\rm sw}$. (b) Gray scale plot of the signal depth versus the plane of $V_{\rm p}$ and $V_{\rm sw}$. White letters and symbols indicate the corresponding tunneling process with the changes in the electron number. The data are for a different sample from those in (a). []{data-label="fig_charge_detection_signal1"}](sig1.eps){width="\linewidth"} Figure \[fig\_charge\_detection\_signal1\](a) shows an example of the charge detection signal as a function of the QD gate voltage $V_{\rm QD}$. The signal appears as dips around the points where, as illustrated in Fig.\[fig\_charge\_detection\_signal1\](a), the energy (chemical potential) levels in the QD matches with the Fermi energy in t-QPC. As also illustrated in Fig.\[fig\_charge\_detection\_signal1\](a), the width of the signal in $V_{\rm QD}$ is the same as the amplitude of the square wave $V_{\rm sw}$. The series of dips ended at $V_{\rm QD}=-1.2$ V, which indicates $N_{\rm D}$ was zero for the lower $V_{\rm QD}$. And we can thus assign the regions in $V_{\rm QD}$ between the dips to $N_{\rm D}$. With increasing $V_{\rm sw}$, the dip width widened with flat bottoms. Hence the signal formed wedge-like regions in the plane of $V_{\rm QD}$-$V_{\rm sw}$ as shown in Fig.\[fig\_charge\_detection\_signal1\](b), where the signal is plotted in a gray-scale. The data here were obtained in the open condition of t-QPC, that is, the electrons in the source were not spin-polarized. If there was some change in the tunneling parameters it should appear as a change in the signal level.[@/content/aip/journal/apl/84/23/10.1063/1.1757023] The flatness over the bottoms of the signal regions thus means there was no significant change in the tunneling parameters within the section of $V_{\rm QD}$ in Fig.\[fig\_charge\_detection\_signal1\](b). On the other hand in the overlapped regions of neighboring wedges, clear deepening of the signal due to double-electron tunneling process was observed. We thus can assign single or double electron tunneling process to every wedge-shaped region in Fig.\[fig\_charge\_detection\_signal1\](b) as noted with white characters and symbols. Here we represent the tunneling process with the change in $N_{\rm D}$, such as 0$\rightarrow$1 for the process, in which $N_{\rm D}$ changes from 0 to 1. Note that the tunneling rate was strongly reduced and the average time for the tunneling was elongated to 50 s, which corresponds to 3 aA and the net (DC) current was zero. In Fig.\[fig\_comparison\_polarization\], we compare the signal levels for t-QPC opened widely (left panels) and for t-QPC conductance equals $G_{\rm q}/2$ (right panels). The data in the left panels (a) and (b) are very similar to those in Fig.\[fig\_charge\_detection\_signal1\](b) besides the shifts due to the difference in the sample characteristics. The signal depths for the processes 0$\rightarrow$2 and 1$\rightarrow$3 are the same within the noise level manifesting that the spin-polarization in wide-opened QPC is almost zero. When t-QPC was pinched for the conductance to decrease and to be quantized at $G_{\rm q}/2$, the pattern changed as in the right panels in Fig.\[fig\_comparison\_polarization\](a), (b), in which an apparent shallowing of the signal for 0$\rightarrow$2 double electron tunneling region was observed. From the qualitative discussion in the previous section, this should come from finite spin-polarization in t-QPC. Since the electric current through t-QPC was kept to 0 while the measurement, the polarization was not due to the non-equilibrium electric current induced by the boundary conditions (the chemical potentials of the connected reservoirs). Though the SOI does not break the time-reversal symmetry in the total Hamiltonian, it may generate magnetic dipoles resulting in local spin-polarization, which probably has been detected in the present experiment. ![\[fig\_spinpolsig\] (a) Gray scale plots of the signal depth for open (left) and 0.5 plateau (right) conditions of t-QPC on the plane $V_{\rm sw}$ versus $V_{\rm p}$. White broken lines in the right panel are just for eyes because the contrast is comparatively weak. The shifts in the abscissae and in the ordinates are due to the electrostatic coupling between the QD and the Schottky gate of t-QPC. (b) Signal profiles along the horizontal white lines in the gray scale plots in (a). (c) Signal intensities at transitions 0$\rightarrow$2 and 1$\rightarrow$3 as a function of period in the square wave driving the QD. []{data-label="fig_comparison_polarization"}](spinpolsig.eps){width="\linewidth"} To make the analysis more quantitative, we measured the signal intensities for these two regions ($I_{0\rightarrow2}$, $I_{1\rightarrow3}$) as a funtion of the square wave integraion time $\tau_{\rm sw}$ as shown in Fig.\[fig\_comparison\_polarization\](c). When t-QPC was open (the left panel), $I_{0\rightarrow2}$, $I_{1\rightarrow3}$ were at the same level but a small difference due to that in the tunneling rate. On the other hand for 0.5 plateau (the right panel), $I_{1\rightarrow3}$ was ovbiously larger than $I_{0\rightarrow2}$ and the saturation $\tau_{\rm sw}$ was faster for $I_{1\rightarrow3}$ reflecting the spin-polarization. Applying a simple phenomenological rate equation analysis given in Appendix A, we can fit the data of $I_{0\rightarrow2}(t_{\rm sw})$ $I_{1\rightarrow3}(t_{\rm sw})$ with $P$, $\tau_{\rm sf}$, $\tau_{\rm t}$ as fitting parameters, which can thus be obtained through the analysis. Though the fitting is successful shown as the solid curves, there are some distributions in the data and we need to pay attention to the errors in the fitted values. We can also obtain the longitudinal spin relaxation time $\tau_{\rm sf}$ in the dot as 200 $-$100$+$250 s. Note that this is so called $T_1$ in the terms of magnetic resonance[@Slichter199603]. In weak magnetic fields, that is with finite Zeeman energy, extremely long $T_1$’s reaching ms region have been reported for electron spins in few-electron quantum dots[@fujisawa2002allowed; @johnson2005triplet; @PhysRevLett.91.196802; @PhysRevLett.98.126601; @doi:10.1143/JPSJ.75.054702]. On the other hand, rapid shrinkage of the relaxation time due to fluctuating nuclear spins has generally been observed with decreasing the energy separation between corresponding spin states to zero. In the present case, the relaxation time of s order is expected referring to preceding results because no external magnetic field is applied and the energy difference is zero. We need to consider, however, that electrons captured by the QD are highly polarized, and dynamic nuclear polarization (DNP)[@PhysRevLett.100.067601] should take place inside the QD through the Overhauser effect.[@PhysRevLett.99.096804] It is not surprising then, that $\tau_{\rm sf}$ is much longer than s order because fluctuation of nuclear spins in the QD is strongly suppressed by the DNP. From the fitting we obtained $P$ as 0.7 $\pm$ 0.1 for the right panel in Fig.\[fig\_comparison\_polarization\](c). Note that the measured polarization is for the electrons in the vicinity of sampling gate to the QD and naturally lower than 1 even if the outgoing electrons from t-QPC is perfectly polarized. ![(a) Conductance of a t-QPC as a function of the gate voltage. The following data of spin-poalrization has been taken for this sample. Vertical arrows 1 and 3 indicate the positions where the data in (c) and (d) have been taken respectively. (b) Zero-bias spin polarization as a function of t-QPC conductance. Data $1.5G_{\rm q}\sim2.0G_{\rm q}$ are missing due to experimental limitation. Doubling of the data points at $G_{\rm qpc}=G_{\rm q}/2$ are due to the difference in the results at arrow 1 and 2 in (a), in the former of which we have the Kondo-like conductance enhancement. (c) Spin-polarization in t-QPC on 0.5 conductance plateau as a function of the source-drain bias voltage ($V_{\rm sd}$). The vertical arrow indicates the zero-bias, where a characteristic decrease in $P$ is observed. (d) Results of the same measurement as (c) on 1.0 conductance plateau. \[fig\_sdbiasdep\] ](sdbiasdep.eps){width="\linewidth"} High spin polarization $P$ around $G_{\rm qpc}\sim G_{\rm q}/2$ at zero-bias has now been established. We then proceed to measurement of $P$ under other conditions of t-QPC. The conductance of t-QPC $G_{\rm qpc}$ in the sample for this purpose is shown in Fig.\[fig\_sdbiasdep\](a) as a function of the gate voltage. Quantization at $G_{\rm qpc}\sim G_{\rm q}/2$ is again observed though a peak appears at the gate voltage indicated by the arrow numbered as 1. Figure \[fig\_sdbiasdep\](a) displays measured $P$ as a function of $G_{\rm qpc}$, which shows an unexpected second peak around $G_{\rm qpc}\sim G_{\rm q}$. By some unknown reason, the measurement was unstable in the section from $1.5G_{\rm q}$ to $2G_{\rm q}$ unfortunately and we could not get reliable data. The doubling of data around $G_{\rm qpc}\sim G_{\rm q}/2$ is due to the difference in $V_{\rm qpc}$ as explained shortly. Note that the sampling of electrons to the QD is through a QPC with conductance much lower than $G_{\rm q}/2$. Hence the sampling itself does not have any function of spin-polarization, which fact is consistent with the present measurement method. An advantage of side-QD probe is tunability in the position of energy window as reported for the detection of double quasi-Fermi levels in a non-equilibrium quantum wire.[@:/content/aip/journal/apl/93/11/10.1063/1.2987424] The present measurement can therefore be applied to the case of finite source-drain bias voltage $V_{\rm sd}$ of t-QPC with tuning the window to local chemical potentials. To investigate the polarization mechanism, we thus measure the variation of $P$ versus $V_{\rm sd}$ on 0.5 and 1.0 conductance plateaus. In this measurement, we tuned the center of sensitivity in our detector QD to the electrode of the opposite side. If the source-drain separation of quasi-Fermi levels is larger than the resolution, we can thus selectively detect the polarization of, [*e.g.*]{}, electrons passing through the QPC though in the present case the double tunneling energy window is set to be about 2 meV for sufficient signal to noise ratio and clear decomposition was not obtained. Figure \[fig\_sdbiasdep\](c) and (d) show the results for 0.5 and 1.0 plateaus respectively exhibiting quite surprising differences in the response to $V_{\rm sd}$. While the polarization decreases with bias voltage and almost vanishes around $V_{\rm sd}\sim 2$ mV on 0.5 plateau, it increases with $V_{\rm sd}$ up to 2 mV on 1.0 plateau. The difference manifests that the spin polarization has been developed through different mechanisms on these plateaus. The data in Fig.\[fig\_sdbiasdep\](c) is for the gate voltage indicated with arrow 1 in Fig.\[fig\_sdbiasdep\](a). The small dip in $P$ around zero-bias indicated by the perpendicular arrow in Fig.\[fig\_sdbiasdep\](c) can be observed only around this gate voltage and disappears for the gate position indicated by arrow 2 in Fig.\[fig\_sdbiasdep\](a). The zero bias spin-polarization $P$ for this is hence around 0.7 resulting in the data doubling of $P$ at $G_{\rm qpc}\sim G_{\rm q}/2$ in Fig.\[fig\_sdbiasdep\](b). Hence possible theoretical models for the polarization at 0.5 plateau should be based on spin filtering action, in which a spin-dependent potential at t-QPC blocks either the transport of electrons with up or down spin. In one of such proposals, the transverse confinement potential of QPC works as magnetic field gradient to traversing electron spins (magnetic moments) through the RSOI and causes Stern-Gerlach type force, which works as a spin-dependent potential.[@PhysRevB.72.041308; @kohda2012spin] In another proposal, RSOI in combination with the electron-electron Coulomb interaction forms spin-dependent effective potential.[@PhysRevB.80.155440] They are in qualitative agreement with the present experimental results in that finite $V_{\rm sd}$ would overcome such spin-dependent potential resulting in decrease of $P$. In Fig.\[fig\_kondo\](a), we show the differential conductance of t-QPC as a function of $V_{\rm sd}$. As indicated by the arrow in Fig.\[fig\_kondo\](a), at zero-bias a conductance peak structure appears corresponding to the dip in $P$ shown in Fig.\[fig\_sdbiasdep\](c). Besides the zero-bias peak, $G_{\rm qpc}$ increases with $|V_{\rm sd}|$ due to surmounting the barrier of QPC.[@PhysRevB.62.10950] The increment in $G_{\rm qpc}$ is in the range from 0.5 to 2 mV in $V_{\rm sd}$ in accordance with the decrement in $P$ supporting the above inference. We cannot go to further judgment though, which theory is more plausible, the both being consistent with the present result. The zero-bias conductance peak shown in Fig.\[fig\_kondo\_like\_peak\](a) is reminiscent of the Kondo peak often observed around 0.7 anomaly in QPCs with small SOI.[@PhysRevLett.88.226805] The peak is observed only at the gate position indicated as 1 in Fig.\[fig\_sdbiasdep\](a). The width of the peak is roughly 100 V, which corresponds to about 1 K in temperature. The peak conductance increases with lowering the temperature as shown in Fig.\[fig\_kondo\_like\_peak\](b). It is possible to fit the well-known Kondo temperature dependence[@0953-8984-6-13-013] to the data with the Kondo temperature $T_{\rm K}\sim 1$ K as indicated by the blue curve. The fridge condition prevented us from further confirmation of the Kondo effect and we thus call the conductance peak as ”Kondo-like" peak henceforth. The spin polarization in Fig.\[fig\_sdbiasdep\](c) shows a dip structure around zero-bias corresponding to the conductance peak indicating that the Kondo resonance enhances the spin scattering and disturbs the polarization of traversing electrons. ![\[fig\_kondo\] (a) Bias voltage dependence of the QPC conductance on a 0.5 plateau. (b) Temperature dependence of the peak conductance shown in (a). []{data-label="fig_kondo_like_peak"}](kondo.eps){width="\linewidth"} At small biases, the polarization recovers the value as high as 0.7. The result means the Kondo-like effect has narrower bias window than the spin filter effect manifesting that there is no direct relation between the origins of these two effects but the competition through the spin scattering. There is some asymmetry in Fig.\[fig\_kondo\_like\_peak\](a) for the zero-bias. This is presumably because the detector energy window is tuned to the electrode in the opposite side. A small bias voltage gives imperfect separation of two quasi-Fermi levels and would enhance the polarization detected when it is not as large as to overcome the spin-dependent potential. On the other hand at 1.0 plateau, the spin filtering mechanisms cannot be applied because there is no significant conductance shift from the original quantized value. Instead some spin rotation process should be considered to elucidate the spin polarization. To our knowledge there is only one such theoretical proposal,[@doi:10.1143/JPSJ.74.1934] in which the SOI and the transverse confinement potential cause level anti-crossings between dispersion branches with different spins resulting in spin rotation of traversing electrons. The direction dependent spin rotation generates spin polarizations on both ends of the QPC with inverse polar directions, one of which would be detected in our measurement because the electron exchange point is slightly off from the narrowest point of the QPC. From our measurement, then, a magnetic quadrupole would be formed around the QPC. The polarization by the spin rotation mechanism should not change very much as far as the quasi-Fermi level exists above the level anti-crossing point. The small and almost symmetric enhancement of polarization with the finite bias then again would come from separation in the quasi-Fermi levels and tuning of the energy window to the electrode across the QPC. As the separation increases, the spin polarization detector gradually focuses on the electrons which have passed the top of the QPC potential. Summary ======= In summary, we have measured the spin polarization of electrons in the vicinity of quantum point contacts with strong spin-orbit interaction by using side-coupled quantum dot spin detectors. The polarization is as high as 0.7 on the conductance plateaus of 0.5 and 1.0 times conductance quantum 2$e^2/h$. The polarization decreases with finite biasing on the former while it increases on the latter indicating that a spin filtering effect is working in the former and a spin rotation is working in the latter. This work was supported by Grant-in-Aid for Scientific Research on Innovative Area, “Nano Spin Conversion Science" (Grant No.26103003), also by Grant No.25247051 and by Special Coordination Funds for Promoting Science and Technology. [26]{}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 [****,  ()]{} [****,  ()](\doibase 10.1103/PhysRevB.72.115321) [****,  ()](\doibase 10.1143/JPSJ.74.1934),  [****,  ()](\doibase 10.1103/PhysRevB.84.035323) @noop [****,  ()]{} @noop [****, ()]{} [****,  ()](\doibase 10.1103/PhysRevB.80.155440) [****,  ()](\doibase 10.1007/s11664-008-0559-4) [****,  ()](\doibase 10.1103/PhysRevB.68.035315) [****,  ()](\doibase 10.1143/JPSJ.76.084706),  [****,  ()](\doibase 10.1103/PhysRevB.72.115331) [****,  ()](\doibase http://dx.doi.org/10.1063/1.1757023) [**](http://amazon.co.jp/o/ASIN/3540501576/),  ed. (, ) @noop [****,  ()]{} @noop [****,  ()]{} [****, ()](\doibase 10.1103/PhysRevLett.91.196802) [****,  ()](\doibase 10.1103/PhysRevLett.98.126601) [****,  ()](\doibase 10.1143/JPSJ.75.054702),  [****,  ()](\doibase 10.1103/PhysRevLett.100.067601) [****,  ()](\doibase 10.1103/PhysRevLett.99.096804) [****,  ()](\doibase http://dx.doi.org/10.1063/1.2987424) [****,  ()](\doibase 10.1103/PhysRevB.72.041308) [****,  ()](\doibase 10.1103/PhysRevB.62.10950) [****,  ()](\doibase 10.1103/PhysRevLett.88.226805) [****,  ()](http://stacks.iop.org/0953-8984/6/i=13/a=013) @noop [****,  ()]{} Appendix {#appendix .unnumbered} ======== Let $N_{\rm t}\equiv N_\uparrow+N_\downarrow$ be the total number of electrons energetically available and spatially close enough to tunnel into the QD. Then from , $N_\uparrow=(1+P)N_{\rm t}/2$, $N_\downarrow=(1-P)N_{\rm t}/2$. 0$\leftrightarrow$2 process {#leftrightarrow2-process .unnumbered} --------------------------- In the bias region of 2$\leftrightarrow$0 and in the charge up process of the QD, the tunneling channels for up and down spins can be treated independently due to the Pauli principle and the rate equations for $q_\sigma(t)$ (averaged charge for spin $\sigma$ electrons on the QD) are written as $$\left\{ \begin{aligned} \delta q_\uparrow(t)&=[(e-q_\uparrow(t))N_\uparrow\gamma_0+(q_\downarrow-q_\uparrow)\gamma_f]\delta t,\\ \delta q_\downarrow(t)&=[(e-q_\downarrow(t))N_\downarrow\gamma_0+(q_\uparrow-q_\downarrow)\gamma_f]\delta t \end{aligned} \right. \label{eq_02_charge_rate}$$ for the evolutions $\delta q_{\uparrow\downarrow}$ in infinitesimal time $\delta t$, where $\gamma_f$ is the spin relaxation rate in the QD. For simplicity, the tunneling probability is fixed to $\gamma_0$. In eq., the crossing terms of $q_\uparrow q_\downarrow$ cancel out. These rate equations are readily solved to give $q(t)=q_\uparrow(t)+q_\downarrow(t)$ as $$q(t)/e= 2+A^+\exp(\lambda^+t) + A^-\exp(\lambda^-t),$$ where $A^\pm$ are the integration constants and $$\lambda^\pm\equiv-(N_t\gamma_0+\gamma_f)\pm\xi,\quad\xi\equiv\sqrt{(PN_t\gamma_0)^2+\gamma_f^2}. \label{eq_lambda_def}$$ The discharging process 2$\rightarrow$0 can be viewed as the charging process of “hole" and the rate equations are $$\left\{ \begin{aligned} \delta q_\uparrow(t)&=[-q_\uparrow(t)N_\uparrow\gamma_0+(q_\downarrow-q_\uparrow)\gamma_f]\delta t,\\ \delta q_\downarrow(t)&=[-q_\downarrow(t)N_\downarrow\gamma_0+(q_\uparrow-q_\downarrow)\gamma_f]\delta t. \end{aligned} \right. \label{eq_02_discharge_rate}$$ Eq. are essentially the same differential equations as eq.. Imposing the boundary condition of the square-wave gate-voltage with period 2$\tau$, duty ration 50%, the result $$\begin{gathered} \frac{q(t)}{e}= 2-\left(\dfrac{\gamma_f}{\xi}+1\right)\dfrac{\exp(\lambda^+t)}{\exp(\lambda^+\tau)+1}\\ +\left(\dfrac{\gamma_f}{\xi}-1\right)\dfrac{\exp(\lambda^-t)}{\exp(\lambda^-\tau)+1}, \quad 0\leq t <\tau,\end{gathered}$$ $$\begin{gathered} \phantom{\frac{q(t)}{e}}=\left(\dfrac{\gamma_f}{\xi}+1\right)\dfrac{\exp[\lambda^+(t-\tau)]}{\exp(\lambda^+\tau)+1}\\ -\left(\dfrac{\gamma_f}{\xi}-1\right)\dfrac{\exp[\lambda^-(t-\tau)]}{\exp(\lambda^-\tau)+1}, \quad \tau \leq t <2\tau\end{gathered}$$ \[eq\_02\_semi\_final\] is obtained. Because the signal $S$ is obtained from the lock-in technique, we need to perform the integral calculation $$S\propto \int_0^{2\tau}q(t)\cos\left(2\pi\frac{t}{\tau}\right)dt. \label{eq_lock_in_signal}$$ This can be analytically done[@/content/aip/journal/apl/84/23/10.1063/1.1757023] though we avoid to write down a long tedious expression. 1$\leftrightarrow$3 process {#leftrightarrow3-process .unnumbered} --------------------------- Addition of an energy level makes the calculation in this process much more complicated. Let the average charge of the upper $p$-orbital as $q_3$. For simplicity we use the fact that the energy relaxation from the $p$-state to the lower $s$-state is very fast[@zibik2009long]. Then we write down the rate equations for the charging process on the assumptions: 1) there is a tunneling process with a constant rate to the lower $s$-state through the upper $p$-state; 2) the charge of the upper $p$-state does not increase until the lower $s$-state is full; as $$\begin{aligned} \dfrac{dq_3}{dt} &=\dfrac{q_\uparrow q_\downarrow}{e}(e-q_3)N_t\gamma_3,\label{eq_upper_prob}\\ \dfrac{dq_\uparrow}{dt} & =(e-q_\uparrow)N_\uparrow\gamma_2+(q_\downarrow-q_\uparrow)\gamma_f,\\ \dfrac{dq_\downarrow}{dt} & =(e-q_\downarrow)N_\downarrow\gamma_2+(q_\uparrow-q_\downarrow)\gamma_f.\end{aligned}$$ \[eq\_probs\] Even with this approximation, eq. contains non-linear terms though the other two differential equations in can be independently solved. Hence with substituting the solutions to we can obtain the analytical results. $$\begin{aligned} q_\uparrow&=e+A_{11}\exp(\lambda_3^+t)+A_{12}\exp(\lambda_3^-t),\\ q_\uparrow&=e+A_{21}\exp(\lambda_3^+t)+A_{22}\exp(\lambda_3^-t),\end{aligned}$$ where $A_{ij}$ are constants and $\lambda_3^\pm$ can be obtained from $\lambda^\pm$ in replacing $\gamma_0$ with $\gamma_2$. And the first equation in is now written as $$\begin{gathered} -\frac{edq_3}{q_3+e}=(e+A_{11}\exp(\lambda_3^+t)+A_{12}\exp(\lambda_3^-t)\\ \times (e+A_{21}\exp(\lambda_3^+t)+A_{22}\exp(\lambda_3^-t))dt.\end{gathered}$$ The RHS can be readily integrated and LHS is $-\log|1+q_3/e|$ though, again we avoid writing down the tedious solution. In the discharging process, with reversing the above idea for approximation, the discharging rate from the upper level is a constant and $$q_3(t)=q_{30}\exp(-\lambda_{3{\rm d}}t).$$ Then the other two are $$\begin{aligned} \dfrac{dq_\uparrow}{dt}&=(1-q_3/e)N_\uparrow\gamma_2q_\uparrow +(q_\downarrow-q_\uparrow)\gamma_f,\\ \dfrac{dq_\downarrow}{dt}&=(1-q_3/e)N_\downarrow\gamma_2q_\downarrow+(q_\uparrow-q_\downarrow)\gamma_f.\end{aligned}$$ Writing down the solutions of the above equations requires special functions. Boundary condition fitting and the signal calculation on should be done numerically. Practical fitting process {#practical-fitting-process .unnumbered} ------------------------- Fitting with the above obtained results are too time consuming. Instead we first adopt empirical formulas to obtain the estimation of the tunneling rates. For [0$\leftrightarrow$2 process]{} $$\begin{gathered} S\propto\left[ \left\{ \left(1+\frac{\gamma_f}{\xi} \right)\frac{\pi^2}{\lambda^{+2}\tau^2+\pi^2}\left(1+\exp(\tau\lambda^+)\right)-2 \right\}\right.\\ +\left.\left\{ \left(1-\frac{\gamma_f}{\xi} \right)\frac{\pi^2}{\lambda^{-2}\tau^2+\pi^2}\left(1+\exp(\tau\lambda^-)\right)-2\right\} \right].\end{gathered}$$ For [1$\leftrightarrow$3 process]{} $$\begin{gathered} S\propto\left[ \left\{ \left(1+\frac{\gamma_f}{\xi} \right)\frac{\pi^2}{\lambda_3^{+2}\tau^2+\pi^2}\left(1+\exp(\tau\lambda_3^+)\right)-2 \right\}\right.\\ +\left.\left\{ \left(1-\frac{\gamma_f}{\xi} \right)\frac{\pi^2}{\lambda_3^{-2}\tau^2+\pi^2}\left(1+\exp(\tau\lambda_3^-)\right)-2\right\} \right]\\ +\chi_2\left[ \frac{\pi^2}{\lambda_{3{\rm d}}^2\tau^2+\pi^2}-1 \right], \label{eq_4_13cds}\end{gathered}$$ where $\chi_2$ is a fitting parameter determined by the boundary condition. The obtained values are checked with substitution to the original equations.

ArXiv

--- abstract: '[In the framework of warped extra dimension models addressing the gauge hierarchy problem, we consider the Randall-Sundrum (RS) scenario under the usual hypothesis of a bulk custodial symmetry. It is shown in detail that there can exist large corrections to the Higgs boson Vacuum Expectation Value (VEV) induced by mixings of the gauge bosons with their Kaluza–Klein (KK) excitations. The connection with electroweak precision tests is developed. A noteworthy result is that the correct treatment of the Higgs VEV leads to an increase of the lower limit at $95 \% C.L.$ on KK masses that can reach $+30 \%$ (from usually accepted values). For a Higgs mass $120 \leq m_h \leq 150$ GeV, the obtained limit \[from updated precision data\] on the first KK gauge boson mass lies in the range $3.3$ - $4.0$ TeV. The VEV corrections also play a central role in the corrections to the Higgs couplings. We find possibly substantial RS corrections to the various Higgs couplings able to affect its phenomenology, starting with a Higgs discovery at LHC more challenging than in the Standard Model (SM). The deviations to the Higgs production/decay rates found in RS will be testable at ILC as well as at LHC. Such RS signatures could even be used at ILC to discriminate among several models beyond the SM. Finally, the possibility of a light Higgs boson ($m_h \simeq 99$ GeV) interpreting the excess at $2.3 \sigma$ observed at LEP2 is pointed out. ]{}' --- =16.5pt [Charles Bouchart, Grégory Moreau]{} [*Laboratoire de Physique Théorique, CNRS and Université Paris–Sud,\ Bât. 210, F–91405 Orsay Cedex, France.*]{} Introduction {#intro} ============ The Standard Model (SM) represents a successful description of the ElectroWeak (EW) interactions. Nevertheless, the instability of the EW scale under radiative corrections is a strong indication for a new physics underlying the SM. The attractive scenario proposed by Randall and Sundrum (RS) [@RS] to stabilize the EW scale is based on a 5–dimensional (5D) theory where the extra dimension is warped and compactified. The non–factorisable metric there is of type Anti–de Sitter and the space-time, which is thus a slice of $AdS_5$, has two 4–dimensional boundaries: the ultra–violet (UV) boundary at the Planck scale and the infra–red (IR) brane with an exponentially suppressed scale in the vicinity of the TeV scale. The Higgs boson has to be localized at this so–called TeV–brane if the EW scale is to be stabilized by such a geometrical structure.\ In contrast, letting propagate the other SM fields in the bulk [@RSloc] allows to suppress higher dimensional operators, potentially troublesome with respect to Flavor–Changing Neutral Current effects, by energy scales larger than the TeV scale. This feature has also the advantage to possibly generate the fermion mass hierarchy and flavor structure by a simple geometrical mechanism [@RSloc; @RSmass; @RSmassBIS] (even including the small neutrino mass scale [@RSnu]). In addition, this RS version with bulk matter allows for the unification of gauge couplings at high–energies [@UNI-RS]. More generically speaking, this RS version often turns out to constitute a suitable framework for model building issues. From a general point of view, those RS models with a fundamental Higgs boson, the alternative models of gauge–Higgs unification [@GHunif] and the Higgsless models [@Higgsless] can be thought of as warped extra dimension models constituting dual descriptions, through the $AdS/CFT$ correspondence [@AdSCFT], of 4D strongly coupled gauge theories (in the limit of a large number of colors) predicting the effective Higgs field as a composite state (see e.g. [@MHCM]). Within this new paradigm about the Higgs field sector, one has to take care of the realization of the ElectroWeak Symmetry Breaking (EWSB). In particular, additional contributions to the SM observables arise e.g. for the EW gauge boson masses $m_{W,Z}$ and, in turn, for the Fermi coupling constant $G_F$. More precisely, within the RS scenario, these tree–level $m_{W,Z}$ corrections, with respect to the pure Higgs boson Vacuum Expectation Value (VEV) contribution, originate from the mixing between zero–mode gauge bosons and their Kaluza–Klein (KK) excitations. Even for relatively heavy KK states, this gauge boson mixing, induced by mass mixing terms also driven by the Higgs VEV, could be quite large since the KK boson wave functions are generally peaked at the TeV–brane where is confined the Higgs boson. Therefore significant corrections to the SM Higgs VEV can be expected, with various important phenomenological implications. In the present paper, we first compute the Higgs VEV in the RS context from EW data, including SM loop corrections, and find that significant deviations w.r.t. the SM Higgs VEV can be reached in large regions of the allowed parameter space. This computation allows us to treat precisely the oblique parameters $S$,$T$,$U$ [@STU] which depend on the Higgs VEV, and hence to analyze the constraints from precision EW observables. A new aspect that matters for these EW Precision Tests (EWPT) is that the correlation between $S$ and $T$ through the RS parameters is modified by the new VEV dependence on these parameters. These EWPT, based on updated experimental data, are studied under the theoretical assumption of a bulk gauge custodial symmetry ${\rm SU(2)_L\! \times\! SU(2)_R\! \times\! U(1)_X}$ which allows to reduce the final bound on the mass of the first KK gauge boson excitation $M_{KK}$ (strictly, the KK photon mass) from $\sim 10$ TeV [@Burdman] down to roughly a few TeV [@ADMS], avoiding then the little hierarchy problem. However, an important result here is that this EWPT analysis leads to more severe bounds on $M_{KK}$ when the obtained VEV deviations are taken into account: the increase of this limit, at $95 \% C.L.$, can reach $\sim +1$ TeV in some extreme situations. Such increases of the limit on $M_{KK}$ are crucial for the next–coming LHC search of resonances due to the direct production of KK excitations of gluons or EW gauge bosons [@LHCboson] in RS–like models. Those increases of the possible $M_{KK}$ values make the most promising tests of KK resonances, which already suffer from low cross sections, even more challenging. Another possibility for discovering signatures of RS models at next colliders is to observe deviations of the Higgs boson couplings from their values predicted by the SM. Such deviations may also of course play a primordial role in the experimental search for the Higgs boson that will probably represents the first goal of the LHC. Hence, we have computed the Higgs boson couplings, based on the VEV determination mentioned above. We find that the deviations of the Higgs couplings from the SM values can be large and also that major parts of those are due to the VEV corrections in RS. The Higgs couplings to pairs of fermions and EW gauge bosons as well as the effective couplings at loop level to two gluons or two photons are calculated, taking into account the studied EW precision constraints on $M_{KK}$. In particular, we make the connection between the possible RS solution [@RSAFB] of the anomaly on the forward–backward $b$–quark asymmetry $A^b_{FB}$ observed in $e^+e^-$ collisions [@AFB-SM] and the cross section $\sigma(g g \to h)$ for the main Higgs production process at the LHC: the gluon–gluon fusion mechanism which proceeds through heavy (KK) quark triangular loops. Those past and future collider observables $A^b_{FB}$ and $\sigma(g g \to h)$ are indeed connected through the choice of localization in the bulk, for the bottom and top quarks, that fixes their couplings and own masses.\ The obtained reduced production rates at LHC can render the Higgs boson search more difficult in the RS model than in the SM, especially during the first phases at low energies and luminosities. Interestingly, the obtained Higgs production rates at LEP2 are also significantly reduced w.r.t. SM so that the well–known excess at $2.3 \sigma$ in the $Z + b's$ channel can be precisely interpreted by a light Higgs boson ($m_h \simeq 99$ GeV) produced within the RS context - thus differently than the way it could also be interpreted within the NMSSM via a reduced $B(h\to b \bar b)$ branching ratio [@Gunion]. This interpretation is interestingly correlated to the $A^b_{FB}$ solution through the $b$–quark localizations in the bulk. Such a light Higgs boson lies in a mass range quite difficult to explore at the LHC.\ Concerning the effects in the Higgs boson couplings, the possibly large corrections obtained combined with the expected LHC performance in measurement precisions, for the various Higgs production and decay channels, will allow indirect detections/exclusions of RS models in the Higgs sector. The future linear collider like ILC will even be capable of precisely measuring the smallest Higgs coupling deviations arising in RS scenarios, and, discriminate between several models. We will illustrate quantitatively those phenomenological results for small Higgs boson masses of $120$ GeV, $150$ GeV and the entire range of $g_{Z'}$ coupling constant values \[associated to the extra $Z'$ boson issued from the enhanced bulk gauge symmetry\]. At this stage, one has to mention related works about the Higgs boson couplings within the $AdS/CFT$ paradigm. First, the gluon–gluon–Higgs amplitude has been computed in the context of a 5D gauge–Higgs unification as a dual realization of 4D composite Higgs scenarios [@Adam]. $\sigma(g g \to h)$ has also been evaluated in warped extra dimension models [@Gautam] (based on a bulk custodial symmetry [@RSggh; @LyonGroup]) without a precise consideration of EWPT. The effective Higgs coupling to two photons has been compared among various extra dimensional models for some fixed sets of parameters [@Gautam; @LyonGroup]. Besides, the rate deviations from SM predictions, for various Higgs discovery channels at LHC, have been computed within the strongly–interacting light Higgs models [@SILH]: the LHC will be able to explore values of the cut–off scale $4 \pi f$ (where the model becomes strongly coupled) up to $5-7$ TeV. Finally, the deviations of the Higgs couplings to EW gauge bosons have been computed in the framework of the so–called gaugephobic Higgs boson scenarios [@Lillie]. There, the Higgs couplings can be so suppressed that it can be definitely out of reach for the LHC (in case of a bulk Higgs) [@Caccia]. The organization of the paper is as follows. In Section \[VEV\], the corrections to the Higgs VEV arising in the RS framework are computed. The connections with EWPT and the Higgs boson phenomenology are studied in the next sections. In Section \[EWPT\], an analysis of the oblique parameters is performed in order to derive the induced bounds on the KK scale. Then in Section \[Higgs\], after calculating the Higgs couplings, the various tests of the Higgs boson are discussed at the different colliders (LEP, Tevatron, LHC and ILC). Finally, variations on the main scenario studied are briefly discussed in Section \[variants\]. Higgs boson VEV {#VEV} =============== Theoretical framework {#formalism} --------------------- [**Energy scales:**]{} Within the RS model, the gravity scale on the Planck–brane is $M_{\rm Planck}= 2.44\times 10^{18}$ GeV, whereas the effective scale on the TeV–brane $M_{\star}=e^{-\pi kR_{c}} M_{\rm Planck}$ is suppressed by the warp factor which depends on the curvature radius of the $AdS$ space $1/k$ and on the compactification radius $R_c$. For a product $k R_{c} \simeq 11$, $M_{\star}\!=\!{\cal O}(1)$ TeV allowing to address the gauge hierarchy problem. We will take $k R_{c} \simeq 10.11$ so that the maximum value of $M_{KK} \simeq 2.45 k e^{-\pi kR_{c}}$, fixed by the theoretical consistency bound $k<0.105 M_{\rm Planck}$, is $\sim 10$ TeV in agreement with the the typical EWPT limits at a few TeV and the range of $M_{KK}$ values considered here.\ The parameters noted $c_f$ fix the 5D masses $\pm c_f k$, affected to each fermion $f$, and thus control the fermion localizations in the bulk. Those satisfy $\vert c_f \vert \! = \! {\cal O}(1)$ to avoid the introduction of new fundamental scales.\ \ [**Gauge symmetry breaking:**]{} The SM gauge group is recovered after the breaking of the ${\rm SU(2)_R}$ group into ${\rm U(1)_R}$, by boundary conditions and possibly also by a small breaking of ${\rm SU(2)_R}$ in the bulk effectively parametrized by the $\widetilde W^\pm$ mass $\tilde M$ (the $\widetilde W_\mu^\pm$ boson associated to ${\rm SU(2)_R}$ without zero–mode). Then the breaking ${\rm U(1)_R \times \rm U(1)_X} \to {\rm U(1)_Y}$ occurs via a VEV on the UV brane: the state $\widetilde W^3$, associated to ${\rm U(1)_R}$, mixes with $\widetilde B$, associated to ${\rm U(1)_X}$, to give the SM hypercharge $B$ boson, the orthogonal linear combination being the extra $Z'$ boson. The $Z'$ has no zero–mode and its first KK mass is close to $M_{KK}$: $M'_{KK} \simeq 2.40 k e^{-\pi kR_{c}}$. VEV modification {#system} ---------------- As in the SM, the Higgs VEV value is determined by the Fermi constant $G_F$. Nevertheless, the corrections to $G_F$ due to the mixing of the SM boson $W^\pm$ with its KK excitations introduce a dependence on the bare EW gauge coupling constants $g$ and $g'$ [^1]. So within the RS context two other observables must be used to fix the values of the Higgs VEV noted $\tilde v$ and the two bare parameters $g,g'$. The natural choice is to use the EW gauge boson masses $m_W$ and $m_Z$. The divergent parts of the one–loop EW loop corrections to respectively $G_F$ and $m_{Z,W}$ will affect the $hVV$ coupling ($V=Z^0,W^\pm$), via $\tilde v$, $g$, $g'$, but will be canceled out by the SM quantum corrections to the coupling itself: respectively the Higgs boson self–energy and the vertex irreducible correction [@SMHiggsCorr]. The inputs $G_F$ and $m_Z$ are among the most accurately measured quantities and serve as excellent reference points for EWPT as will be described later. The tiny experimental error on $m_W$ (see updated value in [@mWEXP]) turns out to not affect significantly the $\tilde v$ value obtained that way. The $Z$ boson mass defined as the pole of its propagator reads as [^2], $$m_Z^2 = (g^2+g'^2)\frac{\tilde v^2}{4}+(g^2+g'^2) \ \Pi_{33}(m_Z^2) + \delta^{\rm SM}m^2_Z, \label{MZ2}$$ where $\Pi_{33}(q^2)$ is the vacuum polarization amplitude taking into account the RS–type corrections, more precisely the KK gauge mixing effect [@ADMS]: $$\Pi_{33}(q^2) = \pi R_{c} \bigg ( \frac{\tilde v^2}{4} \bigg ) ^2 \ \bigg [ (g^2+g'^2) \ (G^{5D}_{q(++)}-G^{(0)}_{q(++)}) + g_{Z'}^2 \ \cos^4 \theta' \ G^{5D}_{q(-+)} \bigg ]. \label{Pi33}$$ $G^{5D}_{q(++)}$ ($G^{5D}_{q(-+)}$) is the 5D propagator for the $W^3$,$B$ ($Z'$) KK excitations. $(++)$ ($(-+)$) indicates Neumann (Dirichlet) and Neumann boundary conditions on the Planck–brane and TeV–brane, respectively. The massless pole is subtracted from the $W^3$ and $B$ towers. The new mixing angle is given by $\sin \theta' \equiv \tilde g' / g_{Z'}$ with $g_{ Z^\prime}^2 = \tilde g^2 + \tilde g^{\prime 2}$, where $\tilde g$ and $\tilde g'$ are respectively the ${\rm SU(2)_ R}$ and ${\rm U(1)_{X}}$ couplings; the coupling $g'$ of the SM ${\rm U(1)_Y}$ group reads as $g'=\tilde g \tilde g' /g_{Z'}$. One deduces $2 \sin^2 \theta' = 1 \pm \sqrt{1- (2 g'/g_{Z'}) ^2}$.\ $\delta^{\rm SM}m^2_Z$ is the SM one–loop correction [@SMHiggsCorr]: $$\delta^{\rm SM}m^2_Z = \frac{3 g_Z^2}{16 \pi^2} \sum_{f=t,b} m_f^2 \bigg ( \frac{\mu}{m_f} \bigg ) ^{2\epsilon} \ \bigg [ \frac{1}{2\epsilon} + \frac{\pi^2\epsilon}{24} + {\cal O}(\epsilon^2) \bigg ], \label{DSMZ}$$ $D=4-2\epsilon$ being the space-time dimension in dimensional regularization and $\mu$ the ’t Hooft renormalization scale. For the SM corrections, we restrict to top–quark loops as those constitute the dominant EW corrections to productions and decays of a light Higgs boson ($m_h \ll 2 m_t$). Indeed, we will consider Higgs masses below $150$ GeV leading to the lowest EWPT limits on $M_{KK}$ as preferred by little hierarchy arguments.\ Similarly, one has, $$m_W^2 = g^2 \frac{\tilde v^2}{4} + g^2 \ \Pi_{11}(m_W^2) + \delta^{\rm SM}m^2_W, \label{MW2}$$ with, $$\Pi_{11}(q^2) = \pi R_{c} \bigg ( \frac{\tilde v^2}{4} \bigg ) ^2 \ \bigg [ g^2 \ (G^{5D}_{q(++)}-G^{(0)}_{q(++)}) + \tilde g^2 \ G^{5D}_{q(-+)} \bigg ], \label{Pi11}$$ $$\delta^{\rm SM}m^2_W = \frac{3 g^2}{16 \pi^2} \bigg ( \frac{\mu}{m_t} \bigg ) ^{2\epsilon} \ \bigg [ \frac{m_b^2+m_t^2}{2\epsilon} + \frac{m_b^2+m_t^2}{4} + \frac{m_b^4 \ ln(m_b^2 / m_t^2)}{2(m_t^2-m_b^2)} + {\cal O}(\epsilon) \bigg ]. \label{DSMW}$$ Taking the limit $q^2 \to 0$, one obtains $$\frac{1}{4\sqrt{2}G_F} = \frac{\tilde v^2}{4} + \Pi_{11}(0) + \delta^{\rm SM}G_F^{-1}, \label{GF}$$ with at first order in $1 / k \pi R_{c}$ $$\Pi_{11}(0) = - k \pi R_{c} \frac{g^2}{32} \bigg ( \frac{\tilde v^2}{ke^{-\pi kR_c}} \bigg )^2 \ \bigg [ 1-\frac{1}{k \pi R_{c}} \bigg ] - k \pi R_{c} \frac{\tilde g^2}{32} \bigg ( \frac{\tilde v^2}{ke^{-\pi kR_c}} \bigg )^2 \ \bigg [ 1-\frac{\tilde M^2}{4 k^2} \bigg ], \label{Pi110}$$ and, $$\delta^{\rm SM}G_F^{-1} = \frac{6}{(8 \pi)^2} \bigg [ \frac{m_b^2+m_t^2}{2} + \frac{m_b^2 \ ln(m_b^2 / m_t^2)}{(1-m_b^2/m_t^2)} \bigg ]. \label{DSMG}$$ Numerically, the SM loop corrections are small compared to the RS tree–level ones, as we are going to see. We do not consider higher order corrections involving both loops and KK excitations. Hence, the quantities $\tilde v$, $g$ and $g'$ can be determined within given RS scenarios by solving the system formed by the three equations: Eq.(\[MZ2\]), Eq.(\[MW2\]) and Eq.(\[GF\]). We have checked that the computation of the gauge boson masses through Eq.(\[MZ2\])-(\[MW2\]) is equivalent to their calculation within the perturbation approach in the limit of the inclusion of a large number of KK excitations \[see the description of the KK gauge boson mass matrix in Section \[couplings\]\]. The obtained value for the coupling constant $g$ is $g=0.653$ in both the SM and RS models, which means with and without the corrections due to KK mixing. $g'$ takes similar values in the SM and RS models: $g'=0.356$ and $g'=0.357$ respectively. The numerical results for $\tilde v$ with characteristic values of the parameters $M_{KK}$ and $g_{Z'}$ are shown in Table \[MainTable\]. The motivation for the chosen values of $M_{KK}$ will become clear in Section \[EWPT\]. The allowed range for the $g_{Z'}$ value is defined as follows. The minimum $g_{Z'}$ value is equal to $2 g' \simeq 0.72$ for consistency reasons about the $\theta'$ mixing angle. On the other side, the perturbativity condition for the $Z'$ boson coupling is that $2 k \pi R_c g_{Z'}^2 Q_{Z'}^2 / 16 \pi^2$ must be smaller than unity according to the naive dimensional argument [@MHCM; @LHCboson]. The reason being that the effective 4D coupling of $Z'$ is increased by a factor as large as $\sqrt{2 \pi k R_c}$ for fields at the TeV–brane. In general even for reproducing the correct top mass, the top quark field is not exactly located on the TeV–brane but has only a peaked profile there, so that the real overlap factor should be significantly smaller than $\sqrt{2 k \pi R_c}$. We will consider the coupling constant constraint $g_{Z'} < 2 \sqrt{2\pi / k R_c} \simeq 1.57$, keeping in mind that the new charge $Q_{Z'}$ is an additional source for suppressing the $Z'$ coupling given e.g. the ${\rm SU(2)_R}$ fermionic representations addressing the $A^b_{FB}$ anomaly [@RSrep].\ The first observation from Table \[MainTable\] is that the RS induced deviations to the Higgs VEV are large compared to the pure SM quantum corrections: the Higgs VEV is at the reference value of $v=246$ GeV in the SM at tree–level and at $v_{SM}=245$ GeV after including the one–loop EW corrections in Eq.(\[GF\]) for $m_t=173.1$ GeV [@mtopEXP]. Because of the negative sign of $\Pi_{11}(0)$, the variation of the Higgs VEV induced in RS is positive as show the $\tilde v$ values in Table \[MainTable\] which are significantly larger than $v_{SM}$. This increase of the VEV is higher with the reduction of $M_{KK}$ (e.g. at fixed $g_{Z'}=1.57$) or with the enhancement of $g_{Z'}$ is explained by a larger KK gauge mixing effect in Eq.(\[MZ2\])-(\[MW2\])-(\[GF\]). The increase of the KK mixing with $g_{Z'}$ will be clear from the texture of the KK gauge boson mass matrix (see Appendix \[MassMat\]). ------------------------------------------------------------------- --------------------------------------------------------------------- ------------------------------------------------------------------- [**A\]**]{} $ \ m_h = 120$ GeV, $g_{Z'} = 1.57 \ \ $ [**B\]**]{} $ \ m_h = 120$ GeV, $g_{Z'} = 0.72 \ \ $ [**C\]**]{} $ \ m_h = 150$ GeV, $g_{Z'} = 1.57 \ \ $ $M_{KK} = 4025$ GeV $M_{KK} = 3370$ GeV $M_{KK} = 4095$ GeV $\tilde v = 322$ GeV $\tilde v = 257$ GeV $\tilde v = 311$ GeV $g^{RS}_{hZZ}/g^{SM}_{hZZ} = 57.3\%$ $ g^{RS}_{hZZ}/g^{SM}_{hZZ} = 87.2\%$ $ g^{RS}_{hZZ}/g^{SM}_{hZZ} = 60.1\%$ $g^{RS}_{hWW}/g^{SM}_{hWW} = 57.5\%$ $g^{RS}_{hWW}/g^{SM}_{hWW} = 87.4\%$ $g^{RS}_{hWW}/g^{SM}_{hWW} = 60.3\%$ $\lambda^{RS}_{\tau}/\lambda^{SM}_{\tau} = 76.2\%$ $\lambda^{RS}_{\tau}/\lambda^{SM}_{\tau} = 95.5\%$ $\lambda^{RS}_{\tau}/\lambda^{SM}_{\tau} = 79.1\%$ $\lambda^{RS}_{b}/\lambda^{SM}_{b} = [71 , 75]\%$ $\lambda^{RS}_{b}/\lambda^{SM}_{b} = [90 , 93] \%$ $\lambda^{RS}_{b}/\lambda^{SM}_{b} = [74 , 78]\%$ $g^{RS}_{hgg}/g^{SM}_{hgg} = [77.6 , 80.8]\%$ $g^{RS}_{hgg}/g^{SM}_{hgg} = [96.2 , 99.1]\%$ $g^{RS}_{hgg}/g^{SM}_{hgg} = [80.1 , 83.3]\%$ $g^{RS}_{h\gamma\gamma}/g^{SM}_{h\gamma\gamma} = [74.9 , 75.1]\%$ $g^{RS}_{h\gamma\gamma}/g^{SM}_{h\gamma\gamma} = [100.6 , 100.8]\%$ $g^{RS}_{h\gamma\gamma}/g^{SM}_{h\gamma\gamma} = [77.1 , 77.3]\%$ ------------------------------------------------------------------- --------------------------------------------------------------------- ------------------------------------------------------------------- : Numerical results for the lower limit at $95.45 \% C.L.$ on $M_{KK}$ (from EWPT) and for the Higgs boson VEV within the RS scenario, with three characteristic sets of parameters \[called A, B and C\]. $\tilde M/k = 0.11$, $0.41$ and $0.12$ respectively (optimized values leading to the minimum $M_{KK}$ limits) for the points A, B and C. Note that $\tilde v$ is obtained for the indicated allowed $M_{KK}$ value, and reciprocally, the $M_{KK}$ limit corresponds to the given Higgs VEV. For these 3 points, the values for the ratios (RS over SM) of the effective Higgs couplings to EW gauge bosons ($Z$, $W$, $\gamma$), gluons ($g$) and fermions ($\tau$, $b$) are also given. The Yukawa coupling ratios for the other light fermions are identical to the $\tau$ lepton one given here.[]{data-label="MainTable"} EW precision tests {#EWPT} ================== The mixings with KK gauge boson excitations induce modifications of the EW gauge boson propagators, the oblique corrections, that can be parametrized by the three $S_{\rm RS}$,$T_{\rm RS}$,$U_{\rm RS}$ quantities [@STU] (we define those such that they vanish in the absence of KK mixing). Here, we are interested in variations of the constraints from considerations on $S$,$T$ due to the corrections on the Higgs VEV. There exist also constraints on the $W,Y$ parameters [@BPRS] [^3]. Compared to $S$,$T$, these parameters are proportional to higher derivatives of the vacuum polarization amplitudes, involved in higher order terms of the $\Pi_V(q^2)$ expansion, that are not considered within the present RS models. $W$ and $Y$ allow to take into account the relevant LEP2 observables, namely, the differential cross sections for $e^+e^- \to f \bar f$ which we already partially include in this analysis: the process $e^+e^- \to b \bar b$ is included in the $A^b_{FB}$ solution above the $Z$ pole studied in Section \[couplings\]. The corrections to EW observables in the third quark generation sector \[$b$ and $t$\] are treated separately through a fit independent from the oblique parameters. In contrast, for the light SM fermions, the associated parameters $c_{\rm light}$ are taken larger than $0.5$ to generate small masses [@RSmass] and to minimize their couplings to KK gauge boson excitations (and thus corrections to EW observables). Then the fermion–Higgs higher–dimensional operators, obtained after having integrated out heavy KK modes, get a special form which allows one to redefine their effects into purely oblique corrections [@ADMS]. The effects from the effective 4–fermion operators are negligible [@ADMS] for $c_{\rm light}>0.5$ and $M_{KK} \geq 3$ TeV. For $c_{\rm light} \geq 0.5$, the oblique parameter $S_{\rm RS}$ reads as (not writing terms suppressed e.g. by $1 / k \pi R_c$ factors [@ADMS; @DavSonPer] which are however included numerically), $$S_{\rm RS} \simeq 2 \pi \bigg ( \frac{2.45 \ \tilde v}{M_{KK}} \bigg ) ^2 - \pi^2 k R_c \ \frac{g^2 + g^{\prime 2} + g_{Z'}^2 \cos^4 \theta'}{16} \bigg ( \frac{2.45 \ \tilde v}{M_{KK}} \bigg ) ^4 , \label{SRS}$$ where the first term is the contribution from fermion–Higgs higher–dimensional operators. The other term comes from the gauge–Higgs sector and is at order $(\tilde v/M_{KK})^4$ thus smaller than the first one. Here, we insist on the fact that $g$ and $g'$ are the bare coupling constants, that are computed as described in Section \[system\]. It was shown recently that the KK mixing induces one–loop contributions (considering only the Higgs sector [@BurdmanBIS]) to the $S$ parameter which are not finite and are cut–off dependent (or depend on an energy scale). The physical $m_h$ dependence in $S$ should be approached with care in such a 5D theory as it can be affected by a renormalization procedure. Therefore here, given our goal of studying the Higgs VEV correction effect, we do not consider higher order corrections to $S$ involving both loops and KK excitations.\ the oblique parameter $T_{\rm RS}$ is given by, $$\alpha_0 T_{\rm RS} \equiv \frac{g^2+g'^2}{m_Z^2} (\Pi_{11}(0)-\Pi_{33}(0)) \ \simeq k \pi R_c \ \frac{g^2+g'^2}{m_Z^2} \ \frac{\tilde g^2}{32} \ \frac{\tilde M^2}{4k^2} \ \bigg ( \frac{2.45 \ \tilde v^2}{M_{KK}} \bigg ) ^2 , \label{TRS}$$ where $\alpha_0 = \alpha (q^2=0)$ is the QED fine structure constant given e.g. in [@PDG] and $\tilde M$ is the $\widetilde W^\pm$ mass originating from the small bulk breaking of ${\rm SU(2)_R}$. This expression has the interest of illustrating the custodial protection of $T_{\rm RS}$. Other scenarios, including in particular the case where ${\rm SU(2)_R}$ remains unbroken in the bulk, are discussed later.\ The parameter $U_{\rm RS}$ is non–vanishing only at the order $(\tilde v/M_{KK})^4$ in the KK expansion, in contrast with $S_{\rm RS}$ and $T_{\rm RS}$. Hence, for the relevant values of RS parameters, the $U_{\rm RS}$ values obtained are totally negligible compared to $S_{\rm RS}$, $T_{\rm RS}$. We thus fix $U_{\rm RS}$ at zero in an extremely good approximation. The corrections to EW observables measured up to the $m_Z$ scale can be expressed in function of the three variables $S_{\rm RS}$,$T_{\rm RS}$,$U_{\rm RS}$ (we keep $U_{\rm RS}$ at this stage for the completeness of given formulas). We concentrate on the experimental measurements of $m_W$, $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$ [^4] and the partial $Z$ width into charged leptons $\Gamma_{\ell \ell}$ as those are the most precise and crucial in constraining the plan $\{ T,S \}$ [@PDG]. The theoretical expression for the observable $m_{W^\pm}$ reads as $$m_W^2 = m_W^2 \vert_{ref} + \frac{\alpha_0 \ c^2}{c^2-s^2} m_Z^2 \left ( - \frac{1}{2} \ S_{\rm RS} + c^2 \ T_{\rm RS} + \frac{c^2-s^2}{4 s^2} \ U_{\rm RS} \right ) \label{mWexp}$$ where $m_{W^\pm} \vert_{ref}$ represents the value calculated as accurately as possible within the pure SM and $s$ ($c$) stands for $\sin \theta_W$ ($\cos \theta_W$), $\theta_W$ being the EW mixing angle. Here we emphasize that the strict expression for $m_W$ involves the bare value of the EW mixing angle $\theta_W$. The expression for $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$ is $$\sin^2\theta_{{\rm eff}}^{{\rm lept}} = \sin^2\theta \vert_{ref} + \frac{\alpha_0}{c^2-s^2} \left ( \frac{1}{4} \ S_{\rm RS} - s^2 c^2 \ T_{\rm RS} \right ) . \label{sinexp}$$ The partial $Z$ width reads as, $$\Gamma_{\ell \ell} = \Gamma_{\ell \ell} \vert_{ref} + \alpha_0 \ \Gamma_{\ell \ell}^0 \bigg ( - \frac{2(1-4 \sin^2 \theta_0)}{(1+(1-4 \sin^2 \theta_0)^2)} \frac{S_{\rm RS}}{c^2-s^2} + \bigg \{ 1 + \frac{8(1-4 \sin^2 \theta_0)}{(1+(1-4 \sin^2 \theta_0)^2)} \frac{s^2c^2}{c^2-s^2} \bigg \} T_{\rm RS} \bigg ) , \label{Gllexp}$$ with, $$\Gamma_{\ell \ell}^0 = \frac{m_Z^3 \ G_F}{3\sqrt{2}\pi} (2\sin^4 \theta_0-\sin^2 \theta_0+\frac{1}{4}).$$ $\sin \theta_0$ involves the electroweak mixing angle obtained in the improved Born approximation, namely by taking into account only the well–known QED running of $\alpha$ up to $m_Z$: $$\sin^2 \theta_0 = \frac{1}{2} \bigg [ 1 - \sqrt{ 1 - 4 \frac{\pi \alpha}{\sqrt{2} G_F m_Z^2} } \bigg ] = 0.2310$$ where $\alpha = \alpha (q^2=m_Z^2) = \alpha_0/(1 - \Delta \alpha)$ [@HiggsReviewI] with $\Delta \alpha = \Delta \alpha_{lept} + \Delta \alpha_{had} + \Delta \alpha_{top}$, $\Delta \alpha_{lept}=0.0315$ [@alpha] and $\Delta \alpha_{top} = - 0.00007$ [@alpha]. By virtue of the theorem obtained in Ref. [@thm], only SM fermions contribute to the $\alpha$ running and hence to the leptonic contribution $\Delta \alpha_{lept}$, the hadronic contribution $\Delta \alpha_{had}$ and the top quark one $\Delta \alpha_{top}$.\ At the moment, there is a small controversy on the precise experimental value of $\Delta \alpha_{had}(m_Z^2)$. This value enters in our EWPT analysis through $\sin^2 \theta_0$ as well as through the SM loop calculations of the fitted observables (see later). Recent measurements of $\Delta \alpha_{had}$ [@PDG] via the $e^+e^-$ annihilation cross sections give the results $0.02758 \pm 0.00035$ [@Daee58]; $0.027594 \pm 0.000219$ [@Daee59]; $0.02768 \pm 0.00022$ [@Daee68], whereas the measurement gives $\Delta \alpha_{had}=0.02782 \pm 0.00016$ using the hadronic $\tau$ decays [@Datau82]. The consensus is still to use the $e^+e^-$ data, which are close to each other. We thus use $\Delta \alpha_{had} = 0.02768 \pm 0.00022$, keeping in mind that the result from $\tau$ data is higher. If $\tau$ data are used in the calculation of $a_\mu^{SM}$, the muon magnetic anomaly discrepancy decreases (from $3.1 \sigma$ for $e^+e^-$-based results [@DavierLAST]) down to $1.8 \sigma$ (see recent updates in [@Datau82; @DavierLASTBIS]). While almost solving $\Delta a_\mu$, the $\tau$ data, which raise the value of $\Delta \alpha_{had}$, lead to an EWPT upper bound on $m_h$ of $133$ GeV [@Datau82] leaving a narrow window for the SM Higgs mass, given the direct LEP2 lower bound of $114.4$ GeV [@DirectMH]. It is interesting to note that such a tension does not appear, within the RS framework, where both the EWPT limit can be higher (there is possibly a positive contribution to $T$) and the LEP2 limit smaller \[see later\]. For example, with $M_{KK}=4.1$ TeV, $g_{Z'}=1.51$ and $\tilde M/k = 0.14$, the EWPT limit at $95 \% C.L.$ on $m_h$ is $150$ GeV for $\Delta \alpha_{had}=0.02782$. As shows e.g. Eq.(\[sinexp\]), the accurate measurements of EW observables translate into limits in the plan $T_{\rm RS}$ versus $S_{\rm RS}$. These limits depend on the SM expectation e.g. noted $\sin^2\theta \vert_{ref}$ for $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$. In general, the precise predictions of EW observable values calculated within the SM from QCD/EW corrections [@MwSM; @GllSM; @sinSM] depend in turn on the top and Higgs masses as well as on the strong coupling constant and the photon vacuum polarization $\Delta \alpha$.\ In Fig.(\[fig:ST\]) are presented the limits in the plan $\{ T_{\rm RS},S_{\rm RS} \}$ \[for $m_h = 120$ GeV\] corresponding to values of $m_W = 80.399 \pm 0.025$ GeV (combined LEP2 and Tevatron Run II data) [@mWEXP], $\Gamma_{\ell \ell} = 83.985 \pm 0.086$ MeV (single lepton channel) [@alpha; @LEPex] and $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$ within $1 \sigma$ deviation from their experimental central value. The experimental value used here for $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$ is a combination of the 5 values resulting from the 5 asymmetry measurements: $A^\ell_{FB}(m_{Z})$, ${\cal A}_\ell (P_\tau)$, ${\cal A}_\ell (SLD)$, $A^c_{FB}(m_{Z})$ and $Q^{had}_{FB}$ [@alpha; @LEPex].\ In Fig.(\[fig:ST\]) are also shown the contour levels in $\{ T_{\rm RS},S_{\rm RS} \}$ \[$m_h = 120$; $150$ GeV\] associated to the confidence levels at $68.27 \% C.L.$ and $95.45 \% C.L.$. Those result from a $\chi^2$–analysis of the fit between the RS theoretical predictions for the considered observables and their respective experimental value. In particular, the $95.45 \%$ level corresponds to $\chi^2 / d.o.f.= 12.85 / 6$. ![Contour levels in the plan $\{ T_{\rm RS},S_{\rm RS} \}$ at $68.27 \% C.L.$ and $95.45 \% C.L.$ for the fit of $m_{W}$, $\Gamma_{\ell \ell}$ and $\sin^2\theta_{{\rm eff}}^{{\rm lept}}$: the two ellipses in dashed–line are for $m_h = 120$ GeV, whereas the blue one in plain–line is for $m_h = 150$ GeV. Also shown, for $m_h = 120$ GeV, are the $1\sigma$ bands for the experimental values of the considered EW observables. We use the exact results from the two–loop calculations for $m_{W} \vert_{ref}$ (valid for $100$ GeV $ \lesssim m_h< 1$ TeV) [@MwSM], $\Gamma_{\ell \ell} \vert_{ref}$ ($75$ GeV $<m_h< 350$ GeV) [@GllSM] and $\sin^2\theta \vert_{ref}$ ($10$ GeV $<m_h< 1$ TeV) [@sinSM]. The other quantities, on which these SM predictions depend, are fixed at $m_t=173.1 \pm 1.3$ GeV (in agreement with the recent Tevatron data [@mtopEXP]), $\alpha_s=0.1204 \pm 0.0009$ [@alpha; @LEPex; @Seb]. The plot is obtained for $g_{Z'}=1.57$ and $U_{\rm RS} \simeq 0$. The three red, black, green points are associated to the theoretical predictions for the coordinates $S_{\rm RS}$,$T_{\rm RS}$ with $\tilde M/k=0.13$; $0.11$; $0.10$ respectively. Those correspond to $m_h = 120$ GeV and $M_{KK}=3980$, $4025$, $8000$ GeV from right to left – as indicated. While the blue point is for $m_h = 150$ GeV, $M_{KK}=4095$ GeV and $\tilde M/k=0.12$, the star corresponds to $m_h = 120$ GeV, $M_{KK}=4025$ GeV and $\tilde M/k=0.11$ but using $v_{SM}=245$ GeV (rather than $v_{RS}= \tilde v = 322$ GeV). The cross is located at the origin.[]{data-label="fig:ST"}](RSTsmall.pdf){width="12.cm"} The correlation between $S_{\rm RS}$ and $T_{\rm RS}$ through $M_{KK}$, $g_{Z'}$ is modified by the new dependence of the Higgs VEV ($\tilde v$ entering Eq.(\[SRS\])-(\[TRS\])) on these parameters and on $\tilde M/k$ (via Eq.(\[MZ2\])-(\[Pi110\])). In particular, the lower limit at $95.45 \% C.L.$, from the global EW fit, on $M_{KK}$ as a function of $\tilde M/k$ possesses a minimum depending on $g_{Z'}$ and $m_h$. This feature is illustrated by the fact that, if e.g. $g_{Z'}=1.57$ and $m_h=120$ GeV, the lower limit is $M_{KK}=4025$ GeV for $\tilde M/k = 0.11$ [^5] (theoretical point in black exactly on the $95.45 \% C.L.$ contour in the plan $\{ T_{\rm RS},S_{\rm RS} \}$ of Fig.(\[fig:ST\])) whereas it is higher than $4025$ GeV for $\tilde M/k = 0.10$ (green point on the $95.45 \% C.L.$ contour for $8000>M_{KK}>4025$ GeV) as well as $\tilde M/k = 0.13$ (red points). The lower limit on $M_{KK}$ at $95.45 \% C.L.$ (for an optimized $\tilde M/k$ value corresponding to the minimum of that limit) increases with $\tilde v$ (making $S_{\rm RS}$ and $T_{\rm RS}$ larger) and hence with $g_{Z'}$, as shows Table \[MainTable\] for $m_h=120$ GeV. Finally, the effect of the increase of $m_h$ on the $M_{KK}$ limit is an enhancement (as shows Table \[MainTable\]) as it shifts the $95.45 \% C.L.$ ellipsoidal contour according to Fig.(\[fig:ST\]), so that a larger $M_{KK}$ is required to decrease $\tilde v$ and directly the $S_{\rm RS}$ coordinate of the theoretical point. Fig.(\[fig:ST\]) also illustrates that the behavior of the theoretical predictions for $S_{\rm RS}$ and $T_{\rm RS}$ in the limit of high $M_{KK}$ is a convergence to zero. By consequence, for too large $M_{KK}$’s in the case e.g. $m_h=150$ GeV, the theoretical points go out of the $95.45 \% C.L.$ ellipse as this one does not contain the origin. There exist thus an upper limit at $95.45 \% C.L.$ on $M_{KK}$ due to EWPT which is $M_{KK}<4490$ GeV for $g_{Z'}=1.57$, $\tilde M/k = 0.12$ (even at this upper limit $\tilde v =286$ GeV i.e. there is still a significant RS contribution). An important result is the impact of the Higgs VEV corrections on the EWPT constraint on $M_{KK}$. Let us consider for instance the point A of parameter space given in Table \[MainTable\] and represented in Fig.(\[fig:ST\]) as the black point sitting on the $95.45 \% C.L.$ limit: replacing now the ‘should–be’ Higgs VEV $\tilde v = 322$ GeV by the SM value $v_{SM}=245$ GeV, both the $S_{\rm RS}$ and $T_{\rm RS}$ values are reduced and the black point becomes the black star on the plot. To shift this star back to the $95.45 \% C.L.$ ellipsoidal contour, one can decrease $M_{KK}$ (at fixed $g_{Z'}$ and $m_h$) which would increase $S_{\rm RS}$ and $T_{\rm RS}$. Indeed, for $g_{Z'}=1.57$, $m_h=120$ GeV and the optimized $\tilde M/k = 0.14$, we find numerically that the $95.45 \% C.L.$ lower limit is reduced from $M_{KK}=4025$ GeV down to $M_{KK}=3055$ GeV if one assumes instead a lower VEV fixed at $v_{SM}=245$ GeV. In other words, the increase of the VEV due to included RS corrections has to be compensated by an increase of $M_{KK}$. In conclusion, the obtained significant increase of the Higgs VEV in RS \[see the effect by comparing the black point and star in Fig.(\[fig:ST\])\] can lead to a large enhancement of the $M_{KK}$ lower limit from EWPT, for given sets of RS parameters (it never translates into a reduction of the limit). Higgs boson phenomenology {#Higgs} ========================= Higgs boson couplings {#couplings} --------------------- In order to discuss quantitatively the Higgs boson phenomenology at past, present and future high–energy colliders, let us first derive the Higgs couplings within the RS framework. ### Couplings to EW gauge bosons {#hVVsec} After EWSB, the neutral gauge bosons ($Z^0$ and $Z'$) and their KK excitations get masses through their couplings to the Higgs boson which has a bidoublet structure under the custodial symmetry group ${\rm SU(2)_L \! \times \! SU(2)_R}$. These mass terms can be written, in the 4D Lagrangian including the first three KK excitations, as $${\cal L}_{\rm mass}^{\rm n} \! = \! \frac{1}{2} ( Z^0_\mu \ \ Z^{(1)}_\mu \ \ Z'_\mu \ \ \dots \ \ Z^{(3)}_\mu \ \ Z'^{(3)}_\mu ) \ {\cal M}^2_0 \ ( Z^{0\mu} \ \ Z^{(1)\mu} \ \ Z'^\mu \ \ \dots \ \ Z^{(3)\mu} \ \ Z'^{(3)\mu} )^T$$ where the squared mass matrix ${\cal M}^2_0$ is given in Eq.(\[NeutMassMat\]). Numerically, we indeed include the KK tower up to the third excitations $Z^{(3)}$ and $Z'^{(3)}$ (included). This effective truncation of the tower induces an error of less than the percent on the $hVV$ couplings computed in this section. Similarly, the charged gauge bosons ($W^\pm$ and $\widetilde W^\pm$) get KK–type and VEV–induced masses ([*c.f.*]{} Eq.(\[ChargMassMat\])): $${\cal L}_{\rm mass}^{\rm c} \! = \! (W^+_\mu \ \ W^{+(1)}_\mu \ \ \widetilde W^{+}_\mu \ \ \dots \ \ W^{+(3)}_\mu \ \ \widetilde W^{+(3)}_\mu) \ {\cal M}^2_\pm \ ( W^{-\mu} \ \ W^{-(1)\mu} \ \ \widetilde W^{-\mu} \ \ \dots \ \ W^{-(3)\mu} \ \ \widetilde W^{-(3)\mu} )^T .$$ The mass matrix ${\cal M}^2_0$ is diagonalized by a $7 \times 7$ unitary matrix $U$, via the basis transformation $(Z_{\mu} \ \ Z^1_{\mu} \ \ Z^2_{\mu} \ \ \dots )^T \!=\!\ U \ (Z^0_\mu \ \ Z^{(1)}_\mu \ \ Z'_\mu \ \ \dots )^T$ : $$\mathcal{M'}^2_0 \equiv U \ {\cal M}^2_0 \ U^\dagger \ = \ {\rm diag}~(m^2_Z,m^2_{Z^1},\dots,m^2_{Z^6}) , \label{MnEIGEN}$$ $m_Z$ having to be associated with the experimental value for the $Z$ boson mass (the unitary matrix is chosen such that $m_Z\!<\!m_{Z_1}\!<\!m_{Z_2}\!<\!\dots$). The $Z_{\mu}$ and $Z^i_{\mu}$ \[$i=1,2,\dots$\] components are the mass eigenstates. The mass matrix ${\cal M}^2_\pm$ is diagonalized by a $7 \times 7$ unitary matrix $V$ through, $$\mathcal{M'}^2_\pm \equiv V \ {\cal M}^2_\pm \ V^\dagger \ = \ {\rm diag}~(m^2_W,m^2_{W^1},\dots,m^2_{W^6}) . \label{McEIGEN}$$ The values of $\tilde v$, $g$ and $g'$ are obtained by solving the system formed by Eq.(\[GF\]) and Eq(\[MnEIGEN\])-(\[McEIGEN\]) \[first eigenvalue\]. The $\tilde v$ values obtained that way do not differ significantly from the ones derived using the method developed in Section \[system\]. From the obtained values of $\tilde v$, $g$, $g'$, one can now deduce the rotation matrices $U$ and $V$ needed to compute the couplings. In the weak basis, the 4D Higgs couplings to neutral gauge bosons are then $${\cal L}_{\rm coupling}^{\rm n} \! = \ \frac{h}{\tilde v} \ ( Z^0_\mu \ \ Z^{(1)}_\mu \ \ Z'_\mu \ \ \dots \ \ Z^{(3)}_\mu \ \ Z'^{(3)}_\mu ) \ {\cal C}_0 \ ( Z^{0\mu} \ \ Z^{(1)\mu} \ \ Z'^\mu \ \ \dots \ \ Z^{(3)\mu} \ \ Z'^{(3)\mu} )^T$$ the matrix ${\cal C}_0$ being given in Eq.(\[NeutCouplMat\]). For the charged gauge bosons \[[*c.f.*]{} Eq.(\[ChargCouplMat\])\], $${\cal L}_{\rm coupling}^{\rm c} \! = \ 2 \ \frac{h}{\tilde v} \ (W^+_\mu \ \ W^{+(1)}_\mu \ \ \widetilde W^{+}_\mu \ \ \dots \ \ W^{+(3)}_\mu \ \ \widetilde W^{+(3)}_\mu) \ {\cal C}_\pm \ ( W^{-\mu} \ \ W^{-(1)\mu} \ \ \widetilde W^{-\mu} \ \ \dots \ \ W^{-(3)\mu} \ \ \widetilde W^{-(3)\mu} )^T .$$ Moving to the mass basis, the neutral gauge boson interactions are described by the Lagrangian: $$\begin{aligned} {\cal L'}_{\rm coupling}^{\rm n} \ = \ \frac{h}{\tilde v} \ (Z_{\mu} \ \ Z^1_{\mu} \ \ Z^2_{\mu} \ \ \dots ) \ {\cal C'}_0 \ (Z^{\mu} \ \ Z^{1\mu} \ \ Z^{2\mu} \ \ \dots ) ^T , \ \mbox{where,} \ {\cal C'}_0 \ = \ U \ {\cal C}_0 \ U^\dagger . \label{ZinterMASS} \end{aligned}$$ Similarly, $$\begin{aligned} {\cal L'}_{\rm coupling}^{\rm c} \ = \ 2 \ \frac{h}{\tilde v} \ (W_{\mu} \ \ W^1_{\mu} \ \ W^2_{\mu} \ \ \dots ) \ {\cal C'}_\pm \ (W^{\mu} \ \ W^{1\mu} \ \ W^{2\mu} \ \ \dots ) ^T , \ \mbox{with,} \ {\cal C'}_\pm \ = \ V \ {\cal C}_\pm \ V^\dagger . \label{WinterMASS} \end{aligned}$$ Therefore, the $hZZ$ and $hWW$ effective dimensionful coupling constants calculated within the RS framework are respectively the $(1,1)$–matrix elements: $g^{\rm RS}_{hZZ}=({\cal C'}_0 \vert_{11}) / \tilde v$ and $g^{\rm RS}_{hWW}=2 \ ({\cal C'}_\pm \vert_{11}) / \tilde v$. In contrast, within the pure SM case, the Higgs coupling constants are $g^{\rm SM}_{hZZ}= ( \{g^2+g'^2\} v^2_{SM} / 4 + \delta^{\rm SM}g_{hZZ}) / v_{SM}$ and $g^{\rm SM}_{hWW}= 2 \ ( g^2 v^2_{SM} / 4 + \delta^{\rm SM}g_{hWW}) / v_{SM}$ where $g$,$g'$ are calculated from Eq.(\[MZ2\])-(\[MW2\]), of course without the RS corrections, and the $\delta$’s parts represent the SM loop corrections \[given in Eq.(\[DSMirrZ\])-(\[DSMirrW\])\]. The obtained values of the ratios $g^{\rm RS}_{hVV}/g^{\rm SM}_{hVV}$ are given in Table \[MainTable\] for three characteristic points of the parameter space. These parameter sets respect the EWPT constraints. In particular, the Higgs boson VEV’s $\tilde v$ corresponding to these sets, and involved in $g^{\rm RS}_{hVV}$, lead to $S_{\rm RS}$ and $T_{\rm RS}$ values within the $95.45 \% C.L.$ contour levels. The obtained $Z$ and $W$ coupling strengths are nearly identical because of the approximate custodial symmetry. The first conclusion about these numerical results is that the RS corrections of the Higgs boson couplings to EW gauge bosons can be quite strong. Secondly, the given $\tilde v$ values in the same table show that the role of the Higgs VEV modification in the RS corrections to $g_{hVV}$ is major. For the example of point A, the total relative correction on the Higgs coupling to $Z$ is of $\delta g_{hZZ}/g_{hZZ}=(g^{\rm RS}_{hZZ}-g^{\rm SM}_{hZZ})/g^{\rm SM}_{hZZ}=- 42.7\%$ and an included correction of $- 23.9\%$ is explained by the VEV: assuming [*only*]{} a variation in the denominator through the VEV (both RS and SM couplings are inversely proportional to the Higgs VEV), the correction is indeed $\delta g_{hZZ}/g_{hZZ}=v_{SM}/\tilde v - 1$. The numbers in Table \[MainTable\] confirm that $g^{\rm RS}_{hZZ}$ gets more suppression w.r.t. SM as $\tilde v$ increase. The remaining part of the whole negative RS correction to $g_{hZZ}$ comes from the direct KK gauge boson mixing effect on the Higgs coupling (numerator part). The $g^{\rm RS}_{hZZ}$ suppression is more effective for a larger KK gauge mixing effect on the coupling \[i.e. larger $g_{Z'}$ or smaller $M_{KK}$\] as illustrated in the same table and Fig.(\[fig:ghVV\]). In Fig.(\[fig:ghVV\]), we plot the ratio $g^{\rm RS}_{hZZ}/g^{\rm SM}_{hZZ}$ for RS parameters respecting the EWPT constraints (sets A and B of Table \[MainTable\] with larger $M_{KK}$’s). It shows that RS corrections to the $hVV$ vertex remain important in wide regions of the allowed parameter space. ![Value of the Higgs–$ZZ$ coupling ratio $g^{\rm RS}_{hZZ}/g^{\rm SM}_{hZZ}$ as a function of $M_{KK}$ \[in GeV\] for $g_{Z'} = 0.72$ and $1.57$. The two minimum values of $M_{KK}$, where the curves stop respectively, are equal to the $M_{KK}$ lower limits from EWPT obtained with $m_h = 120$ GeV (points A and B of Table \[MainTable\]).[]{data-label="fig:ghVV"}](gHZZsmall.pdf){width="10.cm"} ### Yukawa couplings {#couplYuk} **Yukawa couplings with KK mixings:** The fermions mass matrices need to be studied as the Yukawa couplings induce mixing with the KK excitations. Moreover, the Yukawa terms have to take an invariant form under the custodial symmetry as the Higgs field is embedded into a bidoublet of the ${\rm SU(2)_L\! \times\! SU(2)_R}$ gauge symmetry. As a consequence, fermions are promoted to higher gauge multiplets, the simplest realization being the right-handed fermion singlets promoted to ${\rm SU(2)_R}$ isodoublets, and new exotic quarks are expected. Since the ${\rm SU(2)_R}$ symmetry is broken by boundary conditions, these exotic quarks have $(-+)$ boundary conditions (BC), so that those have no zero–mode, in contrast with the SM fermions which have $(++)$ BC. Therefore, the first KK excitation of those exotic quarks can be relatively low depending on the value of the $c_f$ parameter. Defining $\mathcal{M}_f$ as the mass matrix of a specific chiral fermion $\Psi_{L/R}$ in the interaction basis (see e.g. Appendix \[FermMat\]), this matrix can be diagonalized by unitary matrices $U_{L/R}$, through the transformation $\Psi'_{L/R}=U_{L/R}\Psi_{L/R}$ : $$\begin{aligned} \mathcal{M}'_f \equiv U_L \mathcal{M}_f U^{\dagger}_R = {\rm diag}~(m_{f_{1}} ,m_{f_{2}} ,\dots ) \label{MfDIAGO}\end{aligned}$$ where $m_{f_{1}}$ corresponds to the measured value of the SM fermion mass (the unitary matrices being defined such that $m_{f_{1}} < m_{f_{2}} < . . .$). The $\Psi'_{L/R}$ components are the mass eigenstates. Then, the Higgs couplings are given by the interaction Lagrangian $$\begin{aligned} {\cal L}_{int} = {h\over \tilde{v}}\bar{\Psi}'_L \mathcal{C}'_f \Psi'_R + \mbox{h.c.} \label{CfDIAGO}\end{aligned}$$ where $\mathcal{C}'_f=U_L\mathcal{C}_f U^{\dagger}_R$ with $\mathcal{C}_f$ being the same matrix as $\mathcal{M}_f$ but with the KK masses set to zero (see for instance Appendix \[FermCoupl\]). Hence, the 4D effective Higgs coupling to an eigenstate $f_i$ (i.e. a component of $\Psi'$) is given by $\lambda_{f_i}^{\rm RS} = (\mathcal{C}'_{f}|_{ii})/\tilde{v}$ in contrast to the usual $m_f/v_{SM}$ Yukawa value within the SM.\ \ **Light fermion sector:** For light fermions (other than the $b$- and $t$-quarks), the $c_{\rm light}$ parameters controlling each fermion multiplet localization along the fifth dimension need to be higher than $0.5$ in order to reproduce the correct set of masses. As a consequence, wave function overlaps between light modes, and their KK excitations or exotic KK partners, induce negligible mixings (of order $(m_{f}/M_{KK})^2$) over the zero mode contribution. Hence, one can safely neglect the KK fermion contributions, inducing a 4D effective Yukawa coupling for light fermions being $\lambda_{f_{\rm light}}^{\rm RS} \simeq m_{f_{\rm light}}/\tilde{v}$. In the end, the deviation of the Yukawa coupling in the RS scenario with respect to the SM for light fermions simplifies to the deviation of the Higgs boson VEV determined in Section \[system\]. It gives: $$\begin{aligned} {{\lambda_f^{\rm RS}}\over{\lambda_f^{\rm SM}}} \simeq {{v_{SM}}\over{\tilde{v}}} \quad\mbox{for every light fermions (other than the $b$- and $t$-quarks).} \label{lightYUK}\end{aligned}$$\ **Third quark generation sector:** What differs greatly between the light fermion sector and the third quark generation sector is the $c_{b,t}$ parameters controlling the bottom/top localization. Indeed, in order to reproduce the top and bottom masses, one needs $c_{b,t}$ parameters smaller than $0.5$. Having so, in particular, the wave function overlaps between the zero–modes and their KK excitations over the Higgs field on the TeV–brane can become quite large, inducing possibly large KK mixing effects in the Yukawa couplings. The KK mixing depends on the mass matrix and hence on the quark representations under the gauge symmetry. In the sector of third generation quarks, the crucial EW constraints come from the precise measurements of the observables $A^b_{FB}$ and $R_b$ (ratio of the partial decay width into $b \bar b$ for the $Z^0$ boson). To be complete in our study of the Higgs boson phenomenology with regard to all the EWPT constraints, we will follow the previous work [@RSrep] where we have determined quark representations able to address the $A^b_{FB}$ anomaly (i.e. improve significantly the global fit of $R_b$ and the $A^b_{FB}$ values at the $Z^0$ pole and outside resonance). Doing so, we consider the proposed choice of multiplets under the ${\rm SU(2)_L\! \times\! SU(2)_R\! \times\! U(1)_X}$ group for the bottom and top quarks: $$\begin{aligned} && \{Q_{1L}\} \equiv (2,2)_{2/3} = \left ( \begin{array}{cc} q_{(5/3)L}' & t_{1L} \\ t_L' & b_{1L} \end{array} \right ) \quad\quad\quad\quad \{t_R^c\} \equiv (1,3)_{2/3} = \left ( \begin{array}{ccc} q^{c\prime}_{(5/3)R} & t^c_R & b^{c\prime}_R \end{array} \right ) \nonumber\\ && \{Q_{2L}\} \equiv (2,3)_{-5/6} = \left ( \begin{array}{ccc} t_{2L} & b_L' & q_{(-4/3)L}'' \\ b_{2L} & q_{(-4/3)L}' & q_{(-7/3)L}' \end{array} \right ) \quad \{b_R^c\} \equiv (1,2)_{-5/6} = \left ( \begin{array}{ccc} b^c_R & q^{c\prime}_{(-4/3)R} \end{array} \right ) \label{Model}\end{aligned}$$ where the left bottom and top SM-like multiplets arise from a mixing mechanism between the two above left-handed multiplets (parametrized by a mixing angle $\theta$) [^6]. The heavy quarks indicated with one or several primes (and an electric charge) are the mentioned fields with $(-+)$ BC \[the so–called ‘custodians’\]. For the rest of the paper, all numerical values are based on this quark model. In the field basis $\Psi^t_L \equiv (b^{(0)}_L, b^{(1)}_L, b^{c(1)}_L, b''^{(1)}_L, b'^{c(1)}_L)^t$, $\Psi^t_R \equiv (b^{c(0)}_R, b^{(1)}_R, b^{c(1)}_R, b''^{(1)}_R, b'^{c(1)}_R)^t$ where we have introduced the charge conjugated fields (indicated by the superscript $c$) in order to use only left-handed SM fields, the effective 4D bottom quark mass matrix and Yukawa coupling matrix induced by EWSB are given in Appendix \[FermMat\] and \[FermCoupl\]. Similar structure can be found for the top quark mass and Yukawa coupling matrices.\ \ **Parameter space:** Our choice of parameters in order to derive the deviation of the Yukawa couplings relies on a few considerations we discuss now. First, numerically we consider only the first two fermionic custodian excitations, which gives rise to heaviest eigenvalues of $\mathcal{M}_f$ around $\sim 2 M_{KK}$ in agreement with the NDA estimation of the cut–off scale of the effective field theory related to the perturbativity of Yukawa couplings: $\Lambda_{IR} \sim 2 M_{KK}$. Secondly, in order to derive the correct value of Yukawa couplings in the RS model, we need to implement in Eq.(\[MfDIAGO\]) the good value of the relevant fermion mass $m_{f_{1}}$ and the exact unitary matrices $U_{L/R}$. In particular, the whole $b$- and $t$-quark mixings with the first two quark families are treated effectively through the parameters $\epsilon^{mixing}_{b,t}$ appearing in front of the first element of the mass matrices $\mathcal{M}_{b,t}$ (see Appendix \[FermMat\]). As the Cabibbo-Kobayashi-Maskawa (CKM) quark mixing matrix $V_{CKM}= U^{up}_L U^{down \dagger}_L$ is close to unity, the simplest case corresponds to both rotation matrices for up and down fields being also close to unity. We thus estimate $\epsilon^{mixing}_{b,t}$ to be roughly of order $\cos \theta_{12} \cos \theta_{13} \sim \cos \theta_{12} \cos \theta_{23} \sim 0.97$ where $\theta_{ij}$ are the CKM mixing angles encoding its hierarchical structure. We then use the allowed range $\epsilon^{mixing}_{b,t} = [0.95 , 1.05]$ for the numerical analysis. A more important effect comes from the mass running and is also implemented into $\epsilon_{b,t} = \epsilon_{b,t}^{mixing} \times \epsilon_{b,t}^{running}$. The parameters $c_{b,t}$ and the dimensionful Yukawa coupling constants $\lambda^{5D}_{b,t}$, entering $\mathcal{M}_{b,t}$, are parameters appearing in the 5D Lagrangian and have thus to be considered e.g. at the effective 5D scale. Then, one has to consider the running of quark masses from $M_{KK}$ typically down to the EWSB scale $\Lambda_{EW} \sim m_Z$ where the EWPT and light Higgs phenomenology are studied in the present paper. The approximate effect of such a running can be estimated from the mass values at the two extreme scales: $$\begin{aligned} &&m_b(m_Z) = [2.8 , 3.0] \mbox{GeV} ,\quad m_b(10 \mbox{TeV}) = [2.1 , 2.3] \mbox{GeV} \nonumber\\ &&m_t(m_Z) = [168 , 180] \mbox{GeV} ,\quad m_t(10 \mbox{TeV}) = [140 , 148] \mbox{GeV} \nonumber\end{aligned}$$ where we have used the SM one loop renormalization group equations [@ref22] to run the quark masses given in Ref. [@ref23] from the scale $m_Z$ to $10$ TeV. At that point, one can neglect KK loop contributions to the running which would correspond to higher order corrections. The running reduces at most the quark masses by about $21 \%$-$42 \%$ ($13 \%$-$28 \%$) for the bottom (top) mass and is taken into account through the considered allowed range for the whole factors $\epsilon_{b,t}$: $$\begin{aligned} \epsilon_b = [1.15 , 1.45] \quad,\quad \epsilon_t = [1.10 , 1.35] . \nonumber $$ Anyway, the precise variation of the $\epsilon_{b,t}$ value has no important effects on the final Yukawa coupling, as the rotation matrices $U_{L/R}$ in Eq.(\[MfDIAGO\]) are systematically such that the smallest mass eigenvalue is equal to $m_{f_{1}}$, namely the measured fermion mass. Nevertheless, for completeness, we include these small possible variations of $\epsilon_{b,t}$. Those lead to several possible values of the parameters $c_{b,t}$, $\lambda^{5D}_{b,t}$ and the mixing angle $\theta$ which reproduce the correct $m_b(m_Z)$, $m_t(m_Z)$ and address the $A^b_{FB}$ anomaly \[the $A^b_{FB}$ solution is fixed by $\mathcal{M}_b$ which determines $b$-$b^{KK}$ mixings\]. Here we are clearly thinking for a given value of $M_{KK}$ which also enters $\mathcal{M}_{b,t}$ through the dependencies of KK fermion masses. Once all parameters are fixed, the Yukawa couplings are extracted through the method given in Eq.(\[CfDIAGO\]).\ \ **Results and discussion:** The obtained values of the ratios $\lambda^{\rm RS}_{b,t}/\lambda^{\rm SM}_{b,t}$ are given in Table \[MainTable\] for three characteristic points of the parameter space respecting the EWPT constraints in the light fermion and gauge boson sector. For these three fixed values of $M_{KK}$, the parameters $c_{b,t}$, $\lambda^{5D}_{b,t}$ and $\theta$ are varied as described above according to $m_{b,t}(m_Z)$ and $A^b_{FB}$. Those variations give rise to a certain range of 4D Yukawa coupling constants, for which the extremal values are given in Table \[MainTable\]. The first conclusion about these numerical results is that the RS corrections of the Higgs couplings to fermions can be quite strong as was the case for the Higgs coupling to the EW gauge bosons (up to $-23.8\%$ for the example of point A). However, in contrast with the EW gauge boson couplings, comparing here RS corrections to the $b$- and $t$-quarks with light fermions, the table shows that the role of the Higgs VEV modification in the RS corrections to $\lambda_{b,t}$ is major in regard of KK mixing effects. At the same time, one can see that the KK mixing corrections also tend to decrease the coupling of the Higgs boson to a few more percents (up to $-5.5\%$ for the example of point B) which is what one can naively expect: for a given fermion mass value, the higher the KK mixing component is, the lower the Yukawa coupling is (‘direct’ mass contribution). ### Effective coupling to gluons **Gluon fusion in the SM:** The dominant production mode of the SM Higgs boson at the LHC is the reaction $gg \to h$ called the gluon–gluon fusion mechanism and mediated by triangular loops of SM quarks (noted as $Q$’s here), as illustrated in Fig. (\[fig:Dhgg\]). In the SM, mainly heavy quarks, namely the top quark and to a lesser extent the bottom quark, contribute to the amplitude. The decreasing Higgs form factor with rising loop mass is counterbalanced by the linear growth of the Higgs coupling with the quark mass (Yukawa coupling). ![Feynman diagram for the loop-induced Higgs-gluon-gluon vertex. KK represents the exchanged zero–modes and KK towers of quarks.[]{data-label="fig:Dhgg"}](ghh-cg.pdf){width="6.cm"} To lowest order, the partonic cross section reads as $$\begin{aligned} \sigma_h^{SM} &=& {{\alpha_s^2 m_h^2}\over{576 \pi v_{SM}^2}} \left| \sum_Q A^h_{1/2}(\tau_Q) \right|^2 \label{SMhgg}\end{aligned}$$ and it must be multiplied by $\delta(\hat{s}-m_h^2)$ where $\hat{s}$ is the $gg$ invariant energy squared. The form factor for spin-${1\over2}$ particles is given by [@HiggsReviewI]: $$\begin{aligned} A^h_{1/2}(\tau) = {3\over2}[\tau + (\tau-1)f(\tau)]\tau^{-2} , \quad \mbox{where} \quad f(\tau) = \left\{ \begin{array}{ll} \arcsin^2{\sqrt{\tau}} & \tau \leq 1\\ -{1\over4}\left[ln \ {{{1+\sqrt{1-\tau^{-1}}}\over{1-\sqrt{1-\tau^{-1}}}}}-i\pi\right]^2 & \tau > 1 \end{array} \right. \label{FFhalf}\end{aligned}$$ The form factor $A^h_{1/2}(\tau_Q)$ with $\tau_Q={{m_h^2}/{4m_Q^2}}$ is normalized such that for $m_h \ll m_Q$, it reaches unity while it approaches zero in the chiral limit $m_Q\rightarrow0$.\ \ **The $g^{\rm RS}_{hgg}$ coupling:** We now extend this result to the RS scenario. First, new loop contributions appear from the various KK excitations of usual SM quarks $q$, and the ones coming from possible exotic quarks $q'$ with $(-+)$ or $(+-)$ BC, extending the sum in Eq.(\[SMhgg\]) as a consequence over all possible KK quarks coupled to the Higgs field (see Fig. (\[fig:Dhgg\]) on the effective $hgg$ coupling). Secondly, the decrease in the Yukawa couplings evaluated in previous section will tend to reduce the $hgg$ effective coupling through the $hqq$ vertex appearing in the loop. In contrast, the $gqq$ ($gq'q'$) vertex couples universally all (KK) quarks with SM-like strength, due to the flat profile of the gluon fields along the fifth dimension. Within this setup the production cross section is simply generalized to: $$\begin{aligned} \sigma_h^{\rm RS} = {{\alpha_s^2 m_h^2}\over{576 \pi v_{SM}^2}} \left| \sum_{\{q\}} {{\lambda^{\rm RS}_q v_{SM}}\over{m_q}} A^h_{1/2}(\tau_q) + \sum_{\{q'\}} {{\lambda^{\rm RS}_{q'} v_{SM}}\over{m_{q'}}} A^h_{1/2}(\tau_{q'}) \right|^2 ,\end{aligned}$$ since e.g. $\lambda^{\rm RS}_q / \lambda^{\rm SM}_q = \lambda^{\rm RS}_q v_{SM} / m_q$. The sum over $\{q\}$ includes all SM quarks and their KK partner towers, whereas the sum over $\{q'\}$ includes all possible custodian quarks of KK type depending on the model under consideration. $\lambda^{RS}_q$ $(\lambda^{RS}_{q'})$ denotes the Yukawa coupling of the corresponding quark $q$ $(q')$ to the Higgs field in the mass eigenbasis. $m_q$ $(m_{q'})$ is the mass of the corresponding quark $q$ $(q')$ running in the loop. It is convenient to consider the ratio $\mathcal{R}_{hgg}=\sigma^{\rm RS}_h/\sigma^{\rm SM}_h$ of the $gg \to h$ cross sections in the RS and SM models which can be rewritten, $$\begin{aligned} \mathcal{R}_{hgg} = \left( {{v_{SM}}\over \tilde{v}} \right)^2 \left| {{\sum_{\{q\}} {{\lambda^{^{RS}}_q\tilde{v}}\over{m_q}} A^h_{1/2}(\tau_q) + \sum_{\{q'\}} {{\lambda^{^{RS}}_{q'}\tilde{v}}\over{m_{q'}}} A^h_{1/2}(\tau_{q'})}\over{\sum_Q A^h_{1/2}(\tau_Q)}} \right|^2 . \label{eq:hgg}\end{aligned}$$ Then the higher order QCD corrections, which are known to be rather large [@NLOhgg], are essentially the same for all quark species and, thus, drop in this ratio. From this ratio, one can also deduce the effective $hgg$ loop-coupling deviation from the SM prediction by the straightforward relation: $$\begin{aligned} {{g^{\rm RS}_{hgg}}\over{g^{\rm SM}_{hgg}}} \equiv \sqrt{\mathcal{R}_{hgg}} .\end{aligned}$$ We can use the formulas demonstrated in Appendix \[FermionSumRule\], more precisely Eq.(\[SUMq\]) and Eq.(\[SUMqprime\]), in order to simplify our expression. The ratio $\mathcal{R}_{hgg}^{1/2}$ then simplifies to: $$\begin{aligned} {{g^{\rm RS}_{hgg}}\over{g^{\rm SM}_{hgg}}} = {v_{SM}\over \tilde{v}} \left| {{\sum_Q \left( 1 + {{\lambda^{^{RS}}_Q\tilde{v}}\over{m_Q}} [ A^h_{1/2}(\tau_Q) - 1 ] \right)}\over{\sum_Q A^h_{1/2}(\tau_Q)}} \right|\end{aligned}$$ where the sum appearing in the numerator has been reduced to the SM quarks only. It is a remarkable feature that the contributions coming from all KK partners simplify to properties over the corresponding zero–mode. In the case of exotic quarks, the contribution vanishes due to the absence of zero–mode, even if these new quarks do couple to the Higgs field. For light quarks, $A^h_{1/2}(\tau_Q) \rightarrow 0$ quickly and the ratio $\lambda^{\rm RS}_Q / m_Q \rightarrow 1 / \tilde{v}$ as can be deduced from Eq.(\[lightYUK\]). Hence, their contributions to $\mathcal{R}_{hgg}^{1/2}$ tend to vanish as was already the case in the SM. For heavy fermions ($b$- and $t$-quarks), overlaps between the zero–mode profile and KK wave functions can be quite large, as we have seen in the previous section, so that $\lambda^{\rm RS}_Q / m_Q \neq 1 / \tilde{v}$. Besides, considering light Higgs masses of $120$ GeV and $150$ GeV, one check numerically that $|A^h_{1/2}(\tau_b)| \simeq 0.1$ which is negligible, but only as a first approximation, relatively to $|A^h_{1/2}(\tau_t)| \simeq 1.05$. In the end, the deviation in the coupling simply reads: $$\begin{aligned} {{g^{\rm RS}_{hgg}}\over{g^{\rm SM}_{hgg}}} \simeq {{v_{SM}}\over \tilde{v}} \left| {{ 2 - x_b (1 - A_b) - x_t (1 - A_t)}\over{A_b + A_t}} \right| \label{CompleteForm}\end{aligned}$$ where $A_Q \equiv A^h_{1/2}(\tau_Q)$ and $x_Q \equiv \tilde{v} \lambda^{\rm RS}_Q/m_Q$. Note that $x_Q\in[0,1]$. The limit $x_Q\rightarrow 1$ corresponds to the pure Higgs mass case when there is no mixing between the fermion zero–mode and its KK partners. $x_Q$ tends to decrease as this mixing gets stronger. For a light Higgs mass, so that ${{m_h^2}/{4m_t^2}} \ll 1$, in the limit where $|A^h_{1/2}(\tau_t)| \rightarrow 1$ and neglecting the bottom quark contribution: $|A^h_{1/2}(\tau_b)| \rightarrow 0$, our relation gets the following really simple structure, $$\begin{aligned} {{g^{\rm RS}_{hgg}}\over{g^{\rm SM}_{hgg}}} \approx {{v_{SM}}\over \tilde{v}} (2-x_b) . \label{SimpleForm}\end{aligned}$$ We present this final approximated relation to help the reader in getting an intuition on the main behavior of RS corrections to the Higgs coupling, but numerically, the complete formula (\[CompleteForm\]) is used. As discussed in part \[system\], $v_{SM} < \tilde{v}$, systematically, which tends to reduce the production cross section of the Higgs field at LHC. This consequence is simply due to the fact that for a higher value of the Higgs VEV, one needs a smaller value of the Yukawa coupling to reproduce a given fermion mass. The second term in Eq.(\[SimpleForm\]) is less trivial and encodes the whole $b$- and $t$-quark KK towers contribution minus the pure $b$-quark contribution. This contribution increases the deviation of the $g^{\rm RS}_{hgg}$ w.r.t. $g^{\rm SM}_{hgg}$ as the mixing between the bottom zero–mode with its KK excitations grows. On the contrary, when $x_b$ decreases, the contribution from the bottom KK quarks becomes more and more crucial to the $hgg$ coupling.\ \ **Results and discussion:** We have derived the values of the ratio $g^{\rm RS}_{hgg} / g^{\rm SM}_{hgg}$ from Eq.(\[CompleteForm\]) and given them in Table \[MainTable\] for the same characteristic points of parameter space respecting all EWPT constraints (including the $A^b_{FB}$ solution). The allowed variations of fundamental parameters give rise to some intervals of values for the Yukawa couplings and masses of the KK fermion towers. Based on those intervals, we have determined the maximum and minimum amplitudes for the loop-induced observable $g^{\rm RS}_{hgg} / g^{\rm SM}_{hgg}$. Once more, we remark that the RS corrections to the effective Higgs boson coupling to two gluons are possibly quite strong (up to $-22.4\%$ for point A). Furthermore, comparing these RS corrections (obeying Eq.(\[SimpleForm\]) in a good approximation) with the case of light fermion coupling to the Higgs boson ([*c.f.*]{} Eq.(\[lightYUK\])), one concludes again on the major role of the Higgs VEV modification. Simultaneously, one can see that the KK mixing corrections combined with the new contributions from exchanges of KK states in the loop \[synthesized in the $(2-x_b)$ factor effect on Eq.(\[SimpleForm\])\] tend to counter the effect of the Higgs VEV deviation, but at a smaller rate. ### Effective coupling to photons **$\gamma\gamma$ channel in the SM:** For low Higgs masses, the dominant decay mode $h \to b\bar{b}$ is swamped by a large QCD background and the Higgs boson can be searched for through more promising loop-induced decays. The decay channel into two photons is the most important one and is mediated by triangular loops of charged fermions as well as massive vector bosons: see Fig. (\[fig:Dhgaga\]). ![Feynman diagrams for the loop-induced Higgs-photon-photon vertex. KK stands for the exchanged zero–modes and KK towers of $W$ gauge boson \[left\] or fermions \[right\].[]{data-label="fig:Dhgaga"}](gahh-cg.pdf){width="14.cm"} The decay width of the Higgs in two photons reads [@HiggsReviewI] $$\begin{aligned} \Gamma_{h \to \gamma\gamma}^{SM} = {{\alpha^2 m_h^3}\over{256 \pi^3 v_{SM}^2}} \left| A^h_{1}(\tau_W) + {4\over3}\sum_f N_c Q_f^2 A^h_{1/2}(\tau_f) \right|^2\end{aligned}$$ where $N_c$ is the number of color states (3 for quarks, 1 for leptons) and $Q_f$ is the electric charge of the fermion in the loop. The form factor for spin-$1/2$ particles, $A^h_{1/2}$, is the one from Eq.(\[FFhalf\]), and the form factor for spin-${1}$ particles is given by: $$\begin{aligned} A^h_{1}(\tau) = -[2\tau^2 + 3\tau + 3(2\tau-1)f(\tau)]\tau^{-2} .\end{aligned}$$ The form factor $A^h_{1}(\tau_V)$ with $\tau_V={{m_h^2}/{4m_V^2}}$ is defined such that for large masses of the boson in the loop, $m_h \ll m_V$, it reaches $A^h_{1}(\tau_V \rightarrow 0) = -7$.\ \ **The $g^{\rm RS}_{h\gamma\gamma}$ coupling:** Here we extend this result to RS case. This extension is similar to the one of the gluon fusion mechanism from previous section. New loop contributions appear from the various exchanged KK excitations of usual SM fermions, the exchanged custodians, but also the exchanged KK EW gauge bosons (see in Fig. (\[fig:Dhgaga\]) the diagram for the induced $h\gamma\gamma$ vertex). We have to take into account as well the deviations in the Yukawa couplings and in the Higgs coupling to the $W$ boson. Finally, the $ff\gamma$, like the $WW\gamma$, vertex couples with an SM-like universal strength all fermion excitations, respectively all $W$ bosons of the tower, due to the flat wave function of the electromagnetic field. The decay width becomes thus, $$\begin{aligned} \Gamma_{h \to \gamma\gamma}^{\rm RS} = {{\alpha^2 m_h^3}\over{256 \pi^3}}{1\over{v_{SM}^2}} \left| \sum_{n} {{g^{\rm RS}_{W^n} v_{SM}}\over{2m_{W^n}^2}} A^h_{1}(\tau_{W^n}) + {4\over3}\sum_{\{f\}} {{\lambda^{\rm RS}_f v_{SM}}\over{m_f}} N_c Q_f^2 A^h_{1/2}(\tau_f) \right|^2\end{aligned}$$ as $g^{\rm RS}_{W^n} / g^{\rm SM}_{hWW} = g^{\rm RS}_{W^n} v_{SM} / 2 m_{W^n}^2$ with $g^{\rm RS}_{W^n} = g^{\rm RS}_{hW^nW^n} = 2 ({\cal C'}_\pm \vert_{nn}) / \tilde v$, following the notations of Section \[hVVsec\]. We recall that $m_{W^n}$ denotes the physical $W$ state eigenmasses and we mention that the KK sum over $n$ includes the zero–mode, namely the observed $W$ boson. We still consider the ratio $\mathcal{R}_{h\gamma\gamma} = \Gamma_{h\to\gamma\gamma}^{\rm RS} / \Gamma_{h \to\gamma\gamma}^{\rm SM}$: $$\begin{aligned} \mathcal{R}_{h\gamma\gamma} = \left( {v_{_{SM}}\over \tilde{v}} \right)^2 \left| {3\over 4}{\sum_{n} {{g^{\rm RS}_{W^n}\tilde v}\over{2m_{W^n}^2}} A^h_{1}(\tau_{W^n}) + \sum_{\{f\}} {{\lambda^{\rm RS}_f \tilde{v}}\over{m_f}} N_c Q_f^2 A^h_{1/2}(\tau_f)}\over{{3\over 4}A^h_{1}(\tau_W) + \sum_f N_c Q_f^2 A^h_{1/2}(\tau_f)} \right|^2 \label{eq:hgaga}\end{aligned}$$ from which one can deduce the $h\gamma\gamma$ coupling deviation, $$\begin{aligned} {{g^{\rm RS}_{h\gamma\gamma}}\over{g^{\rm SM}_{h\gamma\gamma}}} \equiv \sqrt{\mathcal{R}_{h\gamma\gamma}}\end{aligned}$$ Combining now the formulas derived in Appendix \[FermionSumRule\] and Appendix \[BosonSumRule\], one easily find in a similar way for the bosonic part: $$\begin{aligned} {{g^{\rm RS}_{h\gamma\gamma}}\over{g^{\rm SM}_{h\gamma\gamma}}} \simeq {v_{SM}\over \tilde{v}} \left| {{43\over 4} - {9\over4 } x_W (A_W + 7) + x_b (1 - A_b) + 4 x_t (1 - A_t)}\over{{9\over4}A_W + A_b + 4 A_t} \right|\end{aligned}$$ with $x_W \equiv \tilde{v} g^{\rm RS}_{hWW}/2m_W^2$. Similarly to the case of gluon fusion Higgs production, contributions from light quarks and charged leptons vanish, whereas the KK tower contributions for the $W$ boson, quarks and charged leptons can be rewritten leaving only in the formula the explicit dependence on the $W$ boson, bottom and top quark zero–modes.\ \ **Results and discussion:** The extremal values of ${g^{\rm RS}_{h\gamma\gamma}}/{g^{\rm SM}_{h\gamma\gamma}}$ are also shown in Table \[MainTable\]. The $x_{W,b,t}$ values were obtained as discussed in previous sections. Due to the additional $W$ mode effects, the deviation of this effective coupling can reach higher RS corrections (up to $-25.1\%$ e.g. with point A) than $g_{hgg}$. The principal RS deviation comes again from the modified value of the Higgs field VEV which is a general result of the present work. The Higgs boson at colliders {#pheno} ---------------------------- ### LEP2 The four LEP collaborations (ALEPH, DELPHI, L3 and OPAL) have used the collected $e^+e^-$ collision data at center-of-mass energies between $91$ GeV and $210$ GeV to search for the Higgs boson through the Higgs–strahlung production mechanism $e^+e^- \to Zh$ [@LEP114]. By exploring the main $b \bar b$ and $\tau^+ \tau^-$ channels, an upper limit at $95 \% C.L.$ has been put on the products of normalized generic squared couplings and branching ratios: $(g_{hZZ}/g^{\rm SM}_{hZZ})^2 \times B(h \to b \bar b)$ and $(g_{hZZ}/g^{\rm SM}_{hZZ})^2 \times B(h \to \tau^+ \tau^-)$ for Higgs masses between $10$ GeV and $120$ GeV. Our point C at $m_h=150$ GeV is thus not excluded by this 2D constraint. It is also the case for points A and B at $m_h=120$ GeV which correspond to RS values of the products of normalized squared couplings and branching ratios smaller than unity (in both channels) since $(g^{\rm RS}_{hZZ}/g^{\rm SM}_{hZZ})^2 < 1$ as we have shown. Furthermore, assuming even a lighter Higgs boson of say $\sim 99$ GeV \[see next paragraph for motivations of this precise choice\] is still allowed within the RS scenario. First, such a Higgs mass is allowed at $95 \% C.L.$ by EWPT together with e.g. $g_{Z'}=1.57$, $M_{KK}=3985$ GeV and $\tilde M/k = 0.09$ (in this section, we will discuss this new set of parameters denoted as the point D). Secondly, this Higgs mass satisfies the lower bound coming from the considerations on the vacuum stability [@HiggsReviewI]. Finally, the above LEP2 upper constraints from both channels at $m_h \simeq 99$ GeV are also respected for the point D, in contrast with the SM, as then $(g^{\rm RS}_{hZZ}/g^{\rm SM}_{hZZ})^2=0.30$ is much smaller than one so that the Higgs production rate is sufficiently reduced. Interestingly, such a light Higgs boson would be difficult to discover at LHC, and the high precision performances of $e^+e^-$ linear colliders would probably be needed. The combined data of the four LEP collaborations result in an excess of events at $2.3$ standard deviations in the $Z + b's$ channel [@LEP114]. We claim that it can be nicely fitted by a light Higgs boson of $\sim 99$ GeV produced within the RS scenario and decaying into $b \bar b$. Indeed, for the point D, we obtain $\tilde v =337$ GeV and the value $(g^{\rm RS}_{hZZ}/g^{\rm SM}_{hZZ})^2 \times B^{\rm RS}(h \to b \bar b) \simeq 0.231$ which is (smaller but) close to the observed limit of $\sim 0.236$. This correct value of the product is reached within RS thanks to the suppression of $g_{hZZ}$ and to the precise amount of $B(h \to b \bar b)$. The amount of $B(h \to b \bar b)$ is fixed by the 4D effective $b$–quark Yukawa couplings, and hence by its wave functions along the fifth dimension, $M_{KK}$ as well as the 5D Yukawa coupling constants. In turn e.g. the $b$–quark Yukawa couplings determine the $b$ mixing with its KK excitations, and thus the $A^b_{FB}$ prediction. The point D is chosen such that it corresponds to certain $b$–quark wave functions giving rise to the 4D Yukawa coupling ($\lambda^{\rm RS}_b/\lambda^{\rm SM}_b=0.70$) which simultaneously gives the wanted $B(h \to b \bar b)$ and $A^b_{FB}$ \[+ $R_b$, $m_{b,t}$\] values. We thus underline the non–trivial result that the RS model can both address the LEP1-2 anomaly on $A^b_{FB}$ and explain the LEP2 excess of $Z + b's$ events, with common sets of fundamental parameters in the $b$–quark sector in particular. ### Tevatron Run II Based on collected data at $\sqrt{s}=1.96$ TeV with ${\cal L}=0.9-4.2$ fb$^{-1}$, the combined CDF and D0 analyzes allow to put upper limits at $95 \% C.L.$ on the rates of Higgs production (and decays) for $100$ GeV $< m_h < 200$ GeV [@RunIIHiggs]. Both experiments have performed dedicated searches in different channels. At high mass, like for our point at $m_h=150$ GeV, all the sensitivity comes from the channel $gg \to h \to WW$. Within the RS scenario (point C), this Higgs rate has a reduction, relatively to the SM, in the interval: $$\begin{aligned} \frac{ \delta [\sigma_h B(h\to WW)] }{ \sigma_h B(h\to WW) } = \frac{ \sigma^{\rm RS}_h }{ \sigma^{\rm SM}_h }\frac{ B^{\rm RS}(h\to WW) }{ B^{\rm SM}(h\to WW) } - 1 = [-44.8,-39.0] \ \% \label{DSBsSB} \end{aligned}$$ and $B^{\rm RS}(h\to WW) / B^{\rm SM}(h\to WW)$ is found to be in the range $[0.85,0.88]$ (including pure SM radiative corrections only). This is obtained for the ranges of values for the Higgs couplings and for the ratio $\sigma^{\rm RS}_h / \sigma^{\rm SM}_h = (g^{\rm RS}_{hgg}/g^{\rm SM}_{hgg})^2$ taken from Table \[MainTable\]. The EWPT constraints are thus satisfied. In conclusion, for the point C, the Higgs rate is below the rate for the SM case which is itself well under the exclusion limit [@RunIIHiggs] so that the considered Higgs mass of $150$ GeV is clearly allowed by the Tevatron constraints. Similarly, the RS points A and B are clearly permitted by these Tevatron limits. For $m_h=120$ GeV, the dominant Higgs discovery channel is $q \bar q \to Wh \to l \nu b \bar b$. For example with the parameter set A, the reduction of this Higgs rate is $\delta [\sigma_{Wh} B(h\to b \bar b)] / \sigma_{Wh} B(h\to b \bar b) = [-66.4,-65.1] \% $ as $\sigma^{\rm RS}_{Wh} / \sigma^{\rm SM}_{Wh} = (g^{\rm RS}_{hWW}/g^{\rm SM}_{hWW})^2 = 0.33$ and the $B^{\rm RS}(h\to b \bar b) / B^{\rm SM}(h\to b \bar b)$ ratio is inside $[1.01,1.05]$ ([*c.f.*]{} Table \[MainTable\]). The reductions of the other channels w.r.t. SM are $\delta [\sigma_{Zh} B(h\to b \bar b)] / \sigma_{Zh} B(h\to b \bar b) = [-66.6,-65.3] \%$, $\delta [\sigma_h B(h\to WW)] / \sigma_h B(h\to WW) = [-62.7,-56.5] \%$ and $\delta [\sigma_h B(h\to \tau \bar \tau)] / \sigma_h B(h\to \tau \bar \tau) = [-34.9,-24.2] \%$. Therefore, once more, for point A, all the Higgs rates are below the rates in the SM being themselves under the exclusion limit at $95 \% C.L.$ [@RunIIHiggs] so that the other considered Higgs mass of $120$ GeV is also allowed by the recent Tevatron data. In the pure SM, these Tevatron results exclude the range $160$ GeV $\leq m_h \leq 170$ GeV at the $95 \% C.L.$ [@RunIIHiggs]. To reconsider this mass exclusion within the RS case, one should compute the EWPT limits at the various experimental points in this Higgs mass interval and then calculate precisely the corrections to Higgs rates. The Higgs rates within the RS scenario depend on the chosen parameter values. In particular for the minimum $M_{KK}$ values allowed by EWPT, the Higgs rates in RS are expected to be decreased by more than $\sim 30\%$ w.r.t. SM at $m_h = 165$ GeV \[given the numerical result in Eq.(\[DSBsSB\]) for the close mass $m_h = 150$ GeV\] so that these rates reach regions allowed by Tevatron constraints in the range $160$ GeV $\leq m_h \leq 170$ GeV, making such Higgs masses realistic again. ### LHC [**Higgs boson search:**]{} The large deviations to the gluon–gluon fusion mechanism (producing the Higgs boson) and to the couplings (fixing its branching ratios) can affect the Higgs search at the Large Hadron Collider (LHC). For instance with a SM Higgs mass of $150$ GeV, the significance predicted by the ATLAS collaboration is at $\sim 12.5$ with a center-of-mass energy of $14$ TeV and an integrated luminosity of ${\cal L}=10$ fb$^{-1}$ [@ATLASsig]. The significance is defined typically as ${\cal S} = \sigma^{\rm SM}_h B^{\rm SM}_h {\cal L} \epsilon_{sig} / \sqrt{ \sigma_{back} {\cal L} \epsilon_{back} }$ in the narrow–width approximation \[still valid when incorporating the RS corrections in the low/intermediate Higgs mass range\] where $\epsilon_{sig}$ ($\epsilon_{back}$) includes the experimental efficiency for the signal (background), $\sigma^{SM}_h$ is the cross section for the Higgs production process within the SM, $B^{SM}_h$ its branching ratios and $\sigma_{back}$ the background cross section. The main Higgs production process is the gluon–gluon fusion mechanism, its considered decay channel is $h \to VV$ and the dominant background is the EW gauge boson production. Moving to the RS case, the significance would vary according to $$\begin{aligned} \frac{ \delta {\cal S} }{ {\cal S} } \simeq \frac{ \delta [ \sigma_h B_h ] }{ \sigma_h B_h } = \frac{ \sigma^{\rm RS}_h }{ \sigma^{\rm SM}_h }\frac{ B^{\rm RS}_h }{ B^{\rm SM}_h } - 1. \label{Sig} \end{aligned}$$ The background is assumed to be identical since the deviations of gauge boson couplings have severe constraints from EWPT. We do not include in our discussion the dependence of experimental efficiencies on rates and channels. For instance with the point C of Table \[MainTable\], Eq.(\[Sig\]) gives $\delta {\cal S} / {\cal S}$ between $\sim - 44 \%$ and $\sim - 39 \%$ since $B^{\rm RS}(h\to VV) / B^{\rm SM}(h\to VV)$ lies in the interval $[0.85,0.87]$. This result is based on range of values for the various Higgs couplings and $\sigma^{\rm RS}_h / \sigma^{\rm SM}_h = (g^{\rm RS}_{hgg}/g^{\rm SM}_{hgg})^2$ taken in Table \[MainTable\] so that the EWPT constraints are well respected. We end up with a significance between ${\cal S}_{\rm RS} = {\cal S} (1 + \delta {\cal S} / {\cal S}) \simeq 12.5 (1 - 0.44) = 7.0$ and ${\cal S}_{\rm RS} \simeq 12.5 (1 - 0.39) = 7.6$ so that the Higgs discovery at $5\sigma$ remains possible within RS, even if the significance is greatly reduced. Initially, the LHC is expected to run at a lower center-of-mass energy of $\sqrt{s}=10$ TeV at which production rates are reduced by about a factor of two (from those at $\sqrt{s}=14$ TeV) [@LHC2FC]. Assuming such an energy, $\delta {\cal S} / {\cal S} \simeq (\sigma^{\rm RS}_h B^{\rm RS}_h) / ( \sqrt{2} \sigma^{\rm SM}_h B^{\rm SM}_h ) - 1$ and the significance would further decrease down to ${\cal S}_{\rm RS} \simeq 4.8 - 5.3$ at ${\cal L}=10$ fb$^{-1}$ rendering a Higgs boson discovery even more challenging [^7]. In the SM, much lower luminosities are required at $10$ TeV for detecting a Higgs boson of $150$ GeV (see preliminary studies from both CMS and ATLAS [@1OTeVHiggs]). The present results yield indicative estimates of the LHC sensitivity for a chosen point of the RS parameter space. Clearly, for a larger $M_{KK}$ e.g. the suppression of $\sigma_h$ would be soften making the Higgs boson search easier.\ \ [**RS signature search:**]{} From the argumentation developed in Appendix \[Consider\], we have obtained the condition for having a $68\%$ probability to observe at least a $1\sigma$ deviation (relatively to the experimental uncertainty), due to RS effects, between the SM prediction for a certain quantity ${\cal Q}_{\rm SM}$ and its measured central value. Applying this condition (Eq.(\[final\])) to the product of Higgs production and decay rates $\sigma_h B_h$, we get, $$\begin{aligned} \frac{\sigma_h B_h \vert_{\rm RS}}{\sigma_h B_h \vert_{\rm SM}} < \frac{1-\delta [\sigma_h B_h]/\sigma_h B_h}{1+\delta [\sigma_h B_h]/\sigma_h B_h} , \label{finalapplied} \end{aligned}$$ where for $\delta(\sigma_h B_h)/\sigma_h B_h$ we take the relative experimental accuracy according to the prospects at LHC in the measurement of rates for specific individual channels (with $\int {\cal L} dt =30$ fb$^{-1}$ as could be collected after several years of run) [@Duhrssen]. For example, considering the point C of parameter space as above, we find that this condition is fulfilled for the five channels \[combining all theoretical predictions and experimental sensitivities\]: $$\frac{\sigma(qq \to hqq) \vert_{\rm RS}}{\sigma(qq \to hqq) \vert_{\rm SM}} \ \frac{B(h\to WW) \vert_{\rm RS}}{B(h\to WW) \vert_{\rm SM}} = [0.30, 0.32] < \frac{1-0.13}{1+0.13} = 0.76$$ $$\frac{\sigma(qq \to hqq) \vert_{\rm RS}}{\sigma(qq \to hqq) \vert_{\rm SM}} \ \frac{B(h\to ZZ) \vert_{\rm RS}}{B(h\to ZZ) \vert_{\rm SM}} = [0.30, 0.31] < \frac{1-0.49}{1+0.49} = 0.34$$ $$\frac{\sigma(gg \to h) \vert_{\rm RS}}{\sigma(gg \to h) \vert_{\rm SM}} \ \frac{B(h\to WW) \vert_{\rm RS}}{B(h\to WW) \vert_{\rm SM}} = [0.55, 0.60] < \frac{1-0.12}{1+0.12} = 0.78$$ $$\frac{\sigma(gg \to h) \vert_{\rm RS}}{\sigma(gg \to h) \vert_{\rm SM}} \ \frac{B(h\to ZZ) \vert_{\rm RS}}{B(h\to ZZ) \vert_{\rm SM}} = [0.54, 0.60] < \frac{1-0.24}{1+0.24} = 0.61$$ $$\begin{aligned} \frac{\sigma(qq \to Wh) \vert_{\rm RS}}{\sigma(qq \to Wh) \vert_{\rm SM}} \ \frac{B(h\to WW) \vert_{\rm RS}}{B(h\to WW) \vert_{\rm SM}} = [0.31, 0.32] < \frac{1-0.42}{1+0.42} = 0.40 \label{finalnumeric} \end{aligned}$$ $\sigma(qq \to hqq)$ being the cross section for the Weak Boson Fusion (WBF) mechanism. Taking the ratio of squared amplitudes allow to include implicitly the SM loop corrections and the calculated branching ratios also include those corrections. The cross sections for the WBF mechanism, the $Wh$ production and the gluon–gluon fusion process are all significantly reduced w.r.t. SM due to the large decrease of the effective couplings $g_{hVV}$ and $g_{hgg}$ within RS (studied in Table \[MainTable\]). The branching ratio for the decay channel $h\to VV$ is also reduced in RS since $g_{hVV}$ is more reduced than the bottom Yukawa coupling $\lambda_b$ (as discussed above) and the two channels $h\to VV^*$,$h\to b \bar b$ dominate for $m_h = 150$ GeV. By the way, the variations of the values given in Eq.(\[finalnumeric\]) are due in particular to the variation of $\lambda^{\rm RS}_b$ given in Table \[MainTable\]. The significant reductions of the cross sections and branching ratios appearing in Eq.(\[finalnumeric\]) together with the expected LHC performances on $\delta [\sigma_h B_h]/\sigma_h B_h$ allow to satisfy the condition (\[finalapplied\]), or in other words make the RS corrections visible at LHC. For instance with the first channel, the condition for having a $95\%$ probability to observe at least a $2\sigma$ deviation, due to RS effects, is even fulfilled: $$\frac{\sigma(qq \to hqq) \vert_{\rm RS}}{\sigma(qq \to hqq) \vert_{\rm SM}} \ \frac{B(h\to WW) \vert_{\rm RS}}{B(h\to WW) \vert_{\rm SM}} = [0.30, 0.32] < \frac{1-2 \times 0.13}{1+2 \times 0.13} = 0.58$$ In conclusion, the possibly large RS corrections to the Higgs couplings induce deviations w.r.t. SM of the Higgs production and decay rates which could be detected at LHC. It means that the existence of extra dimensions, and more particularly of the RS model, could be probed at the LHC via investigations on the Higgs sector only. The above conclusion is true for the effective Higgs couplings to gluons and to $W$, $Z$ bosons. In contrast, although the RS corrections to the effective $h\gamma\gamma$ coupling can also be large, those cannot be individually tested at LHC for $\int {\cal L} dt =30$ fb$^{-1}$. This is due to the smallness of the branching $B(h\to \gamma\gamma)$ which degrades the experimental sensitivity on the $\gamma\gamma$ channel. As a matter of fact, let us consider the mass $m_h = 120$ GeV at which the promising CMS sensitivity to this channel is optimum [@CMSsig]. For the point A of parameter space, where the deviation to $g_{h\gamma\gamma}$ coming from KK effects is maximal, one gets, $$\begin{aligned} \frac{g^{\rm RS}_{h\gamma\gamma}}{g^{\rm SM}_{h\gamma\gamma}} = [0.74, 0.75] > \frac{1-\delta g_{h\gamma\gamma}/g_{h\gamma\gamma}}{1+\delta g_{h\gamma\gamma}/g_{h\gamma\gamma}} = 0.52, \label{finalgamma} \end{aligned}$$ where $\delta g_{h\gamma\gamma}/g_{h\gamma\gamma} = 0.31$ is the relative error bar that can be obtained on the Higgs coupling by combining several measurements [@DirkZerwas]. It means that the condition (Eq.(\[final\])) on the observability of a $1\sigma$ deviation between $g^{\rm SM}_{h\gamma\gamma}$ and its measured central value is not fulfilled. This constitutes a conservative result w.r.t. the theoretical parameter spanning. From the experimental point of view, the uncertainty on the measured central value makes the observation of a deviation on $g_{h\gamma\gamma}$ even more difficult. ### ILC The two dominant Higgs production reactions at future $e^+e^-$ Linear Colliders (LC) are the Higgs–strahlung mechanism and the $WW$ fusion process [@HiggsReviewI]. For an intermediate center-of-mass energy of $350 - 500$ GeV and an integrated luminosity of ${\cal L}=500$ fb$^{-1}$ (typically after one or two years of run), the experimental accuracy expected in the measurement of the $e^+e^- \to Zh$ cross section is $\pm 2.5\%$ for $m_h = 120$ GeV [@TeslaTDR]. More recent studies dedicated to the International LC (ILC) performances [@ILC-FR] conclude that this accuracy can be improved down to $\pm 2\%$ with only $\sqrt{s}=250$ GeV and ${\cal L}=250$ fb$^{-1}$. For Higgs masses of $140-160$ GeV, the accuracy is a bit worst. The cross section determination is independent of the Higgs decay modes: it is determined through a method analyzing the mass spectrum of the system recoiling against the $Z$ boson. It is not the case for the process $e^+e^- \to h \nu \nu$ which is experimentally analyzed via the decay $h \to b \bar b$ and is measured with a weaker precision. We do not consider this process here. The separation of the Higgs–strahlung mechanism and the $WW$ fusion process is partially controllable by properly choosing the beam polarization configurations. Let us consider RS points A and B as the results will be only slightly modified at $m_h = 150$ GeV due to the stronger EWPT limits on $M_{KK}$ and the weaker experimental sensitivity. Before presenting the numerical results, we mention that below $\sqrt{s}=500$ GeV the direct effect due to the exchange of KK $Z$ excitations in $e^+e^- \to Zh$ is negligible compared to the $Z^0$–$Z^{KK}$ mixing effect on the $g_{hZZ}$ coupling. The RS corrections to the $Z$ boson mass, width and coupling to $e^+e^-$ are also negligible due to the strong EWPT constraints. Assuming the parameter set A/B, the ILC experiment would measure a cross section $\sigma(e^+e^- \to Zh)$ in the interval $[\sigma_-,\sigma_+]$ where, according to Eq.(\[intermed\]), $$\begin{aligned} \sigma_\pm = \sigma^{\rm RS} \frac{1 \pm 2\delta\sigma/\sigma}{1 \mp 2\delta\sigma/\sigma} = \sigma^{\rm SM} \frac{\sigma^{\rm RS}}{\sigma^{\rm SM}} \frac{1 \pm 2\delta\sigma/\sigma}{1 \mp 2\delta\sigma/\sigma}, \label{intermedILC} \end{aligned}$$ $\delta\sigma/\sigma$ being the experimental accuracy. Having replaced $\delta\sigma/\sigma$ by $2\delta\sigma/\sigma$, the probability to obtain a measure in this interval is of $95\%$ like the $C.L.$ for this measurement. Taking $\sigma^{\rm RS} / \sigma^{\rm SM} = ( g^{\rm RS}_{hZZ} / g^{\rm SM}_{hZZ} )^2$ (SM radiative corrections are compensated) and the dependence $\delta\sigma/\sigma = \sqrt{\sigma^{\rm SM} / \sigma^{\rm RS}} \times 2 \%$ [@ILC-FR], we find for the $[\sigma_-,\sigma_+]$ ranges: $$[0.28,0.37] \ \times \ \sigma^{\rm SM} \ \ \ \mbox{(point A), \ and,} \ \ \ [0.69,0.83] \ \times \ \sigma^{\rm SM} \ \ \ \mbox{(point B).}$$ These intervals would translate into possible Higgs couplings measurements in the following ranges (to be compared with the starting theoretical values for $g^{\rm RS}_{hZZ} / g^{\rm SM}_{hZZ}$ in Table \[MainTable\]): $$0.53 < \frac{g^{\rm exp}_{hZZ}}{g^{\rm SM}_{hZZ}} < 0.61 \ \ \ \mbox{(point A), \ and,} \ \ \ 0.83 < \frac{g^{\rm exp}_{hZZ}}{g^{\rm SM}_{hZZ}} < 0.91 \ \ \ \mbox{(point B).}$$ These precise ‘reconstructions’ of the couplings illustrate the ILC capability of clearly discriminating between pure SM couplings and couplings affected by RS corrections, in the Higgs–gauge sector. This clear distinction is explained by the possible large RS corrections combined with the high ILC sensitivity on the cross section. Note that for $m_h = 120$ GeV, the ILC can test the whole range of possible $g^{\rm RS}_{hZZ}$ values when $g_{Z'}$ spans the entire allowed interval $0.72-1.57$ (point B - point A), assuming the smallest $M_{KK}$ value allowed by EWPT. In other words, even the regions with small $g_{hZZ}$ deviations (case of point B) can be tested. These deviations would constitute indirect experimental signatures of the RS scenario. Let us finish this part by making a comment on the possibility at ILC of discriminating between different models beyond the SM, using the Higgs–gauge sector. Under the hypothesis $m_h = 120$ GeV, the theoretical $g^{\rm RS}_{hZZ} / g^{\rm SM}_{hZZ}$ value cannot go below $0.57$ \[see Table \[MainTable\]\] and the found experimental value at ILC could not be below $0.53$ (at $95\% C.L.$). Hence a measurement at the ILC of the coupling constant $g_{hZZ}$ completely below $0.53 \times g^{\rm SM}_{hZZ}$ would exclude the RS model. For instance in the Minimal Supersymmetric SM, such a coupling $g^{\rm MSSM}_{hZZ} / g^{\rm SM}_{hZZ} = \sin (\beta-\alpha) \leq 0.53$ is realizable for a CP–odd boson $A^0$ mass $m_A \lesssim 130$ GeV and $\tan \beta \simeq 30$ [@HiggsReviewII]. The Higgs boson pair production ------------------------------- Let us assume for a while that the Higgs boson has been discovered and that its mass has been measured at LHC or ILC. Then an approach similar to the one adopted throughout this paper could be followed. First, one would have access to the quartic Higgs self–coupling, denoted $\lambda_h$, via the measured Higgs mass, $$\begin{aligned} m_h^2 = 2 \ \lambda_h \tilde v^2 + \delta^{\rm SM} m^2_h , \label{HiggsMass} \end{aligned}$$ where $\delta^{\rm SM} m^2_h$ includes the SM quantum corrections. Then one would be able to deduce the triple Higgs coupling strength: $$\begin{aligned} g_{hhh} = 3 \ \lambda_h \tilde v + \delta^{\rm SM} g_{hhh} , \label{HiggsTriple} \end{aligned}$$ as well as the $hhVV$ coupling strength: $$\begin{aligned} g_{hhVV} = \frac{g^2_V}{2} + \delta g_{hhVV} , \label{HiggsDouble} \end{aligned}$$ with $g^2_V=\{ g^2+g'^2 , g^2 \}$ and $\delta g_{hhVV}$ taking into account both SM radiative corrections and KK gauge mixing effects. Therefore, the theoretical values for the rates of the (more challenging) Higgs boson pair production could be computed within the RS model. Indeed, this pair production proceeds e.g. at LHC through [*(i)*]{} the usual single Higgs production diagrams with the final Higgs leg being connected to two other Higgs fields via the $g_{hhh}$ coupling, [*(ii)*]{} a ‘double’ Higgs–strahlung process $qq\to hhV$ with two radiated Higgs bosons involving either two times the $g_{hVV}$ vertex \[plus the $t$–channel contribution\] or one time $g_{hhVV}$ and [*(iii)*]{} the Vector Boson Fusion (VBF) mechanism $qq\to V^*V^* qq\to hh qq$ with the $g_{hhVV}$ coupling instead of the $g_{hVV}$ one. At ILC, the Higgs pair production occurs similarly trough [*(i)*]{} the usual Higgs production diagrams with the final Higgs connected to two Higgs fields via $g_{hhh}$, [*(ii)*]{} a double Higgs–strahlung reaction off $Z$ bosons $e^+e^-\to hhZ$ involving two times the $g_{hZZ}$ coupling or one time $g_{hhZZ}$ and [*(iii)*]{} the VBF mechanism $e^+e^-\to V^*V^* \ell\ell \to hh \ell\ell$ \[$\ell = e^\pm,\nu$\] involving $g_{hhVV}$ couplings. In conclusion, because of the theoretical relation between the Higgs mass and its couplings (via the $\lambda_h$, $\tilde v$ values), the $m_h$, $m_V$, $G_F$ measurements would permit to predict the $g^{\rm RS}_{hVV}$, $g^{\rm RS}_{hhVV}$, $g^{\rm RS}_{hhh}$ coupling constants and hence potentially to test the RS model via the Higgs pair production. This new test would be more difficult, in particular due to the smaller rates, but complementary to the single Higgs production at LHC or ILC. RS variants {#variants} =========== In this last part, we discuss qualitatively the implications of the Higgs VEV corrections on the EWPT and Higgs couplings that we have treated in detail above, but within the other versions of warped extra dimension models. First, an alternative scenario to the one we have considered is that ${\rm SU(2)_R}$ remains unbroken in the bulk ($\tilde M = 0$). Then the dominant contribution to $T_{\rm RS}$ comes from the exchange of (excited) $t$ and $b$ quarks at the one–loop level. The estimation of this radiatively generated $T_{\rm RS}$ relies on a sum over fermion/boson KK towers which depends on the choice of quark representations. In such a case, $\tilde v$ would no more depend on $\tilde M$ but the correlation between $S_{\rm RS}$ and $T_{\rm RS}$ through would still be modified by the dependence of the Higgs VEV corrections on these parameters. Indeed $T_{\rm RS}$ would still depend on $\tilde v$, $M_{KK}$ and $g_{Z'}$. Besides, the increase of $\tilde v$ w.r.t. $v_{SM}=245$ GeV would imply an enhancement of $S_{\rm RS}$ which tends to increase the EWPT lower limit on $M_{KK}$, as we also find above. Another possibility within the RS framework is to include large kinetic terms for the gauge fields on the IR brane [@kinetics], without bulk custodial symmetry. Such terms repel the KK mode wave functions from the brane so that the KK gauge mixing effect, coming from the coupling of KK gauge fields to the Higgs boson located at the IR brane, is reduced. Hence the Higgs VEV corrections are expected to be significantly weaker than here, and their effects on EWPT constraints as well as on Higgs couplings softer. In the case of gauge–Higgs unification scenarios, where the extended bulk gauge symmetry generally contains the custodial group ${\rm SU(2)_L\! \times\! SU(2)_R\! \times\! U(1)_X}$, the KK gauge mixing effect \[induced by EWSB\] should be of the same order as in the present paper since the Higgs profile is peaked on the TeV–brane (instead of being exactly confined as here). By consequence, comparable Higgs VEV corrections are expected and in turn similar EWPT bounds on $M_{KK}$. Finally, we mention the so–called gaugephobic Higgs models where the Higgs VEV on the brane can be much larger than here, forcing then the lightest $W$ and $Z$ modes to move further from the brane [@Lillie; @Caccia]. In the limit of an infinite VEV, where the couplings between the Higgs and gauge bosons vanish, one recovers the Higgsless gauge boundary conditions. In such models, to maintain compatibility with EWPT, one must render the corrections to the $S$ parameter small. For that purpose, one has to allow all the light fermions to be spread in the bulk [@EWPTHless]. When their profile becomes approximately flat, their wave function being then orthogonal to the KK gauge boson ones, the contributions to $S$ can be made arbitrarily small. Nevertheless, with such a fermion universality, one clearly looses the beauty of generating the mass hierarchy and flavor structure via the simple geometrical mechanism of wave function overlapping. Conclusion {#conclu} ========== Within the RS framework, the corrections to the Higgs boson VEV induced by the KK gauge mixing can be large. Those imply an enhancement of the EWPT lower limit at $95.45 \% C.L.$ on $M_{KK}$ that can be larger than $+30 \%$ for $m_h=120$ GeV. Another important role of these RS corrections to the VEV is played in the calculation of the Higgs couplings. We find that the Higgs couplings can be significantly reduced w.r.t. SM rendering the Higgs production detection at LHC more delicate. The Higgs rate suppressions also allow to pass the LEP2 constraint with $m_h \simeq 99$ GeV, a mass for which the Higgs discovery at LHC would be tricky. Finally, the large deviations to the Higgs couplings due to extra dimensions provide an indirect way of testing the RS model: the LHC precision in light Higgs rate measurements would allow to explore KK boson mass ranges above $4$ TeV (in agreement with EWPT) through Higgs production/decay channels involving the couplings $g_{hgg}$ and $g_{hVV}$. With the higher accuracies expected at ILC, even more clear signatures of the RS scenario may arise in the precise measurement of $g_{hZZ}$ deviations.\ \ **Acknowledgments:** The authors are grateful to H. Bachacou, S. Gopalakrishna, C. Grojean, P. Lutz, M. Passera and F. Richard for useful discussions. We also thank M. Calvet for her contribution to the manuscript. This work is supported by the HEPTOOLS network and the A.N.R. [*LFV-CPV-LHC*]{} under project NT09\_508531. **Appendix** Gauge boson mass matrices {#MassMat} ========================= The neutral gauge boson mass matrix reads as (writing only the first KK mode contributions), $$\begin{aligned} {\cal M}^2_0 = \left ( \begin{array}{ccc} g_Z^2\frac{\tilde v^2}{4} + \delta^{\rm SM}m^2_Z & g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & - g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{g_{Z'}}{g_Z} \cos^2 \theta' \\ g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & M^2_{KK} + g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) & - g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}}{g_Z} \cos^2 \theta' \\ - g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{g_{Z'}}{g_Z} \cos^2 \theta' & - g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}}{g_Z} \cos^2 \theta' & M'^2_{KK} + g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}^2}{g_Z^2} \cos^4 \theta' \end{array} \right ) , \nonumber\\ \label{NeutMassMat} \end{aligned}$$ with $ g_Z^2 = g^2+g'^2$. The increasing factor $\sqrt{2 k \pi R_c}$ is the ratio of the $Z^{(1)}$ over $Z^0$ wave function amounts at the TeV–brane (where is stuck the Higgs boson). We have checked that numerically this ratio is not significantly different for the three first KK states (independently of the BC: $(++)$ or $(-+)$). However, the sign of the n[*th*]{} KK gauge wave function at the IR boundary goes like $(-1)^{n-1}$. The three first KK masses are respectively $M_{KK} \simeq (2.45;5.57;8.70) k e^{-\pi kR_{c}}$ and $M'_{KK} \simeq (2.40;5.52;8.65) k e^{-\pi kR_{c}}$. The charged gauge boson mass matrix is (note the presence of the $\tilde M^2$ term): $$\begin{aligned} {\cal M}^2_\pm = \left ( \begin{array}{ccc} g^2\frac{\tilde v^2}{4} + \delta^{\rm SM}m^2_W & g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & - g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{\tilde g}{g} \\ g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & M^2_{KK} + g^2\frac{\tilde v^2}{4} (2 k \pi R_c) & - g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g}{g} \\ - g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{\tilde g}{g} & - g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g}{g} & ( M'_{KK} + \frac{\tilde M^2}{4k} e^{-\pi kR_{c}} )^2 + g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g^2}{g^2} \end{array} \right ) . \nonumber\\ \label{ChargMassMat} \end{aligned}$$ Gauge boson couplings {#CouplMat} ===================== The neutral gauge boson couplings to the Higgs boson are given by the following matrix (see Section \[couplings\]), $$\begin{aligned} {\cal C}_0 = \left ( \begin{array}{ccc} g_Z^2\frac{\tilde v^2}{4} + \delta^{\rm SM}g_{hZZ} & g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & - g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{g_{Z'}}{g_Z} \cos^2 \theta' \\ g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) & - g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}}{g_Z} \cos^2 \theta' \\ - g_Z^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{g_{Z'}}{g_Z} \cos^2 \theta' & - g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}}{g_Z} \cos^2 \theta' & g_Z^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{g_{Z'}^2}{g_Z^2} \cos^4 \theta' \end{array} \right ) \label{NeutCouplMat} \end{aligned}$$ where the irreducible quantum correction to the $hZZ$ vertex is, $$\delta^{\rm SM}g_{hZZ} = - \frac{3 (g^2+g'^2)}{16 \pi^2} \frac{m_b^2+m_t^2}{2} . \label{DSMirrZ}$$ We do not include explicitly the SM radiative corrections due to the Higgs boson self–energy as those constitute a common factor in the RS and SM coupling $hZZ$, thus disappearing in their ratio that we will compute here. The charged gauge boson couplings to the Higgs boson are alternatively determined by $$\begin{aligned} {\cal C}_\pm = \left ( \begin{array}{ccc} g^2\frac{\tilde v^2}{4} + \delta^{\rm SM}g_{hWW} & g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & - g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{\tilde g}{g} \\ g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} & g^2\frac{\tilde v^2}{4} (2 k \pi R_c) & - g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g}{g} \\ - g^2\frac{\tilde v^2}{4} \sqrt{2 k \pi R_c} \frac{\tilde g}{g} & - g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g}{g} & g^2\frac{\tilde v^2}{4} (2 k \pi R_c) \frac{\tilde g^2}{g^2} \end{array} \right ) \label{ChargCouplMat} \end{aligned}$$ the irreducible quantum correction to the $hWW$ vertex being: $$\delta^{\rm SM}g_{hWW} = - \frac{3 g^2}{16 \pi^2} \bigg ( \frac{m_b^2+m_t^2}{4} - \frac{m_b^4 \ ln(m_b^2 / m_t^2)}{2(m_t^2-m_b^2)} \bigg ) . \label{DSMirrW}$$ Bottom and top quark mass matrices {#FermMat} ================================== In the field basis $\Psi^t_L \equiv (b^{(0)}_L, b^{(1)}_L, b^{c(1)}_L, b''^{(1)}_L, b'^{c(1)}_L)^t$, $\Psi^t_R \equiv (b^{c(0)}_R, b^{(1)}_R, b^{c(1)}_R, b''^{(1)}_R, b'^{c(1)}_R)^t$, the 4D bottom quark mass matrix, up to the first KK modes (numerical analysis includes first and second KK modes) for our model is, $$\begin{aligned} {\cal M}_b = \left ( \begin{array}{ccccc} \epsilon_b \tilde{v}_b c_{\theta} f^{(0)*}_{c_2} f^{(0)}_{c_{b_R}} & 0 & \tilde{v}_b c_{\theta} f^{(0)*}_{c_2} f^{(1)}_{c_{b_R}} & 0 & \sqrt{2} \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} g^{(1)}_{c_{t_R}} \\ \tilde{v}_b c_{\theta} f^{(1)*}_{c_2} f^{(0)}_{c_{b_R}} & s^2_{\theta} m^{(1)}_{c_1} + c^2_{\theta} m^{(1)}_{c_2} & \tilde{v}_b c_{\theta} f^{(1)*}_{c_2} f^{(1)}_{c_{b_R}} & 0 & \sqrt{2} \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} g^{(1)}_{c_{t_R}} \\ 0 & 0 & m^{(1)}_{c_{b_R}} & 0 & 0 \\ {1\over\sqrt{2}} \tilde{v}_b g^{(1)*}_{c_2} f^{(0)}_{c_{b_R}} & 0 & {1\over\sqrt{2}} \tilde{v}_b c_{\theta} g^{(1)*}_{c_2} f^{(1)}_{c_{b_R}} & m'^{(1)}_{c_2} & 0 \\ 0 & 0 & 0 & 0 & m'^{(1)}_{c_{t_R}} \end{array} \right ) . \label{BotMassMat}\end{aligned}$$ In the field basis $\Phi^t_L \equiv (t^{(0)}_L, t^{(1)}_L, t^{c(1)}_L, t'^{(1)}_L)^t$, $\Phi^t_R \equiv (t^{c(0)}_R, t^{(1)}_R, t^{c(1)}_R, t'^{(1)}_R)^t$, the top quark mass matrix up to the first KK states (same comment for the numerical analysis) is given by: $$\begin{aligned} {\cal M}_t = \left ( \begin{array}{cccc} \epsilon_t \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} f^{(0)}_{c_{t_R}} & 0 & \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} f^{(1)}_{c_{t_R}} & 0 \\ \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} f^{(0)}_{c_{t_R}} & s^2_{\theta} m^{(1)}_{c_1} + c^2_{\theta} m^{(1)}_{c_2} & \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} f^{(1)}_{c_{t_R}} & 0 \\ 0 & 0 & m^{(1)}_{c_{t_R}} & 0 \\ \tilde{v}_t s_{\theta} g^{(1)*}_{c_1} f^{(0)}_{c_{t_R}} & 0 & \tilde{v}_t s_{\theta} g^{(1)*}_{c_1} f^{(1)}_{c_{t_R}} & m'^{(1)}_{c_1} \end{array} \right ) . \label{TopMassMat}\end{aligned}$$ $\epsilon_{b,t}$ is described in Section \[couplYuk\]. Besides, in our notations, $\tilde{v}_{b,t} = \lambda^{5D}_{b,t} \tilde{v}/\sqrt{2}kR_c$, $m^{(n)}_c$ ($m'^{(n)}_c$) is the $n$–[*th*]{} KK mass for $(++)$ ($(-+)$) BC fields, $c_{\theta}=\cos \theta$ ($s_{\theta}=\sin \theta$) and $\theta$ is the effective angle of the mixing between the two left multiplets. $f^{(n)}_c$ and $g^{(n)}_c$ are respectively the fermion wave functions along the $5$–[*th*]{} dimension with $(++)$ and $(-+)$ BC, whose values are taken at the position of the TeV–brane, $x_5 = \pi R_c$ (where the Higgs boson is confined). For instance (see e.g. [@AgaSer] for excited profiles): $$\begin{aligned} f^{(0)}_c (x_5) \equiv \sqrt{{(1-2c)k R_c}\over{e^{(1-2c)\pi k R_c}-1}}e^{({1\over2}-c)k x_5}.\end{aligned}$$ The zeroes in the bottom mass matrix originate from the fact that the fields $b^{c(1)}_L$, $b'^{c(1)}_L$, $b^{(1)}_R$ and $b''^{(1)}_R$ (with $n=1,2$) have Dirichlet BC on the TeV–brane and thus, do not couple to the Higgs boson. For the top mass matrix, it originates from the fact that $t^{c(1)}_L$, $t^{(1)}_R$ and $t'^{(1)}_R$ have Dirichlet BC on the TeV-brane. Bottom and top quark Coupling Matrices {#FermCoupl} ====================================== The bottom quark Yukawa coupling matrix reads as, $$\begin{aligned} {\cal C}_b = \left ( \begin{array}{ccccc} \epsilon_b \tilde{v}_b c_{\theta} f^{(0)*}_{c_2} f^{(0)}_{c_{b_R}} & 0 & \tilde{v}_b c_{\theta} f^{(0)*}_{c_2} f^{(1)}_{c_{b_R}} & 0 & \sqrt{2} \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} g^{(1)}_{c_{t_R}} \\ \tilde{v}_b c_{\theta} f^{(1)*}_{c_2} f^{(0)}_{c_{b_R}} & 0 & \tilde{v}_b c_{\theta} f^{(1)*}_{c_2} f^{(1)}_{c_{b_R}} & 0 & \sqrt{2} \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} g^{(1)}_{c_{t_R}} \\ 0 & 0 & 0 & 0 & 0 \\ {1\over\sqrt{2}} \tilde{v}_b g^{(1)*}_{c_2} f^{(0)}_{c_{b_R}} & 0 & {1\over\sqrt{2}} \tilde{v}_b c_{\theta} g^{(1)*}_{c_2} f^{(1)}_{c_{b_R}} & 0 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{array} \right ) , \label{CBotMassMat}\end{aligned}$$ It is obtained from the bottom mass matrix defined in previous appendix, but with the KK mass terms set to zero. Hence, it can be defined through the relation: $$\begin{aligned} {\cal C}_b \equiv \tilde{v}{{\partial\mathcal{M}_b}\over{\partial \tilde{v}}} , \label{exPartial}\end{aligned}$$ as may be useful. The dimensionful top quark Yukawa couplings can be derived from the matrix: $$\begin{aligned} {\cal C}_t = \left ( \begin{array}{cccc} \epsilon_t \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} f^{(0)}_{c_{t_R}} & 0 & \tilde{v}_t s_{\theta} f^{(0)*}_{c_1} f^{(1)}_{c_{t_R}} & 0 \\ \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} f^{(0)}_{c_{t_R}} & 0 & \tilde{v}_t s_{\theta} f^{(1)*}_{c_1} f^{(1)}_{c_{t_R}} & 0 \\ 0 & 0 & 0 & 0 \\ \tilde{v}_t s_{\theta} g^{(1)*}_{c_1} f^{(0)}_{c_{t_R}} & 0 & \tilde{v}_t s_{\theta} g^{(1)*}_{c_1} f^{(1)}_{c_{t_R}} & 0 \end{array} \right ) . \label{CTopMassMat}\end{aligned}$$ Fermion Sum Rule {#FermionSumRule} ================ In this Appendix, we demonstrate analytically some useful theoretical relations which allow to take into account the full KK quark tower in the calculation of the gluon–gluon fusion mechanism amplitude, within the RS scenario. This generalizes results obtained in the case of a bulk Higgs boson within the framework of gauge–Higgs unification [@Adam]. These relations also allow to implement the full KK charged fermion tower when calculating the loop-induced $h\gamma\gamma$ coupling. For that, we use the same conventions/notations as in the Section \[couplYuk\]: $\mathcal{M}_f$ is defined as the mass matrix of a particular fermion $\Psi_{L/R}$ in the interaction basis. We have obtained $\mathcal{C}'_f/\tilde v$ to be the Yukawa matrix in the mass eigenbasis, where $\mathcal{C}'_f=U_L\mathcal{C}_f U^{\dagger}_R$; $\mathcal{C}_f$ being the same matrix as $\mathcal{M}_f$ but with the KK masses set to zero. The matrices $\mathcal{M}_f$ and $\mathcal{C}'_f$ are thus linked through the relation (see for instance Eq.(\[exPartial\])): $$\begin{aligned} \mathcal{C}'_f \equiv U_L . \tilde v{{\partial\mathcal{M}_f}\over{\partial \tilde v}} . U^{\dagger}_R\end{aligned}$$ Let us define here, for simplicity about the subscript notation, $m_i \equiv \mathcal{M}'_{f}|_{ii}$ and $\lambda^{\rm RS}_i \equiv (\mathcal{C}'_f|_{ii})/\tilde v$ for a given fermion within RS. Then we have $$\begin{aligned} \sum_i{{\lambda^{\rm RS}_i\,\tilde v}\over{m_i}} \equiv \mbox{Tr}\left(\mathcal{M'}_{f}^{-1}.{\mathcal{C'}_{f}}\right) \quad\mbox{where}\quad \mathcal{M'}_{f}^{-1}.\mathcal{M'}_{f} = 1.\end{aligned}$$ and can rewrite this trace as (accordingly to Eq.(\[MfDIAGO\])): $$\begin{aligned} \sum_i{{\lambda^{\rm RS}_i\,\tilde v}\over{m_i}}= \mbox{Tr}\left( U_R \mathcal{M}_f^{-1} U^{\dagger}_L. U_L \tilde v{\partial\mathcal{M}_f\over{\partial \tilde v}} U^{\dagger}_R \right) = \mbox{Tr}\left( \tilde v{\partial\mathcal{M}_f\over{\partial \tilde v}}.\mathcal{M}_f^{-1} \right)\end{aligned}$$ $$\begin{aligned} \sum_i{{\lambda^{\rm RS}_i\,\tilde v}\over{m_i}} = \tilde v{\partial\over{\partial \tilde v}} \mbox{Tr}\left( ln \ \mathcal{M}_f \right) = \tilde v{\partial\over{\partial \tilde v}} \ ln \left( \mbox{Det}\mathcal{M}_f \right)\end{aligned}$$ Applying this result to SM fermion mass matrices such as the ones in the Appendix \[FermMat\] (including even possibly the entire KK tower contribution), $$\begin{aligned} \sum_i{{\lambda^{\rm RS}_i\,\tilde v}\over{m_i}} = \tilde v{\partial\over{\partial \tilde v}} \sum_i \ ln \ \mathcal{M}_f|_{ii} = \tilde v{\partial\over{\partial \tilde v}} \ ln \ \mathcal{M}_f|_{11} = 1 \label{eq:FermRule}\end{aligned}$$ and for exotic fermions without zero–mode, as for instance the $q_{(-7/3)}'$ field appearing in a multiplet of Eq.(\[Model\]), $$\begin{aligned} \sum_i{{\lambda^{\rm RS}_i\,\tilde v}\over{m_i}} = \tilde v{\partial\over{\partial \tilde v}} \sum_i \ ln \ \mathcal{M}_f|_{ii} = 0 . \label{eq:FermRulePrime}\end{aligned}$$ .5cm\ \ In both Eq.(\[eq:hgg\]) and Eq.(\[eq:hgaga\]), the fermionic contribution to the amplitude reads, for each independent KK tower, $\sum_{\{f\}} {{\lambda^{\rm RS}_f\tilde{v}}\over{m_f}} A^h_{1/2}(\tau_f)$, up to irrelevant global color and electric charge factors. The sum over $\{f\}$ denotes the sum for a corresponding fermion $f$, of its zero–mode and all KK excitations it couples to through the Higgs field. Using the properties of the spin–$1/2$ form factor, one can set that for all KK excitations, $A^h_{1/2}(\tau_{f^{KK}}) \simeq 1$ with a high precision (up to 1 per 1000). Having so, one can separate the contribution in the tower of the first fermion eigenstate (generally mainly composed by the zero–mode component) noted $f^0$ from the heavier eigenstates (mainly made of KK modes) noted $f^{KK}$’s for simplicity here: $$\begin{aligned} &&\sum_{\{f\}} {{\lambda^{\rm RS}_f\tilde{v}}\over{m_f}} A^h_{1/2}(\tau_f) \equiv {{\lambda^{\rm RS}_{f^0}\tilde{v}}\over{m_{f^0}}} A^h_{1/2}(\tau_{f^0}) + \sum_{KK} {{\lambda^{\rm RS}_{f^{KK}}\tilde{v}}\over{m_{f^{KK}}}} A^h_{1/2}(\tau_{f^{KK}}) = {{\lambda^{\rm RS}_{f^0}\tilde{v}}\over{m_{f^0}}} A^h_{1/2}(\tau_{f^0}) + \sum_{KK} {{\lambda^{\rm RS}_{f^{KK}}\tilde{v}}\over{m_{f^{KK}}}} \nonumber\\ &&\sum_{\{f\}} {{\lambda^{\rm RS}_f\tilde{v}}\over{m_f}} A^h_{1/2}(\tau_f) = {{\lambda^{\rm RS}_{f^0}\tilde{v}}\over{m_{f^0}}} A^h_{1/2}(\tau_{f^0}) + \left( 1 - {{\lambda^{\rm RS}_{f^0}\tilde{v}}\over{m_{f^0}}} \right) = 1 + {{\lambda^{\rm RS}_{f^0}\tilde{v}}\over{m_{f^0}}} \left( A^h_{1/2}(\tau_{f^0}) - 1 \right) , \label{SUMq}\end{aligned}$$ where we have used the relation (\[eq:FermRule\]). It is a remarkable feature that the KK sum can be reflected in a few properties from the lightest mode. Note that for an exotic fermion $f'$, with no zero–mode, the sum even cancels: $$\begin{aligned} \sum_{\{f'\}} {{\lambda^{\rm RS}_{f'}\tilde{v}}\over{m_{f'}}} A^h_{1/2}(\tau_{f'}) \equiv \sum_{KK} {{\lambda^{\rm RS}_{f^{KK}}\tilde{v}}\over{m_{f^{KK}}}} A^h_{1/2}(\tau_{f^{KK}}) = \sum_{KK} {{\lambda^{\rm RS}_{f^{KK}}\tilde{v}}\over{m_{f^{KK}}}} = 0 \label{SUMqprime}\end{aligned}$$ according to the relation (\[eq:FermRulePrime\]). Boson Sum Rule {#BosonSumRule} ============== Here are shown some theoretical sum rules for the bosons, similarly to the fermion case of previous appendix, but which apply now to the computation of the effective $h\gamma\gamma$ vertex. The coupling matrix ${\cal C}'_{\pm}$ and mass matrix ${\cal M}^2_{\pm}$ for the charged gauge bosons are linked via, $$\begin{aligned} {\cal C}'_{\pm} \equiv V . \tilde v^2{{\partial{\cal M}^2_{\pm}}\over{\partial \tilde v^2}} . V^{\dagger}\end{aligned}$$ Reminding $m_{W^n}^2 = \mathcal{M'}^2_\pm|_{nn}$ and $g^{\rm RS}_{hW^nW^n} = 2 (\mathcal{C'}_\pm|_{nn})/\tilde v$ from Section \[hVVsec\], the loop part from the gauge boson KK tower can then be written as \[see Eq.(\[McEIGEN\])\]: $$\begin{aligned} &&\sum_{n} {{g^{\rm RS}_{W^n} \tilde v}\over{2m_{W^n}^2}} = \sum_{n} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} = \mbox{Tr}\left([\mathcal{M'}^2_{\pm}]^{-1}.{\mathcal{C'}_{\pm}}\right) = \mbox{Tr}\left(V [\mathcal{M}^2_{\pm}]^{-1} V^{\dagger}.V \tilde v^2{{\partial\mathcal{M}^2_{\pm}}\over{\partial \tilde v^2}} V^{\dagger}\right) \nonumber\\ &&\sum_{n} {{g^{\rm RS}_{W^n} \tilde v}\over{2m_{W^n}^2}} = \mbox{Tr}\left(\tilde v^2{{\partial\mathcal{M}^2_{\pm}}\over{\partial \tilde v^2}}.[\mathcal{M}^2_{\pm}]^{-1}\right) = \tilde v^2{{\partial}\over{\partial \tilde v^2}}\mbox{Tr}\left(ln \ \mathcal{M}^2_{\pm}\right) = \tilde v^2{{\partial}\over{\partial \tilde v^2}} \ ln \left( \mbox{Det}\mathcal{M}^2_{\pm} \right) .\end{aligned}$$ Obviously, this relation also holds for the neutral gauge boson KK tower, and one can check that $$\begin{aligned} \mbox{Det}\mathcal{M}^2_{\pm} = g^2{{\tilde{v}^2}\over{4}} \prod_{n} \ M_{KK}^{(n)2} \ \bar M_{KK}^{\prime (n)2},\end{aligned}$$ where $\bar M_{KK}^{\prime(n)2} = ( M_{KK}^{\prime(n)} + \frac{\tilde M^2}{4k} e^{-\pi kR_{c}} )^2$, so that, at the end: $$\begin{aligned} \sum_{n} {{g^{\rm RS}_{W^n} \tilde v}\over{2m_{W^n}^2}} = \tilde{v}^2{\partial\over{\partial \tilde{v}^2}} \ ln (g^2{{\tilde{v}^2}\over{4}}) = 1 .\end{aligned}$$ .5cm\ \ Applying the above result to the EW gauge boson KK tower contribution in the $h\gamma\gamma$ effective coupling allows one to rewrite: $$\begin{aligned} &&\sum_{n} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} A^h_{1}(\tau_{W^{n}}) \ \equiv {{g^{\rm RS}_{_{hWW}}\tilde{v}}\over{2m_W^2}} A^h_{1}(\tau_W) + \sum_{n\geq1} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} A^h_{1}(\tau_{W^{n}}) \nonumber\\ &&\sum_{n} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} A^h_{1}(\tau_{W^{n}}) = {{g^{\rm RS}_{_{hWW}}\tilde{v}}\over{2m_W^2}} A^h_{1}(\tau_W) - 7 \sum_{n\geq1} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} = {{g^{\rm RS}_{_{hWW}}\tilde{v}}\over{2m_W^2}} A^h_{1}(\tau_W) - 7 \left( 1 - {{g^{\rm RS}_{_{hWW}}\tilde{v}}\over{2m_W^2}} \right) \nonumber\\ &&\sum_{n} {{\mathcal{C'}_\pm|_{nn}}\over{\mathcal{M'}^2_\pm|_{nn}}} A^h_{1}(\tau_{W^{n}}) = - 7 + {{g^{\rm RS}_{_{hWW}}\tilde{v}}\over{2m_W^2}} \left( 7 + A^h_{1}(\tau_W) \right) ,\end{aligned}$$ where we have used the properties of the spin–$1$ form factor, so one can set that for all KK excitations, $A^h_{1}(\tau_{W^{n}}) = -7$ again at a high accuracy level (up to 1 per 1000). Experimental sensitivity on Higgs observables {#Consider} ============================================= Given the expected relative precision $\delta{\cal Q}/{\cal Q}$ on a quantity ${\cal Q}$ to be measured in the Higgs boson sector, one can predict that there is a probability of $68\%$ that the measurement will be at most given by, $$\begin{aligned} {\cal Q}_{\rm exp}^{\rm max} = {\cal Q}_{\rm RS} + \frac{\delta{\cal Q}}{{\cal Q}} {\cal Q}_{\rm exp}^{\rm max} , \label{DQsQ} \end{aligned}$$ assuming that the exact value ${\cal Q}_{\rm RS}$ is the one obtained theoretically in the RS model. Then an experimental error of $(\delta{\cal Q}/{\cal Q}){\cal Q}_{\rm exp}^{\rm max}$ would be taken. Hence, the largest experimental value that can be expected at a $68 \% C.L.$ is $$\begin{aligned} {\cal Q}_{\rm exp}^{\rm limit} = {\cal Q}_{\rm RS} + \ 2 \times \ \frac{\delta{\cal Q}}{{\cal Q}} {\cal Q}_{\rm exp}^{\rm max} . \label{Qup} \end{aligned}$$ From Eq.(\[DQsQ\]) and Eq.(\[Qup\]), we deduce $$\begin{aligned} {\cal Q}_{\rm exp}^{\rm limit} = {\cal Q}_{\rm RS} + \ 2 \times \ \frac{\delta{\cal Q}}{{\cal Q}} \frac{{\cal Q}_{\rm RS}}{1-\delta{\cal Q}/{\cal Q}} = {\cal Q}_{\rm RS} \frac{1+\delta{\cal Q}/{\cal Q}}{1-\delta{\cal Q}/{\cal Q}}. \label{intermed} \end{aligned}$$ The condition to observe experimentally an RS negative correction to ${\cal Q}$, or in other words to measure a ${\cal Q}$ value smaller than the SM one: ${\cal Q}_{\rm SM}$, reads as $$\begin{aligned} {\cal Q}_{\rm SM} > {\cal Q}_{\rm exp}^{\rm limit} . \label{condition} \end{aligned}$$ Assuming $\delta{\cal Q}/{\cal Q}<1$, one obtains the final condition from combining Eq.(\[intermed\])-(\[condition\]), $$\begin{aligned} \frac{{\cal Q}_{\rm RS}}{{\cal Q}_{\rm SM}} < \frac{1-\delta{\cal Q}/{\cal Q}}{1+\delta{\cal Q}/{\cal Q}} . \label{final} \end{aligned}$$ [999]{} L. Randall and R. Sundrum, Phys. Rev. Lett. **83** (1999) 3370. See also, M. Gogberashvili, Int. J. Mod. Phys. **D11** (2002) 1635. T. Gherghetta and A. Pomarol, Nucl. Phys. **B586** (2000) 141. S. J. Huber and Q. Shafi, Phys. Lett. **B498** (2001) 256; **B512** (2001) 365; Phys. Lett. B544 (2002) 295; Phys. Lett. B583 (2004) 293; S. Chang *et al.*, Phys. Rev. **D73** (2006) 033002; G. Moreau and J. I. Silva-Marcos, JHEP **0601** (2006) 048; **0603** (2006) 090; K. Agashe *et al.*, Phys. Rev. **D71** (2005) 016002; Phys. Rev. Lett. **93** (2004) 201804; Phys. Rev. **D75** (2007) 015002; Phys. Rev. **D74** (2006) 053011; T. Lari *et al.*, Report of Working Group 1 of the CERN Workshop “Flavour in the era of the LHC”, Geneva, Switzerland, November 2005 – March 2007, Eur. Phys. J. **C57** (2008) 183; M. Raidal *et al.*, Report of Working Group 3 of the CERN Workshop “Flavour in the era of the LHC”, Geneva, Switzerland, November 2005 – March 2007, Eur. Phys. J. **C57** (2008) 13. Y. Grossman and M. Neubert, Phys. Lett. **B474** (2000) 361; T. Appelquist *et al.*, Phys. Rev. **D65** (2002) 105019; T. Gherghetta, Phys. Rev. Lett. **92** (2004) 161601; G. Moreau, Eur. Phys. J. **C40** (2005) 539. A. Pomarol, Phys. Rev. Lett. **85** (2000) 4004; L. Randall and M. D. Schwartz, JHEP **0111** (2001) 003; Phys. Rev. Lett. **88** (2002) 081801; K. Agashe, A. Delgado and R. Sundrum, Annals Phys. **304** (2003) 145. See e.g. M. Carena, E. Ponton, J. Santiago and C. E. M. Wagner, Nucl. Phys. **B759** (2006) 202; Phys. Rev. **D76** (2007) 035006. C. Csáki, C. Grojean, H. Murayama, L. Pilo and J. Terning, Phys. Rev. **D69** (2004) 055006; C. Csáki, C. Grojean, L. Pilo and J. Terning, Phys. Rev. Lett. **92** (2004) 101802; G. Cacciapaglia, C. Csáki, C. Grojean and J. Terning, Phys. Rev. **D70** (2004) 075014. J. M. Maldacena, Adv. Theor. Math. Phys. **2** (1998) 231; Int. J. Theor. Phys. **38** (1999) 1113; S. S. Gubser, I. R. Klebanov and A. M. Polyakov, Phys. Lett. **B428** (1998) 105; E. Witten, Adv. Theor. Math. Phys. **2** (1998) 253; N. Arkani-Hamed, M. Porrati and L. Randall, JHEP **0108** (2001) 017. R. Contino, Y. Nomura and A. Pomarol, Nucl. Phys. **B671** (2003) 148; K. Agashe, R. Contino and A. Pomarol, Nucl. Phys. **B719** (2005) 165; K. Agashe and R. Contino, Nucl. Phys. **B742** (2006) 59; R. Contino, L. Da Rold and A. Pomarol, Phys. Rev. **D75** (2007) 055014. M. E. Peskin and T. Takeuchi, Phys. Rev. Lett. **65** (1990) 964; Phys. Rev. **D46** (1992) 381. C. Csáki *et al.*, Phys. Rev. **D66** (2002) 064021; G. Burdman, Phys. Rev. **D66** (2002) 076003; J. Hewett, F. Petriello and T. Rizzo, JHEP **0209** (2002) 030. K. Agashe, A. Delgado, M. J. May and R. Sundrum, JHEP **0308** (2003) 050; A. Delgado and A. Falkowski, JHEP **0705** (2007) 097. K. Agashe *et al.*, Phys. Rev. **D76** (2007) 11505; [arXiv:0810.1497 \[hep-ph\]]{}; B. Lillie, J. Shu and T. M. P. Tait, Phys. Rev. **D76** (2007) 115016; B. Lillie, L. Randall and L.-T. Wang, JHEP **0709** (2007) 074; F. Ledroit, G. Moreau and J. Morel, JHEP **0709** (2007) 071; M. Guchait, F. Mahmoudi and K. Sridhar, Phys. Lett. **B666** (2008) 347; A. Djouadi, G. Moreau and R. K. Singh, Nucl. Phys. **B797** (2008) 1. A. Djouadi, G. Moreau and F. Richard, Nucl. Phys. **B773** (2007) 43; A. Djouadi *et al.*, [arXiv:0906.0604 \[hep-ph\]]{}. A. Djouadi, J. Kühn and P.M. Zerwas, Z. Phys. **C46** (1990) 411. R. Dermisek and J. F. Gunion, Phys. Rev. **D76** (2007) 095006; U. Ellwanger, J. F. Gunion and C. Hugonie, JHEP **0507** (2005) 041; [arXiv:hep-ph/0111179]{}. A. Falkowski, Phys. Rev. **D77** (2008) 055018. G. Bhattacharyya and T. S. Ray, Phys. Lett. **B675** (2009) 222. A. Djouadi and G. Moreau, Phys. Lett. **B660** (2008) 67. G. Cacciapaglia, A. Deandrea and J. Llodra-Perez, [arXiv:0901.0927 \[hep-ph\]]{}. G. F. Giudice, C. Grojean, A. Pomarol and R. Rattazzi, JHEP **0706** (2007) 045. B. Lillie, JHEP **0602** (2006) 019. G. Cacciapaglia, C. Csáki, G. Marandella and J. Terning, JHEP **0702** (2007) 036. M. Veltman, Nucl. Phys. **B123** (1977) 89; M. S. Chanowitz, M. A. Furman and I. Hinchliffe, Phys. Lett. **B78** (1978) 285; A. Djouadi and C. Verzegnassi, Phys. Lett. **B195** (1987) 265; J. J. van der Bij and F. Hoogeveen, Nucl. Phys. **B283** (1987) 477; A. Djouadi, Nuovo Cim. **A100** (1988) 357; B. Kniehl, Nucl. Phys. **B347** (1990) 86; R. Barbieri *et al.*, Phys. Lett. **B288** (1992) 95; A. Djouadi, P. Gambino and B. A. Kniehl, Nucl. Phys. **B523** (1998) 17. Tevatron Electroweak Working Group, [arXiv:0808.0147 \[hep-ex\]]{}. Tevatron Electroweak Working Group, for the CDF Collaboration, the D0 Collaboration, [arXiv:0903.2503 \[hep-ex\]]{}. C. Bouchart and G. Moreau, Nucl. Phys. **B810** (2009) 66. R. Barbieri, A. Pomarol, R. Rattazzi and A. Strumia, Nucl. Phys. **B703** (2004) 127. H. Davoudiasl, G. Perez and A. Soni, Phys. Lett. **B665** (2008) 67. G. Burdman and L. Da Rold, JHEP **0811** (2008) 025. Particle Data Group, J. Phys. **G33** (2006) 1 ; Phys. Lett. **B667** (2008) 1. A. Djouadi, Phys. Rept. **457** (2008) 1. ALEPH Collaboration, CDF Collaboration, D0 Collaboration, DELPHI Collaboration, L3 Collaboration, OPAL Collaboration, SLD Collaboration, LEP Electroweak Working Group, Tevatron Electroweak Working Group, SLD electroweak heavy flavour groups, [arXiv:0811.4682 \[hep-ex\]]{}; [arXiv:0712.0929 \[hep-ex\]]{}. T. Appelquist and J. Carazzone, Phys. Rev. **D11** (1975) 2856. H. Burkhardt and B. Pietrzyk, Phys. Rev. **D72** (2005) 057501. F. Jegerlehner, Nucl. Phys. Proc. Suppl. **181** (2008) 135. K. Hagiwara *et al.*, Phys. Lett. **B649** (2007) 173. M. Davier, S. Eidelman, A. Hocker and Z. Zhang, Eur. Phys. J. **C27** (2003) 497; W. J. Marciano, Invited talks at SSI2004 SLAC Summer Institute, [arXiv:hep-ph/0411179]{}; M. Passera, W. J. Marciano and A. Sirlin, AIP Conf. Proc. **1078** (2009) 378. M. Davier *et al.*, [arXiv:0908.4300 \[hep-ph\]]{}. M. Davier *et al.*, [arXiv:0906.5443 \[hep-ph\]]{}. ALEPH, DELPHI, L3 and OPAL Collaborations, Phys. Lett. **B565** (2003) 61. M. Awramik, M. Czakon, A. Freitas and G. Weiglein, Phys. Rev. **D69** (2004) 053006. G. Degrassi and P. Gambino, Nucl. Phys. **B567** (2000) 3. M. Awramik, M. Czakon and A. Freitas, JHEP **0611** (2006) 048. The LEP Collaborations (ALEPH, DELPHI, L3 and OPAL), the LEP Electroweak Working Group and the SLD Heavy flavor Group, [*A combination of preliminary Electroweak measurements and constraints on the Standard Model*]{}, Phys. Rep. **427** (2006) 257, [http://lepewwg.web.cern.ch/LEPEWWG]{}. M. Davier, S. Descotes-Genon, A. Hocker, B. Malaescu and Z. Zhang, Eur. Phys. J. **C56** (2008) 305. ATLAS Collaboration, [arXiv:0901.0512 \[hep-ex\]]{}. G. L. Bayatian *et al.* \[CMS Collaboration\], *“CMS technical design report, volume II: Physics performance”*, J. Phys.**G34** (2007) 995. S. Dawson *et al.*, Working Group 1 (Higgs) Summary, Proceedings of the Workshop *“From the LHC to Future Colliders (LHC2FC)”*, February 2009, CERN. ;\ [https://twiki.cern.ch/twiki/bin/view/CMS/HiggsAnalysisSummary2009FebCombinedLimits]{} H. Fusaoka and Y. Koide, Phys. Rev. **D57** (1998) 3986. H. Fritzsch and Z-Z. Xing, Prog. Part. Nucl. Phys. **45** (2000) 1. See e.g. M. Spira *et al.*, Nucl. Phys. **B453** (1995) 17. G. Abbiendi *et al.*, Phys. Lett. **B565** (2003) 61. Tevatron New Phenomena, Higgs working group, for the CDF collaboration, D0 collaboration, [arXiv:0903.4001 \[hep-ex\]]{}. M. Dührssen, ATL-PHYS-2003-030, available from [http://cdsweb.cern.ch]{}. R. Lafaye, T. Plehn, M. Rauch, D. Zerwas and M. Duehrssen, [arXiv:0904.3866 \[hep-ph\]]{}. TESLA Technical Design Report Part III: Physics at an $e^+e^-$ Linear Collider, [arXiv:hep-ph/0106315]{}. F. Richard, private communication; Hengne Li, PhD Thesis, To appear. A. Djouadi, Phys. Rept. **459** (2008) 1. M. Carena, A. Delgado, E. Ponton, T. M. P. Tait and C. E. M. Wagner, Phys. Rev. **D68** (2003) 035010; Phys. Rev. **D71** (2005) 015010. G. Cacciapaglia, C. Csáki, C. Grojean and J. Terning, Phys. Rev. **D71**, (2005) 035015. K. Agashe and G. Servant, JCAP **0502** (2005) 002. [^1]: We have checked numerically that the main RS corrections to $G_F$ are oblique, the direct corrections to the leptonic vertex $Wl\nu$ being negligible in comparison. [^2]: There is no $\Pi_{3Q}$ neither $\Pi_{QQ}$ terms in the $m_Z^2$ expression as there is no KK mixing involving the photon. [^3]: The present model does not belong to the category of universal theories considered in [@BPRS]. [^4]: $\theta_{{\rm eff}}^{{\rm lept}}$ denotes the [*Weinberg angle*]{} modified by radiative corrections and the exponent “lept” means that it is the value which can be obtained directly from lepton asymmetry measurements. [^5]: The $\tilde M$ values considered are at worst an order of magnitude close to $k$, so that no new energy scale is introduced in the RS scenario, but remain smaller than $k$ so that a small bulk breaking of the custodial–isospin is guaranteed. [^6]: The precise description for the mentioned mechanism is left to the attention of the reader in our previous paper [@RSrep]. [^7]: By the time the ATLAS and CMS collaborations will have studied $10$ fb$^{-1}$ of data, the Tevatron should have been completed. In case Tevatron runs in 2011, a total collect of $10$ fb$^{-1}$ analyzed data per experiment \[CDF and D0\] is also expected. The latest projections by Tevatron groups conclude from such performances that a $3\sigma$ sensitivity could be achieved for a SM–like Higgs boson in the range $150$ GeV $\leq m_h \leq 180$ GeV [@LHC2FC]. In contrast, the present results for the considered RS parameter set show that with $10$ fb$^{-1}$ the Higgs boson signature observation at center-of-mass energies much lower than $10$ TeV – like at Tevatron – are compromised.

ArXiv

--- abstract: | In standard linear quadratic (LQ) control, the first step in investigating infinite-horizon optimal control is to derive the stabilization condition with the optimal LQ controller. This paper focuses on the stabilization of an It$\hat{o}$ stochastic system with indefinite control and state weighting matrices in the cost functional. A generalized algebraic Riccati equation (GARE) is obtained via the convergence of the generalized differential Riccati equation (GDRE) in the finite-horizon case. More importantly, the necessary and sufficient stabilization conditions for indefinite stochastic control are obtained.\ One of the key techniques is that the solution of the GARE is decomposed into a positive semi-definite matrix that satisfies the singular algebraic Riccati equation (SARE) and a constant matrix that is an element of the set satisfying certain linear matrix inequality conditions. Using the equivalence between the GARE and SARE, we reduce the stabilization of the general indefinite case to that of the definite case, in which the stabilization is studied using a Lyapunov functional defined by the optimal cost functional subject to the SARE. address: 'School of Control Science and Engineering, Shandong University, Jinan, Shandong, P.R.China 250061.' author: - Hongdan Li - Qingyuan Qi - Huanshui Zhang bibliography: - 'autosam.bib' title: 'Stabilization Control for It$\hat{o}$ Stochastic System with Indefinite State and Control Weight Costs' --- , , Indefinite stochastic linear quadratic control; generalized differential Riccati equation; convergence; stabilization. Introduction ============ The linear quadratic (LQ) control pioneered by Kalman [@1Kalman:60] is a classical yet important problem in both theory and engineering applications. In 1968, Wonham [@2Wonham:68] investigated stochastic LQ problems, and this topic has since been studied by many researchers [@6Zhang:2017], [@5Davis:77]. Most results were obtained under the common assumption that the state weighting matrices are non-negative definite and the control weighting matrices are positive definite.\ However, in [@7Chen:98], this common assumption was changed, i.e., stochastic LQ problems with well-posed indefinite control weighting matrices were considered. This phenomenon is known as an indefinite stochastic problem, and it has a deeply uncertain nature; for a more detailed discussion and many examples, see [@7Chen:98]. As they have a wide range of applications, from portfolio selection to pollution control, an increasing number of researchers have studied indefinite stochastic problems. For example, [@8Rami:2001] relaxed the positive definiteness constraint in [@7Chen:98] and solved the indefinite LQ problem by introducing a generalized differential Riccati equation (GDRE). Under the assumption that the system is stabilizable, [@9Wu:2002] showed that the solvability of indefinite stochastic LQ problems in the infinite horizon is equivalent to the existence of a static stabilizing solution to the generalized algebraic Riccati equation (GARE). With regard to discrete cases, [@10Ferrante:2015] derived an equivalent condition for the discrete-time indefinite optimal control problem in finite horizon. The LQ problem for discrete time-invariant systems with arbitrary terminal weight was explored in [@11Bilardi:2007]. [@12Ni:2017] discussed an indefinite stochastic LQ problem with state transmission delay and multiplicative noise. Moreover, because indefinite stochastic LQ problems can be understood as the dual versions of robust filtering problems, some indefinite LQ results have been applied to filtering problems [@13Zorzi:2017], [@14Zorzi:2017].\ The aforementioned papers mainly studied the LQ optimal control problem, whereas relatively few studies have concentrated on the stabilization problem for an indefinite stochastic problem. Nevertheless, the optimal control problem in the infinite horizon case is a worthy topic of research when the system is stabilizable. For instance, [@15Zhang:2018] and [@16Qi:2017] investigated stabilization control for linear discrete- and continuous-time mean-field systems, respectively; in [@17Rami:2000], [@9Wu:2002] and [@18Li:2003], the basic assumption that the system is mean-square stabilizable was imposed throughout. As stated in [@8Rami:2001], indefinite stabilization in the infinite-horizon case is a crucial issue. In this paper, the indefinite stabilization problem for an It$\hat{o}$ stochastic system is investigated.\ The main contributions of this paper are as follows. First, we consider the convergence of the GDRE involving a matrix pseudo-inverse and two additional equality/inequality constraints, which is a weaker requirement than in previous work [@19Rami:2001]. In fact, it is natural to relax the matrix invertibility constraint, because one cannot generally know in advance whether a singularity will occur. Second, in [@19Rami:2001], the asymptotic behavior of the GDRE was only investigated under a strict positive-definite constraint, and the corresponding stabilization results for the system were not given. In contrast, this paper discusses mean-square stabilization for the It$\hat{o}$ stochastic system. The key technique is to decompose the solution of the GARE into a positive semi-definite matrix that satisfies the singular algebraic Riccati equation (SARE) and a constant matrix that is an element of the set satisfying certain linear matrix inequality (LMI) conditions. In view of the equivalence between the GARE and the SARE, the stabilization of the general indefinite case can be reduced to that of the definite case, where the stabilization is studied using a Lyapunov functional defined with the optimal cost functional subject to the SARE.\ The remainder of this paper is organized as follows. Some useful preliminary results are given in Section 2. In Section 3, we discuss the convergence of the GDRE and the mean-square stabilization problem. A two-dimensional numerical example is presented in Section 4, and a summary is provided in Section 5. Preliminaries ============= Let $\mathcal{R}^{n}$ be the family of $n$-dimensional vectors; $M'$ is the transpose of $M$ and $M^{\dagger}$ is the Moore-Penrose pseudo-inverse of $M$; $M > 0 (\geq0)$ denotes a symmetric matrix that is strictly positive-definite (positive semi-definite); $\mathbf{Ker}(M)$ is the kernel of a matrix $M$. $(\Omega,\mathcal{F},P,\mathcal{F}_{t} | t\geq0 )$ is a complete stochastic basis such that $\mathcal{F}_{ 0}$ contains all $P$-null elements of $\mathcal{F}$, and the filtration is generated by the standard Brownian motion $\{w(t)\}_{t\geq0}$. $$\begin{aligned} L_{\mathcal{F}}^{2}(0,T;\mathcal{R}^{m})&=& \Big\{\varphi(t)_{t\in[0,T]}\ \text {is an} \ \mathcal{F}_{t} \ adapted \ stochastic \\ && process \ s.t.\ \ E\int_{0}^{T}\|\varphi(t)\|^{2}dt<\infty\Big\}.\end{aligned}$$\ We consider the following linear It$\hat{o}$ stochastic system: $$\begin{aligned} %方程组开始 \left\{ %方程组的左边包括大括号\{ \begin{array}{lll} %设定列阵的格式：{lll}是三个L，表示三列的对齐方式为Left对齐 dx(t)=[Ax(t)+Bu(t)]dt+[Cx(t)+Du(t)]dw(t), \\ %$――分隔列的标记，\\――表示换行 x(0)=x_{0},%$同上 \end{array} %方程列阵的结束 \right. \label{f01} % 方程组的右边无符号，利用“.“来标示\end{aligned}$$ where $x(t) \in \mathcal{R}^{n} $ is the state; $u(\cdot)$, the admissible control, is any element in $\mathcal{U}_{ad}\equiv L_{\mathcal{F}}^{2}(0,T;\mathcal{R}^{m})$; and $w(t)$ is the one-dimensional standard Brownian motion. $x_{0} \in \mathcal{R}^{n}$ is the initial value, and $A,B,C,D$ are constant matrices with compatible dimensions.\ The associated quadratic cost functional with an infinite horizon is $$\begin{aligned} J(u(\cdot))=\frac{1}{2}E\int_{0}^{\infty}\left[x'(t)Q x(t)+u'(t)R u(t)\right]dt,\label{f02}\end{aligned}$$ where $Q, R$ are symmetric matrices with compatible dimensions.\ [**Problem 1**]{}   Find the optimal control $u(t)=Kx(t)$ with constant matrix gain $K$ that stabilizes (\[f01\]) while minimizing (\[f02\]).\ To solve Problem 1, we define the finite-horizon control as follows:\ Consider system (\[f01\]) with the following cost functional in the finite horizon: $$\begin{aligned} J(u(\cdot))&=&\frac{1}{2}E\int_{0}^{T}\left[x'(t)Q x(t)+u(t)'R u(t)\right]dt\nonumber\\ &&+x'(T)P(T)x(T), \label{f03}\end{aligned}$$ where $Q, R, P(T)$ are symmetric matrices with compatible dimensions.\ The following result shows the necessary and sufficient condition for the optimal control problem in the finite horizon.\ [**Lemma 1 [@8Rami:2001]**]{}   The following conditions are equivalent.\ (1)   There exists an optimal control $u(t)\in \mathcal{U}_{ad}$ that minimizes (\[f03\]) subject to (\[f01\]);\ (2)   The following GDRE admits a solution for $t\in [0,T]$: $$\begin{aligned} \left\{ \begin{array}{lll} \dot{P}(t)+A'P(t)+P(t)A+C'P(t)C+Q\\ -(P(t)B+C'P(t)D)(R+D'P(t)D)^{\dagger} \\ \times(B'P(t)+D'P(t)C)=0, \\ (R+D'P(t)D)(R+D'P(t)D)^{\dagger} \\ \times(B'P(t)+D'P(t)C) =B'P(t)+D'P(t)C,\\ R+D'P(t)D\geq 0, \end{array} \right.\label{f04}\end{aligned}$$ with terminal values $P(T)$. In this case, the optimal controller is $u(t)=K(t)x(t)$, where $K(t)=-(R+D'P(t)D)^{\dagger}(B'P(t)+D'P(t)C)$. Moreover, the value functional is $\inf_{u(\cdot)}J(u(\cdot))=x_{0}'P(0)x_{0}$.\ Based on the above result in the finite time horizon, we will show the main results in the next section. Main results ============ In this section, we discuss the convergence of the GDRE (\[f04\]) and the mean-square stabilization problem of system (\[f01\]).\ Define the GARE as follows: $$\begin{aligned} \left\{ \begin{array}{lll} A'P+PA+C'PC+Q-(PB+C'PD)\\ \times(R+D'PD)^{\dagger}(B'P+D'PC)=0,\\ (R+D'PD)(R+D'PD)^{\dagger}(B'P+D'PC)\\ =B'P+D'PC, \\ R+D'PD\geq 0. \end{array} \right.\label{f06}\end{aligned}$$\ [**Remark 1** ]{} Note that the defined Riccati equation (\[f06\]) is different from that in [@19Rami:2001], where $R+D'PD$ must be positive definite.\ Consider the following set $\mathcal{P}$, which involves the LMI $$\begin{aligned} \mathcal{P}\triangleq \Bigg\{ \begin{smallmatrix} \hat{P}=\hat{P}'\Bigg| &&\left[ \begin{smallmatrix} A'\hat{P}+\hat{P}A+C'\hat{P}C+Q & \hat{P}B+C'\hat{P}D\\ B'\hat{P}+D'\hat{P}C & R+D'\hat{P}D\\ %第二行元素 \end{smallmatrix} \right]\geq 0,\\ &&\mathbf{Ker}(R+D'\hat{P}D)\subseteq (\mathbf{Ker} B \cap \mathbf{Ker} D) \end{smallmatrix}\Bigg\}\end{aligned}$$\ [**Definition 1**]{}    A solution to the GARE (\[f06\]) is called a maximal solution, denoted by $P_{max}$, if $$\begin{aligned} P_{max}\geq \hat{P}, \quad \forall \hat{P}\in\mathcal{P}.\end{aligned}$$ For convenience, we denote $P(t)$ as $P(t,T)$ with the terminal time $T$ in the GDRE (\[f04\]) and the terminal value $P(T,T)=\hat{P}$ with $\hat{P}\in\mathcal{P}$.\ [**Theorem 1**]{}   Assume $\mathcal{P}\neq \emptyset$. If system (\[f01\]) is mean-square stabilizable, then $P(t,T)$ is convergent as $t\rightarrow-\infty$ and the limit of $P(t,T)$ denoted by $\bar{P}$ is the maximal solution to the GARE (\[f06\]).\ [ *Proof*:]{}   Let $\hat{P}\in\mathcal{P}$ and define $$\begin{aligned} \left\{ \begin{array}{lll} Q_{\hat{P}}=A'\hat{P}+\hat{P}A+C'\hat{P}C+Q,\\ L_{\hat{P}}=\hat{P}B+C'\hat{P}D,\\ R_{\hat{P}}=R+D'\hat{P}D. \end{array} \right.\label{f07}\end{aligned}$$ Thus, from $\left[ \begin{array}{cc} Q_{\hat{P}}& L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}}\\ %第二行元素 \end{array} \right]\geq 0$ and Schur’s Lemma, we have $$R_{\hat{P}}\geq 0,\quad Q_{\hat{P}}-L_{\hat{P}}R_{\hat{P}}^{\dagger}L_{\hat{P}}'\geq 0, \quad L_{\hat{P}}(I-R_{\hat{P}}R_{\hat{P}}^{\dagger})=0.$$\ On this basis, i.e., $\left[ \begin{array}{cc} Q_{\hat{P}}& L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}}\\ %第二行元素 \end{array} \right]\geq 0$ and $R_{\hat{P}}\geq 0$, we consider the following singular differential Riccati equation (SDRE) with terminal values $Z_{\hat{P}}(T, T)=0$: $$\begin{aligned} &&\dot{Z}_{\hat{P}}(t,T)+A'Z_{\hat{P}}(t,T)+Z_{\hat{P}}(t,T)A+C'Z_{\hat{P}}(t,T)C +Q_{\hat{P}}\nonumber\\ &&-(Z_{\hat{P}}(t,T)B+C'Z_{\hat{P}}(t,T)D+L_{\hat{P}}) (R_{\hat{P}}+D'Z_{\hat{P}}(t,T)D)^{\dagger}\nonumber\\ &&\times(B'Z_{\hat{P}}(t,T)+D'Z_{\hat{P}}(t,T)C+L_{\hat{P}}')=0,\label{f08}\end{aligned}$$ with the following regular condition $$\begin{aligned} &&[I-(R_{\hat{P}}+D'Z_{\hat{P}}(t,T)D)(R_{\hat{P}}+D'Z_{\hat{P}}(t,T)D)^{\dagger}]\nonumber\\ &&\times(B'Z_{\hat{P}}(t,T)+D'Z_{\hat{P}}(t,T)C+L_{\hat{P}}')=0.\label{f09}\end{aligned}$$ Using the classical ordinary differential equation theory, it is easy to see that there exists a positive semi-definite solution to (\[f08\]). Denote by $Z_{\hat{P}}(\cdot, T)$ the solution of (\[f08\]) with $Z_{\hat{P}}(T, T)=0$.\ Next, we mainly investigate whther $Z_{\hat{P}}(\cdot, T)$ also satisfies (\[f09\]). From the singular value decomposition of $R_{\hat{P}}+D'Z_{\hat{P}}D$ (see Theorem 2.6.3 in [@23Horn:90]), there exists an orthogonal matrix $U$ satisfying $$\begin{aligned} (R_{\hat{P}}+D'Z_{\hat{P}}D)^{\dagger}=U\left[ \begin{array}{cc} \Lambda^{-1} & \textbf{0} \\ \textbf{0}& \textbf{0} \end{array} \right]U',\label{f10}\end{aligned}$$ where $\Lambda$ is an invertible matrix that has the same dimensions as the rank of $R_{\hat{P}}+D'Z_{\hat{P}}D$. Now, decompose $U$ as $\left[ \begin{array}{cc} U_{1} & U_{2} \end{array} \right]$. The orthogonality of $U$ implies that the columns of the matrix $U_{2}$ form a basis of $\mathbf{Ker}(R_{\hat{P}}+D'Z_{\hat{P}}D)$. The positive semi-definite nature of matrices $R_{\hat{P}}, Z_{\hat{P}}$ implies that $\mathbf{Ker}(R_{\hat{P}}+D'Z_{\hat{P}}D)\subseteq \mathbf{Ker}(R_{\hat{P}})$. In view of $\mathbf{Ker}(R_{\hat{P}})\subseteq \mathbf{Ker}(B)\cap \mathbf{Ker}(D)$, we have that $$\begin{aligned} &&(B'Z_{\hat{P}}+D'Z_{\hat{P}}C+L_{\hat{P}})'[I-(R_{\hat{P}}+D'Z_{\hat{P}}D)\nonumber\\ &&\times(R_{\hat{P}}+D'Z_{\hat{P}}D)^{\dagger}]\nonumber\\ &=&(B'Z_{\hat{P}}+D'Z_{\hat{P}}C+L_{\hat{P}})'U_{2}U_{2}'\nonumber\\ &=&[(Z_{\hat{P}}+\hat{P})B+C'(Z_{\hat{P}}+\hat{P})D]U_{2}U_{2}'\nonumber\\ &=&\textbf{0}.\label{f010}\end{aligned}$$ Therefore, $Z_{\hat{P}}(\cdot, T)$ is a solution of the SDRE (\[f08\])–(\[f09\]).\ Considering the following cost functional corresponding to system (\[f01\]): $$\begin{aligned} \tilde{J}_{T}(u(\cdot))=E\int_{0}^{T}\left[ \begin{array}{cc} x(t)\\ u(t) \end{array} \right]'\left[ \begin{array}{cc} Q_{\hat{P}} &L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}} \end{array} \right]\left[ \begin{array}{cc} x(t)\\ u(t) \end{array} \right]dt,\label{f11}\end{aligned}$$ from $\mathcal{P}\neq \emptyset$, we know that $\tilde{J}_{T}(u(\cdot))\geq 0$.\ Applying Lemma 1, we find that $$\begin{aligned} \tilde{V}_{T}\triangleq\inf\limits_{u(\cdot)}\tilde{J}_{T}(u(\cdot)) =x_{0}'Z_{\hat{P}}(0,T)x_{0}.\label{f12}\end{aligned}$$ Note the time-invariance of (\[f08\]) with respect to $T$, with terminal time $T$ and terminal value $Z_{\hat{P}}(T,T)=0$. Thus, for any $t_{1}<t_{2}\leq T$ and for all $x_{0}\neq 0$, we have $$\begin{aligned} &\tilde{V}_{T-t_{1}}=x_{0}'Z_{\hat{P}}(0,T-t_{1})x_{0}=x_{0}'Z_{\hat{P}}(t_{1}, T)x_{0}\\ \geq&\tilde{V}_{T-t_{2}}=x_{0}'Z_{\hat{P}}(0,T-t_{2})x_{0}=x_{0}'Z_{\hat{P}}(t_{2}, T)x_{0}.\end{aligned}$$ Because $x_{0}$ is arbitrary, we have that $Z_{\hat{P}}(t_{1}, T)\geq Z_{\hat{P}}(t_{2}, T)$. Similarly, when $t\leq T_{1}<T_{2}$, $$\begin{aligned} &\tilde{V}_{T_{1}-t}=x_{0}'Z_{\hat{P}}(0,T_{1}-t)x_{0}=x_{0}'Z_{\hat{P}}(t, T_{1})x_{0}\\ \leq&\tilde{V}_{T_{2}-t}=x_{0}'Z_{\hat{P}}(0,T_{2}-t)x_{0}=x_{0}'Z_{\hat{P}}(t, T_{2})x_{0}.\end{aligned}$$ That is, $Z_{\hat{P}}(t, T_{1})\leq Z_{\hat{P}}(t, T_{2})$. Therefore, $Z_{\hat{P}}(t, T)$ is monotonically increasing with respect to $T$ and is monotonically decreasing with respect to $t$.\ Next, we show that $Z_{\hat{P}}(t, T)$ is bounded. The mean-square stabilizability of system (\[f01\]) with the controller $u(t)=Kx(t)$ yields $$\begin{aligned} x_{0}'Z_{\hat{P}}(0,T)x_{0}\leq \tilde{J}_{T}&\leq& E\int_{0}^{\infty}\left[ \begin{array}{cc} x(t)\\ u(t) \end{array} \right]'\left[ \begin{array}{cc} Q_{\hat{P}} &L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}} \end{array} \right]\left[ \begin{array}{cc} x(t)\\ u(t) \end{array} \right]dt\\ &=&E\int_{0}^{\infty}x(t)'\left[ \begin{array}{cc} I\\ K \end{array} \right]'\left[ \begin{array}{cc} Q_{\hat{P}} &L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}} \end{array} \right]\left[ \begin{array}{cc} I\\ K \end{array} \right]\\ &&\cdot x(t)dt\\ &\leq& \lambda_{max}E\int_{0}^{\infty}x'(t)x(t)dt<+\infty,\end{aligned}$$ where $\lambda_{max}$ denotes the maximum eigenvalue of a matrix $\left\{\left[ \begin{array}{cc} I\\ K \end{array} \right]'\left[ \begin{array}{cc} Q_{\hat{P}} &L_{\hat{P}}\\ L_{\hat{P}}' & R_{\hat{P}} \end{array} \right]\left[ \begin{array}{cc} I\\ K \end{array} \right]\right\}$. With an arbitrary $x_{0}$, we find that $Z_{\hat{P}}(t,T)$ is bounded. Hence, we have that $Z_{\hat{P}}(t,T)$ is convergent, i.e., $$\begin{aligned} \lim\limits_{t\rightarrow-\infty}Z_{\hat{P}}(t,T) =\lim\limits_{T\rightarrow+\infty}Z_{\hat{P}}(0,T-t)\triangleq \bar{Z}_{\hat{P}},\end{aligned}$$ where $\bar{Z}_{\hat{P}}$ is a positive semi-definite constant matrix that is independent of $t$.\ Let $t\rightarrow -\infty$ in the SDRE (\[f08\])–(\[f09\]). We have that $\bar{Z}_{\hat{P}}$ is a solution of the following SARE: $$\begin{aligned} \left\{ \begin{array}{lll} A'\bar{Z}_{_{\hat{P}}}+\bar{Z}_{_{\hat{P}}}A+C'\bar{Z}_{_{\hat{P}}}C+Q_{_{\hat{P}}}\\ -(\bar{Z}_{_{\hat{P}}}B+C'\bar{Z}_{_{\hat{P}}}D+L_{_{\hat{P}}})(R_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}D)^{\dagger}\\ \times(B'\bar{Z}_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}C+L_{_{\hat{P}}}')=0,\\ (R_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}D)(R_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}D)^{\dagger}\\ \times(B'\bar{Z}_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}C+L_{_{\hat{P}}}')\\ =B'\bar{Z}_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}C+L_{_{\hat{P}}}', \end{array} \right.\label{f13}\end{aligned}$$\ Define $P(t,T)=Z_{\hat{P}}(t,T)+\hat{P}$. It is easy to verify that $P(t,T)$ is a solution of the GDRE (\[f04\]), is monotonically decreasing with respect to $t$, and is bounded. Therefore, there exists a constant matrix $\bar{P}$ satisfying $$\begin{aligned} \bar{P}=\lim\limits_{t\rightarrow-\infty}P(t,T)=\bar{Z}_{\hat{P}}+\hat{P}.\end{aligned}$$ Clearly, $\bar{P}$ is a solution of the GARE (\[f06\]). Moreover, for arbitrary $\hat{P}$ and $\bar{Z}_{\hat{P}}\geq 0$, it is easy to verify that $\bar{P}\geq\hat{P}$, i.e., $\bar{P}$ is the maximal solution to the GARE (\[f06\]). This completes the proof.\ [**Remark 2**]{}   The above proof implies that the solvability of the GARE (\[f06\]) is equivalent to the solvability of the SARE (\[f13\]).\ To facilitate the proof of the subsequent result, we need the following remark.\ [**Remark 3**]{}   If the SARE (\[f13\]) has a solution $\bar{Z}_{\hat{P}}$, then $$\begin{aligned} A'\bar{Z}_{_{\hat{P}}}+\bar{Z}_{_{\hat{P}}}A+C'\bar{Z}_{_{\hat{P}}}C+Q_{_{\hat{P}}} -M_{_{\hat{P}}}'\Omega_{_{\hat{P}}}^{\dagger}M_{_{\hat{P}}}=0,\label{f14}\end{aligned}$$ where $M_{_{\hat{P}}}=B'\bar{Z}_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}C+L_{_{\hat{P}}}', \Omega_{_{\hat{P}}}=R_{_{\hat{P}}}+D'\bar{Z}_{_{\hat{P}}}D$. In view of the following formula: $$\begin{aligned} M_{_{\hat{P}}}'\Omega_{_{\hat{P}}}^{\dagger}M_{_{\hat{P}}}=-M_{_{\hat{P}}}'K_{_{\hat{P}}}-K_{_{\hat{P}}}'M_{_{\hat{P}}}-K'_{_{\hat{P}}}\Omega_{_{\hat{P}}}K_{_{\hat{P}}},\end{aligned}$$ in which $K_{_{\hat{P}}}=-\Omega_{_{\hat{P}}}^{\dagger}M_{_{\hat{P}}}$, (\[f14\]) can be further rewritten as follows: $$\begin{aligned} \bar{A}'\bar{Z}_{_{\hat{P}}}+\bar{Z}_{_{\hat{P}}}\bar{A} +\bar{C}'\bar{Z}_{_{\hat{P}}}\bar{C}+\bar{Q}_{_{\hat{P}}}=0,\label{f15}\end{aligned}$$ where $$\begin{aligned} \bar{A}=A+BK_{_{\hat{P}}},\quad \bar{C}=C+DK_{_{\hat{P}}}, \quad \nonumber\\ \bar{Q}_{_{\hat{P}}}=Q_{_{\hat{P}}}+L_{_{\hat{P}}}K_{_{\hat{P}}} +K_{_{\hat{P}}}'L'_{_{\hat{P}}}+K_{_{\hat{P}}}'R_{_{\hat{P}}}K_{_{\hat{P}}}.\label{f16}\end{aligned}$$ Using Extended Schur’s Lemma in [@20Albert:69], we have $$\begin{aligned} \bar{Q}_{_{\hat{P}}}&\geq& L_{_{\hat{P}}}R_{_{\hat{P}}}^{\dagger}L_{_{\hat{P}}}'+L_{_{\hat{P}}}R_{_{\hat{P}}}^{\dagger}R_{_{\hat{P}}}K_{_{\hat{P}}}+K_{_{\hat{P}}}'R_{_{\hat{P}}}R_{_{\hat{P}}}^{\dagger}L_{_{\hat{P}}}'\nonumber\\ &&+K_{_{\hat{P}}}'R_{_{\hat{P}}}R_{_{\hat{P}}}^{\dagger}R_{_{\hat{P}}}K_{_{\hat{P}}}\nonumber\\ &=&(L_{_{\hat{P}}}'+R_{_{\hat{P}}}K_{_{\hat{P}}})'R_{_{\hat{P}}}^{\dagger} (L_{_{\hat{P}}}'+R_{_{\hat{P}}}K_{_{\hat{P}}})\geq0.\label{f17}\end{aligned}$$\ [**Definition 2**]{}   Consider the following stochastic system: $$\begin{aligned} \left\{ \begin{array}{lll} dx(t)=Ax(t)dt+Cx(t)dw(t), \\ \mathcal{Y}(t)=Q^{\frac{1}{2}}x(t). \end{array} \right.\label{f18}\end{aligned}$$ $(A, C, Q^{\frac{1}{2}})$ is said to be exact detectable if, for any $T\geq0$, $$\begin{aligned} \mathcal{Y}(t)=0, a.s., \forall t\in[0,T]\quad \Rightarrow \quad\lim\limits_{t\rightarrow +\infty}E(x'(t)x(t))=0.\end{aligned}$$\ [**Assumption 1** ]{}   $(A, C, Q_{\hat{P}}^{\frac{1}{2}})$ is exact detectable.\ [**Lemma 2**]{}   If $(A, C, Q_{\hat{P}}^{\frac{1}{2}})$ is exact detectable, as in Assumption 1, then $(\bar{A}, \bar{C}, \bar{Q}_{\hat{P}}^{\frac{1}{2}})$ is exact detectable, where $\bar{A}, \bar{C}, \bar{Q}_{\hat{P}}$ are defined in (\[f16\]).\ [*Proof*:]{}   The proof is similar to that of Proposition 1 in [@21Zhang:2004], so we omit it here.\ [**Theorem 2** ]{}   If $\mathcal{P}\neq\emptyset$ and Assumption 1 is satisfied, then the closed-loop system (\[f01\]) is mean-square stabilizable if and only if the GARE (\[f06\]) has a solution $\bar{P}$ that is also the maximal solution to (\[f06\]). In this case, the optimal stabilizing controller is given by $$\begin{aligned} u(t)=Kx(t),\label{f19}\end{aligned}$$ where $$\begin{aligned} K=-(R+D'\bar{P}D)^{\dagger}(B'\bar{P}+D'\bar{P}C),\label{f20}\end{aligned}$$ and the optimal cost functional is $J^{\ast}=E(x'_{0}\bar{P}x_{0})$.\ [*Proof*:]{}   “*Sufficiency*": We will show that, under the conditions that $\mathcal{P}\neq\emptyset$ and Assumption 1 holds, when the GARE (\[f06\]) has a solution $\bar{P}$, the closed-loop system (\[f01\]) is mean-square stabilizable.\ Let $\hat{P}\in\mathcal{P}$. From Remark 2, when the GARE (\[f06\]) has a solution $\bar{P}$, the SARE (\[f13\]) has a positive semidefinite solution $\bar{Z}_{\hat{P}}$, i.e., $\bar{Z}_{\hat{P}}\geq0$. Moreover, $\bar{P}=\bar{Z}_{\hat{P}}+\hat{P}$. Next we will show that system (\[f01\]) with $u(t)=Kx(t)$ where $K$ is defined by (\[f20\]), i.e., $$\begin{aligned} dx(t)=(A+BK)x(t)dt+(C+DK)x(t)dw(t)\label{f22}\end{aligned}$$ is mean square stabilizable. In view of the relationship $K=K_{\hat{P}}$ which is defined in Remark 3, we can see that the stabilization for the system (\[f01\]) with $u(t)=Kx(t)$ is equivalent to the stabilization for the system (\[f01\]) with $u(t)=K_{\hat{P}}x(t)$. We define the Lyapunov function candidate as $$\begin{aligned} V(t,x(t))=E[x'(t)\bar{Z}_{\hat{P}}x(t)], \quad t\geq0.\label{f23}\end{aligned}$$ In view of $\bar{Z}_{\hat{P}}\geq0$, the limits of $V(t,x(t))$ exists which can be similarly obtained from the proof of Theorem 4 in [@6Zhang:2017]. And there exists an orthogonal matrix $U_{\hat{P}}$ such that $$\begin{aligned} U_{\hat{P}}'\bar{Z}_{\hat{P}}U_{\hat{P}}=\left[ \begin{array}{cc} 0 & 0\\ 0& \bar{Z}_{\hat{P}_{2}} \end{array} \right], \quad \bar{Z}_{\hat{P}_{2}}>0.\label{f25}\end{aligned}$$\ From (\[f15\]), we have $$\begin{aligned} &&U_{\hat{P}}'\bar{A}'U_{\hat{P}}U_{\hat{P}}'\bar{Z}_{\hat{P}}U_{\hat{P}}+U_{\hat{P}}'\bar{Z}_{\hat{P}}U_{\hat{P}}U_{\hat{P}}'\bar{A}U_{\hat{P}}\nonumber\\ &&+U_{\hat{P}}'\bar{C}'U_{\hat{P}}U_{\hat{P}}'\bar{Z}_{\hat{P}}U_{\hat{P}} U_{\hat{P}}'\bar{C}U_{\hat{P}}+U_{\hat{P}}'\bar{Q}_{\hat{P}}U_{\hat{P}}=0.\label{f26}\end{aligned}$$\ Now we assume $$\begin{aligned} U_{\hat{P}}'\bar{A}U_{\hat{P}}&=&\left[ \begin{array}{cc} \bar{A}_{11}&\bar{A}_{12}\\ \bar{A}_{21}&\bar{A}_{22} \end{array} \right],\quad U_{\hat{P}}'\bar{C}U_{\hat{P}}=\left[ \begin{array}{cc} \bar{C}_{11}&\bar{C}_{12}\\ \bar{C}_{21}&\bar{C}_{22} \end{array} \right],\nonumber\\ U_{_{\hat{P}}}'\bar{Q}_{_{\hat{P}}}U_{_{\hat{P}}}&=&\left[ \begin{array}{cc} \bar{Q}_{\hat{P}_{11}}&\bar{Q}_{\hat{P}_{12}}\\ \bar{Q}_{\hat{P}_{12}}'&\bar{Q}_{\hat{P}_{22}} \end{array} \right],\label{f27}\end{aligned}$$ and by a simple calculation we can obtain $$\begin{aligned} &&\begin{bmatrix} \begin{smallmatrix} \bar{C}_{21}'\bar{Z}_{\hat{P}_{2}}\bar{C}_{21}+\bar{Q}_{\hat{P}_{11}} & \bar{A}_{21}'\bar{Z}_{\hat{P}_{2}}+\bar{C}_{21}'\bar{Z}_{\hat{P}_{2}}\bar{C}_{22} +\bar{Q}_{\hat{P}_{12}}\\ \bar{Z}_{\hat{P}_{2}}\bar{A}_{21}+\bar{C}_{22}'\bar{Z}_{\hat{P}_{2}}\bar{C}_{21} +\bar{Q}_{\hat{P}_{12}} & \bar{A}_{22}'\bar{Z}_{\hat{P}_{2}}+\bar{Z}_{\hat{P}_{2}}\bar{A}_{22} +\bar{C}'_{22}\bar{Z}_{\hat{P}_{2}}\bar{C}_{22}+\bar{Q}_{\hat{P}_{22}} \end{smallmatrix} \end{bmatrix}\nonumber\\ &&=0.\label{f28}\end{aligned}$$\ From Remark 3, we have $\bar{Q}_{\hat{P}}\geq0$, with $\bar{Z}_{\hat{P}_{2}}>0$, it is easy to obtain $\bar{Q}_{\hat{P}_{11}}=0$, $\bar{C}_{21}=0$. Next we will illustrate $\bar{Q}_{\hat{P}_{12}}=0$. Actually, for any $x=U_{\hat{P}}\left[ \begin{array}{cc} x_{1}\\ x_{2} \end{array} \right]$, where $x_{2}$ has the same dimension with $\bar{Q}_{\hat{P}_{12}}$, it has $$\begin{aligned} x'\bar{Q}_{\hat{P}}x&=&\left[ \begin{array}{cc} x_{1}\\ x_{2} \end{array} \right]'U_{\hat{P}}'\bar{Q}_{\hat{P}}U_{\hat{P}}\left[ \begin{array}{cc} x_{1}\\ x_{2} \end{array} \right]\nonumber\\ &=&x_{2}'\bar{Q}_{\hat{P}_{12}}'x_{1}+x_{1}'\bar{Q}_{\hat{P}_{12}}x_{2} +x_{2}'\bar{Q}_{\hat{P}_{22}}x_{2}.\label{f29}\end{aligned}$$ If $\bar{Q}_{\hat{P}_{12}}\neq0$, we can always find $x_{1}$, $x_{2}$, such that $x'\bar{Q}_{\hat{P}}x<0$ which is contrary with $\bar{Q}_{\hat{P}}\geq0$. Therefore, $\bar{Q}_{\hat{P}_{12}}=0$, $\bar{A}_{21}=0$ and $\bar{Q}_{\hat{P}_{22}}\geq0$.\ Plugging the above results into (\[f28\]), we have $$\begin{aligned} \bar{A}_{22}'\bar{Z}_{\hat{P}_{2}}+\bar{Z}_{\hat{P}_{2}}\bar{A}_{22} +\bar{C}_{22}'\bar{Z}_{\hat{P}_{2}}\bar{C}_{22}+\bar{Q}_{\hat{P}_{22}}=0.\label{f30}\end{aligned}$$\ Denote $U_{\hat{P}}'x(t)=\bar{x}(t)=\left[ \begin{array}{cc} \bar{x}^{(1)}(t)\\ \bar{x}^{(2)}(t) \end{array} \right]$, and the dimension of $\bar{x}^{(2)}(t)$ coincides with the rank of $\bar{Z}_{\hat{P}_{2}}$. Thus, (\[f22\]) can be rewritten as $$\begin{aligned} U_{\hat{P}}'dx(t)=U_{\hat{P}}'\bar{A}U_{\hat{P}}U_{\hat{P}}'x(t)dt +U_{\hat{P}}'\bar{C}U_{\hat{P}}U_{\hat{P}}'x(t)dw(t)\label{f31}\end{aligned}$$ i.e., $$\begin{aligned} d\bar{x}^{(1)}(t)&=&[\bar{A}_{11}\bar{x}^{(1)}(t)+\bar{A}_{12}\bar{x}^{(2)}(t)]dt\nonumber\\ &&+[\bar{C}_{11}\bar{x}^{(1)}(t)+\bar{C}_{12}\bar{x}^{(2)}(t)]dw(t),\label{f32}\\ d\bar{x}^{(2)}(t)&=&\bar{A}_{22}\bar{x}^{(2)}(t)dt+\bar{C}_{22}\bar{x}^{(2)}(t)dw(t).\label{f33}\end{aligned}$$\ Firstly, we show the stability of $(\bar{A}_{22}, \bar{C}_{22})$.\ Applying It$\hat{o}$’s formula to $E[x'(t)\bar{Z}_{\hat{P}}x(t)]$ and taking integral from 0 to $T$, it holds that $$\begin{aligned} &&E[x'(T)\bar{Z}_{\hat{P}}x(T)]-E[x'_{0}\bar{Z}_{\hat{P}}x_{0}]\nonumber\\ &=&-E\int_{0}^{T}[x'(t)Q_{\hat{P}}x(t)+x'(t)L_{\hat{P}}u(t)+u'(t)L'_{\hat{P}}x(t)\nonumber\\ &&+u'(t)R_{\hat{P}}u(t)]dt\nonumber\\ &=&-E\int_{0}^{T}x'(t)\bar{Q}_{\hat{P}}x(t)dt\leq0.\label{f34}\end{aligned}$$ Thus, it holds that $$\begin{aligned} &&\int_{0}^{T}E[\bar{x}^{(2)'}(t)\bar{Q}_{\hat{P}_{22}}\bar{x}^{(2)}(t)]dt =\int_{0}^{T}E[x'(t)\bar{Q}_{\hat{P}}x(t)]dt\nonumber\\ &=&E[x'_{0}\bar{Z}_{\hat{P}}x_{0}]-E[x'(T)\bar{Z}_{\hat{P}}x(T)]\nonumber\\ &=&E[\bar{x}^{(2)'}_{0}\bar{Z}_{\hat{P}_{2}}\bar{x}^{(2)}_{0}] -E[\bar{x}^{(2)'}(T)\bar{Z}_{\hat{P}_{2}}\bar{x}^{(2)}(T)].\label{f35}\end{aligned}$$ Following from Lemma 3 in [@16Qi:2017], we know that system $(\bar{A}_{22}, \bar{C}_{22}, \bar{Q}^{\frac{1}{2}}_{\hat{P}_{22}})$ is exact observable with $\bar{Z}_{\hat{P}_{2}}>0$.\ Since $$\begin{aligned} \int_{0}^{T}E[\bar{x}^{(2)'}(t)\bar{Q}_{\hat{P}_{22}}\bar{x}^{(2)}(t)]dt =E[\bar{x}^{(2)'}_{0}\bar{F}_{\hat{P}}(0,T)\bar{x}^{(2)}_{0}],\label{f36}\end{aligned}$$ where $\bar{F}_{\hat{P}}(t,T)$ satisfies the following differential equation: $$\begin{aligned} -\dot{\bar{F}}_{\hat{P}}(t,T)&=&\bar{Q}_{\hat{P}_{22}}+\bar{A}_{22}'\bar{F}_{\hat{P}}(t,T) +\bar{F}_{\hat{P}}(t,T)\bar{A}_{22}\nonumber\\ &&+\bar{C}_{22}'\bar{F}_{\hat{P}}(t,T)\bar{C}_{22},\label{f37}\end{aligned}$$ with final condition $\bar{F}_{\hat{P}}(T,T)=0$. From $\bar{Q}_{\hat{P}_{22}}\geq0$, we know $\bar{F}_{\hat{P}}(t,T)\geq0$ for $t\in[0,T]$.\ Now we will show $\bar{F}_{\hat{P}}(0,T)>0$. If not, there exists a nonzero $y$ such that $E[y'\bar{F}_{\hat{P}}(0,T)y]=0$. Then we choose the initial state be $y$, (\[f36\]) can be reduced to $$\begin{aligned} \int_{0}^{T}E[\bar{x}^{(2)'}(t)\bar{Q}_{\hat{P}_{22}}\bar{x}^{(2)}(t)]dt =E[y'\bar{F}_{\hat{P}}(0,T)y]=0,\label{f38}\end{aligned}$$ which implies that $\bar{Q}^{\frac{1}{2}}_{\hat{P}_{22}}\bar{x}^{(2)}(t)=0$, for any $t\in[0,T]$. With the exact observability of system $(\bar{A}_{22}, \bar{C}_{22}, \bar{Q}^{\frac{1}{2}}_{\hat{P}_{22}})$, we can obtain that $y=0$, which is contrary with $y\neq0$. Therefore, we have $\bar{F}_{\hat{P}}(0,T)>0$.\ Via a time shift of $t$, combining (\[f36\]), we have $$\begin{aligned} &&\int_{t}^{t+T}E[\bar{x}^{(2)'}(s)\bar{Q}_{\hat{P}_{22}}\bar{x}^{(2)}(s)]ds\nonumber\\ &=&E[\bar{x}^{(2)'}(t)\bar{F}_{\hat{P}}(0,T)\bar{x}^{(2)}(t)]\nonumber\\ &=&V_{2}(t,\bar{x}^{(2)}(t))-V_{2}(t+T,\bar{x}^{(2)}(t+T)),\label{f39}\end{aligned}$$ where $V_{2}(t,x(t))=E[x'(t)\bar{Z}_{\hat{P}_{2}}x(t)]$ and similar to the proof of the convergence of $V(t,x(t))$, we know $\lim\limits_{t\rightarrow+\infty}V_{2}(t,x(t))$ exists. Taking limitation on both sides of (\[f39\]), we have that $$\begin{aligned} \lim\limits_{t\rightarrow+\infty}E[\bar{x}'^{(2)}(t)\bar{x}^{(2)}(t)]=0 \label{f40} \end{aligned}$$ ,i.e., system $(\bar{A}_{22}, \bar{C}_{22}, \bar{Q}^{\frac{1}{2}}_{\hat{P}_{22}})$ is mean square stabilizable.\ Next we will illustrate the stability of $(\bar{A}_{11}, \bar{C}_{11})$. Let $\bar{x}^{(2)}(0)=0$, from (\[f33\]) we know $\bar{x}^{(2)}(t)=0$, $t\geq0$. Now (\[f32\]) can be rewritten as $$\begin{aligned} d\bar{x}^{(1)}(t)=\bar{A}_{11}\bar{x}^{(1)}(t)dt+\bar{C}_{11} \bar{x}^{(1)}(t)dw(t).\label{f41}\end{aligned}$$ Under the condition of $\bar{x}^{(2)}(t)=0$, it is easy to see that $$\begin{aligned} E[x'(t)\bar{Q}_{\hat{P}}x(t)]=E[\bar{x}'^{(2)}(t)\bar{Q}_{\hat{P}_{22}} \bar{x}^{(2)}(t)]=0.\label{f42}\end{aligned}$$ Hence, from the exact detectability of $(\bar{A}, \bar{C}, \bar{Q}^{\frac{1}{2}}_{\hat{P}})$ and $\bar{x}^{(2)}(t)=0$ we know that $$\begin{aligned} &&\lim\limits_{t\rightarrow+\infty}E[\bar{x}'^{(1)}(t)\bar{x}^{(1)}(t)]\nonumber\\ &=&\lim\limits_{t\rightarrow+\infty}{E[\bar{x}'^{(1)}(t)\bar{x}^{(1)}(t)] +E[\bar{x}'^{(2)}(t)\bar{x}^{(2)}(t)]}\nonumber\\ &=&\lim\limits_{t\rightarrow+\infty}E[\bar{x}'(t)\bar{x}(t)] =\lim\limits_{t\rightarrow+\infty}E[x'(t)x(t)]=0,\label{f43}\end{aligned}$$ which means $(\bar{A}_{11}, \bar{C}_{11})$ is mean square stable.\ Secondly, we will show system (\[f01\]) with controller $u(t)=Kx(t)$ is stabilizable in the mean square sense. In fact, we denote $\bar{\mathcal{A}}=\left[ \begin{array}{cc} \bar{A}_{11} &0 \\ 0&\bar{A}_{22} \end{array} \right], \bar{\mathcal{C}}=\left[ \begin{array}{cc} \bar{C}_{11} &0 \\ 0&\bar{C}_{22} \end{array} \right]$. Thus (\[f32\]) and (\[f33\]) can be rewritten as below $$\begin{aligned} d\bar{x}(t)&=&\big\{\bar{\mathcal{A}}\bar{x}(t)+\left[ \begin{array}{cc} \bar{A}_{12}\\ 0 \end{array} \right]\bar{u}(t)\big\}dt\nonumber\\ &&+\big\{\bar{\mathcal{C}}\bar{x}(t)+\left[ \begin{array}{cc} \bar{C}_{12}\\ 0 \end{array} \right]\bar{u}(t)\big\}dw(t),\label{f44}\end{aligned}$$ where $\bar{u}(t)$ is the solution to equation (\[f33\]) with initial condition $\bar{u}(0)=\bar{x}^{(2)}(0)$. The stability of $(\bar{A}_{11}, \bar{C}_{11})$ and $(\bar{A}_{22}, \bar{C}_{22})$ as shown above indicates that $(\bar{\mathcal{A}}, \bar{\mathcal{C}})$ is stable in the mean square sense. From (\[f40\]) it is easy to see that $\lim\limits_{t\rightarrow+\infty}E[\bar{u}'(t)\bar{u}(t)]=0$ and $\int_{0}^{\infty}E[\bar{u}'(t)\bar{u}(t)]dt<+\infty$. Applying the corresponding results in [@22Hinrichsen:98], we have that there exists a constant $c_{0}$ satisfying $$\begin{aligned} \int_{0}^{\infty}E[\bar{x}'(t)\bar{x}(t)]dt<c_{0}\int_{0}^{\infty} E[\bar{u}'(t)\bar{u}(t)]dt<+\infty.\label{f45}\end{aligned}$$ Hence, $\lim\limits_{t\rightarrow+\infty}E[\bar{x}'(t)\bar{x}(t)]=0$ can be verified from (\[f45\]) and $$\begin{aligned} \lim\limits_{t\rightarrow+\infty}E[x'(t)x(t)] =\lim\limits_{t\rightarrow+\infty}E[\bar{x}'(t)\bar{x}(t)]=0.\label{f46}\end{aligned}$$ In conclusion, system (\[f01\]) can be stabilizable with controller $u(t)=Kx(t)$ in the mean square sense.\ Finally, we will give the optimal controller and the associated cost functional. Applying It$\hat{o}$’s formula to $x'(t)\bar{P}x(t)$ with the GARE (\[f06\]), we have $$\begin{aligned} 2J_{T}&=&E\int_{0}^{T}\bigg\{x'(t)[A'\bar{P}+\bar{P}A+C'\bar{P}C]x(t)+u'(t)[B'\bar{P}\\ &&+D'\bar{P}C]x(t)+x'(t)[\bar{P}B+C'\bar{P}D]u(t)\\ &&+u'(t)D'\bar{P}Du(t)\bigg\}dt+E(x'_{0}\bar{P}x_{0})-E[x'(T)\bar{P}x(T)]\\ &=&E\int_{0}^{T}[u(t)-Kx(t)]'(R+D'\bar{P}D)[u(t)-Kx(t)]dt\\ &&+E(x'_{0}\bar{P}x_{0})-E[x'(T)\bar{P}x(T)].\end{aligned}$$ With the mean square stabilizable of system (\[f01\]), let $T\rightarrow+\infty$ in the above equation, the infinite cost functional can be reformulated as $$\begin{aligned} J&=&E(x'_{0}\bar{P}x_{0})+E\int_{0}^{\infty}[u(t)-Kx(t)]'(R+D'\bar{P}D)\nonumber\\ &&\cdot[u(t)-Kx(t)]dt.\label{f47}\end{aligned}$$ Thus the optimal controller is $u(t)=Kx(t)$ where $K$ as (\[f20\]). In addition, the optimal cost functional is $J^{\ast}=E(x'_{0}\bar{P}x_{0})$.\ “*Necessity*": On the other hand, we should illustrate that if $\mathcal{P}\neq\emptyset$, when the closed-loop system (\[f01\]) is mean square stabilizable, then the GARE (\[f06\]) has a solution $\bar{P}$, which is also the maximal solution to the GARE (\[f06\]). In fact, the existence of solutions to the GARE (\[f06\]) have been obtained in Theorem 1.\ The proof is complete.\ [**Remark 4**]{}    The key technique of sufficient proof is that in terms of the equivalence between the GARE and the SARE, the stabilization of the general indefinite case is reduced to the definite one whose stabilization is studied by Lyapunov functional defined with the optimal cost functional. An example ========== In this section, we give an example to illustrate the result obtained by Theorem 2.\ Consider two-dimensional system (\[f01\]) and cost functional (\[f02\]) with the following parameters $A=\left[ \begin{array}{cc} 0.01 &0\\ 0&-0.1 \end{array} \right] , B=\left[ \begin{array}{cc} 0.2\\ 0 \end{array} \right] , C=\left[ \begin{array}{cc} -0.1&0\\ 0&0.1 \end{array} \right], D=\left[ \begin{array}{cc} 0.6\\ 0 \end{array} \right], Q=\left[ \begin{array}{cc} 0.5&0\\ 0&-1 \end{array} \right], R=-0.05$ with initial state $x_{0}=\left[ \begin{array}{cc} -0.01\\ 0.1 \end{array} \right] $, noting that $Q, R$ are indefinite matrices. By solving the LMI in set $\mathcal{P}$ with Matlab tool, the solution to GARE (\[f06\]) with the given parameters can be obtained as $\bar{P}=\left[ \begin{array}{cc} 20.143&0\\ 0&-5.2632 \end{array} \right] $. Therefore, the gain matrix $K$ in (\[f20\]) can be easily obtained as $K=\left[ \begin{array}{cc} -0.2891&0 \end{array} \right] $. Furthermore, Assumption 1 is also satisfied due to $Q_{\bar{P}}=\left[ \begin{array}{cc} 0.9230 &0\\ 0&0.1053 \end{array} \right] $. Accordingly, by Theorem 2 we know that the system (\[f01\]) with the above parameters is mean square stabilizable. The simulation result is shown in Fig. 1. It can be seen that the state $x(t)$ (Fig. 3 ) is stabilized with the optimal controller (Fig. 2), as expected. ![Simulations for the state trajectory $E[x'(t)x(t)]$.[]{data-label="fig:digit"}](fig008.eps){width="42.00000%"} ![Optimal control.[]{data-label="fig:digit"}](fig006.eps){width="42.00000%"} ![Optimal states.[]{data-label="fig:digit"}](fig004.eps){width="42.00000%"} Conclusions =========== In this paper, we mainly discussed the stabilization problem for It$\hat{o}$ stochastic system, whose control and state weighting matrices in the cost functional are indefinite. The convergence of the GDRE which involves a matric pseudo-inverse and two additional equality/inequality constraints was studied. And the infinite horizon optimal controller was obtained accordingly. Finally, in terms of the equivalence between the GARE and the SARE, the stabilization of the general indefinite case was reduced to the definite one whose stabilization is studied by Lyapunov functional defined with the optimal cost functional subject to the SARE. The contents in this paper are an extension and improvement of the previous works [@19Rami:2001]. [99]{} R. E. Kalman (1960). Contribution to the Theory of Optimal Control. [*Boletin Sociedad Matematica Mexicana,*]{} 5(2), 102–119. W. M. Wonham (1968). On a matrix Riccati equation of stochastic control. [*SIAM Journal on Optimization and Control,*]{} 6(4), 681–697. M. H. A. Davis (1977). Linear Estimation and Stochastic Control. [*Chapman and Hall, London*]{}. H. Zhang and J. Xu, (2017). Control for It$\hat{o}$ stochastic systems with input delay. [*IEEE Transactions on Automatic Control,* ]{} 62(1), 350–365. S. Chen, X. Li, and X. Zhou (1998). Stochastic linear quadratic regulators with indefinite control weight costs. [*SIAM Journal on Optimization and Control,* ]{} 36(5), 1685–1702. M. A. Rami, J. B. Moore, and X. Zhou (2001). Indefinite stochastic linear quadratic control and generalized differential Riccati equation. [*SIAM Journal on Optimization and Control,*]{} 40, 1296–1311. H. Wu, and X. Zhou (2002). Characterizing all optimal controls for an indefinite stochastic linear quadratic control problem. [*IEEE Transactions on Automatic Control*]{}, 47(7), 1119–1122. A. Ferrante, and L. Ntogramatzidis (2015). A note on finite-horizon LQ problems with indefinite cost. [*Automatica*]{}, 52, 290–293.. G.  Bilardi and A. Ferrante (2007). The role of terminal cost/reward in finite-horizon discrete-time LQ optimal control. [*Linear Algebra and its Applications*]{}, 425, 323–344. Y. Ni, C. K. F. Yiu, H. Zhang, and J. Zhang (2017). Delayed Optimal Control of Stochastic LQ Problem. [*SIAM Journal on Optimization and Control*]{}, 55(5), 3370–3407. M. Zorzi (2017). On the Robustness of the Bayes and Wiener Estimators under Model Uncertainty. [*Automatica* ]{}, 83(9), 133–140. M. Zorzi (2017). Convergence analysis of a family of robust Kalman filters based on the contraction principle. [*SIAM Journal on Optimization and Control*]{}, 55(5), 3116-3131. H. Zhang, Q. Qi, and M. Fu (2018). Optimal Stabilization Control for Discrete-time Mean-field Stochastic Systems. [*IEEE Transactions on Automatic Control*]{}, DOI 10.1109/TAC.2018.2813006. Q. Qi, and H. Zhang (2017). Stabilization Control for Linear Continuous-time Mean-field Systems. arXiv preprint arXiv: 1608.06475. M. A. Rami, and X. Zhou (2000). Linear matrix inequalities, Riccati equations, and indefinite stochastic linear quadratic control. [*IEEE Transactions on Automatic Control*]{}, 45(6), 1131–1142. X. Li, X. Zhou, and M. A. Rami(2003). Indefinite stochastic linear quadratic control with Markovian jumps in infinite time horizon. [*Journal of Global Optimization*]{}, 27, 149–175. M. A. Rami, X. Chen, J. B. Moore, and X. Zhou (2001). Solvability and asymptotic behavior of generalized Riccati equations arising in indefinite stochastic LQ controls. [*IEEE Transactions on Automatic Control*]{}, 46(3), 428–440. A. Albert (1969). Conditions for positive and nonnegative definiteness in terms of pseudo-inverse. [*SIAM Journal on Optimization and Control*]{}, 17, 434–440. W. Zhang, and B. S. Chen (2004). On stabilizability and exact observability of stochastic systems with their applications. [*Automatica*]{}, 40, 87–94. D. Hinrichsen, and A. J. Pritchard (1998). Stochastic $H_{\infty}$. [*SIAM Journal on Optimization and Control*]{}, 36(5), 1504–1538. R. A. Horn, and C. R. Johnson (1990). Matrix Analysis. Cambridge University Press, Cambridge.

ArXiv

--- abstract: 'The Debye-Hückel approximation to the free-energy of a simple fluid is written as a functional of the pair correlation function. This functional can be seen as the Debye-Hückel equivalent to the functional derived in the hyper-netted chain framework by Morita and Hiroike, as well as by Lado. It allows one to obtain the Debye-Hückel integral equation through a minimization with respect to the pair correlation function, leads to the correct form of the internal energy, and fulfills the virial theorem.' author: - 'R. Piron$^1$[^1] and T. Blenski$^2$' bibliography: - 'draft.bib' title: 'Free-energy functional of the Debye-Hückel model of simple fluids' --- Introduction ============ The Debye-Hückel (DH) model originated in the physics of electrolytes [@DebyeHuckel23], that is, fluids of charged particles interacting through a Coulomb potential. It is the linearized version of the classical “Hartree-like” mean-field model, which is often called Non-Linear Debye-Hückel (NLDH) model or Poisson-Boltzmann model in the context of electrolytes. These two models can be extended to arbitrary interaction potentials, leading respectively to the DH and NLDH models of classical fluids, with the corresponding integral equations. Although it is based on very strong assumptions, the DH model gives physical insight into the screening of interaction potentials and the decay of correlation functions, both in the framework of simple and multicomponent fluids. The DH model is valid in the low-coupling limit, that is, in the limit of small interactions compared to the kinetic energy of the particles. More sophisticated models of classical fluids exist, as for instance the Hyper-Netted Chain (HNC) [@Morita58] or Percus-Yevick [@PercusYevick58] models, which account for part of the correlations. However, some theoretical studies also proceed by introducing corrections, using the low-coupling DH limit as a starting point (see [@Abe59; @DeWitt65]). It was also recently shown that in the DH model, the energy and virial routes are thermodynamically consistent for any potential [@Santos09]. For these reasons, the DH model is of permanent theoretical interest. Moreover, in a number of applications, modified versions of the DH model are used, as for example in the physics of electrolytes [@Nordholm84; @Penfold90] or in plasma physics [@KidderDeWitt61; @Vieillefosse81]. The DH model therefore also has some practical interest. In the case of a long-range attractive potential, the linearization performed in the DH model allows one to circumvent the “classical catastrophe” of collapsing particles. To some extent, this explains why this model is so common in plasma physics as well as in the physics of electrolytes. Apart from the historical derivation as a linearized mean-field theory of a charged-particle mixture [@DebyeHuckel23], the DH model of simple fluids can be obtained from the so-called “Percus trick” [@Percus64] or from a diagrammatic analysis [@Mayer50; @Abe59; @DeWitt65]. In the research on variational models of atoms in plasmas, there is a serious need to account for ion and electron correlations and their impact on the atomic structure and dynamics [@Liberman79; @Blenski07b; @Blenski07a; @Piron11; @Piron11b; @Blenski13; @Piron13; @Caizergues14; @Caizergues16; @Starrett12; @Starrett12b; @Chihara16]. In this framework, it is also useful to have a variational derivation of simple fluid models, using expressions of the free-energy as a functional of the ion-ion pair correlation function. Such a variational expression can be used in a more general theory that also includes the ion electronic structure. A free-energy functional is available in the HNC theory [@MoritaHiroike60; @Lado73]. However, in the well-known DH theory, an expression of the free-energy as a functional of the pair correlation function has not been yet proposed. The purpose of this paper is to present a brief derivation of such an expression, which can be seen as the DH equivalent to the HNC free-energy functional of Morita and Hiroike [@MoritaHiroike60], also derived by Lado [@Lado73]. Debye-Hückel integral equation ============================== The DH integral equation for a homogeneous simple fluid with the given pair potential $u(r)$ writes (see for example [@Percus64], Eq. (6.5)): $$\label{eq_DH_int_r} h(r)=-\beta u(r) - {\varrho}\beta\int d^3r'\left\{h(r')u(|{\mathbf}{r}-{\mathbf}{r}'|)\right\}$$ where ${\varrho}$ is the particle density of the fluid, $\beta$ the inverse temperature, and $h(r)=g(r)-1$, $g(r)$ being the usual pair distribution function. For some particular pair potentials such as the Coulomb ($1/r$) or Yukawa ($e^{-\alpha r}/r$) potentials, this equation can be recast as a differential equation (Poisson equation in the Coulomb case, Helmholtz equation in the Yukawa case). In a diagrammatic analysis, Eq.  corresponds to the sum of all $u$-bond chain diagrams. Eq.  can also be seen as a Ornstein-Zernicke relation with the simplest closure $c(r)=-\beta u(r)$, where $c(r)$ is the direct correlation function. The Fourier transform $\mathcal{F}_{{\mathbf}{k}}$ of a function $\mathcal{F}({\mathbf}{r})$ is defined as: $$\mathcal{F}_{{\mathbf}{k}}=\int d^3r\left\{\mathcal{F}({\mathbf}{r})e^{i{\mathbf}{k}.{\mathbf}{r}}\right\}$$ For a potential that has a Fourier transform, Eq.  can be rewritten in the Fourier space as the algebraic relation: $$\label{eq_DH_int_k} h_k=-\beta u_k - {\varrho}\beta h_k u_k$$ which has the direct formal solution: $$\label{eq_us} h_{\text{eq},k}=-\beta u_{s,k}\text{ ; }u_{s,k}\equiv\frac{u_k}{1+{\varrho}\beta u_{k}}$$ Here, $h_{\text{eq},k}$ denotes the DH equilibrium correlation function and $u_{s,k}$ denotes the screened potential, both in the Fourier space. For any long-range potential (i.e. Coulombic at infinity), $\lim_{k\rightarrow 0} u_k\sim k^{-2}$, which implies $u_{k=0}=-{\varrho}^{-1}$. This is the case in any model that can be written as an Ornstein-Zernicke equation with a direct correlation function that includes the interaction potential. In the following, the pair potential $u(r)$ will be replaced by $u^\xi(r)\equiv\xi u(r)$, $\xi$ being a charge parameter, and the corresponding $\xi$ indices will label $h_\text{eq}^\xi$, $u^\xi$ and $u_s^\xi$. Expression of the free-energy functional ======================================== We consider a system which is immersed in a rigid, homogeneous, “neutralizing” background. Such background do not impact on the equations-of-motion or on the DH integral equation. It however leads to renormalized energies by accounting for the contribution: $$\frac{E_{bg}}{V}=-\frac{{\varrho}^2}{2}\int d^3r\left\{u^\xi(r)\right\}$$ which cancels the divergences of energies in case of a long-range potential. In case of a short-range potential, one can disregard any background by removing this contribution. In the following, we will address the configurational part of the renormalized free-energy and give its expression as a functional of the pair correlation function. In the particular case of the Coulomb potential: $u^\xi(r)\equiv\xi/r$, it is easy to check that if one minimizes w.r.t. $h(r)$ the following functional $A^\xi\left\{{\varrho},\beta,h(r)\right\}$: $$\begin{aligned} \frac{A^\xi}{V}=\frac{2{\varrho}^2}{3\beta} \int d^3r &\left\{h(r)\left(\frac{h(r)}{2}+\frac{\beta\xi}{r}\right.\right. \nonumber\\ &\left.\left.+\frac{{\varrho}\beta\xi}{2}\int d^3r'\left\{\frac{h(r')}{|{\mathbf}{r}-{\mathbf}{r}'|}\right\}\right)\right\}\end{aligned}$$ one gets Eq. . This functional also yields all the well-known equilibrium results for the renormalized configurational internal energy $U_\text{eq}^\xi$, free-energy $A_\text{eq}^\xi$, and pressure $P_\text{eq}^\xi$ (see, for example, [@LandauStatisticalPhysics] 78), namely: $$\begin{aligned} &\frac{U_\text{eq}^\xi}{V}\label{eq_U_DHOCP} =\frac{\partial}{\partial \beta}\left(\frac{\beta A_\text{eq}^\xi}{V}\right) =\left.\frac{\partial}{\partial \beta}\left(\frac{\beta A^\xi}{V}\right)\right|_\text{eq}=-\frac{\xi{\varrho}}{2}\sqrt{4\pi{\varrho}\beta\xi}\\ &\frac{A_\text{eq}^\xi}{V}=\frac{2}{3}\frac{U_\text{eq}^\xi}{V}\label{eq_A_DHOCP}\\ &P_\text{eq}^\xi ={\varrho}^2\frac{\partial}{\partial {\varrho}}\left(\frac{A_\text{eq}^\xi}{{\varrho}V}\right)\label{eq_P_DHOCP} ={\varrho}^2\frac{\partial}{\partial {\varrho}}\left.\left(\frac{A^\xi}{{\varrho}V}\right)\right|_\text{eq} =\frac{1}{3}\frac{U_\text{eq}^\xi}{V}\end{aligned}$$ where $|_\text{eq}$ means that the functional is taken at the DH equilibrium, that is, for $h(r)=h_\text{eq}^\xi(r)$ from Eq. . We stress that Eqs ,, hold for a One-component Classical Plasma (OCP) with the contribution of the neutralizing background included. A slightly more complicated functional can be guessed as well in the Yukawa-OCP case: $u^{\xi}(r)=\xi\exp(-\alpha r)/r$. The present derivation aims at giving a general form of the free-energy functional for a simple fluid with any pair potential $u(r)$. As in [@Lado73], we use Debye’s charging method and build an expression of the renormalized configurational free-energy functional $A^\xi\left\{{\varrho},\beta,u(r),h(r)\right\}$ from the equilibrium relation: $$\label{eq_charging_A} \frac{A_\text{eq}^\xi\left\{{\varrho},\beta,u(r)\right\}}{V} =\frac{{\varrho}^2}{2}\int_0^\xi\frac{d\xi'}{\xi'}\int d^3r\left\{h_\text{eq}^{\xi'}(r)u^{\xi'}(r)\right\}$$ In the case of the exact equilibrium, Eq.  is an exact relation. We consider the case of $|_\text{eq}$ denoting the DH equilibrium, with $h(r)=h_\text{eq}^\xi(r)$ from Eq. , and require Eq.  to hold. Let us assume that the free-energy functional $A^\xi\left\{{\varrho},\beta,u(r),h(r)\right\}$ can be written in the form that follows: $$\frac{A^\xi}{V}=\int\frac{d^3k}{(2\pi)^3}\left\{f_k^\xi h_k\left(\frac{h_k}{2} + \beta u_k^\xi + \frac{{\varrho}\beta}{2}h_k u_k^\xi\right)\right\}$$ where $f_k^\xi\equiv f({\varrho},\beta,\xi,k,u_k^\xi)$ is independent of $h_k$. Minimization of $A^\xi$ w.r.t. $h(r)$ then lead to Eq. , i.e. to the DH equation. Differentiating w.r.t. $\xi$, we obtain: $$\begin{aligned} \frac{\xi}{V}\frac{\partial A_\text{eq}^\xi}{\partial \xi} =& \left.\frac{\xi}{V}\frac{\partial A^\xi}{\partial \xi}\right|_\text{eq}\\ =&\int\frac{d^3k}{(2\pi)^3}\left\{\xi\frac{\partial f_k^\xi}{\partial \xi} h_k\left(\frac{h_k}{2} + \beta u_k^\xi + \frac{{\varrho}\beta}{2}h_k u_k^\xi\right)\right.\nonumber\\ &+\left.\left.f_k^\xi h_k\left(\beta \xi\frac{\partial u_k^\xi}{\partial \xi} + \frac{{\varrho}\beta}{2}h_k \xi\frac{\partial u_k^\xi}{\partial \xi}\right)\right\}\right|_\text{eq}\end{aligned}$$ Using $u_k^\xi=\xi u_k$, and substituting the formal equilibrium solution of Eq. , we get: $$\begin{aligned} \frac{\xi}{V}\frac{\partial A_\text{eq}^\xi}{\partial \xi} =&\frac{-\beta^2}{2}\int\frac{d^3k}{(2\pi)^3}\left\{ \frac{\partial}{\partial \xi} \left(\xi f_k^\xi\right) u_{s,k}^\xi u_k^\xi + f_k^\xi u_{s,k}^{\xi\,2} \right\}\end{aligned}$$ On the other hand, from Eqs.  and we can write: $$\label{eq_charging_A_bis} \frac{\xi}{V}\frac{\partial A_\text{eq}^\xi}{\partial \xi}=\frac{-\beta{\varrho}^2}{2}\int \frac{d^3k}{(2\pi)^3}\left\{u_{s,k}^{\xi} u^{\xi}_k\right\}$$ Requiring Eq.  (or equivalently Eq. ) to hold, results in the equation: $$\begin{aligned} \frac{\partial}{\partial \xi}\left(\xi f_k^\xi\right) + \frac{\xi f_k^\xi}{\xi+{\varrho}\beta u_k \xi^2}=\frac{{\varrho}^2}{\beta}\end{aligned}$$ A solution to this differential equation for $\xi f_k^\xi$ can be looked after in the form of a series. This leads to the following expression for $f_k^\xi$: $$\label{eq_sol_fkxi} f_k^\xi=\frac{{\varrho}^2}{\beta}F({\varrho}\beta u_k^\xi)$$ where the function $F$ is defined as the series: $$F(x)=1-\sum_{n=0}^\infty\left\{\frac{(-x)^n}{(n+1)(n+2)}\right\}$$ The latter can be identified with the function: $$F(x)=\left(1+\frac{1}{x}\right)\left(1-\frac{\ln(1+x)}{x}\right)$$ As can be seen in Eq. , the only dependence of $f_k^\xi$ on $k$ and $\xi$ is via the potential $u_k^\xi$. Finally, the DH free-energy functional reads: $$\begin{aligned} \label{eq_explicit_functional} \frac{A^\xi}{V}=\frac{{\varrho}^2}{\beta}\int\frac{d^3k}{(2\pi)^3}\left\{ \left(1+\frac{1}{{\varrho}\beta u_k^\xi}\right)\left(1-\frac{\ln(1+{\varrho}\beta u_k^\xi)}{{\varrho}\beta u_k^\xi}\right) \right.\nonumber\\\left. h_k\left(\frac{h_k}{2}+\beta u_k^\xi+\frac{{\varrho}\beta}{2}h_ku_k^\xi\right)\right\}\end{aligned}$$ which is the main result of the present paper. It is worth to note the appearance of the logarithm function, both in the DH and in the HNC free-energy functionals [@MoritaHiroike60; @Lado73], which is related to the fact that the Ornstein-Zernicke relation is explicitly fulfilled in both approaches. Internal energy, pressure and virial theorem ============================================ We first consider the renormalized configurational internal energy, as defined by the thermodynamics: $$\frac{U^\xi_\text{eq}}{V}=\frac{\partial}{\partial \beta}\left(\frac{\beta A^\xi_\text{eq}}{V}\right)=\left.\frac{\partial}{\partial \beta}\left(\frac{\beta A^\xi}{V}\right)\right|_\text{eq}$$ As can be seen in Eq. , $\beta A^\xi/V$ is a functional of $h(r)$, ${\varrho}$, and of the product $\beta \xi u(r)$. As a consequence, we can write: $$\frac{\partial}{\partial \beta}\left(\frac{\beta A^\xi}{V}\right)=\xi\frac{\partial}{\partial \xi}\left(\frac{A^\xi}{V}\right)$$ We then have, in virtue of Eq. : $$\label{eq_U_eq} \frac{U^\xi_\text{eq}}{V}=\left.\xi\frac{\partial}{\partial \xi}\left(\frac{A^\xi}{V}\right)\right|_\text{eq} =\int d^3r\left\{h_\text{eq}^{\xi}(r)u^{\xi}(r)\right\}$$ which corresponds to the exact expression of the internal energy, with $h_\text{eq}^{\xi}(r)$ taken to be the DH approximation to the equilibrium correlation function. The result Eq.  can also be obtained in a more pedestrian way, by making the explicit differentiation of the free-energy expression Eq.  and then substituting $h_{\text{eq},k}^\xi$. We do not reproduce this calculation here. The virial configurational pressure as a functional of ${\varrho}$,$u(r)$,$h(r)$ is: $$\label{eq_def_virialP} P_v^\xi=-\frac{1}{3}\frac{{\varrho}^2}{2}\int d^3r\left\{h(r) {\mathbf}{r}.{\mathbf}{\nabla}_{\mathbf}{r}u^\xi(r)\right\}$$ For a potential that decays as $1/r$ or faster, using the Fourier representation and integrating by part, we can write from Eq. : $$P_v^\xi=\int d^3r\left\{h(r)u^{\xi}(r)\right\}+\frac{1}{3}\frac{{\varrho}^2}{2}\int \frac{d^3k}{(2\pi)^3}\left\{h_k {\mathbf}{k}.{\mathbf}{\nabla}_{\mathbf}{k}u^\xi_k\right\}$$ Using Eq. , the equilibrium virial pressure becomes: $$P_{v,\text{eq}}^\xi=\frac{U^\xi_\text{eq}}{V} -\frac{{\varrho}^2}{2}\int_0^\infty \frac{dk}{(2\pi)^3}\left\{\frac{4\pi}{3}k^3\frac{\beta u_{k}^\xi}{1+{\varrho}\beta u_{k}^\xi} \frac{\partial u_k^\xi}{\partial k}\right\}$$ Finally, we obtain: $$P_{v,\text{eq}}^\xi= \frac{U^\xi_\text{eq}}{V}+\frac{1}{2\beta}\int \frac{d^3k}{(2\pi)^3}\left\{{\varrho}\beta u^\xi_k-\ln\left(1+{\varrho}\beta u^\xi_k\right)\right\}$$ Then, differentiating w.r.t. $\xi$, we get: $$\label{eq_diff_Pvir} \xi\frac{\partial P_{v,\text{eq}}^\xi}{\partial \xi} =\frac{-{\varrho}^2\beta}{2}\int \frac{d^3k}{(2\pi)^3}\left\{\left(u^{\xi}_{s,k}\right)^2\right\}$$ We now consider the equilibrium pressure, as it is defined by the thermodynamics: $$P_\text{eq}^\xi={\varrho}^2\frac{\partial}{\partial {\varrho}}\left(\frac{A_\text{eq}^\xi}{{\varrho}V}\right)$$ Using Eq. , we obtain: $$P_\text{eq}^\xi ={\varrho}^2\frac{\partial}{\partial {\varrho}} \left( \frac{-\beta{\varrho}}{2} \int_0^\xi \frac{d\xi'}{\xi'} \int \frac{d^3k}{(2\pi)^3}\left\{u_{s,k}^{\xi'} u^{\xi'}_k\right\} \right)$$ Performing the differentiation w.r.t. ${\varrho}$, we end up with: $$P_\text{eq}^\xi =\frac{-{\varrho}^2\beta}{2}\int_0^\xi \frac{d\xi'}{\xi'}\int \frac{d^3k}{(2\pi)^3}\left\{\left(u^{\xi}_{s,k}\right)^2\right\}$$ Finally, using Eq. , we have: $$\label{eq_virial} P_\text{eq}^\xi =\int_0^\xi \frac{d\xi'}{\xi'}\left\{\xi'\frac{\partial P_{v,\text{eq}}^{\xi'}}{\partial \xi'}\right\} =P_{v,\text{eq}}^{\xi}$$ Thus, using the exact Eq.  in order to define the free-energy in the approximate DH case ensures that the virial theorem is fulfilled. Again, the result Eq.  can also be obtained in a more pedestrian way, from the explicit differentiation of the free-energy expression of Eq.  and the substitution of $h_{\text{eq},k}^\xi$. We do not reproduce this calculation here either. Conclusion ========== Our Eq.  is an explicit expression of the renormalized configurational free-energy functional in the Debye-Hückel approximation. This expression allows one to obtain the Debye-Hückel equation from a minimization procedure w.r.t. the pair correlation function. In the DH case, as in the HNC case of [@MoritaHiroike60; @Lado73], requiring the exact charging relation to hold in the approximate model allows one to define a free-energy functional that yields the correct expression for the internal energy and fulfills the virial theorem. Work is now in progress in order to generalize the present formalism to multicomponent fluids, as it was done in the HNC case [@Lado73b; @Enciso87]. [^1]: Corresponding author\ Electronic address: [email protected]

ArXiv

--- abstract: 'This paper is concerned with the problem of low-rank plus sparse matrix decomposition for big data. Conventional algorithms for matrix decomposition use the entire data to extract the low-rank and sparse components, and are based on optimization problems that scale with the dimension of the data, which limit their scalability. Furthermore, the existing randomized approaches mostly rely on uniform random sampling, which can be quite inefficient for many real world data matrices that exhibit additional structures (e.g. clustering). In this paper, a scalable subspace-pursuit approach that transforms the decomposition problem to a subspace learning problem is proposed. The decomposition is carried out using a small data sketch formed from sampled columns/rows. Even when the data is sampled uniformly at random, it is shown that the sufficient number of sampled columns/rows is roughly $\calO(r \mu)$, where $\mu$ is the coherency parameter and $r$ the rank of the low-rank component. In addition, efficient sampling algorithms are proposed to address the problem of column/row sampling from structured data. The proposed sampling algorithms can be independently used for feature selection from high-dimensional data. The proposed approach is amenable to online implementation and an online scheme is proposed.' author: - 'Mostafa Rahmani, and George K. Atia,  [^1]' bibliography: - 'IEEEabrv.bib' - 'bibfile.bib' title: --- [Shell : Bare Demo of IEEEtran.cls for Journals]{} Low-Rank Matrix, Subspace Learning, Big Data, Matrix Decomposition, Column Sampling, Sketching Introduction ============ we are given a data matrix $\bD \in \mathbb{R}^{N_1 \times N_2} $, which can be expressed as $$\begin{aligned} \bD= \bL+\bS, \label{eq1}\end{aligned}$$ where $\bL$ is a low rank (LR) and $\bS$ is a sparse matrix with arbitrary unknown support, whose entries can have arbitrarily large magnitude. Many important applications in which the data under study can be naturally modeled using (\[eq1\]) were discussed in [@lamport2]. The cutting-edge Principal Component Pursuit approach developed in [@lamport2; @lamport1], directly decomposes $\bD$ into its LR and sparse components by solving the convex program $$\begin{aligned} \begin{aligned} & \underset{\dot{\bL},\dot{\bS}}{\min} & & \lambda\|\dot{\bS}\|_1 + \|\dot{\bL} \|_* \\ & \text{subject to} & & \dot{\bL} + \dot{\bS} = \bD \: \\ \end{aligned} \label{eq2}\end{aligned}$$ where $\|.\|_1$ is the $\ell_1$-norm, $\|.\|_*$ is the nuclear norm and $\lambda$ determines the trade-off between the sparse and LR components [@lamport1]. The convex program (\[eq2\]) can precisely recover both the LR and sparse components if the columns and rows subspace of $\bL$ are sufficiently incoherent with the standard basis and the non-zero elements of $\bS$ are sufficiently diffused [@lamport1]. Although the problem in (\[eq2\]) is convex, its computational complexity is intolerable with large volumes of high-dimensional data. Even the efficient iterative algorithms proposed in [@lamport53; @lamport23] have prohibitive computational and memory requirements in high-dimensional settings. **Contributions:** This paper proposes a new randomized decomposition approach, which extracts the LR component in two consecutive steps. First, the column-space (CS) of $\bL$ is learned from a small subset of the columns of the data matrix. Second, the row-space (RS) of $\bL$ is obtained using a small subset of the rows of $\bD$. Unlike conventional decomposition that uses the entire data, we only utilize a small data sketch, and solve two low-dimensional optimization problems in lieu of one high-dimensional matrix decomposition problem (\[eq2\]) resulting in significant running time speed-ups. To the best of our knowledge, it is shown here for the first time that the sufficient number of randomly sampled columns/rows scales linearly with the rank $r$ and the coherency parameter of $\bL$ even with uniform random sampling. Also, in contrast to the existing randomized approaches [@new1; @new2; @lamport8], which use blind uniform random sampling, we propose a new methodology for efficient column/row sampling. When the columns/rows of $\bL$ are not distributed uniformly in the CS/RS of $\bL$, which prevails much of the real world data, the proposed sampling approach is shown to achieve significant savings in data usage compared to uniform random sampling-based methods that require remarkable portions of the data. The proposed sampling algorithms can be independently used for feature selection from high-dimensional data. In the presented approach, once the CS is learned, each column is decomposed efficiently and independently using the proposed randomized vector decomposition method. Unlike most existing approaches, which are batch-based, this unique feature enables applicability to online settings. The presented vector decomposition method can be independently used in many applications as an efficient vector decomposition algorithm or for efficient linear decoding [@lamport17; @minaee2015screen; @minaee2015screenj]. Notation and definitions ------------------------ We use bold-face upper-case letters to denote matrices and bold-face lower-case letters to denote vectors. Given a matrix $\bL$, $\|\bL\|$ denotes its spectral norm, $\|\bL\|_F $ its Frobenius norm, and $\| \bL \|_\infty$ the infinity norm, which is equal to the maximum absolute value of its elements. In an $N$-dimensional space, $\be_i$ is the $i^{\text{th}}$ vector of the standard basis (i.e., the $i^{\text{th}}$ element of $\be_i$ is equal to one and all the other elements are equal to zero). The notation $A = [\:]$ denotes an empty matrix and the matrix $$\bA = [ \bA_1 \: \bA_2 \: ... \: \bA_n]$$ is the column-wise concatenation of the matrices $\{ \bA_i\}_{i=1}^n$. Random sampling refers to sampling without replacement. Background and Related Work =========================== Exact LR plus sparse matrix decomposition ----------------------------------------- The incoherence of the CS and RS of $\bL$ is an important requirement for the identifiability of the decompostion problem in (\[eq1\]) [@lamport2; @lamport1]. For the LR matrix $\bL$ with rank $r$ and compact SVD $\bL = \bU \mathbf{\Sigma} \bV^T$ (where $\bU \in \mathbb{R}^{N_1 \times r}$, $\mathbf{\Sigma} \in \mathbb{R}^{r \times r}$ and $\bV \in \mathbb{R}^{N_2 \times r}$), the incoherence condition is typically defined through the requirements [@lamport1; @lamport2] $$\begin{aligned} \begin{aligned} & \underset{i}{\max} \|\bU^T\be_i\|_2^2 \leq \frac{\mu r}{N_1} \: , \: \underset{i}{\max} \|\bV^T\be_i\|_2^2 \leq \frac{\mu r}{N_2} \: \\ & \text{and} \: \: \|\bU\bV^T\|_{\infty} \leq \sqrt{\frac{\mu r}{N_2 N_1}} \label{eq9} \end{aligned}\end{aligned}$$ for some parameter $\mu$ that bounds the projection of the standard basis $\{\be_i\}$ onto the CS and RS. Other useful measures for the coherency of subspaces are given in [@lamport3] as, $$\begin{aligned} \gamma(\bU) \hspace{-0.5mm}= \hspace{-.5mm}\sqrt{N_1} \max_{i,j} |\bU(i,j)|, \gamma(\bV) \hspace{-0.5mm}= \hspace{-0.5mm}\sqrt{N_2} \max_{i,j} |\bV(i,j)|, \label{eq10}\end{aligned}$$ where $\gamma(\bU)$ and $\gamma(\bV)$ bound the coherency of the CS and the RS, respectively. When some of the elements of the orthonormal basis of a subspace are too large, the subspace is coherent with the standard vectors. Actually, it is not hard to show that $\max \left( \gamma(\bV),\gamma(\bU)\right) \leq \sqrt{\mu}$. The decomposition of a data matrix into its LR and sparse components was analyzed in [@lamport1; @lamport2], and sufficient conditions for exact recovery using the convex minimization (\[eq2\]) were derived. In [@lamport2], the sparsity pattern of the sparse matrix is selected uniformly at random following the so-called Bernoulli model to ensure that the sparse matrix is not LR with overwhelming probability. In this model, which is also used in this paper, each element of the sparse matrix can be non-zero independently with a constant probability. Without loss of generality (w.l.o.g.), suppose that $N_2 \leq N_1$. The following lemma states the main result of [@lamport2]. Suppose that the support set of $\bS$ follows the Bernoulli model with parameter $\rho$. The convex program (\[eq2\]) with $\lambda=\frac{1}{\sqrt{N_1}}$ yields the exact decomposition with probability at least $1-c_1 {N_1}^{-10}$ provided that $$\begin{aligned} r \leq \rho_r N_2 {\mu}^{-1} \left(\log(N_1)\right)^{-2} \quad , \quad \rho \leq {\rho}_s \label{eq12}\end{aligned}$$ where $\rho_s$, $c_1$ and $\rho_r$ are numerical constants. \[lm1\] The optimization problem in (\[eq2\]) is convex and can be solved using standard techniques such as interior point methods [@lamport1]. Although these methods have fast convergence rates, their usage is limited to small-size problems due to the high complexity of computing a step direction. Similar to the iterative shrinking algorithms for $\ell_1$-norm and nuclear norm minimization, a family of iterative algorithms for solving the optimization problem (\[eq2\]) were proposed in [@lamport23; @lamport53]. However, they also require working with the entire data. For example, the algorithm in [@lamport23] requires computing the SVD of an $N_1 \times N_2 $ matrix in every iteration. Randomized approaches --------------------- Owing to their inherent low-dimensional structures, the robust principal component analysis (PCA) and matrix decomposition problems can be conceivably solved using small data sketches, i.e., a small set of random observations of the data [@lamport16; @new3; @rahmani2015analysis; @new2; @lamport8; @lamport14]. In [@lamport16], it was shown based on a simple degrees-of-freedom analysis that the LR and the sparse components can be precisely recovered using a small set of random linear measurements of $\bD$. A convex program was proposed in [@lamport16] to recover these components using random matrix embedding with a polylogarithmic penalty factor in sample complexity, albeit the formulation also requires solving a high-dimensional optimization problem. The iterative algorithms which solve (\[eq2\]) have complexity $\calO(N_1 N_2 r)$ per iteration since they compute the partial SVD decomposition of $N_1 \times N_2$ dimensional matrices [@lamport23]. To reduce complexity, GoDec [@zhou2011godec] uses a randomized method to efficiently compute the SVD, and the decomposition algorithm in [@mu2011accelerated] minimizes the rank of $\mathbf{\Phi} \bL$ instead of $\bL$, where $\mathbf{\Phi}$ is a random projection matrix. However, these approaches do not have provable performance guarantees and their memory requirements scale with the full data dimensions. Another limitation of the algorithm in [@mu2011accelerated] is its instability since different random projections may yield different results. The divide-and-conquer approach in [@new1] (and a similar algorithm in [@liu2011solving]), can achieve super-linear speedups over full-scale matrix decomposition. This approach forms an estimate of $\bL$ by combining two low-rank approximations obtained from submatrices formed from sampled rows and columns of $\bD$ using the generalized Nyström method [@goreinov1997theory]. Our approach also achieves super-linear speedups in decomposition, yet is fundamentally different from [@new1] and offers several advantages for the following reasons. First, our approach is a *subspace-pursuit approach* that focuses on subspace learning in a structure-preserving data sketch. Once the CS is learned, each column of the data is decomposed independently using a proposed randomized vector decomposition algorithm. Second, unlike [@new1], which is a batch approach that requires to store the entire data, the structure of the proposed approach naturally lends itself to online implementation (c.f. Section \[sec:online\]), which could be very beneficial for settings where the data comes in on the fly. Third, while the analysis provided in [@new1] requires roughly $\calO (r^2 \mu^2 \max (N_1 , N_2))$ random observations to ensure exact decomposition with high probability (whp), we show that the order of sufficient number of random observations depends linearly on the rank and the coherency parameter even if uniform random sampling is used. Fourth, the structure of the proposed approach enables us to leverage efficient sampling strategies for challenging and realistic scenarios in which the columns and rows of $\bL$ are not uniformly distributed in their respective subspaces, or when the data exhibits additional structures (e.g. clustering structures) (c.f. Sections \[sec:eff\_sampling\],\[sec:alg\_eff\_sampling\]). In such settings, the uniform random sampling used in [@new1] requires significantly larger amounts of data to carry out the decomposition. Structure of the Proposed Approach and Theoretical Result ========================================================= In this section, the structure of the proposed randomized decomposition method is presented. A step-by-step analysis of the proposed approach is provided and sufficient conditions for exact decomposition are derived. Theorem \[thm:main\_result\] stating the main theoretical result of the paper is presented at the end of this section. The proofs of the lemmas and the theorem are deferred to the appendix. Let us rewrite (\[eq1\]) as $$\begin{aligned} \bD=\bU\bQ+\bS, \label{eq17}\end{aligned}$$ where $\bQ=\boldsymbol{\Sigma} \bV $. The representation matrix $\bQ \in \mathbb{R}^{r \times N_2} $ is a full row rank matrix that contains the expansion of the columns of $\bL$ in the orthonormal basis $\bU$. The first step of the proposed approach aims to learn the CS of $\bL$ using a subset of the columns of $\bD$, and in the second step the representation matrix is obtained using a subset of the rows of $\bD$. Let $\:\calU$ denote the CS of $\bL$. Fundamentally, $\calU$ can be obtained from a small subset of the columns of $\bL$. However, since we do not have direct access to the LR matrix, a random subset of the columns of $\bD$ is first selected. Hence, the matrix of sampled columns $\bD_{s1}$ can be written as $\bD_{s1}=\bD \bS_{1}$, where $\bS_1 \in \mathbb{R}^{N_2 \times m_1} $ is the column sampling matrix and $m_1$ is the number of selected columns. The matrix of selected columns can be written as $$\begin{aligned} \bD_{s1}=\bL_{s1} + \bS_{s1}, \label{eq19}\end{aligned}$$ Suppose $m_1$ columns are sampled uniformly at random from the matrix $\bL$ with rank $r$. If $$\begin{aligned} m_1 \ge r\gamma^2 (\bV) \max \left( c_2 \log r , c_3 \log\left(\frac{3}{\delta}\right) \right), \label{eq27}\end{aligned}$$ then the selected columns of the matrix $\bL$ span the columns subspace of $\bL$ with probability at least $(1-\delta)$ where $c_2$ and $c_3$ are numerical constants. \[lm3\] Thus, if $\gamma (\bV)$ is small (i.e., the RS is not coherent), a small set of randomly sampled columns can span $\calU$. According to Lemma \[lm3\], if $m_1$ satisfies (\[eq27\]), then $\bL$ and $\bL_{s1}$ have the same CS whp. The following optimization problem (of dimensionality $N_1 m_1$) is solved to decompose $\bD_{s1}$ into its LR and sparse components. $$\begin{aligned} \begin{aligned} & \underset{\dot{\bL}_{s1},\dot{\bS}_{s1}}{\min} & & \frac{1}{\sqrt{N_1}}\| \dot{\bS}_{s1} \|_1 + \| \dot{\bL}_{s1} \|_* \\ & \text{subject to} & & \dot{\bL}_{s1} + \dot{\bS}_{s1} = \bD_{s1} . \\ \end{aligned} \label{eq20}\end{aligned}$$ Thus, the columns subspace of the LR matrix can be recovered by finding the columns subspace of $ \bL_{s1} $. Our next lemma establishes that (\[eq20\]) yields the exact decomposition using roughly $m_1 = \calO (\mu r)$ randomly sampled columns. To simplify the analysis, in the following lemma it is assumed that the CS of the LR matrix is sampled from the random orthogonal model [@lamport4], i.e., the columns of $\bU$ are selected uniformly at random among all families of $r$-orthonormal vectors. Suppose the columns subspace of $\bL$ is sampled from the random orthogonal model, $\bL_{s1}$ has the same column subspace of $\bL$ and the support set of $\bS$ follows the Bernoulli model with parameter $\rho$. In addition, assume that the columns of $\bD_{s1}$ were sampled uniformly at random. If $$\begin{aligned} m_1 \ge {\frac{r}{\rho}_r} \mu^{'} (\log N_1)^2 \: \: \: \text{and} \: \: \: \rho \leq \rho_s \:, \label{eq28}\end{aligned}$$ then (\[eq20\]) yields the exact decomposition with probability at least $1-c_8 N_1^{-3}$, where $$\begin{aligned} \mu^{'} \hspace{-1.5mm} = \hspace{-.5mm} \max \hspace{-.75mm}\left( \hspace{-.75mm}\frac{c_7 \max (r,\log N_1)}{r} , 6\gamma^2(\bV) , (c_9\gamma(\bV) \log N_1)^2 \hspace{-.75mm}\right) \label{eq30}\end{aligned}$$ and $c_7$, $c_8$ and $c_9$ are constant numbers provided that $N_1$ is greater than the RHS of first inequality of (\[eq28\]). \[lm4\] Therefore, according to Lemma \[lm4\] and Lemma \[lm3\], the CS of $\bL$ can be obtained using roughly $O(r \mu)$ uniformly sampled data columns. Note that $m_1 \ll N_1$ for high-dimensional data as $m_1$ scales linearly with $r$. Hence, the requirement that $N_1$ is also greater than the RHS of first inequality of (\[eq28\]) is by no means restrictive and is naturally satisfied. Suppose that (\[eq20\]) decomposes $ \bD_{s1} $ into its exact components and assume that $\calU$ has been correctly identified. W.l.o.g., we can use $ \bU$ as an orthonormal basis for the learned CS. An arbitrary column $\bd_i$ of $\bD$ can be written as $\bd_i = \bU \bq_i + \bs_i$, where $\bq_i$ and $\bs_i$ are the corresponding columns of $\bQ$ and $\bS$, respectively. Thus, $\bd_i - \bU \bq_i$ is a sparse vector. This suggests that $\bq_i$ can be learned using the minimization $$\begin{aligned} \underset{\hat{\bq_i}}{\min} \|\bd_i-\bU \hat{\bq_i}\|_1 \:, \label{firstrow}\end{aligned}$$ where the $\ell_1$-norm is used as a surrogate for the $\ell_0$-norm to promote a sparse solution [@lamport17; @lamport3]. The optimization problem (\[firstrow\]) is similar to a system of linear equations with $r$ unknown variables and $N_1$ equations. Since $r \ll N_1$, the idea is to learn $\bq_i$ using only a small subset of the equations. Thus, we propose the following vector decomposition program $$\begin{aligned} \underset{\hat{\bq_i}}{\min} \|\bS_2^T \bd_i- \bS_2^T \bU \hat{\bq_i}\|_1 \:, \label{n_eq1} \label{secondrows}\end{aligned}$$ where $\bS_2 \in \mathbb{R}^{N_1 \times m_2} $ selects $m_2$ rows of $\bU$ (and the corresponding $m_2$ elements of $\bd_i$). First, we have to ensure that the rank of $\bS_2^T \bU$ is equal to the rank of $\bU$, for if $\bq^{*}$ is the optimal point of (\[secondrows\]), then $\bU \bq^{*}$ will be the LR component of $\bd_i$. According to Lemma \[lm3\], $m_2 = \calO (r \gamma(\bU))$, is sufficient to preserve the rank of $\bU$ when the rows are sampled uniformly at random. In addition, the following lemma establishes that if the rank of $\bU$ is equal to the rank of $\bS_2^T \bU$, then the sufficient value of $m_2$ for (\[n\_eq1\]) to yield the correct columns of $\bQ$ whp is linear in $r$. Suppose that the rank of $\bS_2^T \bU$ is equal to the rank of $\bL$ and assume that the CS of $\bL$ is sampled from the random orthogonal model. The optimal point of (\[secondrows\]) is equal to $\bq_i$ with probability at least $(1-3\delta)$ provided that $$\begin{aligned} \begin{aligned} &\rho \leq \frac{0.5}{r \beta \left( c_6 \kappa \log \frac{N_1 }{\delta} + 1 \right)}~, \\ &m_2 \ge \max \bigg( \frac{ 2 r \beta (\beta - 2) \log \left( \frac{ 1 }{\delta} \right)}{3(\beta - 1)^2} \left( c_6 \kappa \log \frac{N_1 }{\delta} +1 \right) , \\ &\qquad\qquad\qquad\qquad\qquad\qquad c_5 (\log \frac{N_1 }{\delta})^2 , \sqrt[6]{\frac{3}{\delta}} \bigg) \end{aligned} \label{eq38}\end{aligned}$$ where $\kappa= \frac{\log N_1}{r}$, $c_5$ and $c_6$ are constant numbers and $\beta$ can be any real number greater than one. \[lm8\] Therefore, we can obtain the LR component of each column using a random subset of its elements. Since (\[firstrow\]) is an $\ell_1$-norm minimization, we can write the representation matrix learning problem as $$\begin{aligned} \underset{\hat{\bQ}}{\min} \| \bS_2^T \bD- \bS_2^T \bU \hat{\bQ}\|_1. \label{eq23}\end{aligned}$$ Thus, (\[eq23\]) learns $\bQ$ using a subset of the rows of $\bD$ as $\bS_2^T \bD$ is the matrix formed from $m_2$ sampled rows of $\bD$. As such, we solve two low-dimensional subspace pursuit problems (\[eq20\]) and (\[eq23\]) of dimensions $N_1 m_1$ and $N_2 m_2$, respectively, instead of an $N_1 N_2$-dimensional decomposition problem (\[eq2\]), and use a small random subset of the data to learn $\bU$ and $\bQ$. The table of Algorithm 1 explains the structure of the proposed approach. We can readily state the following theorem which establishes sufficient conditions for Algorithm 1 to yield exact decomposition. In this theorem, it is assumed that the columns and rows are sampled uniformly at random. In the next section, an efficient method for column and row sampling is presented. In addition, a scheme for online implementation is proposed. \[thm:main\_result\] Suppose the columns subspace of the LR matrix is sampled from the random orthogonal model and the support set of $\bS$ follows the Bernoulli model with parameter $\rho$. In addition, it is assumed that Algorithm 1 samples the columns and rows uniformly at random. If for any small $\delta > 0$, $$\begin{aligned} \begin{aligned} m_1 & \ge \max \Bigg ( r\gamma^2(\bV) \max \left( c_2 \log r , c_3 \log \frac{3}{\delta} \right),\\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad {\frac{r}{\rho}_r} \mu^{'} (\log N_1)^2 \hspace{-0.85mm}\Bigg)\\ m_2 & \ge \max \Bigg( r \log N_1 \max \big( c_2^{'} \log r , c_3^{'} \log \frac{3}{\delta} \big), \\ & \qquad \frac{ 2 r \beta (\beta - 2)\log \left( \frac{N_2}{\delta} \right)}{3 (\beta - 1)^2} \left( c_6 \kappa \log \frac{N_1 N_2}{\delta} +1 \right) ,\\ & \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad c_5 (\log \frac{N_1 N_2}{\delta})^2 , \sqrt[6]{\frac{3}{\delta}} \Bigg)\\ \rho &\leq \min \left( \rho_s , \frac{0.5}{r \beta \left( c_6 \kappa \log \frac{N_1 N_2}{\delta} + 1 \right) } \right) \end{aligned} \label{eq25}\end{aligned}$$ where $$\begin{aligned} %\begin{aligned} \mu^{'} &= \max \bigg( \frac{c_7 \max (r,\log N_1)}{r} , 6\gamma^2 (\bV) , %\nonumber\\ \left(c_9 \gamma (\bV) \log N_1 \right)^2 \bigg), \nonumber \\ \kappa &= \frac{\log N_1}{r},\nonumber \label{eq26} %\end{aligned}\end{aligned}$$ $\{c_i\}_{i=1}^9$, $c_2^{'}$ and $c_3^{'}$ are constant numbers and $\beta$ is any real number greater that one, then the proposed approach (Algorithm 1) yields the exact decomposition with probability at least $(1-5\delta - 3r N_1^{-7})$ provided that $N_1$ is greater than the RHS of first inequality of (\[eq28\]). Theorem (\[thm:main\_result\]) guarantees that the LR component can be obtained using a small subset of the data. The randomized approach has two main advantages. First, it significantly reduces the memory/storage requirements since it only uses a small data sketch and solves two low-dimensional optimization problems versus one large problem. Second, the proposed approach has $\calO(\max (N_1 , N_2) \times \max(m_1 , m_2) \times r)$ per-iteration running time complexity, which is significantly lower than $\calO(N_1 N_2 r)$ per iteration for full scale decomposition (\[eq2\]) [@lamport53; @lamport23] implying remarkable speedups for big data. For instance, consider $\bU$ and $\bQ$ sampled from $\calN (0 , 1)$, $r = 5$, and $\bS$ following the Bernoulli model with $\rho = 0.02$. For values of $N_1 = N_2$ equal to 500, 1000, 5000, $10^4$ and $2\times 10^4$, if $m_1 = m_2 = 10r$, the proposed approach yields the correct decomposition with 90, 300, 680, 1520 and 4800 - fold speedup, respectively, over directly solving (\[eq2\]). [**Input**: Data matrix $\bD \in \mathbb{R}^{N_1 \times N_2} $\ **1. Initialization**: Form column sampling matrix $\bS_1 \in \mathbb{R}^{N_2 \times m_1} $ and row sampling matrix $\bS_2 \in \mathbb{R}^{N_1 \times m_2} $.\ **2. CS Learning**\ **2.1** Column sampling: Matrix $\bS_1$ samples $m_1$ columns of the given data matrix, $\bD_{s1} = \bD \bS_1$.\ **2.2** CS learning: The convex program (\[eq20\]) is applied to the sampled columns $\bD_{s1}$.\ **2.3** CS calculation: The CS is found as the columns subspace of the calculated LR component.\ **3. Representation Matrix Learning**\ **3.1** Row sampling: Matrix $\bS_2$ samples $m_2$ rows of the given data matrix, $\bD_{s2} = \bS_2^T \bD$.\ **3.2** Representation matrix learning: The convex problem (\[eq23\]) is applied to the sampled rows to find the representation matrix.\ **Output:** If $\hat{\bU}$ is an orthonormal basis for the learned CS and $\hat{\bQ}$ is the obtained representation matrix, then $\hat{\bL} = \hat{\bU} \hat{\bQ}$ is the obtained LR component. ]{} Efficient Column/Row Sampling ============================= In sharp contrast to randomized algorithms for matrix approximations rooted in numerical linear algebra (NLA) [@ghadim1; @mahoney2011randomized], which seek to compute matrix approximations from sampled data using importance sampling, in matrix decomposition and robust PCA we do not have direct access to the LR matrix to measure how informative particular columns/rows are. As such, the existing randomized algorithms for matrix decomposition and robust PCA[@new1; @new2; @new3; @lamport8] have predominantly relied upon uniform random sampling of columns/rows. In Section \[sec:non\_uniform\], we briefly describe the implications of non-uniform data distribution and show that uniform random sampling may not be favorable for data matrices exhibiting some structures that prevail much of the real datasets. In Section \[sec:eff\_sampling\], we demonstrate an efficient column sampling strategy which will be integrated with the proposed decomposition method. The decomposition method with efficient column/row sampling is presented in Sections \[sec:alg\_eff\_sampling\] and \[sec:both\_col\_row\]. Non-uniform data distribution {#sec:non_uniform} ----------------------------- When data points lie in a low-dimensional subspace, a small subset of the points can span the subspace. However, uniform random sampling is only effective when the data points are distributed uniformly in the subspace. To clarify, Fig. \[fig:innovationfig\] shows two scenarios for a set of data points in a two-dimensional subspace. In the left plot, the data points are distributed uniformly at random. In this case, two randomly sampled data points can span the subspace whp. In the right plot, 95 percent of the data lie on a one-dimensional subspace, thus we may not be able to capture the two-dimensional subspace from a small random subset of the data points. ![Data distributions in a two-dimensional subspace. The red points are the normalized data points.[]{data-label="fig:innovationfig"}](distribution){width="41.00000%"} ![The rank of a set of uniformly random sampled columns for different number of clusters. []{data-label="fig:mumv"}](rank){width="30.00000%"} In practice, the data points in a low-dimensional subspace may not be uniformly distributed, but rather exhibit some additional structures. A prevailing structure in many modern applications is clustered data [@new5; @rahmani2015innovation]. For example, user ratings for certain products (e.g. movies) in recommender systems are not only LR due to their inherent correlations, but also exhibit additional clustering structures owing to the similarity of the preferences of individuals from similar backgrounds (e.g. education, culture, or gender) [@new5; @7389496]. To further show that uniform random sampling falls short when the data points are not distributed uniformly in the subspace, consider a matrix $\bG \in \mathbb{R}^{2000 \times 6150}$ generated as $ \bG = [ \bG_1 \: \bG_2 \: ... \: \bG_n ] \: $. For $1 \leq i \leq \frac{n}{2}$, $\bG_i = \bU_i \bQ_i\: ,$ where $\bU_i \in \mathbb{R}^{2000 \times \frac{r}{n}}$, $\bQ_i \in \mathbb{R}^{\frac{r}{n} \times \frac{200 r}{n}}$. For $n/2+1 \leq i \leq n$, $\bG_i = \bU_i \bQ_i\: ,$ where $\bU_i \in \mathbb{R}^{2000 \times \frac{r}{n}}$, $\bQ_i \in \mathbb{R}^{\frac{r}{n} \times \frac{5 r}{n}}$. The elements of $\bU_i$ and $\bQ_i$ are sampled independently from a normal $\calN(0 , 1)$ distribution. The parameter $r$ is set equal to 60, thus, the rank of $\bG$ is equal to 60 whp. Fig. \[fig:mumv\] illustrates the rank of the randomly sampled columns versus the number of sampled columns for different number of clusters $n$. As $n$ increases, so does the required number of uniformly sampled columns. When $n=60$, it turns out that we need to sample more than half of the columns to span the CS. As such, we cannot evade high-dimensionality with uniform random column/row sampling. In [@new5], it was shown that the RS coherency increases when the columns follow a more clustered distribution, and vice-versa. These observations match our theoretical result in Lemma \[lm3\], which established that the sufficient number of randomly sampled columns depends linearly on the coherency parameter of the RS. Efficient column sampling method {#sec:eff_sampling} -------------------------------- Column sampling is widely used for dimensionality reduction and feature selection [@lamport14; @lamport26; @new4]. In the column sampling problem, the LR matrix (or the matrix whose span is to be approximated with a small set of its columns) is available. Thus, the columns are sampled based on their importance, measured by the so-called leverage scores [@ghadim1], as opposed to blind uniform sampling. We refer the reader to [@lamport14; @new4] and references therein for more information about efficient column sampling methods. [**Input:** Matrix $\bA$.\ **1. Initialize**\ **1.1** The parameter $C$ is chosen as an integer greater than or equal to one. The algorithm finds $C$ sets of linearly dependent columns.\ **1.2** Set $\calI = \emptyset $ as the index set of the sampled columns and set $v = \tau$, $\bB = \bA$ and $\bC = [\:]$.\ \ **2.1** Let $\bb$ be a non-zero randomly sampled column from $\bB$ with index $i_b$. Update $\bC$ and $\calI$ as $\bC = [\bC \: \: \bb]$, $\calI = \{ \calI \: , \: i_b \}$.\ **2.2** **While $v \ge \tau$**\ **2.2.1** Set $\bE = \bP_c \bB \: , $ where $\bP_c$ is the projection matrix onto the complement space of $\spn(\bC)$.\ **2.2.2** Define $\mathbf{f}$ as the column of $\bE$ with the maximum $\ell_2$-norm with index $i_f$. Update $\bC$, $\calI$ and $v$ as $\bC = [\bC \: \: \: \mathbf{f}] \: \: , \: \: \calI = \{ \calI \: , \: i_f \} \: \: \text{and} \:\: v = \| \mathbf{f} \|_2 \: . $\ **2.2 End While**\ **2.3** Set $\bC = [\:]$ and set $\bB$ equal to $\bA$ with the columns indexed by $\calI$ set to zero.\ \ **Output:** The set $\calI$ contains the indices of the selected columns. ]{} ![Visualization of the matrices defined in Section \[sec:alg\_eff\_sampling\]. Matrix $\bD_w$ is selected randomly or using Algorithm 3 described in Section \[sec:both\_col\_row\].[]{data-label="fig: visula 1"}](blck1){width="44.00000%"} Proposed decomposition algorithm with efficient sampling {#sec:alg_eff_sampling} -------------------------------------------------------- In this section, we develop a modified decomposition algorithm that replaces uniform random sampling with the efficient column/row sampling method (Algorithm 2). In Section \[sec:sim\], it is shown that the proposed technique can remarkably reduce the sampling requirement. We consider a setting wherein the data points (the columns of $\bL$) are not uniformly distributed, rather they admit an additional structure (such as clustering), wherefore a small subset of uniformly sampled columns is not likely to span the CS. However, we assume that the *rows* of $\bL$ are distributed well enough, in the sense that they do not much align along any specific directions, such that $C_r r$ rows of $\bL$ sampled uniformly at random span its RS whp, for some constant $C_r$. In Section \[sec:both\_col\_row\], we dispense with this assumption. The proposed decomposition algorithm rests on three key ideas detailed next. ### Informative column sampling The first important idea underlying the proposed sampling approach is to start sampling along the dimension that has the better distribution. For instance, in the example considered in Section \[sec:non\_uniform\], the columns of $\bG$ admit a clustering structure. However, the CS of $\bG$ is a random $r$-dimensional subspace, which means that the rows of $\bG$ are distributed uniformly at random in the RS of $\bG$. Thus, in this case we start with row sampling. The main intuition is that while almost 60 randomly sampled rows of $\bG$ span the RS, a considerable portion of the columns (almost 4000) should be sampled to capture the CS as shown in Fig. \[fig:mumv\]. As another example, consider an extreme scenario where only two columns of $\bG \in \mathbb{R}^{1000 \times 1000}$ are non-zero. In this case, with random sampling we need to sample almost all the columns to ensure that the sampled columns span the CS of $\bG$. But, if the non-zero columns are non-sparse, a small subset of randomly chosen rows of $\bG$ will span its row space. Let $\hat{r}$ denote a known upper bound on $r$. Such knowledge is often available as side information depending on the particular application. For instance, facial images under varying illumination and facial expressions are known to lie on a special low-dimensional subspace [@wright2009robust]. For visualization, Fig. \[fig: visula 1\] provides a simplified illustration of the matrices defined in this section. We sample $C_r \hat{r}$ rows of $\bD$ uniformly at random. Let $\bD_w \in \mathbb{R}^{ (C_r \hat{r}) \times N_2 }$ denote the matrix of sampled rows. We choose $C_r$ sufficiently large to ensure that the non-sparse component of $\bD_w$ is a LR matrix. Define $\bL_w$, assumably with rank $r$, as the LR component of $\bD_w$. If we locate a subset of the columns of $\bL_w$ that span its CS, the corresponding columns of $\bL$ would span its CS. To this end, the convex program (\[eq2\]) is applied to $\bD_w$ to extract its LR component denoted $\hat{\bL}_w$. Then, Algorithm 2 is applied to $\hat{\bL}_w$ to find a set of informative columns by sampling $m_1 \approx C \hat{r}$ columns. In Remark \[rem:how\_to\_choose\_c\], we discuss how to choose $C$ in the algorithm. Define $\hat{\bL}_w^s$ as the matrix of columns selected from $\hat{\bL}_w$. The matrix $\bD_{s1}$ is formed using the columns of $\bD$ corresponding to the sampled columns of $\hat{\bL}_w$. ### CS learning Similar to the CS learning step of Algorithm 1, we can obtain the CS of $\bL$ by decomposing $\bD_{s1}$. However, we propose to leverage valuable information in the matrix $\hat{\bL}_w^s$ in decomposing $\bD_{s1}$. In particular, if $\bD_w$ is decomposed correctly, the RS of $\hat{\bL}_w^s$ would be same as that of $\bL_{s1}$ given that the rank of $\bL_w$ is equal to $r$. Let $\bV_{s1}$ be an orthonormal basis for the RS of $\hat{\bL}_w^s$. Thus, in order to learn the CS of $\bD_{s1}$, we only need to solve $$\begin{aligned} \underset{\hat{\bU}}{\min} \| \bD_{s1} - \hat{\bU} \bV_{s1}^T \|_1 \:. \label{eq:columnspaceler}\end{aligned}$$ \[rem:how\_to\_choose\_c\] Define $\bd_{s1}^i$ as the $i^{\text{th}}$ row of $\bD_{s1}$. According to (\[eq:columnspaceler\]), $\bU^i$ (the $i^{\text{th}}$ row of $\bU$) is obtained as the optimal point of $$\begin{aligned} \underset{\hat{\bU}^i}{\min} \| (\bd_{s1}^i)^T - \bV_{s1} (\hat{\bU}^i)^T \|_1 \: . \label{eq_jobr}\end{aligned}$$ Based on the analysis provided for the proof of Lemma \[lm8\], the optimal point of (\[eq\_jobr\]) is equal to $\bU^i$ if $m_1 \ge r + \eta \| \bS_{s1}^i \|_0 $, where $\bS_{s1}^i$ is the $i^{\text{th}}$ row of $\bS_{s1}$, $\|\bS_{s1}^i\|_0$ the number of non-zero elements of $\bS_{s1}^i$, and $\eta$ a real number which depends on the coherency of the subspace spanned by $\bV_{s1}$. Thus, here $C$ is determined based on the rank of $\bL$ and the sparsity of $\bS$, i.e., $C r - r$ has to be sufficiently greater than the expected value for the number of non-zero elements of the rows of $\bS_{s1}$. We note that the convex algorithm (\[eq2\]) may not always yield accurate decomposition of $\bD_w$ since structured data may not be sufficiently incoherent [@lamport1; @new5] suggesting that the decomposition step can be further improved. Let $\bD_w^s$ be the matrix consisting of the columns of $\bD_w$ corresponding to the columns selected from $\hat{\bL}_w$ to form $\hat{\bL}_w^s$. According to our investigations, an improved $\bV_{s1}$ can be obtained by applying the decomposition algorithm presented in [@lamport19] to $\bD_w^s$ then use the RS of $\hat{\bL}_w^s$ as an initial guess for the RS of the non-sparse component of $\bD_w^s$. Since $\bD_w^s$ is low-dimensional , this extra step is a low complexity operation. ### Representation matrix learning Suppose that the CS of $\bL$ was learned correctly, i.e., the span of the optimal point of (\[eq:columnspaceler\]) is equal to the span of $\bU$. Thus, we use $\bU$ as a basis for the learned CS. Now we leverage the information embedded in $\bU$ to select the informative rows. Algorithm 2 is applied to $\bU^T$ to locate $m_2 \approx C r$ rows of $\bU$. Thus, we form the matrix $\bD_{s2}$ from the rows of $\bD$ corresponding to the selected rows of $\bU$. Thus, the representation matrix is learned as $$\begin{aligned} \underset{\hat{\bQ}}{\min} \| \bD_{s2} - \bU_{s2} \hat{\bQ} \|_1 \: ,\end{aligned}$$ where $\bU_{s2} \in \mathbb{R}^{m_2 \times r}$ is the matrix of the selected rows of $\bU$. Subsequently, the LR matrix can be obtained from the learned CS and the representation matrix. Column/Row sampling from sparsely corrupted data {#sec:both_col_row} ------------------------------------------------ In Section \[sec:alg\_eff\_sampling\], we assumed that the LR component of $\bD_w$ has rank $r$. However, if the rows are not well-distributed, a reasonably sized random subset of the rows may not span the RS of $\bL$. Here, we present a sampling approach which can find the informative columns/rows even when both the columns and the rows exhibit clustering structures such that a small random subset of the columns/rows of $\bL$ cannot span its CS/RS. The algorithm presented in this section (Algorithm 3) can be independently used as an efficient sampling approach from big data. In this paper, we use Algorithm 3 to form $\bD_w$ if both the columns and rows exhibit clustering structures. The table of Algorithm 3, Fig. \[alg3\] and its caption provide the details of the proposed sampling approach and the definitions of the used matrices. We start the cycle from the position marked “I” in Fig. \[alg3\] with $\bD_w$ formed according to the initialization step of Algorithm 3. For ease of exposition, assume that $\hat{\bL}_w = \bL_w$ and $\hat{\bL}_c = \bL_c$, i.e., $\bD_w$ and $\bD_c$ are decomposed correctly. The matrix $\hat{\bL}_w^s$ is the informative columns of $\hat{\bL}_w$. Thus, the rank of $\hat{\bL}_w^s$ is equal to the rank of $\hat{\bL}_w$. Since $\hat{\bL}_w = \bL_w$, $\hat{\bL}_w^s$ is a subset of the rows of $\bL_c$. If the rows of $\bL$ exhibit a clustering structure, it is likely that rank$(\hat{\bL}_w^s)< \text{rank} (\bL_c)$. Thus, rank$({\bL}_w)< \text{rank} (\bL_c)$. We continue one cycle of the algorithm by going through steps 1, 2 and 3 of Fig. \[alg3\] to update $\bD_w$. Using a similar argument, we see that the rank of an updated $\bL_w$ will be greater than the rank of $\bL_c$. Thus, if we run more cycles of the algorithm – each time updating $\bD_w$ and $\bD_c$ – the rank of $\bL_w$ and $\bL_c$ will increase. As detailed in the table of Algorithm 3, we stop if the dimension of the span of the obtained LR component does not change in $T$ consecutive iterations. While there is no guarantee that the rank of $\bL_w$ will converge to $r$ (it can converge to a value smaller than $r$), our investigations have shown that Algorithm 3 performs quite well and the RS of $\bL_w$ converges to the RS of $\bL$ in few steps. We have also found that adding some randomly sampled columns (rows) to $\bD_c (\bD_w)$ can effectively avert converging to a lower dimensional subspace. For instance, some randomly sampled columns can be added to $\bD_c$, which was obtained by applying Algorithm 2 to $\hat{\bL}_w$. [**1. Initialization**\ Form $\bD_{w} \in \mathbb{R}^{C_r \hat{r} \times N_2}$ by randomly choosing $C_r \hat{r}$ rows of $\bD$. Initialize $k = 1$ and set $T$ equal to an integer greater than 1.\ **2. While** $k > 0$\ **2.1 Sample the most informative columns**\ **2.1.1** Obtain $\hat{\bL}_w$ via (\[eq2\]) as the LR component of $\bD_w$.\ **2.1.2** Apply Algorithm 2 to $\hat{\bL}_w$ with $C = C_r$.\ **2.1.3** Form the matrix $\bD_c$ from the columns of $\bD$ corresponding to the sampled columns of $\hat{\bL}_w$.\ **2.2 Sample the most informative rows**\ **2.2.1** Obtain $\hat{\bL}_c$ via (\[eq2\]) as the LR component of $\bD_c$.\ **2.2.2** Apply Algorithm 2 to $\hat{\bL}_c^T$ with $C = C_r$.\ **2.2.3** Form the matrix $\bD_w$ from the rows of $\bD$ corresponding to the sampled rows of $\hat{\bL}_c$.\ **2.3 If** the dimension of the RS of $\hat{\bL}_w$ does not increase in $T$ consecutive iterations, set $k = 0$ to stop the algorithm.\ **2. End While**\ **Output:** The matrices $\bD_w$ and $\hat{\bL}_w$ can be used for column sampling in the first step of the Algorithm presented in Section \[sec:alg\_eff\_sampling\]. ]{} ![ Visualization of Algorithm 3. We run few cycles of the algorithm and stop when the rank of the LR component does not change over $T$ consecutive steps. One cycle of the algorithm starts from the point marked “I” and proceeds as follows. **I**: Matrix $\bD_w$ is decomposed and $\hat{\bL}_w$ is the obtained LR component of $\bD_w$. **II**: Algorithm 2 is applied to $\hat{\bL}_w$ to select the informative columns of $\hat{\bL}_w$. $\hat{\bL}_w^s$ is the matrix of columns selected from $\hat{\bL}_w$. **III**: Matrix $\bD_c$ is formed from the columns of $\bD$ that correspond to the columns of $\hat{\bL}_w^s$. **1**: Matrix $\bD_c$ is decomposed and $\hat{\bL}_c$ is the obtained LR component of $\bD_c$. **2**: Algorithm 2 is applied to $\hat{\bL}_c^T$ to select the informative rows of $\hat{\bL}_c$. $\hat{\bL}_c^s$ is the matrix of rows selected from $\hat{\bL}_c$. **3**: Matrix $\bD_w$ is formed as the rows of $\bD$ corresponding to the rows used to form $\hat{\bL}_c^s$. []{data-label="alg3"}](Untitled){width="42.00000%"} Algorithm 3 was found to converge in a very small number of iterations (typically less than 4 iterations). Thus, even when Algorithm 3 is used to form the matrix $\bD_w$, the order of complexity of the proposed decomposition method with efficient column/row sampling (presented in Section \[sec:alg\_eff\_sampling\]) is roughly $\calO(\max(N_1,N_2) r^2)$. Online Implementation {#sec:online} --------------------- The proposed decomposition approach consists of two main steps, namely, learning the CS of the LR component then decomposing the columns independently. This structure lends itself to online implementation, which could be very beneficial in settings where the data arrives on the fly. The idea is to first learn the CS of the LR component from a small batch of the data and keep tracking the CS. Since the CS is being tracked, any new data column can be decomposed based on the updated subspace. The table of Algorithm 4 details the proposed online matrix decomposition algorithm, where $\bd_t$ denotes the $t^{\text{th}}$ received data column. Algorithm 4 uses a parameter $n_u$ which determines the rate at which the algorithm updates the CS of the LR component. For instance, if $n_u = 20$, then the CS is updated every 20 new data columns (step 2.2 of Algorithm 4). The parameter $n_u$ has to be set in accordance with the rate of the change of the subspace of the LR component; a small value for $n_u$ is used if the subspace is changing rapidly. The parameter $n_s$ determines the number of columns last received that are used to update the CS. If the subspace changes rapidly, the older columns may be less relevant to the current subspace, hence a small value for $n_s$ is used. On the other hand, when the data is noisy and the subspace changes at a slower rate, choosing a larger value for $n_s$ can lead to more accurate estimation of the CS. [**1. Initialization**\ **1.1** Set the parameters $n_u$ and $n_s$ equal to integers greater than or equal to one.\ **1.2** Form $\bD_0 \in \mathbb{R}^{N_1 \times (C_r \hat{r})}$ as $$\bD_0 = [\bd_1 \: \bd_2 \: ... \: \bd_{C_r \hat{r}}].$$ Decompose $\bD_0$ using (2) and obtain the CS of its LR component. Define ${\bU}_o$ as the learned CS, ${\bQ}_o$ the appropriate representation matrix and $\hat{\bS}$ the obtained sparse component of $\bD_0$.\ **1.3** Apply Algorithm 2 to ${\bU}_o^T$ to construct the row sampling matrix $\bS_2$.\ \ **2. For** any new data column $\bd_t$ do\ **2.1** Decompose $\bd_t$ as $$\begin{aligned} \underset{\hat{\bq}_t}{\min} \| \bS_2^T \bd_{t} - \bS_2^T \bU_o \hat{\bq}_t \|_1 \: , \label{eq_q_find}\end{aligned}$$ and update\ $\bQ_o \gets [\bQ_o \: \: \bq_t^{*}]$, $\hat{\bS} \gets [\hat{\bS} \: \: (\bd_t - \bU_o \bq_t^{*})]$, where $\bq_t^{*}$ is the optimal point of (\[eq\_q\_find\]).\ **2.2 If** the remainder of $\frac{t}{n_u}$ is equal to zero, update $\bU_o$ as $$\begin{aligned} \underset{\hat{U}_o}{\min} \| \bD_t - \hat{\bU}_o \bQ_o^t \|_1 \: , \label{eq_online_up}\end{aligned}$$ where $\bQ_o^t$ is the last $n_s \hat{r}$ columns of $\bQ_o$ and $\bD_t$ is the matrix formed from the last $n_s \hat{r}$ received data columns. Apply Algorithm 2 to the new $\bU_o^T$ to update the row sampling matrix $\bS_2$.\ **2. End For**\ **Output** The matrix $\hat{\bS}$ as the obtained sparse matrix, $\hat{\bL} = \bD - \hat{\bS}$ as the obtained LR matrix and $\bU_o$ as the current basis for the CS of the LR component. ]{} Noisy Data ---------- In practice, noisy data can be modeled as $$\begin{aligned} \bD = \bL + \bS + \bN \:,\end{aligned}$$ where $\bN$ is an additive noise component. In [@lamport22], it was shown that the program $$\begin{aligned} \begin{aligned} & \underset{\hat{\bL},\hat{\bS}}{\min} & & \lambda\|\hat{\bS}\|_1 + \|\hat{\bL} \|_* \\ & \text{subject to} & & \big\| \hat{\bL} + \hat{\bS} - \bD \big\|_F \leq \epsilon_n \:, \\ \end{aligned} \label{eq2noisy}\end{aligned}$$ can recover the LR and sparse components with an error bound that is proportional to the noise level. The parameter $\epsilon_n$ has to be chosen based on the noise level. This modified version can be used in the proposed algorithms to account for the noise. Similarly, to account for the noise in the representation learning problem (\[eq23\]), the $\ell_1$-norm minimization problem can be modified as follows: $$\begin{aligned} \underset{\hat{\bQ} , \hat{\bE}}{\min} \| \bS_2^T \bD- \bS_2^T \bU \hat{\bQ} - \hat{\bE} \|_1 \quad \text{subject to} \quad \| \hat{\bE} \|_F \leq \delta_n.\end{aligned}$$ $\hat{\bE} \in \mathbb{R}^{m_2 \times N_2}$ is used to cancel out the effect of the noise and the parameter $\delta_n$ is chosen based on the noise level [@candesstab]. Numerical Simulations {#sec:sim} ===================== In this section, we present some numerical simulations to study the performance of the proposed randomized decomposition method. First, we present a set of simulations confirming our analysis which established that the sufficient number of sampled columns/rows is linear in $r$. Then, we compare the proposed approach to the state-of-the-art randomized algorithm [@new1] and demonstrate that the proposed sampling strategy can lead to notable improvement in performance. We then provide an illustrative example to showcase the effectiveness of our approach on real video frames for background subtraction and activity detection. Given the structure of the proposed approach, it is shown that side information can be leveraged to further simplify the decomposition task. In addition, a numerical example is provided to examine the performance of Algorithm 3. Finally, we investigate the performance of the online algorithm and show that the proposed online method can successfully track the underlying subspace. In all simulations, the Augmented Lagrange multiplier (ALM) algorithm [@lamport23; @lamport2] is used to solve the optimization problem (\[eq2\]). In addition, the $\ell_1$-magic routine [@lamport57] is used to solve the $\ell_1$-norm minimization problems. It is important to note that in all the provided simulations (except in Section \[sec:alternating\_alg\]), the convex program (\[eq2\]) that operates on the entire data can yield correct decomposition with respect to the considered criteria. Thus, if the randomized methods cannot yield correct decomposition, it is because they fall short of acquiring the essential information through sampling. Phase transition plots ---------------------- In this section, we investigate the required number of randomly sampled columns/rows. The LR matrix is generated as a product $\bL= \bU_r \bQ_r$, where $\bU_r \in \mathbb{R}^{N_1 \times r}$ and $\bQ_r \in \mathbb{R}^{r \times N_2}$. The elements of $\bU_r$ and $\bQ_r$ are sampled independently from a standard normal $\mathcal{N}(0,1)$ distribution. The sparse matrix $\bS$ follows the Bernoulli model with $\rho = 0.02$. In this experiment, Algorithm 1 is used and the column/rows are sampled uniformly at random. Fig. \[fig:trans\_plots\] shows the phase transition plots for different numbers of randomly sampled rows/columns. In this simulation, the data is a $1000\times1000$ matrix. For each $(m_1 , m_2)$, we generate 10 random realizations. A trial is considered successful if the recovered LR matrix $\hat{\bL}$ satisfies $\frac{\| \bL - \hat{\bL} \|_F}{\| \bL \|_F} \leq 5 \times 10^{-3}$. It is clear that the required number of sampled columns/rows increases as the rank or the sparsity parameter $\rho$ are increased. When the sparsity parameter is increased to 0.3, the proposed algorithm can hardly yield correct decomposition. Actually, in this case the matrix $\bS$ is no longer a sparse matrix. The top row of Fig. \[fig:trans\_plots\] confirms that the sufficient values for $m_1$ and $m_2$ are roughly linear in $r$. For instance, when the rank is increased from 5 to 25, the required value for $m_1$ increases from 30 to 140. In this experiment, the column and RS of $\bL$ are sampled from the random orthogonal model. Thus, the CS and RS have small coherency whp [@lamport4]. Therefore, the important factor governing the sample complexity is the rank of $\bL$. Indeed, Fig. \[fig:data\_dim\] shows the phase transition for different sizes of the data matrix when the rank of $\bL$ is fixed. One can see that the required values for $m_1$ and $m_2$ are almost independent of the size of the data confirming our analysis. ![Phase transition plots for various rank and sparsity levels. White designates successful decomposition and black designates incorrect decomposition.[]{data-label="fig:trans_plots"}](cv3){width="50.00000%"} ![Phase transition plots for various data matrix dimensions.[]{data-label="fig:data_dim"}](c4){width="50.00000%"} Efficient column/row sampling ----------------------------- In this experiment, the algorithm presented in Section \[sec:alg\_eff\_sampling\] is compared to the randomized decomposition algorithm in [@new1]. It is shown that the proposed sampling strategy can effectively reduce the required number of sampled columns/rows, and makes the proposed method remarkably robust to structured data. In this experiment, $\bD$ is a $2000\times4200$ matrix. The LR component is generated as $$\bL = [ \bG_1 \: \bG_2 \: ... \: \bG_n ] \: .$$ For $1 \leq i \leq \frac{n}{2}$, $$\bG_i = \bU_i \bQ_i\: ,$$ where $\bU_i \in \mathbb{R}^{2000 \times \frac{r}{n}}$, $\bQ_i \in \mathbb{R}^{\frac{r}{n} \times \frac{130 r}{n}}$ and the elements of $\bU_i$ and $\bQ_i$ are sampled independently from a normal distribution $\calN(0 , 1)$. For $n/2+1 \leq i \leq n$, $$\bG_i = 13 \bU_i \bQ_i\: ,$$ where $\bU_i \in \mathbb{R}^{2000 \times \frac{r}{n}}$, $\bQ_i \in \mathbb{R}^{\frac{r}{n} \times \frac{10 r}{n}}$, and the elements of $\bU_i$ and $\bQ_i$ are sampled independently from an $\calN(0 , 1)$ distribution. We set $r$ equal to 60; thus, the rank of $\bL$ is equal to 60 whp. The sparse matrix $\bS$ follows the Bernoulli model and each element of $\bS$ is non-zero with probability 0.02. In this simulation, we do not use Algorithm 3 to form $\bD_w$. The matrix $\bD_w$ is formed from 300 uniformly sampled *rows* of $\bD$. We evaluate the performance of the algorithm for different values of $n$, i.e., different number of clusters. Fig. \[fig:comp\] shows the performance of the proposed approach and the approach in [@new1] for different values of $m_1$ and $m_2$. For each value of $m_1 = m_2$, we compute the error in LR matrix recovery $\frac{\| \bL - \hat{\bL} \|_F}{\| \bL \|_F} $ averaged over 10 independent runs, and conclude that the algorithm can yield correct decomposition if the average error is less than 0.01. In Fig. \[fig:comp\], the values 0, 1 designate incorrect and correct decomposition, respectively. It can be seen that the presented approach requires a significantly smaller number of samples to yield the correct decomposition. This is due to the fact that the randomized algorithm [@new1] samples both the columns and rows uniformly at random and independently. In sharp contrast, we use $\hat{\bL}_w$ to find the most informative columns to form $\bD_{s1}$, and also leverage the information embedded in the CS to find the informative rows to from $\bD_{s2}$. One can see that when $n = 60$, [@new1] cannot yield correct decomposition even when $m_1 = m_2 = 1800$. ![Performance of the proposed approach and the randomized algorithm in [@new1]. A value 1 indicates correct decomposition and a value 0 indicates incorrect decomposition.[]{data-label="fig:comp"}](new4){width="50.00000%"} Vector decomposition for background subtraction ----------------------------------------------- The LR plus sparse matrix decomposition can be effectively used to detect a moving object in a stationary background [@lamport2; @bouwmans2015decomposition]. The background is modeled as a LR matrix and the moving object as a sparse matrix. Since videos are typically high dimensional objects, standard algorithms can be quite slow for such applications. Our algorithm is a good candidate for such a problem as it reduces the dimensionality significantly. The decomposition problem can be further simplified by leveraging prior information about the stationary background. In particular, we know that the background does not change or we can construct it with some pre-known dictionary. For example, consider the video from [@lamport61], which was also used in [@lamport2]. Few frames of the stationary background are illustrated in Fig. \[fig:bkgnd\]. ![Stationary background.[]{data-label="fig:bkgnd"}](fig3){width="40.00000%"} Thus, we can simply form the CS of the LR matrix using these frames which can describe the stationary background in different states. Accordingly, we just need to learn the representation matrix. As such, the background subtraction problem is simplified to a vector decomposition problem. ![Two frames of a video taken in a lobby. The first column displays the original frames. The second and third columns display the LR and sparse components recovered using the proposed approach. []{data-label="fig:decomp"}](cv){width="40.00000%"} Fig. \[fig:decomp\] shows that the proposed method successfully separates the background and the moving objects. In this experiment, 500 randomly sampled rows are used (i.e., 500 randomly sampled pixels) for the representation matrix learning (\[eq23\]). While the running time of our approach is just few milliseconds, it takes almost half an hour if we use (\[eq2\]) to decompose the video file [@lamport2]. Alternating algorithm for column sampling {#sec:alternating_alg} ----------------------------------------- In this section, we investigate the performance of Algorithm 3 for column sampling. The rank of the selected columns is shown to converge to the rank of $\bL$ even when both the rows and columns of $\bL$ exhibit a highly structured distribution. To generate the LR matrix $\bL$ we first generate a matrix $\bG$ as in Section \[sec:non\_uniform\] but setting $r = 100$. Then, we construct the matrix $\bU_g$ from the first $r$ right singular vectors of $\bG$. We then generate $\bG$ in a similar way and set $\bV_g$ equal to the first $r$ right singular vectors of $\bG$. Let the matrix $\bL = \bU_g \bV_g^T$. For example, for $n=100$, $\bL \in \mathbb{R}^{10250 \times 10250}$. Note that the resulting LR matrix is nearly sparse since in this simulation we consider a very challenging scenario in which both the columns and rows of $\bL$ are highly structured and coherent. Thus, in this simulation we set the sparse matrix equal to zero and use Algorithm 3 as follows. The matrix $\bD_c$ is formed using 300 columns sampled uniformly at random and the following steps are performed iteratively:\ 1. Apply Algorithm 2 to $\bD_c^T$ with $C=3$ to sample approximately $3 r$ columns of $\bD_c^T$ and form $\bD_w$ from the rows of $\bD$ corresponding to the selected rows of $\bD_c$.\ 2. Apply Algorithm 2 to $\bD_w$ with $C=3$ to sample approximately $3 r$ columns of $\bD_w$ and form $\bD_c$ from the columns of $\bD$ corresponding to the selected columns of $\bD_c$. Fig. \[fig:iteration\] shows the rank of $\bD_c$ after each iteration. It is evident that the algorithm converges to the rank of $\bL$ in less than 3 iterations even for $n = 100$ clusters. For all values of $n$, i.e., $n\in\{2, 50, 60\}$, the data is a $10250\times10250$ matrix. ![The rank of the matrix of sampled columns.[]{data-label="fig:iteration"}](iteration){width="31.00000%"} Online Implementation {#online-implementation} --------------------- In this section, the proposed online method is examined. It is shown that the proposed scalable online algorithm tracks the underlying subspace successfully. The matrix $\bS$ follows the Bernoulli model with $\rho = 0.01$. Assume that the orthonormal matrix $\bU \in \mathbb{R}^{N_1 \times r}$ spans a random $r$-dimensional subspace. The matrix $\bL$ is generated as follows. **For** $k$ from 1 to $N_2$\ **1.** Generate $\bE \in \mathbb{R}^{N_1 \times r}$ and $\bq \in \mathbb{R}^{r \times 1}$ randomly.\ **2.** $\bL = [\bL \: \: \bU \bq] $ .\ **3.** **If** (mod$(k , n)$ = 0)\ $ \bU = \text{approx-r} ( \bU + \alpha \bE).$\ **End If**\ **End For** The elements of $\bq_i $ and $\bE$ are sampled from standard normal distributions. The output of the function approx-r is the matrix of the first $r$ left singular vectors of the input matrix and mod$(k , n)$ is the remainder of $k / n$. The parameters $\alpha$ and $n$ control the rate of change of the underlying subspace. The subspace changes at a higher rate if $\alpha$ is increased or $n$ is decreased. In this simulation, $n = 10$, i.e., the CS is randomly rotated every 10 new data columns. In this simulation, the parameter $r = 5$ and $N_1 = 400$. We compare the performance of the proposed online approach to the online algorithm in [@citonline]. For our proposed method, we set $C=20$ when Algorithm 2 is applied to $\bU$, i.e., $20 r$ rows of $\bU$ are sampled. The method presented in [@citonline] is initialized with the exact CS and its tuning parameter is set equal to $1/\sqrt{N_1}$. The algorithm [@citonline] updates the CS with every new data column. The parameter $n_u$ of the proposed online method is set equal to 4 (i.e., the CS is updated every 4 new data columns) and the parameter $n_s$ is set equal to $5 r$. Define $\hat{\bL}$ as the recovered LR matrix. Fig. \[fig: online\] shows the $\ell_2$-norm of the columns of $\bL - \hat{\bL}$ normalized by the average $\ell_2$-norm of the columns of $\bL$ for different values of $\alpha$. One can see that the proposed method can successfully track the CS while it is continuously changing. The online method [@citonline] performs well when the subspace is not changing ($\alpha = 0$), however, it fails to track the subspace when it is changing. ![Performance of the proposed online approach and the online algorithm in [@citonline]. []{data-label="fig: online"}](online_fig){width="50.00000%"} Appendix {#appendix .unnumbered} ======== **Proof of Lemma \[lm3\]**\ The selected columns of $\bL$ can be written as $\bL_{s1}=\bL \bS_1$. Using the compact SVD of $\bL$, $\bL_{s1}$ can be rewritten as $\bL_{s1}=\bU \boldsymbol{\Sigma} \bV^T \bS_1$. Therefore, to show that the CS of $\bL_{s1}$ is equal to that of $\bL$, it suffices to show that the matrix $\bV^T \bS_1$ is a full rank matrix. The matrix $\bS_1$ selects $m_1$ rows of $\bV$ uniformly at random. Therefore, using Theorem 2 in [@lamport3], if $$\begin{aligned} m_1 \ge r\gamma^2(\bV) \max \left( c_2 \log r , c_3 \log \frac{3}{\delta} \right), \label{eq42}\end{aligned}$$ then the matrix $\bV^T \bS_1$ satisfies the inequality $$\begin{aligned} \| I - \frac{N_2}{m_1} \bV^T \bS_1 \bS_1^T \bV \|\leq \frac{1}{2} \label{eq43}\end{aligned}$$ with probability at least $(1-\delta)$, where $c_2, c_3$ are numerical constants [@lamport3]. Accordingly, if $\sigma_1$ and $\sigma_r$ denote the largest and smallest singular values of $\bS_1^T \bV$, respectively, then $$\begin{aligned} \frac{m_1}{2 N_2} \leq \sigma_1^2 \leq \sigma_r^2 \leq \frac{3 m_1}{2 N_2} \label{eq44}\end{aligned}$$ Therefore, the singular values of the matrix $\bV^T \bS_1$ are greater than $\sqrt{\frac{m_1}{2 N_2}}$. Accordingly, the matrix $\bV^T \bS_1$ is a full rank matrix. A direct application of Theorem 2 in [@lamport3] would in fact lead to the sufficient condition $$\begin{aligned} m_1 \ge r\gamma^2(\bR) \max \left( c_2 \log r , c_3 \log \frac{3}{\delta} \right), \label{neweq42}\end{aligned}$$ where $\bR \in \mathbb{R}^{N_2 \times N_2}$ denotes the matrix of right singular vectors of $\bL$. The bound in (\[eq42\]) is slightly tighter since it uses the incoherence parameter $\gamma(\bV)\leq \gamma(\bR) \triangleq \sqrt{N_2} \max_{i,j} |\bR(i,j)|$ in (\[neweq42\]), where $\bV$ consists of the first $r$ columns of $\bR$. This follows easily by replacing the incoherence parameter in the step that bounds the $\ell_2$-norm of the row vectors of the submatrix in the proof of ([@lamport3], Theorem 2). **Proof of lemma \[lm4\]**\ The sampled columns are written as $$\begin{aligned} \bD_{s1}=\bD \bS_1= \bL_{s1} + \bS_{s1}. \label{eq45}\end{aligned}$$ First, we investigate the coherency of the new LR matrix $\bL_{s1}$. Define $\textbf{P}_{\bS_1^T \bV}$ as the projection matrix onto the CS of $\bS_1^T \bV$ which is equal to the rows subspace of $\bL_{s1}$. Therefore, the projection of the standard basis onto the rows subspace of $\bL_{s1}$ can be written as $$\begin{aligned} \begin{aligned} &\underset{i}{\max} \| \textbf{P}_{\bS_1^T \bV} \be_i \|_2^2 = \underset{i}{\max} \| \bS_1^T \bV (\bV^T \bS_1 \bS_1^T \bV)^{-1} \bV^T \bS_1 \be_i \|_2^2 \\ & \leq \underset{j}{\max} \| \bS_1^T \bV (\bV^T \bS_1 \bS_1^T \bV)^{-1} \bV^T \be_j \|_2^2\\ &\leq \| \bS_1^T \bV (\bV^T \bS_1 \bS_1^T \bV)^{-1} \|^2 \| \bV^T \be_j \|_2^2 \\ & \leq \frac{\gamma^2 (\bV) r}{N_2} (\frac{\sigma_1^2}{\sigma_r^4})= \frac{\gamma^2 (\bV) r}{N_2} \frac{6 N_2}{m_1}=\frac{(6\gamma^2 (\bV))r}{m_1} \end{aligned} \label{eq46}\end{aligned}$$ where $ (\bS_1^T \bV (\bV^T \bS_1 \bS_1^T \bV)^{-1} \bV^T \bS_1)$ is the projection matrix onto the CS of $\bS_1^T \bV$. The first inequality follows from the fact that $\{ \bS_1 \be_i \}_{i=1}^{m_1}$ is a subset of $\{ \be_j \}_{j=1}^{N_2}$. The second inequality follows from Cauchy-Schwarz inequality and the third inequality follows from (\[eq10\]) and (\[eq44\]). Using lemma 2.2 of [@lamport4], there exists numerical constants $c_7$, $c_8$ such that $$\begin{aligned} \underset{i}{\max} \| \bU^T e_i \|_2^2 \leq \frac{\mu_p r}{N_1} \label{eq47}\end{aligned}$$ with probability at least $1-c_8 N_1^{-3}$ and $\mu_p = \frac{c_7 \max (r,\log N_1)}{r}$.\ In addition, we need to find a bound similar to the third condition of (\[eq9\]) for the LR matrix $\bL_{s1}$. Let $\bL_{s1} = \bU_{s1} \boldsymbol{\Sigma}_{s1} \bV_{s1}^T$ be the SVD decomposition of $\bL_{s1}$. Define $$\begin{aligned} \bH= \bU_{s1} \bV_{s1}^T = \sum_{i=1}^r \bU_{s1}^i (\bV_{s1}^i)^T \label{eq48}\end{aligned}$$ where $\bU_{s1}^i$ is the $i^{\text{th}}$ column of $\bU_{s1}$ and $\bV_{s1}^i$ is the $i^{\text{th}}$ column of $\bV_{s1}$. Given the random orthogonal model of the CS, $\bH$ has the same distribution as $$\begin{aligned} \bH^{'} = \sum_{i=1}^r \epsilon_i \bU_{s1}^i (\bV_{s1}^i)^T \label{eq49}\end{aligned}$$ where $\{ \epsilon_i \}$ is an independent Rademacher sequence. Using Hoeffding’s inequality [@lamport43], conditioned on $\bU_{s1}$ and $\bV_{s1}$ we have $$\begin{aligned} \begin{aligned} & \mathbb{P} \left( |\bH^{'} (i,j)| > t \right) \leq 2 e^{\frac{- t^2}{2 h_{ij}^2}} , \\ & h_{ij}^2=\sum_{k=1}^r (\bU_{s1} (i , k))^2 (\bV_{s1} (j , k))^2 \label{eq50} \end{aligned}\end{aligned}$$ Consider the following lemma adapted from Lemma 2.2 of [@lamport4]. If the orthonormal matrix $\bU$ follows the random orthogonal model, then $$\begin{aligned} \mathbb{P} \left( |\bU (i,j)|^2 \ge 20\frac{\log N_1}{N_1} \right) \leq 3 N_1^{-8} . \label{eq31}\end{aligned}$$ \[lm5\] Therefore, $$\begin{aligned} |\bU_{s1} (i , k)|^2 \leq 20 \frac{ \log N_1}{N_1} \label{eq51}\end{aligned}$$ with probability at least $1-3 N_1^{-8}$. Therefore, we can bound $h_{ij}^2$ as $$\begin{aligned} h_{ij}^2 \leq 20\frac{\log N_1}{N_1} \| \bV_{s1} e_i \|_2^2 \label{eq52}\end{aligned}$$ Using (\[eq46\]), (\[eq52\]) can be rewritten as $$\begin{aligned} h_{ij}^2 \leq 120\frac{\log N_1 \gamma^2 (\bV) r}{N_1 m_1} \label{eq53}\end{aligned}$$ Choose $t = \omega \frac{\gamma (\bV) \sqrt{r}}{\sqrt{N_1 m_1}}$ for some constant $\omega$. Thus, the unconditional form of (\[eq50\]) can be written as $$\begin{aligned} \begin{aligned} &\mathbb{P} \left( |\bH^{'} (i,j)| > \omega \frac{\gamma (\bV) \sqrt{r}}{\sqrt{N_1 m_1} } \right) \leq 2 e^{\frac{- \zeta \omega^2}{\log N_1}} +\\ & \mathbb{P} \left( h_{ij}^2 \ge 120\frac{\log N_1 \gamma^2 (\bV) r}{N_1 m_1} \right) \end{aligned} \label{eq54}\end{aligned}$$ for some numerical constant $\zeta$. Setting $\omega=\zeta^{'} \log N_1$ where $\zeta^{'}$ is a sufficiently large numerical constant gives $$\begin{aligned} \mathbb{P} \left( \| \bH^{'} \|_{\infty} \ge c_9 \log N_1 \frac{\gamma (\bV) \sqrt{r}}{\sqrt{N_1 m_1}} \right) \leq 3 r N_1^{-7} \label{eq55}\end{aligned}$$ for come constant number $c_9$ since (\[eq51\]) should be satisfied for $r N_1$ random variables. Therefore, according to lemma \[lm1\], if $$\begin{aligned} m_1 \ge {\frac{r}{\rho}_r} \mu^{'} (\log N_1)^2 \label{eq56}\end{aligned}$$ $$\begin{aligned} \rho \leq \rho_s \label{eq57}\end{aligned}$$ then, the convex algorithm (\[eq20\]) yields the exact decomposition with probability at least $1-c_8 N_1^{-3}$ where $$\begin{aligned} \mu^{'} = \max \left( \mu_p , 6\gamma^2 (\bV) , (\gamma (\bV) c_9 \log N_1)^2 \right) . \label{eq58}\end{aligned}$$\ **Proof of lemma \[lm8\]**\ Based on (\[eq1\]), the matrix of sampled rows can be written as $$\begin{aligned} \begin{aligned} \bD_{s2}=\bS_2^T \bD=\bS_2^T \bL + \bS_2^T \bS =\bL_{s2} + \bS_{s2} \end{aligned} \label{eq62}\end{aligned}$$ Let $\bL_{s2}=\bU_{s2} \boldsymbol{\Sigma}_{s2} \bV_{s2}^T$ be the compact SVD decomposition of $\bL_{s2}$ and $\bL_{s2}=\bU_{s2}^c \boldsymbol{\Sigma}_{s2}^c (\bV_{s2}^c)^T$ its complete SVD. It can be shown [@lamport17] that (\[secondrows\]) is equivalent to $$\begin{aligned} \begin{aligned} & \underset{\hat{\textbf{z}}_{i}}{\min} & & \| \hat{\textbf{z}}_{i} \|_1 \\ & \text{subject to} & & (\bU_{s2}^{\perp})^T \hat{\textbf{z}}_{i}= (\bU_{s2}^{\perp})^T \bS_{s2}^i \:. \end{aligned} \label{eq63}\end{aligned}$$ where $\bS_{s2}^i$ is the $i^{th}$ column of $\bS_{s2}$ and $\bU_{s2}^{\perp}$ is the last $(m_2-r)$ columns of $\bU_{s2}^c$ which are orthogonal to $\bU_{s2}$. In other words, if $\bq_i^{*}$ is the optimal point of (\[secondrows\]) and $\bz_i^{*} \in \mathbb{R}^{m_2}$ is the optimal point of (\[eq63\]), then $\bz_i^{*} = \bS_2^T (\bd_i - \bU \bq_i^{*})$. Thus, it is enough to show that the optimal point of (\[eq63\]) is equal to $\bS_{s2}^i$. The columns subspace of $\bU_{s2}$ obeys the random orthogonal model. Thus, $\bU_{s2}^{\perp}$ can be modeled as a random subset of $\bU_{s2}^c$. Based on the result in [@lamport3], if we assume that the sign of the non-zero elements of $\bS_{s2}^i$ are uniformly random, then the optimal point of (\[eq63\]) is $\bS_{s2}^i$ with probability at least $(1-\delta)$ provided that $$\begin{aligned} m_2 -r \ge \max \left( c_4 \| \bS_{s2}^i \|_0 \gamma^2 (\bU_{s2}^c) \log \frac{m_2}{\delta} , c_5 \left( \log \frac{m_2}{\delta} \right)^2 \right) \label{eq64}\end{aligned}$$ for some fixed numerical constants $c_4$ and $c_5$. The parameter $\gamma (\bU_{s2}^c) = \sqrt{m_2} \underset{i,j}{\max} |\bU_{s2}^c (i,j)|$ and $\| \bS_{s2}^i \|_0$ is the $l_0$-norm of $\bS_{s2}^i$. In this paper, we do not assume that the sign of the non-zero elements of the sparse matrix $\bS$ is random. However, according to Theorem 2.3 of [@lamport2] (de-randomization technique) if the locations of the nonzero entries of $\bS$ follow the Bernoulli model with parameter $2\rho$, and the signs of $\bS$ are uniformly random and if (\[eq63\]) yields the exact solution who, then it is also exact with at least the same probability for the model in which the signs are fixed and the locations follow the Bernoulli model with parameter $\rho$ [@lamport2]. Therefore, it suffices to provide the sufficient condition for the exact recovery of a random sign sparse vector with Bernoulli parameter $2\rho$. First, we provide sufficient conditions to guarantee that $$\begin{aligned} m_2 -r \ge c_4 \| \bS_{s2}^i \|_0 \gamma^2 (\bU_{s2}^c) \log \frac{m_2}{\delta} \label{eq65}\end{aligned}$$ with high probability. Using Lemma \[lm5\] and the union bound, $$\begin{aligned} \underset{i,j}{\max} |\bU_{s2}^c (i,j)|^2 \leq 20 \frac{\log m_2}{m_2} \label{eq66}\end{aligned}$$ with probability at least $1-3m_2^{-6}$.\ Now, we find the sufficient number of randomly sampled rows, $ m_2$, to guarantee that (\[eq65\]) is satisfied with high probability. It is obvious that $m_2 < N_1$. Define $\kappa = \frac{\log N_1}{r}$. Therefore, it is sufficient to show that $$\begin{aligned} \frac{m_2}{\|\bS_{s2}^i \|_0} \ge r \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right) \label{eq67}\end{aligned}$$ whp, where $c_6=20 c_4 $. Suppose that $$\begin{aligned} \rho \leq \frac{1}{\beta r \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right)} \label{eq68}\end{aligned}$$ where $\beta$ is a real number greater than one. Define $\alpha= r \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right)$. According to (\[eq68\]) and the Chernoff Bound for Binomial random variables [@mcdiarmid1998concentration], we have $$\begin{aligned} \begin{aligned} & \mathbb{P} \left( \|\bS_{s2}^i \|_0 - \frac{m_2}{\beta \alpha} > a \right) \leq \exp \left( \frac{- a^2}{2(\frac{m_2}{\alpha\beta} + \frac{a}{3})} \right) . \end{aligned} \label{eq1n}\end{aligned}$$ If we set $a=\frac{m_2}{\alpha} \left( 1- \frac{1}{\beta} \right)$, then the inequality (\[eq67\]) is satisfied. Therefore, (\[eq1n\]) can be rewritten as $$\begin{aligned} \begin{aligned} & \mathbb{P} \left( \|\bS_{s2}^i \|_0 - \frac{m_2}{\beta \alpha} > \frac{m_2}{\alpha} \left( 1- \frac{1}{\beta} \right) \right) \\ & \leq 2 \exp \left( \frac{- m_2^2 (\beta -1)^2}{\alpha^2 \beta^2} \frac{3\alpha\beta}{2 m_2(\beta+2)} \right). \end{aligned} \label{eq2n}\end{aligned}$$ Therefore, if $$\begin{aligned} m_2 \ge \frac{ 2 r \beta (\beta - 2) \log \left( \frac{1}{\delta} \right)}{3 (\beta - 1)^2} \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right) , \label{eq3n}\end{aligned}$$ then the inequality (\[eq67\]) is satisfied with probability at least $(1-\delta)$. Accordingly, if $$\begin{aligned} \begin{aligned} &\rho \leq \frac{0.5}{r \beta \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right)} \\ &m_2 \ge \max \Bigg( \frac{ 2 r \beta (\beta - 2) \log \left( \frac{1}{\delta} \right)}{3 (\beta - 1)^2} \left( c_6 \kappa \log \frac{N_1}{\delta} +1 \right) , \\ & \quad \quad \quad \quad\quad\quad\quad\quad\quad\quad\quad\quad \sqrt[6]{\frac{3}{\delta}}, ~c_5 \left( \log \frac{N_1}{\delta} \right)^2 \Bigg) \end{aligned} \label{eq75}\end{aligned}$$ then (\[eq63\]) returns the exact sparse vector with probability at least $1-3 \delta$. The factor 0.5 in the numerator of the right hand side of the first inequality of (\[eq75\]) is due to the de-randomization technique [@lamport2] to provide the guarantee for the fixed sign case. **Proof of Theorem \[thm:main\_result\]**\ The proposed decomposition algorithm yields the exact decomposition if:\ 1. The sampled columns of the LR matrix span the the columns subspace of $\bL$. Lemma \[lm3\] provides the sufficient conditions on $m_1$ to guarantee that the columns of $\bL_{s1}$ span $\calU$ with high probability.\ 2. The program (\[eq20\]) yields the correct LR and sparse components of $\bD_{s1}$. Lemma \[lm4\] provides the sufficient conditions on $m_1$ and $\rho$ to guarantee that $\bD_{s1}$ is decomposed correctly with high probability.\ 3. The sampled rows of the LR matrix span the rows subspace of $\bL$. Since it is assumed that the CS of $\bL$ is sampled from the random orthogonal model, according to Lemma \[lm5\] $$\begin{aligned} \mathbb{P} \left(\max_{i,j} |\bU (i,j)|^2 \ge 20\frac{\log N_1}{N_1} \right) \leq 3 r N_1^{-7} \:. \label{eq32}\end{aligned}$$ Therefore, according to Lemma \[lm3\] if $$\begin{aligned} m_2 \ge r \log N_1 \max \left( c_2^{'} \log r , c_3^{'} \log \left(\frac{3}{\delta}\right) \right), \label{eq33}\end{aligned}$$ then the selected rows of the matrix $\bL$ span the rows subspace of $\bL$ with probability at least $(1-\delta - 3 r N_1^{-7})$ where $c_2^{'}$ and $c_3^{'}$ are numerical constants.\ 4. The minimization (\[eq23\]) yields the correct RS. Lemma \[lm8\] provides the sufficient conditions to ensure that (\[secondrows\]) yields the correct representation vector. In order to guarantee the performance of (\[eq23\]), we substitute $\delta$ with $\delta/N_2$ since (\[eq23\]) has to return exact representation for the columns of $\bD$. Therefore, $$\mathbb{P} \left( \text{Incorrect Decomposition} \right) \leq \delta + c_8 N_1^{-3} + \delta + 3r N_1^{-7} + 3\delta.$$ [^1]: This material is based upon work supported by the National Science Foundation under NSF grant No. CCF-1320547 and NSF CAREER Award CCF-1552497. The authors are with the Department of Electrical Engineering and Computer Science, University of Central Florida,Orlando, FL 32816 USA (e-mail: [email protected], [email protected]).

ArXiv

--- abstract: 'We analyze the properties of the edge states of the one-dimensional Kitaev model with long-range anisotropic pairing and tunneling. Tunneling and pairing are assumed to decay algebraically with exponents $\alpha$ and $\beta$, respectively, and $\alpha,\beta>1$. We determine analytically the decay of the edges modes. We show that the decay is exponential for $\alpha=\beta$ and when the coefficients scaling tunneling and pairing terms are equal. Otherwise, the decay is exponential at sufficiently short distances and then algebraic at the asymptotics. We show that the exponent of the algebraic tail is determined by the smallest exponent between $\alpha$ and $\beta$. Our predictions are in agreement with numerical results found by exact diagonalization and in the literature.' author: - 'Simon B. Jäger' - 'Luca Dell’Anna' - Giovanna Morigi title: 'Edge states of the long-range Kitaev chain: an analytical study' --- Introduction ============ Topological phases of matter have been attracting great interest since the discovery of the integer and fractional quantum Hall effect [@Klitzing:1980; @Fractional:Hall:Effect] and of topological phase transitions [@NobelpriceTop:2016]. Topological superconductors [@Hasan:2010; @Bernegiv:2013], in particular, offer promising perspectives for realizing robust quantum devices [@Kitaev2008; @Nayak2008; @Stern:2010] due to the presence of topologically protected states, the so-called edge modes [@Kitaev2001]. In this context, the Kitaev chain is a theoretical model which exhibits topological order [@Kitaev2001] and can be mapped the Ising model with nearest-neighbor interactions [@Greiter:2014]. For open boundaries and in the topological non-trivial phase the Kitaev model supports the existence of zero energy excitations, the Majorana edge modes [@Kitaev2001]. These modes are spatially localized at the chain’s edge and are an indicator for the non-trivial topological nature of this phase [@Ryu:2002]. An extension of the Kitaev model has been recently discussed which describes algebraic decay of the tunneling and/or pairing terms [@Pientka2013; @Klinovaja2013; @Pientka2014; @Neupert2016; @Vodola2014; @Vodola2016; @Alecce2017]. This model is expected to describe experimental realizations of long-range topological superconductors [@Nadj-Perge2014; @Pawlak2016; @Ruby2017]. It has been shown that the long-range interactions leads to a modification of the phase diagram [@Vodola2014; @Vodola2016; @Alecce2017]. Moreover, numerical studies of this model revealed algebraically localized edge states and an algebraic closing of the energy gap [@Vodola2014; @Vodola2016]. When the pairing and tunneling terms are isotropic, instead, exponential localization is recovered independently of the power law exponent, as long as it is larger than unity [@Vodola2016]. In this paper we perform an analytical study of the spatial localization of the Majorana edge states in the long-range Kitaev models. Our analysis includes anisotropic and isotropic pairing and tunneling terms with the same or different algebraic exponents. For this model, we determine analytically the asymptotic scaling of the edge modes’ tails. Our findings are supported by numerical calculations that are in good agreement with the numerical results reported in the literature for specific parameter choices [@Vodola2014; @Vodola2016; @Alecce2017; @Viyuela2015; @Viyuela:2018fpv]. This paper is structured as follows. In Sec. \[sec:1\] we introduce the long-range Kitaev model and the Majorana operators. Here, we discuss the formalism which is the starting point of our analysis. In Sec. \[sec:2\] the basic equation determining the edge states is derived, which allows us to determine their site occupation inside the bulk. We determine a general expression that gives the behavior of the zero eigenmodes away from the edges as a function of the parameters of the model. In section \[sec:3\] we compare our analytical findings with numerical results and in Sec. \[sec:conclusions\] we draw the conclusions. Long-range Kitaev chain {#sec:1} ======================= We consider $N$ polarized Fermions on a lattice with open boundary conditions. Their Hamiltonian has the form of a Kitaev model with long-range interactions: $$\begin{aligned} \hat{H}=-\sum_{n=1}^N\left[\sum_{r=1}^{N-n}\left(j_r^{\alpha}\hat{c}_n^{\dag}\hat{c}_{n+r}+\Delta_r^{\beta}\hat{c}_n^{\dag}\hat{c}_{n+r}^{\dag}+\mathrm{H.c.}\right)+\mu\hat{c}_n^{\dag}\hat{c}_n\right]\,,\label{LRK}\end{aligned}$$ where $\hat{c}_i$ and $\hat{c}_i^{\dag}$ are the Fermionic annihilation and creation operators that fulfill the anticommutation relations $\{\hat{c}_i,\hat{c}_j\}=\{\hat{c}_i^{\dag},\hat{c}_j^{\dag}\}=0$, ${\{\hat{c}_i^{\dag},\hat{c}_j\}=\delta_{i,j}}$, and $\delta_{i,j}$ is the Kronecker delta. Here, $\mu$ is the chemical potential. The other coefficients $j_r^{\alpha}$ and $\Delta_r^{\beta}$ scale the tunneling and pairing terms between two sites at distance $r$. They depend on the distance according to the power law decay given by: $$\begin{aligned} &&j_r^{\alpha}=\frac{J}{N_\alpha}\frac{1}{r^{\alpha}},\\ &&\Delta_r^{\beta}=\frac{\Delta}{N_\beta}\frac{1}{r^{\beta}}\,,\end{aligned}$$ with the exponent $\alpha,\beta>1$. We denote the parameters $J$ and $\Delta$ by tunneling rate and the pairing strength, respectively. The coefficient $N_{\gamma}=\sum_{r=1}^{N}r^{-\gamma}$ warrants normalization. For $\gamma=\alpha,\beta>1$, which is the case we consider, $N_{\gamma}\to\zeta(\gamma)$ for $N\to\infty$ and $\zeta(\gamma)$ is the Riemann zeta function [@Olver:2010]. For sufficiently fast decaying interaction and hopping terms the system possesses two different phases separated by the quantum critical point $\mu_c=2J$ [@Kitaev2001]. In the thermodynamic limit the two topological phases can be distinguished by the bulk topological invariant $w$: For $|\mu|>\mu_{c}$ the ground state is non-degenerate and $w=0$; in the nontrivial phase $|\mu|<\mu_{c}$ the bulk topological invariant is $w=1$, the ground state is doubly degenerate, and can support Majorana edge modes. At finite size $N$ the spectrum is always gapped. In this work we analyze the spatial localization of the Majorana edge modes for a chain with open boundaries and as a function of the exponents $\alpha$ and $\beta$. Bogoliubov-de Gennes Hamiltonian -------------------------------- The Hamiltonian in Eq.  is quadratic and can be cast into a compact form by introducing the $2N$ component vector operator $\hat{\bf C}=(\hat c_1,\hat c_1^\dagger,\ldots,\hat c_N, \hat c_N^\dagger)^T$: $$\begin{aligned} \label{H:C} \hat{H}=\hat{\bf C}^{\dag}\hat{H}_{\mathrm{BdG}}\hat{\bf C}\,\end{aligned}$$ where $\hat H_{\mathrm{BdG}}$ is the Bogoliubov-de Gennes Hamiltonian, and is a $2N\times 2N$ matrix. In order to give its form in a compact way, first we write the vector-operator $\hat{\bf C}$ as $$\begin{aligned} \label{C:n} \hat{\bf C}=\sum_{n=1}^N\left[|1,n\rangle\hat{c}_n+|0,n\rangle\hat{c}_n^{\dag}\right],\end{aligned}$$ where ${|1,n\rangle\hat{c}_n+|0,n\rangle\hat{c}_n^{\dag}=(0,\ldots,0,\hat c_n,\hat c_n^\dagger,0,\ldots,0)^T}$ is a $2N$ component vector whose elements are all zeroes except for the $2n$ and the $2n+1$ components. Using this definition, the Bogoliubov-de Gennes Hamiltonian can be written as $$\begin{aligned} \label{H:BdG} \hat{H}_{\mathrm{BdG}}=-\mu\,\hat{\tau}_z\otimes \mathbb{1}_N-\hat{\tau}_z\otimes \hat{J}^{\alpha}-\hat{\tau}_y\otimes \hat{\Delta}^{\beta}\,.\end{aligned}$$ In Eq.  we have introduced the $N\times N$ matrices $$\begin{aligned} &&\mathbb{1}_N=\sum_{n=1}^N|n\rangle\langle n|\,,\\ &&\hat{J}^\alpha=\sum_{n=1}^{N} \sum_{r=1}^{N-n}j^{\alpha}_r\left(|n\rangle\langle n+r|+|n+r\rangle\langle n|\right)\,,\label{J}\\ &&\hat{\Delta}^\beta=\sum_{n=1}^{N} \sum_{r=1}^{N-n}i\Delta^{\beta}_r\left(|n\rangle\langle n+r|-|n+r\rangle\langle n|\right)\label{D}\,,\end{aligned}$$ and the Pauli matrices $$\begin{aligned} \hat{\tau}_x=&\frac{|1\rangle\langle 0|+|0\rangle\langle 1|}{2}\,,\\ \hat{\tau}_y=&\frac{|1\rangle\langle 0|-|0\rangle\langle 1|}{2i}\,,\\ \hat{\tau}_z=&\frac{|1\rangle\langle 1|-|0\rangle\langle 0|}{2}\,,\end{aligned}$$ with $[\hat{\tau}_l,\hat{\tau}_m]=i\sum_{n}\epsilon_{lmn}\tau_n$, where $\epsilon_{lmn}$ is the Levi-Civita tensor and the indices are $n,m,l\in\{x,y,z\}$. Majorana Fermions ----------------- In this work we are interested in the properties of Majorana edge states. Majorana edge states are eigenstates of Hamiltonian , their eigenvalue vanishes in the thermodynamic limit $N\to\infty$. This property becomes evident when the Hamiltonian in presented in terms of the Majorana operators. The Majorana operators are hermitian Fermionic operators, are here denoted by the operators $\hat{\gamma}_j$ and $\hat{\gamma}_{j+N}$ ($j=1,2,\dots N)$ and are defined by the relations $$\begin{aligned} &\hat{\gamma}_j=\frac{\hat{c}_j+\hat{c}_{j}^\dagger}{\sqrt{2}},\\ &\hat{\gamma}_{N+j}=i\frac{\hat{c}_j^\dagger-\hat{c}_{j}}{\sqrt{2}}\,,\\\end{aligned}$$ with $\{\hat{\gamma}_i,\hat{\gamma}_j\}=\delta_{i,j}$. For convenience, we define the Majorana vector operator $$\begin{aligned} \label{def:gamma} \hat{\boldsymbol{\gamma}}=\sum_{n=1}^N\left[|1,n\rangle\hat{\gamma}_n+|0,n\rangle\hat{\gamma}_{n+N}\right].\end{aligned}$$ The Majorana vector operator is connected to $\hat{\bf C}$ by the relation $$\begin{aligned} \label{C:gamma} \hat{\bf C}=&\frac{1}{\sqrt{2}}\left[\begin{pmatrix} 1 &i\\ 1& -i \end{pmatrix}\otimes\mathbb{1}_N\right]\hat{\boldsymbol{\gamma}}\equiv A\hat\gamma\,,\end{aligned}$$ where the last equality defines the matrix $A=A_2\otimes \mathbb{1}_N$ connecting the two representations, with $$\begin{aligned} A_2=\frac{1}{\sqrt{2}}\begin{pmatrix} 1 &i\\ 1& -i \end{pmatrix}\,.\end{aligned}$$ Using Eq.  we rewrite the Hamiltonian, Eq. , as $$\hat{H}=\hat{\boldsymbol{\gamma}}^{\dag}\hat{H}_{M}\hat{\boldsymbol{\gamma}}\,,$$ where $$\begin{aligned} \hat{H}_{M}&=A^\dagger H_{\rm BdG}A\nonumber\\ &=\mu\hat{\tau}_y\otimes\mathbb{1}_N+\hat{\tau}_y\otimes \hat{J}^{\alpha}+\hat{\tau}_x\otimes \hat{\Delta}^{\beta}\,.\label{H:M}\end{aligned}$$ To obtain this expression we have used that $A_2^\dagger\hat{\tau}_xA_2=\hat\tau_z$, $A_2^\dagger\hat{\tau}_yA_2=-\hat\tau_x$, and $A_2^\dagger\hat{\tau}_zA_2=-\hat\tau_y$. For later convenience we further elaborate on this notation. We define the generalized lowering operators by $$\begin{aligned} \hat{J}^{\alpha}_{-}=&\sum_{n=1}^N\sum_{r=1}^{N-n}j_r^{\alpha}|n\rangle\langle n+r|,\\ \hat{\Delta}^{\beta}_{-}=&\sum_{n=1}^N\sum_{r=1}^{N-n}\Delta_r^{\beta}|n\rangle\langle n+r|\,.\end{aligned}$$ The raising operators are the hermitian conjugate: ${\hat{J}_{+}^{\alpha}=(\hat{J}_{-}^{\alpha})^{\dag}}$, ${\hat{\Delta}_{+}^{\beta}=(\hat{\Delta}_{-}^{\beta})^{\dag}}$. Using these definitions, we rewrite Eq.  and Eq.  as $$\begin{aligned} \hat{J}^{\alpha}=\hat{J}_{+}^{\alpha}+\hat{J}_{-}^{\alpha},\end{aligned}$$ and $$\begin{aligned} \hat{\Delta}^{\beta}=-i\hat{\Delta}_{+}^{\beta}+i\hat{\Delta}_{-}^{\beta},\end{aligned}$$ respectively. The Bogoliubov-de Gennes Hamiltonian for the Majorana vector operators takes then the form $$\begin{aligned} \hat{H}_{M}=\mu\,\hat{\tau}_y\otimes\mathbb{1}_N&+i\left(\hat{\tau}_x\otimes \hat{\Delta}^{\beta}_{-}-i\hat{\tau}_y\otimes \hat{J}_{-}^{\alpha}\right)\nonumber\\ &-i\left(\hat{\tau}_x\otimes \hat{\Delta}^{\beta}_{+}+i\hat{\tau}_y\otimes \hat{J}_{+}^{\alpha}\right)\,.\label{HM}\end{aligned}$$ This is the Hamiltonian form at the basis of our treatment. Majorana edge states in the isotropic Kitaev chain with vanishing chemical potential ------------------------------------------------------------------------------------ We now consider first the case of the isotropic Kitaev chain with $\mu=0$. We thus set $\Delta=J$ in Eq.  and take $\alpha=\beta$, namely, tunneling and pairing decay with the same power-law exponent. The corresponding Hamiltonian reads $$\begin{aligned} \hat{H}_0&=i\left(\hat{\tau}_-\otimes \hat{J}^{\alpha}_{-}\right)-i\left(\hat{\tau}_+\otimes \hat{J}^{\alpha}_{+}\right),\label{H0}\end{aligned}$$ where $\hat{\tau}_{-}=\hat{\tau}_x-i\hat{\tau}_y$ and $\hat{\tau}_{+}=\hat{\tau}_-^{\dag}$ are the spin lowering and raising operators, respectively. The Majorana edge states are the eigenstates of $\hat{H}_0$ with zero eigenenergy: $$\begin{aligned} |e_1\rangle=&|1,1\rangle,\\ |e_2\rangle=&|0,N\rangle\,.\end{aligned}$$ They are the exact eigenstates for any power-law exponent $\alpha$. Independently of the exponent, they are localized at edge of the Kitaev chain, at site $n=1$ or $n=N$, respectively. From definition they correspond to the Majorana modes $\hat{\gamma}_1$ and $\hat{\gamma}_{2N}$. These operators commute with the Hamiltonian $\hat{\boldsymbol{\gamma}}^{\dag}\hat{H}_{0}\hat{\boldsymbol{\gamma}}$. These edge states still exist for finite but small $\mu$ and even for the anisotropic Kitaev chains. The analysis of the properties of the edge states in the anisotropic Kitaev chain is the aim of the next section. Edge modes in the anisotropic Kitaev chain {#sec:2} ========================================== In this section we discuss the properties of the edge modes when the Kitaev chain is anisotropic, namely, when the exponent of the power law decay are not equal, $\alpha\neq \beta$, and/or the coefficients of tunneling and pairing differ, $J\neq \Delta$. We thus seek for eigenmodes $\hat\gamma_E$ that commute with $\hat{H}$, Eq. , in the thermodynamic limit $N\to\infty$: $$\begin{aligned} \label{Eq:Edge} \left[\hat{H},\hat\gamma_E\right]=0\,.\end{aligned}$$ We write the eigenmodes as a linear superposition of the modes $\hat\gamma_j$ with scalar coefficients $h_j$: $$\hat\gamma_E=\sum_{\ell=1}^{2N}h_\ell\,\hat\gamma_\ell={\bf h}^T\hat{\boldsymbol{\gamma}}=\hat{\boldsymbol{\gamma}}^T{\bf h}\,,$$ where we have introduced the vector ${{\bf h}=(h_1,h_2,...,h_{2N})^T}$. Edge modes are characterized by coeffcients $h_\ell$ whose modulus is expected to be maximum close to the edges, which are here given by $\ell=1$ and $\ell=2N$, and which shall decrease with the distance from the closer edge. From Eq.  we can identify some general requirements that these coefficients shall fulfill. Since $\hat{H}$ is quadratic the commutator on the left-hand-side of Eq.  can be rewritten as $$\begin{aligned} \left[\hat{H},{\bf h}^T\hat{\boldsymbol{\gamma}}\right]=-\left[{\bf h}^T\hat{H}_{M}\hat{\boldsymbol{\gamma}}-\hat{\boldsymbol{\gamma}}^T\hat{H}_{M}{\bf h}\right]=2\hat{\boldsymbol{\gamma}}^T\hat{H}_{M}{\bf h}\,,\end{aligned}$$ where in the last equality we have used that $\hat{H}_{M}^T=-\hat{H}_{M}$. Therefore, Eq.  takes the form of the eigenvalue equation $$\begin{aligned} \label{Eq:Edge:h} \hat{H}_{M}{\bf h}=0\,.\end{aligned}$$ For the remainder of this section we want to present how we intend to derive ${\bf h}$. Eigenvalue equation ------------------- Let us first formulate the eigenvalue problem $$\begin{aligned} \label{Eq:Edge:h:2} \hat{H}_{M}|v\rangle=E|v\rangle\,,\end{aligned}$$ where we use Dirac’s notation of operator for matrices and ket (bra) vectors for right (left) eigenvectors. We define the Hamiltonian $$\begin{aligned} \hat{H}_1=\hat{H}_M-\hat{H}_0\,,\end{aligned}$$ such that $\hat{H}_M=\hat{H}_0+\hat{H}_1$. For $\hat{H}_1=0$ eigenstates of $\hat{H}_{M}$ at $E=0$ are the vectors $|v_0\rangle =a|e_1\rangle+b|e_2\rangle$, with $a$ and $b$ scalars. In particular, vectors $|e_1\rangle$ and $|e_2\rangle$ form a basis for the kernel of $\hat{H}_0$ for $J\neq0$. We now consider the perturbation and search for eigenstates with $E\approx 0$, which are localized at the edges. For this purpose we introduce the projector $\hat{\mathcal P}$ into the subspace spanned by $|e_1\rangle$ and $|e_2\rangle$, and the projector $\hat{\mathcal Q}=\mathbb{1}-\hat{\mathcal P}$ onto the orthogonal complement of the kernel of $\hat{H}_0$, such that $\hat{\mathcal Q}+\hat{\mathcal P}={\mathbb{1}}$ is the unity matrix $\mathbb{1}$. We note that $\hat{\mathcal P}$ projects on the edges, while $\hat{\mathcal Q}$ contains the chain’s bulk. We now make some general considerations. By taking the projection of Eq.  on the subspace corresponding to $\hat{\mathcal P}$ and using that $\hat{\mathcal P}\hat H_0=\hat H_0\hat{\mathcal P}=0$ we obtain the relation $$\begin{aligned} \hat{\mathcal P}(\hat H_1-E)|v\rangle=0\,,\label{generalizedEigenvalue}\end{aligned}$$ We note that, since we expect $E\approx 0$, then the scalar product $|\langle v_0|\hat H_1|v\rangle|\ll 1$, thus the states we search for have small overlap with the bulk. We further find the structure of $|v\rangle$ by considering that Eq.  can be rewritten as $$\begin{aligned} E|v\rangle=&\hat H_M|v\rangle\\ =&\left[\hat{\mathcal Q}\hat H_0+\hat{\mathcal Q}(\hat{H}_1-E \mathbb{1})\right]|v\rangle+[\hat{\mathcal Q}E+\hat{\mathcal P}\hat{H}_1]|v\rangle\\ =&\hat{\mathcal Q}\hat H_0\left[\mathbb{1}+(\hat{\mathcal Q}\hat{H}_0)^{-1}\hat{\mathcal Q}(\hat{H}_1-E\mathbb{1})\right]|v\rangle\\ &+[E+\hat{\mathcal P}(\hat{H}_1-E\mathbb{1})]|v\rangle \,.\end{aligned}$$ Using Eq. we can write the eigenstates as a function of $|v_0\rangle$ such that $$\begin{aligned} |v\rangle=\frac{1}{\mathbb{1}+(\hat{\mathcal{Q}}\hat{H}_0)^{-1}\hat{\mathcal{Q}}(\hat{H}_1-E\mathbb{1})}|v_0\rangle\,,\label{v}\end{aligned}$$ which ensures that $\big[\mathbb{1}+(\hat{\mathcal Q}\hat{H}_0)^{-1}\hat{\mathcal Q}(\hat{H}_1-E\mathbb{1})\big]|v\rangle=|v_0\rangle$. Substituting Eq. in Eq. , we can now cast the problem into finding the kernel of operator $\hat \Gamma$: $$\begin{aligned} \hat \Gamma=\hat{\mathcal{P}}(\hat{H}_1-E)\frac{1}{\mathbb{1}+(\hat{\mathcal{Q}}\hat{H}_0)^{-1}\hat{\mathcal{Q}}(\hat{H}_1-E\mathbb{1})}\hat{\mathcal{P}}\,,\label{Gamma}\end{aligned}$$ namely, we shall find the eigenvalues $E$ for which $\hat \Gamma$ has a non-trivial kernel. In particular, for edge state we should obtain $E=0$ in the thermodynamic limit $N\to\infty$. Let us now use the explicit forms of $\hat{H}_0$ and $\hat{H}_1$ to calculate the matrix ${\big[\mathbb{1}+(\hat{\mathcal{Q}}\hat{H}_0)^{-1}\hat{\mathcal{Q}}(H_1-E\mathbb{1})\big]^{-1}}$. This can be performed after considering that $\hat{\mathcal{Q}}$ projects on the bulk of the chain. Therefore, in the thermodynamic limit we can apply the Fourier transform to derive the spectrum and the eigenvectors of $\hat{H}_M$ and $\hat{H}_1$. For this purpose we define the $k$ vectors $$\begin{aligned} |k\rangle=\frac{1}{\sqrt{N}}\sum_{n=1}^Ne^{i(n-1)k}|n\rangle,\end{aligned}$$ with $k=2\pi m/N$ and $m=0,1,\dots,N-1$. Then, $$\begin{aligned} \hat{\mathcal{Q}}\hat{H}_{0}=\sum_{k}\begin{pmatrix} 0&\frac{-iJ}{\zeta(\alpha)}\mathrm{Li}_\alpha(e^{-ik})\\ \frac{iJ}{\zeta(\alpha)}\mathrm{Li}_\alpha(e^{ik})&0 \end{pmatrix}\otimes|k\rangle\langle k|\end{aligned}$$ and $$\begin{aligned} \hat{\mathcal{Q}}\hat{H}_{M}=\sum_{k}\begin{pmatrix} 0&-iF(-k)\\ iF(k)&0 \end{pmatrix}\otimes|k\rangle\langle k|\,,\end{aligned}$$ with $$\begin{aligned} F(k)=\frac{\mu}{2}&+\frac{\Delta}{2\zeta(\beta)}\left[\mathrm{Li}_\beta(e^{ik})-\mathrm{Li}_\beta(e^{-ik})\right]\nonumber\\ &+\frac{J}{2\zeta(\alpha)}\left[\mathrm{Li}_\alpha(e^{ik})+\mathrm{Li}_\alpha(e^{-ik})\right]\,.\label{Fk}\end{aligned}$$ Here, $\mathrm{Li}_{\gamma}(z)=\sum_{r=1}^\infty z^r/r^\gamma$ denotes the Polylogarithm and $|z|\leq1$ [@Olver:2010]. Now the problem of finding the inverse in Eq.  reduces to invert $2\times 2$ matrices. This can be solved analytically and gives, for $E=0$, $$\begin{aligned} \frac{1}{\mathbb{1}+(\hat{\mathcal{Q}}\hat{H}_0)^{-1}\mathcal{Q}H_1}=&\sum_k\begin{pmatrix} {\mathcal M}(k)&0\\ 0&{\mathcal M}(-k) \end{pmatrix}\otimes |k\rangle\langle k|\,,\label{M}\end{aligned}$$ with $$\begin{aligned} {\mathcal M}(k)=\frac{J}{\zeta(\alpha)}\,\frac{\mathrm{Li}_\alpha(e^{ik})}{F(k)}\,.\label{Mk}\end{aligned}$$ Later on we will check the consistency of the assumption that the eigenvectors have vanishing energy by estimating the scaling of $E$ with the chain’s size. Localization of the edge modes along the chain ---------------------------------------------- We can now determine the probability amplitudes $h_{\ell}$ for the eigenstates at zero eigenvalue. This reduces to calculate the projection of the edge state that we will call $|e\rangle$, which is $|v\rangle$ given by Eq.  at $E=0$, onto the bulk position $|n\rangle\equiv |1,n\rangle$. We take $|v_0\rangle=|e_1\rangle\equiv |1,1\rangle$ and calculate the amplitude $$\begin{aligned} \langle n| e\rangle= A(n) $$ where $$\begin{aligned} \label{Cn} A(n)=\frac{1}{2\pi}\int_{0}^{2\pi}dk\,e^{i(n-1)k}{\mathcal M}(k)\,.\end{aligned}$$ The function $|A(n)|^2$ gives the spatial occupation of the chain site $n$ of the eigenmode at zero eigenvalue. By construction, for $\hat{H}_M=\hat{H}_0$ it reduces to the edge state $|e_1\rangle$, namely, $A(n)=\delta_{n,1}$. As it has been used here, the probability amplitude $A(n)$ can be found taking the continuum limit, when the summation $\sum_{k}$ goes over to a continuum of $k$-values in the interval $[0,2\pi)$. Using the residue theorem it can be rewritten as the sum of the integral along the contour, $I(n)$, and of the residues, $R(n)$, as it follows $$\begin{aligned} A(n)= R(n)+I(n). \label{Cnequ}\end{aligned}$$ The integral is taken along the contour illustrated in Fig. \[Fig:contour\] and reads $$\begin{aligned} I(n)=\lim_{\eta\rightarrow 0^+}\left[\int_{\mathcal{C}_\eta}+ \int_{\mathcal{C}_{2\pi-\eta}}+ \int_{\mathcal{C}_M}\right] dk\,\frac{e^{i(n-1)k}}{2\pi}{\mathcal M}(k)\,,\label{In}\end{aligned}$$ where the interval of integration along the imaginary axis and the real axis are $[0,M]$ ($M>0$) and $[\eta,2\pi-\eta]$, respectively. The individual paths are $\mathcal{C}_{\eta}(y)=\eta+iy$ with $y\in[0,M]$, $\mathcal{C}_{M}(y)=y+iM$ with $y\in[\eta,2\pi-\eta]$, and $\mathcal{C}_{2\pi-\eta}(y)=2\pi-\eta+i(M-y)$ with $y\in[0,M]$. The summation over the residues in Eq.  goes over the complex numbers $k_0$ with ${0<\mathrm{Re}(k_0)<2\pi}$ and ${\mathrm{Im}(k_0)>0}$ for which $\mathrm{Res}[e^{i(n-1)k}{\mathcal M}(k),k_0]$ does not vanish and reads $$\begin{aligned} R(n)=i\hspace{-0.15cm}\sum_{k_0:\mathrm{Im}(k_0)>0}\mathrm{Res}\big[e^{i(n-1)k}{\mathcal M}(k),k_0\big]\,.\label{Rn}\end{aligned}$$ Both residues and integral contributions determine the behavior of $|e\rangle$ at the chain bulk. Below we argue that the residues contributes to $A(n)$ with an exponential decay (see Eq. ), while the integral term is different from zero only in the anisotropic Kitaev chain. In this case its contribution is an algebraic decay with the smallest exponent, either $\alpha$ or $\beta$ (see Eq. ). ![Sketch of the contour integration . Here, the blue line indicates the integration path $\mathcal{C}(y)=y$ along the real axis ($y\in[\eta,2\pi-\eta]$), the red line shows the contour. Here, $\mathcal{C}_{\eta}(y)=\eta+iy$ with $y\in[0,M]$, $\mathcal{C}_{M}(y)=y+iM$ with $y\in[\eta,2\pi-\eta]$, and $\mathcal{C}_{2\pi-\eta}(y)=2\pi-\eta+i(M-y)$ with $y\in[0,M]$. \[Fig:contour\]](Contour.pdf){width="0.70\linewidth"} ### Residues By inspection of Eq.  we observe that the only values where $\mathrm{Res}[e^{i(n-1)k}{\mathcal M}(k),k_0]$ does not vanish are the zeros of $F(k)$, Eq. . Therefore we search for the values $k_0$ such that $F(k_0)=0$. Within the area of the contour ${\mathcal M}(k)$ is a meromorphic function and we can take its Laurent series about the root $k_0$ of $F(k)$: $$\begin{aligned} {\mathcal M}(k)=\sum_{\ell=-\infty}^{\infty}a_{\ell}(k_0)(k-k_0)^{\ell}\,,\label{Laurent}\end{aligned}$$ where the $a_{\ell}$ are the coefficients. Moreover, since ${\mathcal M}(k)$ is meromorphic there exists a finite and positive index $L_{k_0}$ such that $a_{\ell}(k_0)=0$ for $\ell<-L_{k_0}$. This index determines the order of the pole of ${\mathcal M}(k)$ at $k_0$. Using Eq.  the residues at $k_0$ can be expressed as $$\begin{aligned} \label{R:fig} \mathrm{Res}\big[e^{i(n-1)k}{\mathcal M}(k),k_0\big]=e^{i(n-1)k_0}P_{k_0}(n)\,,\end{aligned}$$ where $P_{k_0}(n)$ is a polynomial in $n$ which depends on the coefficients of the Laurent expansion as $$\begin{aligned} P_{k_0}(n)=\sum_{\ell=0}^{L_{k_0}-1}\frac{[i(n-1)]^{\ell}}{\ell!}a_{-1-\ell}(k_0),\end{aligned}$$ using the expansion of $e^{i(n-1)(k-k_0)}$ close to $k_0$. Since all poles are isolated, then the behavior of Eq.  in the bulk is dominated by the residue at the point $k_0'$ with the smallest imaginary part, namely ${\rm Im}(k_0')=\xi$ is such that ${\xi\le {\rm Im}(k_0)}$ for all $k_0$. We distinguish two cases, when $\xi>0$ and when instead $\xi=0$. For $\xi>0$ then for $n\gg 1$ the sum over the residues Eq.  behaves as $$\begin{aligned} \label{R:exp} R(n) \sim e^{-\xi n}\tilde P_{k_0'}(n)\,,\end{aligned}$$ where $\tilde P_{k_0'}(n)= iP_{k_0'}(n)e^{i(n-1){\rm Re}(k_0')}$. If $k_0'$ is a pole of order one, then the polynomial $P_{k_0'}(n)$ is simply a constant independent on $n$. When $\xi=0$, a pole lies on the real axis. This occurs at the critical point, where there is no localized mode. This treatment allows us to determine the critical values. To derive the phase diagram it is sufficient to find the zeros of $F(k)$ in Eq.  which lie on the real axis in the interval $k\in[0,2\pi]$. The gap closes at $k=0$ and $k=\pi$. For $k=\pi$ we obtain the threshold $$\begin{aligned} \mu=2J\left(1-2^{1-\alpha}\right).\end{aligned}$$ For $k=0$ we obtain $$\begin{aligned} \mu=-2J,\end{aligned}$$ therefore the topological phase is bounded as follows $$\begin{aligned} -2J\leq\mu\leq 2J\left(1-2^{1-\alpha}\right), \label{Boundaries}\end{aligned}$$ in agreement with Ref. [@Alecce2017]. Note that the boundaries are independent of the power-law exponent of the pairing term, and, for $\alpha\to\infty$, one recovers the standard topological phase, in agreement with the findings in Ref. [@Vodola2014; @Vodola2016; @Alecce2017]. When $\mu$ is inside this interval, the residues contributes with a function which exponentially decays away from the edges. For any $\eta>0$, $k=0$ is excluded, therefore a good candidate for $k_0^\prime$, the pole which contributes more to the sum of residues, can be searched above $k=\pi$. ### Integrals We will now extract the behavior of the integrals in Eq. . For this purpose we use that $\int_{\mathcal{C}_M}dke^{i(n-1)k}{\mathcal M}(k)$ vanishes in the limit $M\to\infty$. In this limit the integral to solve is $$\begin{aligned} I(n)&=-\frac{1}{\pi}\int_{0}^\infty e^{-y(n-1)}{\rm Im}({\mathcal M}(iy))\,.\label{Integral}\end{aligned}$$ Here, ${\mathcal M}(k)$ is given in Eq. , and its imaginary part specifically reads: $$\begin{aligned} \label{Mk:Im} \mathrm{Im}({\mathcal M}(iy))=-\frac{J}{\zeta(\alpha)}\, {\rm Im}(F(iy))\,\frac{\mathrm{Li}_\alpha(e^{-y})}{|F(iy)|^2}\,,\end{aligned}$$ with the function $F(k)$ of Eq. . In order to determine the behavior for $n\gg 1$, we expand $\mathrm{Im}({\mathcal M}(iy))$ in leading order of $y$ using the Taylor expansion of the Polylogarithm [@Olver:2010]: $$\begin{aligned} \mathrm{Li}_\gamma(e^{-y})=&\Gamma(1-\gamma)y^{\gamma-1}+\sum_{k=0}^\infty\frac{\zeta(\gamma-k)}{k!}(-y)^k,\\ \mathrm{Li}_\gamma(e^{y})=&\Gamma(1-\gamma)\cos(\pi(\gamma-1))y^{\gamma-1}+\sum_{k=0}^\infty\frac{\zeta(\gamma-k)}{k!}y^k\nonumber\\&+i\Gamma(1-\gamma)\sin(\pi(\gamma-1))y^{\gamma-1}\,.\end{aligned}$$ where the real part is here only well-defined for $\gamma\notin \mathbb{N}$, while the coefficient of the imaginary part is ${\Gamma(1-\gamma)\sin(\pi(1-\gamma))}={\pi/\Gamma(\gamma)}$. In leading order in the expansion, Eq. is given by $$\begin{aligned} \mathrm{Im}({\mathcal M}(iy))\approx\Lambda\left[\frac{J}{2\zeta(\alpha)\Gamma(\alpha)}y^{\alpha-1}-\frac{\Delta}{2\zeta(\beta)\Gamma(\beta)}y^{\beta-1}\right]\,,\end{aligned}$$ and $\Lambda=J\pi/(\mu/2+J)^2$. Substituting in Eq. we obtain $$\begin{aligned} I(n)&\approx-\left[\frac{J}{2\zeta(\alpha)}n^{-\alpha}-\frac{\Delta}{2\zeta(\beta)}n^{-\beta}\right]\frac{J}{\left(\frac{\mu}{2}+J\right)^2}\,,\label{algebraic}\end{aligned}$$ which is valid for $n\gg 1$. This expression shows that the integral vanishes for $J=\Delta$ and $\alpha=\beta$. In this case the decay of the edge state is dominated by the residues and is purely exponential. Otherwise, when $J\neq\Delta$ or $\alpha\neq\beta$ the behavior of $I(n)$ is dominated by an algebraic decay with the smallest exponent between $\alpha$ and $\beta$. In this case the contribution of the integral determines the decay of the edge mode in the bulk. This result is in agreement with the behavior reported in the specific limits [@Vodola2014; @Vodola2016; @Alecce2017]. Discussion ---------- We have analyzed the properties of the spatial dependence of the zero eigenmodes focusing on their behavior away from the edges. Our study shows that their form is determined by the properties of the integral , whose integrand is determined by the ratio between the coefficients of $\hat H_0$ and of $\hat H_M$ in Fourier space. In particular, the asymptotic scaling is determined by the properties of the coefficients of the Hamiltonian $\hat H_M$, while the properties of the coefficients of $\hat{H}_0$ enter into the scaling factors. The results we obtained did not make use of perturbation theory. Nevertheless they were derived under the assumption that the resulting eigenvector $|v\rangle$ are at zero energy, so that we could set $E=0$ in Eq. , that we called $|e\rangle$. We now verify that the result we obtain is consistent with this assumption. For this purpose we show that the eigenvector $|e\rangle$ we have found fulfills Eq.  at $E=0$. We first determine the scalar product $\langle v_0'|\hat{H}_1|e\rangle$ where $|v_0^\prime\rangle=a |e_1\rangle+b |e_2\rangle$ and $$\begin{aligned} \hat{H}_1|e\rangle=\sum_{k}i\left[F(k)-\frac{J}{\zeta(\alpha)}\mathrm{Li}_{\alpha}(e^{-ik})\right]{\mathcal M}(k)\begin{pmatrix} 0\\ 1 \end{pmatrix}\otimes|k\rangle\,,\end{aligned}$$ where we took $|v_0\rangle=|e_1\rangle$ in Eq. . For $|v_0^\prime\rangle= |e_2\rangle$ the overlap is maximal and reads $$\begin{aligned} &\langle e_2|\hat{H}_1|e\rangle\nonumber\\ &=\int_{0}^{2\pi}dk\frac{ie^{i(N-1)k}}{2\pi}\frac{\left[F(k)-\frac{J}{\zeta(\alpha)}\mathrm{Li}_{\alpha}(e^{ik})\right]\frac{J}{\zeta(\alpha)}\mathrm{Li}_{\alpha}(e^{ik})}{F(k)}.\label{Energyestimate}\end{aligned}$$ This integral can be calculated using the residues theorem as above. It is composed by the sum of an exponential scaling $\exp(-\xi N)$ and an algebraic scaling $N^{-\mathrm{min}\{\alpha,\beta\}}$, where the second contribution vanishes for $\alpha=\beta$ and $\Delta=J$. This provides an estimate how the energy scales for every finite $N$ value and shows that the assumption $E=0$ is correct in the thermodynamic limit. Numerical analysis of the edge modes {#sec:3} ==================================== In this section we investigate numerically the spatial distribution of the edge states and their energy. This is done by determining the smallest positive eigenenergy $E_{\mathrm{edge}}$ of $\hat{H}_{M}$, Eq. , and the corresponding edge state $|e\rangle$, in a chain of $N=8000$ sites. We analyze in particular the probability $$\mathcal P_n=|A(n)|^2=|\langle n|e\rangle|^2$$ that the eigenstate $|e\rangle$ occupies site $n$. Figure \[Fig:tails\] displays the probability $\mathcal P_n$ as a function of $n$ and for different values of $\Delta,J$ and $\alpha,\beta$. The behavior of subplot (a) corresponds to a parameter choice with $\Delta=J$ and $\alpha=\beta$ and displays an exponential decay of the edge mode. Subplots (b)-(d) correspond to different choices of parameters for the anisotropic Kitaev chain, where the algebraic decay at large $n$ is visible. We have fitted the numerical results at the asymptotics using the fitting function $$\mathcal P_n\simeq C \,n^{-\gamma}$$ finding $\gamma$ in good agreement with the analytical result $\gamma=2\,\mathrm{min}\{\alpha,\beta\}$. The effect of the contribution $R(n)$ of the residues in Eq.  is also visible in Fig. \[Fig:tails\]. In order to show that, we find the pole $k_0'$ of ${\mathcal M}(k)$ with the smallest imaginary part, calculate $\xi=i\textrm{Im}(k_0')$, and plot $$\mathcal {P}_n\simeq \mathcal{P}_1 e^{-2 \xi (n-1)}$$ as gray dashed line. While we observe for the isotropic case (subplot (a)) very good agreement in the full range of $n$, we see that the exponential behavior is still in good agreement for small $n$ in the anisotropic case (subplots (b)-(d)). However, for larger values of $n$, in the anisotropic case, we observe eventually that the algebraic tail becomes dominant as expected. ![Probability that the edge state occupies the site $n$, $\mathcal P_n=|\langle n|e\rangle|^2$, as a function of $n$ and for different choices of the parameters. The parameters are given on top of each subplot. The results have been evaluated by numerically diagonalizing the matrix $\hat H_M$, Eq. . The gray dashed lines show the exponential behavior $\mathcal{P}_n=\mathcal{P}_1e^{-2\xi (n-1)}$. The exponent $\xi$ is found by numerically determining the pole $k_0'$ of ${\mathcal M}(k)$ with the smallest imaginary part. The tails of the curves, in the anisotropic cases, fitted by the function $\mathcal{P}_n=Cn^{-\gamma}$, are shown by the red line in subplots (b)-(d), the interval of the fit is from $n=200$ to $n=4000$. We find $\gamma=4.03$ (b), $\gamma=2.98$ (c), and $\gamma=3.98$ (d), with a rounding error to two digits after the decimal point.\[Fig:tails\]](Fig1.pdf){width="1\linewidth"} Figure \[Fig:energy\] displays the scaling of the energy of the eigenmode $|e\rangle$ with the chain size $N$ and for different values of $\Delta,J$, and $\alpha,\beta$. Subplot \[Fig:energy\](a) is calculated for the same parameters as subplot \[Fig:tails\](a) but different sizes and displays an exponential scaling. In the anisotropic Kitaev chain, subplots (b)-(d), the scaling is algebraic, the fitting function at large $N$ gives $$\begin{aligned} E_{\mathrm{edge}} \sim C^{\prime}N^{-\gamma^{\prime}}\label{Fitenergy}\end{aligned}$$ with $\gamma^{\prime}\approx\mathrm{min}\{\alpha,\beta\}$ in good agreement with our analytical result. ![The energy $E_{\mathrm{edge}}$ of the edge state as a function of the system size $N$ and for different choices of the parameters. The parameters are given on top of each subplot. The results have been evaluated by numerically diagonalizing the matrix $\hat H_M$, Eq. . The parts of the curves fitted by the function $C^{\prime}N^{-\gamma^{\prime}}$ are shown by the red line in subplots (b)-(d), the interval of the fit goes from $N=131$ to $N=8000$. We find $\gamma^{\prime}=2.01$ (b), $\gamma^{\prime}=1.55$ (c), and $\gamma^{\prime}=1.98$ (d), with a rounding error to two digits after the decimal point.\[Fig:energy\]](Fig2.pdf){width="1\linewidth"} Conclusions {#sec:conclusions} =========== We have analytically determined the spatial decay of the edge states in the long-range Kitaev chain. The expressions we obtain are general and hold for any choice of the tunneling and pairing rates and of the exponents of the algebraic decay, as long as these are larger than unity. Our result allows to determine the boundaries of the topological non-trivial phase. Within this phase, it predicts that the edge modes are exponentially localized at the chain edges in the isotropic case, namely, when the pairing and tunneling rates are equal and decay with the same exponent. By means of this model we can extrapolate the characteristic length as a function of the parameters. This behavior agrees with the numerical results reported in Refs. [@Vodola2014; @Vodola2016]. Algebraic decay of the edge modes is instead found in the anisotropic case, when either the exponent and/or the rates of tunneling and pairing differ. In this case, the smallest exponent determines the algebraic scaling of the tails, while at shorted distances the decay is exponential. This behavior has been observed for some specific cases  [@Vodola2014; @Vodola2016; @Alecce2017]. Our result is analytical and generalizes these findings. Our approach could be generalized to higher dimensional cases, e.g. two or three dimensional systems [@Zhang:2019]. It could be extended in order to describe the out-of-equilibrium behavior following slow quenches across the critical point Ref. [@Defenu:2019]. The authors acknowledge stimulating discussions with Michael Kaicher, Nicoló Defenu, and Tilman Enss. This work has been supported by the German Research Foundation (the priority program No. 1929 GiRyd), by the European Commission (ITN ColOpt) and by the German Ministry of Education and Research (BMBF) via the QuantERA project NAQUAS. Project NAQUAS has received funding from the QuantERA ERA-NET Cofund in Quantum Technologies implemented within the European Union’s Horizon 2020 program. We acknowledge support from the DARPA and ARO grant W911NF-16-1-0576. [66]{} K. von Klitzing, Rev. Mod. Phys. [**58**]{}, 519 (1986). R. B. Laughlin, Rev. Mod. Phys. [**71**]{}, 863 (1999); H. L. Stormer, Rev. Mod. Phys. [**71**]{}, 875 (1999); D. C. Tsui, Rev. Mod. Phys. [**71**]{}, 891 (1999). J. M. Kosterlitz, Rev. Mod. Phys. [**89**]{}, 040501 (2017); F. D. M. Haldane, Rev. Mod. Phys. [**89**]{}, 040502 (2017). M. Z. Hasan and C. L. Kane, Rev. Mod. Phys. [**82**]{}, 3045 (2010). A. B. Bernevig and T. L. Hughes, [*Topological insulators and topological superconductors*]{} (Princeton University Press, Princeton, NJ, 2013). A. Y. Kitaev, Ann. Phys. [**303**]{}, 2 (2003). C. Nayak, S. H. Simon, A. Stern, M. Freedman, and S. Das Sarma, Rev. Mod. Phys. **80**, 1083 (2008). A. Stern, Nature (London) [**464**]{}, 187 (2010). A. Y. Kitaev, Physics-Uspekhi **44**, 131 (2001). M. Greiter, V. Schnells, and R. Thomale, Ann. Phys. [**351**]{}, 1026 (2014). S. Ryu and Y. Hatsugai, Phys. Rev. Lett. [**89**]{}, 077002 (2002). F. Pientka, L. I. Glazman, and F. von Oppen, Phys. Rev. B **88**, 155420 (2013). J. Klinovaja, P. Stano, A. Yazdani, and D. Loss, Phys. Rev. Lett. **111**, 186805 (2013). F. Pientka, L. I. Glazman, and F. von Oppen, Phys. Rev. B **89**, 180505(R) (2014). T. Neupert, A. Yazdani, and B. A. Bernevig, Phys. Rev. B **93**, 094508 (2016). D. Vodola, L. Lepori, E. Ercolessi, A. V. Gorshkov, and G. Pupillo, Phys. Rev. Lett. **113**, 156402 (2014). D. Vodola, L. Lepori, E. Ercolessi, and G. Pupillo, New J. Phys. **18**, 015001 (2016). A. Alecce and L. Dell’Anna, Phys. Rev. B **95**, 195160 (2017). S. Nadj-Perge, I. K. Drozdov, J. Li, H. Chen, S. Jeon, J. Seo, A. H. MacDonald, B. A. Bernevig, and A. Yazdani, Science **346**, 602 (2014). R. Pawlak, M. Kisiel, J. Klinovaja, T. Meier, S. Kawai, T. Glatzel, D. Loss, and E. Meyer, npj Quantum Inf. **2**, 16035 (2016). M. Ruby, B. W. Heinrich, Y. Peng, F. von Oppen, and K. J. Franke, Nano Lett. **17**, 4473 (2017). O. Viyuela, D. Vodola, G. Pupillo, and M. A. Martin-Delgado, Phys. Rev. B **94** (2016). O. Viyuela, L. Fu, and M. A. Martin-Delgado, Phys. Rev. Lett. **120**, 017001 (2018). F. W. J. Olver, D. W. Lozier, R. F. Boisvert, and C. W. Clark, *NIST Handbook of Mathematical Functions* (Cambridge University Press, New York, 2010). K. L. Zhang, P. Wang, and Z. Song, Sci. Rep. [**9**]{}, 4978 (2019). N. Defenu, G. Morigi, L. Dell’Anna, and Tilman Enss Phys. Rev. B [**100**]{}, 184306 (2019).

ArXiv

--- abstract: 'We present a three-dimensional reconstruction of an eruption that occurred on 3 April 2010 using observations from *SWAP* onboard *PROBA2* and *SECCHI* onboard *STEREO*. The event unfolded in two parts: an initial flow of cooler material confined to a height low in the corona, followed by a flux rope eruption higher in the corona. We conclude that mass off-loading from the first part triggered a rise, and, subsequently, catastrophic loss of equilibrium of the flux rope.' author: - 'Daniel B. Seaton, Marilena Mierla , David Berghmans, Andrei N. Zhukov ,' - Laurent Dolla title: 'SWAP-SECCHI Observations of a Mass-Loading Type Solar Eruption' --- Introduction ============ Though Coronal Mass Ejections (CMEs) have been known to the scientific community for decades, the mechanisms responsible for their initiation remain the subject of much discussion. Numerous authors have proposed models that explain the physical processes that might trigger the destabilization of structures in the low corona and cause eruptions [see the review of @Forbes2006]. In particular, a number of authors have studied the conditions under which a catastrophic loss-of-equilibrium may lead to an eruption [@PriestForbes2002]. @ForbesIsenberg1991 proposed a two-dimensional analytic model describing how a coronal filament can erupt when its magnetic energy reaches a critical value. This loss-of-equilibrium model has been expanded and improved by a series of authors [@ForbesPriest1995; @LinForbes2000; @Reeves2006] who have shown how magnetic reconnection in the wake of such an eruption can accelerate a CME and produce a solar flare. More recently, these two-dimensional, analytical models have been extended to three dimensions using numerical models by authors such as @Roussev2003, @Kliem2004, and @Schrijver2008. In fact, there is essentially universal agreement that magnetic reconnection is the primary mechanism by which the stored magnetic energy necessary to accelerate a CME and generate a solar flare is released. In most models an initial perturbation leads to the catastrophic loss-of-equilibrium of a flux rope and, therefore, an eruption. However, what mechanisms are responsible for that perturbation remains an open question. One of the reasons for this lack of consensus is that there are a number of different ways in which a CME can be triggered. Flux cancellation and magnetohydrodynamic instabilities are common proposals, but some authors have explored additional mechanisms, including the role of mass loading and unloading. @WolfsonDlamini1997 and @WolfsonSaran1998 found that excess mass loading could contribute to the stored energy necessary to lift an eruption against the Sun’s gravity. @Low1996 and several others have suggested that a reduction of mass in a prominence can cause the prominence to rise buoyantly and thus lead to an eruption. @Klimchuk2001, however, rightly argues that given the number of CMEs that occur in the absence of a massive prominence, such a mechanism can only account for some subset of all eruptions. One confounding factor in determining the cause of a CME is that there are relatively few examples where observers can clearly identify what mechanism actually initiated the event. However, in this Letter we present one such observation of an eruption: a hybrid of the mass-loading and loss-of-equilibrium types described above. The event, which occurred on 3 April 2010, triggered the CME that eventually caused the failure of the Galaxy 15 satellite. In our analysis, the event begins when a mass off-load in an underlying filament triggers the slow rise of a coronal flux rope. When the flux rope reaches a critical height, a catastrophic loss-of-equilibrium occurs and the flux rope erupts and produces a CME. Observations {#Observations} ============ Data from the *Solar TErrestrial RElations Observatory (STEREO)* make it possible to use triangulation to derive the three-dimensional location of various solar phenomena: loops [e.g. @Feng2007; @Aschwanden2008; @Rodriguez2009]; polar coronal jets [@Patsourakos2008]; solar filaments and prominences [@Gissot2008; @Bemporad2009; @Gosain2009; @Liewer2009]; and CMEs [see the review of @Mierla2010]. In order to perform three-dimensional reconstructions of the features that make up this event, we must use a pair of images taken from different perspectives. In this case, we used pairs of images taken from two different perspectives of the *STEREO* spacecraft. We used two different passbands from the *Extreme UltraViolet Imager (EUVI)* on *Sun Earth Connection Coronal and Heliospheric Investigation (SECCHI)*: 195 Å (Fe <span style="font-variant:small-caps;">xii</span> at $\log\,T \approx 6.17$) and 304 Å (He <span style="font-variant:small-caps;">ii</span> at $\log\,T \approx 4.9$). In our dataset, these pairs of *EUVI* 195 Å images from the two spacecraft were acquired with a cadence of 5 minutes, while the 304 Å images were acquired with a cadence of 10 minutes. [For a complete discussion of *SECCHI* see @Howard2008]. We also used observations from the *Sun Watcher with Active Pixels and Image Processing (SWAP)* [@Berghmans2006; @DeGroof2008; @Seaton2010], a wide-field extreme-ultraviolet (EUV) solar imager onboard the *PROBA2* spacecraft in Earth orbit. The *SWAP* observations consist of a series of images at 174 Å (Fe <span style="font-variant:small-caps;">ix/x</span> at $\log\,T \approx 6.0$) taken with cadence of roughly 100 seconds. *SWAP* images are $1024 \times 1024$ pixels with a linear pixel size of approximately $3.17 \arcsec$, meaning the instrument has a total field of view of approximately $54 \times 54\;\mathrm{arcmin}^{2}$. (*SWAP’s* field of view is similar in size to *EUVI’s*, in contrast to the higher resolution, but smaller, field of view of *AIA* on *SDO*.) In these images, we focused on a region of interest approximately $950\arcsec \times 950\arcsec$ in angular size, centered on NOAA AR 11059. Though *EUVI* can acquire images in the 171 Å passband, it observed at very low cadence in this channel on 3 April 2010 and did not acquire any images that included the eruption itself. Thus we were unable to use matched sets of images in the same passband from all three spacecraft. Figure \[fig1\] shows an overview of the event as seen by *SWAP*. The first panel shows an image of our region of interest obtained at about UT 08:00, before the eruption began. At this point, the region was dominated by a dark, S-shaped structure at the center of a series of large overlying loops. Overlaying the *SWAP* images of this dark structure on magnetograms from *GONG* revealed that it ran roughly along the magnetic neutral line of an essentially bipolar active region. The north end of the dark feature was anchored in a region of positive polarity, while the south end was anchored in a region of negative polarity. The subsequent panels of Figure \[fig1\] show running-difference images of the evolution of the region as the eruption unfolded, highlighting changes that occurred between pairs of sequential images. In the figure, each subsequent panel is separated by 500 s from the one preceding it, while the individual frames used to compose each panel are separated by 100 s. These images reveal the main features of the eruption: in the second panel we see the initial flow of mass that triggered the event, which we refer to as the mass off-load. In this and the next several frames we can also see the expansion of the overlying loops that followed the initial mass flow. By about UT 09:30 the eruption was well underway and, in the corresponding panels, we see the core of the eruption, which traveled southward through the *SWAP* field of view as the event unfolded. At the time of the event, the *STEREO* spacecraft were separated by $138.63^{\circ}$, with *STEREO-A* leading the earth by $67.43^{\circ}$ and *STEREO-B* trailing by $71.19^{\circ}$, meaning that objects near disk-center in *SWAP* appear on the east limb in images from *STEREO-A* and the west limb in images from *STEREO-B*. From these *SECCHI* images, which have approximately twice the spatial resolution of *SWAP* images, we identified the subregions corresponding to our *SWAP* region-of-interest for analysis. These subregions are shown in the top row of images in Figure \[fig2\], which we will discuss in detail in the next section. Reconstruction and Analysis {#Analysis} =========================== In order to understand the relationships between the individual features that made up this event, it is useful to know where they were in three-dimensional space. For example, *SWAP* images, because they show the eruption from almost directly above, tell us very little about how high above the solar surface any individual element was. Thus we used *SECCHI* images to make a series of three-dimensional reconstructions of the structures involved in the eruption. To make these reconstructions we used a triangulation technique described in @Inhester2006. In this approach, we treat the positions of the two spacecraft used in the reconstruction as two viewpoints or observers. The two positions of these spacecraft and the point in the solar corona to be triangulated define a plane called the epipolar plane. Since every epipolar plane is seen from a point lying inside that plane by both spacecraft, it is reduced to a single line—called the epipolar line—in the respective image projections. Any object identified to be situated on a certain epipolar line in one image must lie on the corresponding epipolar line in the other image. Finding a correspondence between pixels in each image therefore requires only matching structures that appear along the correct epipolar line in each image. Once we identify the feature in each image, we can make the three-dimensional reconstruction by tracking the lines of sight for each feature back into three-dimensional space. Since these lines of sight must lie in the same epipolar plane, their intersection defines a unique location in space. Thus we can establish the three-dimensional position of any feature that appears in both images. We used the *SolarSoft* routine **scc\_measure.pro** to reconstruct features seen in either 195 Å and 304 Å images from *SECCHI*. Once we had determined the three-dimensional locations of points within these structures, we could also determine how they should appear when seen in projection in images from all three instruments—that is, both of the *SECCHI* instruments as well as *SWAP*. We used the 195 Å images to study the structure and height of a large, faint, loop-like feature—presumably the top of the erupting flux rope—that appeared to be associated with the active region. This feature (seen at the beginning of the event) is shown in the top three panels of Figure \[fig2\]. We have plotted the projected positions of the reconstructed points as they appear from the perspective of each spacecraft’s viewpoint on top of the respective images. Using reconstructions for each of the *SECCHI* images in which this feature is visible, we were able to track its maximum height throughout the eruption. The upper panel in Figure \[fig3\] shows these measured heights as a function of time. The figure shows that the loop structure remains relatively stable until the appearance of the mass-offload in the *SECCHI* 195 Å passband around UT 08:20, when it slowly begins to rise. This rise accelerates gradually until UT 09:00, when the entire structure rapidly erupts and the top of the structure leaves the *SECCHI* field of view. Since the front of the outgoing CME appears in *COR1* coronagraph images from *SECCHI* at UT 09:10, it seems likely that this feature forms part of the CME associated with the event. We made two additional reconstructions using 304 Å *SECCHI* images as well. The second row of Figure \[fig2\] shows the first of these, of the blob of material that comprised the initial mass-offload as it travels south from AR 11059. Projections of the reconstructed points onto *SWAP* images confirm that this structure is the same as the initial blob seen in running difference images of the event (i.e., the structure visible in the third panel of Figure \[fig1\]). This blob travels southward for about one hour before it is overtaken by a larger blob of material that is directly associated with the eruption. We show a reconstruction of this second blob in the third set of panels in Figure \[fig2\]. We tracked the southward progress of both of these blobs as seen in projection from *SECCHI-A* and *SWAP*, using a combination of passbands and display techniques, and plotted their southward progress in the bottom panel of Figure \[fig3\]. Using *SWAP* images, we tracked two blobs of material that were clearly visible in running-difference movies. Using *SECCHI* images, we tracked two blobs in both the 304 Å passband and 195 Å| passband (using running-difference images in the latter case). Because *PROBA2* and *STEREO-A* do not orbit the sun in the same plane, the viewing angle for the two spacecraft differs by several degrees. To account for this difference, we applied a correction to the measured, projected positions of features seen in *SWAP* images to match them to those seen in the *SECCHI* images. After plotting the locations of the individual blobs we could detect using the different imagers and bandpasses, we identified three groups of plotted points that appeared to be related. The first, labeled “A” in Figure \[fig3\], is an initial flow of material seen from *SECCHI* that originates close to the center of the AR 11059. The second, labeled “B”, essentially a second pulse of flow that appears behind the initial one, is visible to both imagers. The third, labeled “C”, first appears further south, after the initial flow from the active region seems to have subsided. The blob associated with the points we see in group C appears to be composed mainly of the remnants of the erupting flux rope. Comparing the timing of the appearance of these three flows to the rise of the large loop-like structure in the top panel of Figure \[fig3\] revealed that flow A and the beginning of the slow rise of the loop-like structure occurred at approximately the same time. Flow B occurred just as the slowly rising loop began to accelerate, and is associated with the structures we reconstructed in the second panel of Figure \[fig2\]. Apparently both of these flows are part of the mass off-loading process that precipitated the eruption. Figure \[fig3\] also shows that flow A does not appear to be co-located in the two *SECCHI* channels we studied, while flow B appears in all three observations in essentially the same location. Apparently the first pulse consisted of mostly cool, dense material, seen predominantly in the 304 Å channel—it is only visible in running-difference images in the 195 Å channel—while the subsequent flows were hotter, and therefore more clearly visible in all three channels we studied. In particular, we note that flow B is bright in the view from *SWAP*, which is a clear indication that some heating has occurred. This heating appears to be entirely confined to the filamentary material, since we see no additional brightening in *SWAP* images for several minutes. Once the rising loop began to accelerate, it rapidly left the *EUVI* field-of-view. Minutes later, *GOES* detected the first X-ray emission from the active region, signifying the onset of the flare (this is marked in Figure \[fig3\] by a vertical line). The fact that the X-ray flare began at approximately the same time as the erupting loop structure underwent rapid acceleration is a clear indicator that the magnetic reconnection that released the energy necessary to accelerate the CME and generate flare emission [@Reeves2006] only began later in the event. Thus we conclude that reconnection is only responsible for accelerating the erupting structures, rather than triggering the event. The initial rise apparently occurred in a quasi-equilibrium state and was due to the increased buoyancy of the erupting flux rope because of the mass off-loading that occurred as a result of flows A and B. Only once this equilibrium breaks down completely does the eruption begin in earnest. Flow C, which is associated with the features we reconstructed in the bottom panel of Figure \[fig3\], first appears at a location further south than either of the first two flows and travels southward until it reaches the limb of the sun as seen by *SWAP*, at which point it effectively disappears from view. From our reconstructions, we know that this feature is generally higher in the corona than flow B, and that it does not appear to follow the same trajectory as flow B either. This latter flow appears to be associated with the remains of the erupted flux rope, rather than the initial off-loaded mass flow. Discussion ========== While models that might explain the dynamics and initiation of solar eruptions are numerous [@Klimchuk2001; @Forbes2006], observations that clearly show the mechanism of CME initiation are relatively rare. The event we present in this paper appears to be one of these rare observations: we see clear evidence that the eruption trigger is of the mass-loading type, and the subsequent evolution of the event appears to match the loss-of-equilibrium model. (Another example of such an observation, featuring a different initiation mechanism, is described in [@SterMoore2005].) One question we cannot answer is what caused the initial flow of mass from the active region. Images from both *EUVI* and *SWAP* reveal evidence of instability in the underlying filament in the hours before the eruption, particularly to the south of the active region. However, the active region itself remained essentially unchanged during this time and we see little evidence of heating or impulsive flows—as has been observed in other eruptions like the one discussed in @Bone2009—that might signal that magnetic reconnection was taking place before the large mass flow occurred. The fact that the mass flow appears bright in *SWAP* suggests that some localized heating did occur in the filament during the end of the off-load phase of the eruption, but the mechanism for this heating is impossible to identify. Additionally, neither EUV light curves from *LYRA* onboard *PROBA2* nor X-ray light curves from *GOES* show any increase in brightness in the first 30 minutes after the first mass flow occurs. If reconnection had played a significant role in lifting the flux rope during the first phase of the event, we would expect some evidence of large-scale heating to appear in these light curves. Thus, while we cannot conclude that reconnection played no part in triggering or heating the mass flow, we can say with some certainty that reconnection was not a driver of the early phase of the eruption. That this event appears to be triggered and driven by a combination of processes described by different CME initiation models [e.g. @LinForbes2000; @Low1996] is not surprising. Eruptions can begin in active regions with very different initial magnetic configurations, so it is likely that individual events will have different trigger mechanisms. This fact suggests that there is no silver-bullet model that will describe the initiation of all eruptions; a single event cannot resolve a question as long-standing and complex as what causes eruptions. We are grateful to S. Koutchmy for helpful discussions. Support for this paper came from PRODEX grant No. C90345 managed by the European Space Agency in collaboration with the Belgian Federal Science Policy Office (BELSPO) in support of the PROBA2/SWAP mission, and from the European Commission’s Seventh Framework Programme (FP7/2007-2013) under the grant agreement No. 218816 (SOTERIA project, www.soteria-space.eu). *SWAP* and *LYRA* are projects of the Centre Spatial de Liege and the Royal Observatory of Belgium funded by the Belgian Federal Science Policy Office (BELSPO). We are thankful for the anonymous referee’s thoughtful response, which helped us improve the paper. [*Facilities:*]{} , , [27]{} natexlab\#1[\#1]{} , M. J., [Nitta]{}, N. V., [Wuelser]{}, J., & [Lemen]{}, J. R. 2008, , 680, 1477 , A. 2009, , 701, 298 , D., [et al.]{} 2006, Advances in Space Research, 38, 1807 , L. A., [van Driel-Gesztelyi]{}, L., [Culhane]{}, J. L., [Aulanier]{}, G., & [Liewer]{}, P., 2009, , 259, 31 , A., [Berghmans]{}, D., [Nicula]{}, B., [Halain]{}, J., [Defise]{}, J., [Thibert]{}, T., & [Sch[ü]{}hle]{}, U. 2008, , 249, 147 , L., [Inhester]{}, B., [Solanki]{}, S. K., [Wiegelmann]{}, T., [Podlipnik]{}, B., [Howard]{}, R. A., & [Wuelser]{}, J. 2007, , 671, L205 , T. G., & [Isenberg]{}, P. A. 1991, , 373, 294 , T. G., & [Priest]{}, E. R. 1995, , 446, 377 , T. G., [et al.]{} 2006, Space Science Reviews, 123, 251 , S. F., [Hochedez]{}, J., [Chainais]{}, P., & [Antoine]{}, J. 2008, , 252, 397 , S., [Schmieder]{}, B., [Venkatakrishnan]{}, P., [Chandra]{}, R., & [Artzner]{}, G. 2009, , 259, 13 , R. A., [et al.]{} 2008, Space Science Reviews, 136, 67 , B. 2006, ArXiv Astrophysics e-prints , B., [Titov]{}, V. S., & [T[ö]{}r[ö]{}k]{}, T. 2004, , 413, L23 , J. A. 2001, Space Weather (Geophysical Monograph 125), ed. P. Song, H. Singer, G. Siscoe (Washington: Am. Geophys. Un.), 143 (2001), 125, 143 , P. C., [de Jong]{}, E. M., [Hall]{}, J. R., [Howard]{}, R. A., [Thompson]{}, W. T., [Culhane]{}, J. L., [Bone]{}, L., & [van Driel-Gesztelyi]{}, L. 2009, , 256, 57 , J., & [Forbes]{}, T. G. 2000, , 105, 2375 , B. C. 1996, , 167, 217 Mierla, M., [et al.]{} 2010, Annales Geophysicae, 28, 203 , S., [Pariat]{}, E., [Vourlidas]{}, A., [Antiochos]{}, S. K., & [Wuelser]{}, J. P. 2008, , 680, L73 , E. R., & [Forbes]{}, T. G. 2002, , 10, 313 , K. K. 2006, , 644, 592 , L., [Zhukov]{}, A. N., [Gissot]{}, S., & [Mierla]{}, M. 2009, , 256, 41 , I. I., [Forbes]{}, T. G., [Gombosi]{}, T. I., [Sokolov]{}, I. V., [DeZeeuw]{}, D. L., & [Birn]{}, J. 2003, , 588, L45 , C. J., [Elmore]{}, C., [Kliem]{}, B., [Török]{}, T., & [Title]{}, A. M., , 674, 586 , D. B., [et al.]{} 2010, in preparation , A. C., & [Moore]{}, R. L. 2005, , 630, 1148 , R., & [Dlamini]{}, B. 1997, , 483, 961 , R., & [Saran]{}, S. 1998, , 499, 496   

ArXiv

--- abstract: 'This paper is concerned with identifying contexts useful for training word representation models for different word classes such as adjectives (A), verbs (V), and nouns (N). We introduce a simple yet effective framework for an automatic selection of [*class-specific context configurations*]{}. We construct a context configuration space based on universal dependency relations between words, and efficiently search this space with an adapted beam search algorithm. In word similarity tasks for each word class, we show that our framework is both effective and efficient. Particularly, it improves the Spearman’s $\rho$ correlation with human scores on SimLex-999 over the best previously proposed class-specific contexts by 6 (A), 6 (V) and 5 (N) $\rho$ points. With our selected context configurations, we train on only 14% (A), 26.2% (V), and 33.6% (N) of all dependency-based contexts, resulting in a reduced training time. Our results generalise: we show that the configurations our algorithm learns for one English training setup outperform previously proposed context types in another training setup for English. Moreover, basing the configuration space on universal dependencies, it is possible to *transfer* the learned configurations to German and Italian. We also demonstrate improved per-class results over other context types in these two languages.' author: - | Ivan Vulić$^{\mathbf{1}}$,   Roy Schwartz$^{\mathbf{2, 3}}$,   Ari Rappoport$^{\mathbf{4}}$\ **[Roi Reichart]{}$^{\mathbf{5}}$,   [Anna Korhonen]{}$^{\mathbf{1}}$\ $^{\mathbf{1}}$ Language Technology Lab, DTAL, University of Cambridge\ $^{\mathbf{2}}$ CS & Engineering, University of Washington and $^{\mathbf{3}}$Allen Institute for AI\ $^{\mathbf{4}}$ Institute of Computer Science, The Hebrew University of Jerusalem\ $^{\mathbf{5}}$ Faculty of Industrial Engineering and Management, Technion, IIT\ `{iv250,alk23}@cam.ac.uk` `[email protected]`\ `[email protected]` `[email protected]`** bibliography: - 'acl2017\_refs.bib' title: 'Automatic Selection of Context Configurations for Improved Class-Specific Word Representations' --- Introduction {#s:intro} ============ Dense real-valued word representations (embeddings) have become ubiquitous in NLP, serving as invaluable features in a broad range of tasks [@Turian:2010acl; @Collobert:2011jmlr; @Chen:2014emnlp]. The omnipresent `word2vec` skip-gram model with negative sampling (SGNS) [@Mikolov:2013nips] is still considered a robust and effective choice for a word representation model, due to its simplicity, fast training, as well as its solid performance across semantic tasks [@Baroni:2014acl; @Levy:2015tacl]. The original SGNS implementation learns word representations from local bag-of-words contexts (BOW). However, the underlying model is equally applicable with other context types [@Levy:2014acl]. Recent work suggests that “not all contexts are created equal”. For example, reaching beyond standard BOW contexts towards contexts based on dependency parses [@Bansal:2014acl; @Melamud:2016naacl] or symmetric patterns [@Schwartz:2015conll; @Schwartz:2016naacl] yields significant improvements in learning representations for particular word classes such as [*adjectives (A)*]{} and [*verbs (V)*]{}. Moreover, demonstrated that a subset of dependency-based contexts which covers only coordination structures is particularly effective for SGNS training, both in terms of the quality of the induced representations and in the reduced training time of the model. Interestingly, they also demonstrated that despite the success with adjectives and verbs, BOW contexts are still the optimal choice when learning representations for [*nouns (N)*]{}. In this work, we propose a simple yet effective framework for selecting [*context configurations*]{}, which yields improved representations for verbs, adjectives, [*and*]{} nouns. We start with a definition of our context configuration space (Sect. \[ss:space\]). Our basic definition of a context refers to a single typed (or labeled) dependency link between words (e.g., the `amod` link or the `dobj` link). Our configuration space then naturally consists of all possible subsets of the set of labeled dependency links between words. We employ the universal dependencies (UD) scheme to make our framework applicable across languages. We then describe (Sect. \[ss:pools\]) our adapted beam search algorithm that aims to select an optimal context configuration for a given word class. We show that SGNS requires different context configurations to produce improved results for each word class. For instance, our algorithm detects that the combination of `amod` and `conj` contexts is effective for adjective representation. Moreover, some contexts that boost representation learning for one word class (e.g., `amod` contexts for adjectives) may be uninformative when learning representations for another class (e.g., `amod` for verbs). By removing such dispensable contexts, we are able both to speed up the SGNS training and to improve representation quality. We first experiment with the task of predicting similarity scores for the A/V/N portions of the benchmarking SimLex-999 evaluation set, running our algorithm in a standard SGNS experimental setup [@Levy:2015tacl]. When training SGNS with our learned context configurations it outperforms SGNS trained with the best previously proposed context type [*for each word class*]{}: the improvements in Spearman’s $\rho$ rank correlations are 6 (A), 6 (V), and 5 (N) points. We also show that by building context configurations we obtain improvements on the entire SimLex-999 (4 $\rho$ points over the best baseline). Interestingly, this context configuration is not the optimal configuration for any word class. We then demonstrate that our approach is robust by showing that transferring the optimal configurations learned in the above setup to three other setups yields improved performance. First, the above context configurations, learned with the SGNS training on the English Wikipedia corpus, have an even stronger impact on SimLex999 performance when SGNS is trained on a larger corpus. Second, the transferred configurations also result in competitive performance on the task of solving class-specific TOEFL questions. Finally, we transfer the learned context configurations across languages: these configurations improve the SGNS performance when trained with German or Italian corpora and evaluated on class-specific subsets of the multilingual SimLex-999 [@Leviant:2015arxiv], without any language-specific tuning. Related Work {#s:rw} ============ Word representation models typically train on ([*word, context*]{}) pairs. Traditionally, most models use bag-of-words (BOW) contexts, which represent a word using its neighbouring words, irrespective of the syntactic or semantic relations between them [@Collobert:2011jmlr; @Mikolov:2013nips; @Mnih:2013nips; @Pennington:2014emnlp inter alia]. Several alternative context types have been proposed, motivated by the limitations of BOW contexts, most notably their focus on topical rather than functional similarity (e.g., *coffee:cup* vs. *coffee:tea*). These include dependency contexts [@Pado:2007cl; @Levy:2014acl], pattern contexts [@Baroni:2010cogsci; @Schwartz:2015conll] and substitute vectors [@Yatbaz:2012emnlp; @Melamud:2015naacl]. Several recent studies examined the effect of context types on word representation learning. compared three context types on a set of intrinsic and extrinsic evaluation setups: BOW, dependency links, and substitute vectors. They show that the optimal type largely depends on the task at hand, with dependency-based contexts displaying strong performance on semantic similarity tasks. extended the comparison to more languages, reaching similar conclusions. , showed that symmetric patterns are useful as contexts for V and A similarity, while BOW still works best for nouns. They also indicated that coordination structures, a particular dependency link, are more useful for verbs and adjectives than the entire set of dependencies. In this work, we generalise their approach: our algorithm systematically and efficiently searches the space of dependency-based context configurations, yielding [*class-specific*]{} representations with substantial gains *for all three word classes*. Previous attempts on specialising word representations for a particular relation (e.g., similarity vs relatedness, antonyms) operate in one of two frameworks: (1) modifying the prior or the regularisation of the original training procedure [@Yu:2014acl; @Wieting:2015tacl; @Liu:2015acl; @Kiela:2015emnlpemb; @Ling:2015emnlp]; (2) post-processing procedures which use lexical knowledge to refine previously trained word vectors [@Faruqui:2015naacl; @Wieting:2015tacl; @Mrksic:2017ar]. Our work suggests that the induced representations can be specialised by directly training the word representation model with carefully selected contexts. Context Selection: Methodology {#s:metho} ============================== ![Extracting dependency-based contexts. **Top**: An example English sentence from [@Levy:2014acl], now UD-parsed. **Middle**: the same sentence in Italian, UD-parsed. Note the similarity between the two parses which suggests that our context selection framework may be extended to other languages. **Bottom**: prepositional arc collapsing. The uninformative short-range `case` arc is removed, while a “pseudo-arc” specifying the exact link (`prep:with`) between [*discovers*]{} and [*telescope*]{} is added.](./dep_graphs_en.pdf "fig:"){width="0.95\linewidth"} ![Extracting dependency-based contexts. **Top**: An example English sentence from [@Levy:2014acl], now UD-parsed. **Middle**: the same sentence in Italian, UD-parsed. Note the similarity between the two parses which suggests that our context selection framework may be extended to other languages. **Bottom**: prepositional arc collapsing. The uninformative short-range `case` arc is removed, while a “pseudo-arc” specifying the exact link (`prep:with`) between [*discovers*]{} and [*telescope*]{} is added.](./dep_graphs_it.pdf "fig:"){width="0.95\linewidth"} ![Extracting dependency-based contexts. **Top**: An example English sentence from [@Levy:2014acl], now UD-parsed. **Middle**: the same sentence in Italian, UD-parsed. Note the similarity between the two parses which suggests that our context selection framework may be extended to other languages. **Bottom**: prepositional arc collapsing. The uninformative short-range `case` arc is removed, while a “pseudo-arc” specifying the exact link (`prep:with`) between [*discovers*]{} and [*telescope*]{} is added.](./dep_graphs_arc.pdf "fig:"){width="0.95\linewidth"} \[fig:rex\] The goal of our work is to develop a methodology for the identification of optimal context configurations for word representation model training. We hope to get improved word representations and, at the same time, cut down the training time of the word representation model. Fundamentally, we are not trying to design a new word representation model, but rather to find valuable configurations for existing algorithms. The motivation to search for such training context configurations lies in the intuition that the distributional hypothesis [@Harris:1954] should not necessarily be made with respect to BOW contexts. Instead, it may be restated as a series of statements according to particular word relations. For example, the hypothesis can be restated as: “two adjectives are similar if they modify similar nouns”, which is captured by the `amod` typed dependency relation. This could also be reversed to reflect noun similarity by saying that “two nouns are similar if they are modified by similar adjectives”. In another example, “two verbs are similar if they are used as predicates of similar nominal subjects” (the `nsubj` and `nsubjpass` dependency relations). First, we have to define an expressive context configuration space that contains potential training configurations and is effectively decomposed so that useful configurations may be sought algorithmically. We can then continue by designing a search algorithm over the configuration space. Context Configuration Space {#ss:space} --------------------------- We focus on the configuration space based on dependency-based contexts (DEPS) [@Pado:2007cl; @Utt:2014tacl]. We choose this space due to multiple reasons. First, dependency structures are known to be very useful in capturing functional relations between words, even if these relations are long distance. Second, they have been proven useful in learning word embeddings [@Levy:2014acl; @Melamud:2016naacl]. Finally, owing to the recent development of the Universal Dependencies (UD) annotation scheme [@McDonald:2013acl; @Nivre:2015ud][^1] it is possible to reason over dependency structures in a multilingual manner (e.g., Fig. \[fig:rex\]). Consequently, a search algorithm in such DEPS-based configuration space can be developed for multiple languages based on the same design principles. Indeed, in this work we show that the optimal configurations for English translate to improved representations in two additional languages, German and Italian. And so, given a (UD-)parsed training corpus, for each target word $w$ with modifiers $m_1,\ldots,m_k$ and a head $h$, the word $w$ is paired with context elements $m_1\_r_1,\ldots,m_k\_r_k,h\_r_h^{-1}$, where $r$ is the type of the dependency relation between the head and the modifier (e.g., `amod`), and $r^{-1}$ denotes an inverse relation. To simplify the presentation, we adopt the assumption that all training data for the word representation model are in the form of such $(word, context)$ pairs [@Levy:2014acl; @Levy:2014nips], where $word$ is the current target word, and $context$ is its observed context (e.g., BOW, positional, dependency-based). A naive version of DEPS extracts contexts from the parsed corpus without any post-processing. Given the example from Fig. \[fig:rex\], the DEPS contexts of [*discovers*]{} are: [*scientist\_nsubj*]{}, [*stars\_dobj*]{}, [*telescope\_nmod*]{}. DEPS not only emphasises functional similarity, but also provides a natural implicit grouping of related contexts. For instance, all pairs with the shared relation $r$ and $r^{-1}$ are taken as an $r$-based [*context bag*]{}, e.g., the pairs $\{$[*(scientist, Australian\_amod)*]{}, [*(Australian, scientist\_$amod^{-1}$)*]{}$\}$ from Fig. \[fig:rex\] are inserted into the `amod` context bag, while $\{$[*(discovers, stars\_dobj)*]{}, [*(stars, discovers\_$dobj^{-1}$)*]{}$\}$ are labelled with `dobj`. Assume that we have obtained $M$ distinct dependency relations $r_1,\ldots,r_M$ after parsing and post-processing the corpus. The $j$-th [*individual context bag*]{}, $j=1,\ldots,M$, labelled $r_j$, is a bag (or a multiset) of $(word, context)$ pairs where $context$ has one of the following forms: $v\_r_j$ or $v\_r_j^{-1}$, where $v$ is some vocabulary word. A [*context configuration*]{} is then simply a set of individual context bags, e.g., $R=\{r_i,r_j,r_k\}$, also labelled as $R$: $r_i+r_j+r_k$. We call a configuration consisting of $K$ individual context bags a $K$-set configuration (e.g., in this example, $R$ is a $3$-set configuration).[^2] {width="1.0\linewidth"} \[fig:alggraph\] [latex@errorgobble]{} : [*pool*]{} of those $K\leq M$ candidate individual context bags $\{r_1,\ldots,r_K\}$ for which $E(r_i) >= threshold, i \in \{1, \ldots, M\}$, where $E(\cdot)$ is a fitness function.\ [**build:**]{} $K$-set configuration $R^{Pool} = \{r_1,\ldots,r_K\}$ (1) set of candidate configurations $\mathbf{R} = \{R^{Pool}\}$ ; (2) current level $l=K$ ; (3) best configuration $R_o =\emptyset$ :\ Although a brute-force exhaustive search over all possible configurations is possible in theory and for small pools (e.g., for adjectives, see Tab. \[tab:pool\]), it becomes challenging or practically infeasible for large pools and large training data. For instance, based on the pool from Tab. \[tab:pool\], the search for the optimal configuration would involve trying out $2^{10}-1=1023$ configurations for nouns (i.e., training 1023 different word representation models). Therefore, to reduce the number of visited configurations, we present a simple heuristic search algorithm inspired by beam search [@Pearl:1984]. Class-Specific Configuration Search {#ss:pools} ----------------------------------- Alg. 1 provides a high-level overview of the algorithm. An example of its flow is given in Fig. \[fig:alggraph\]. Starting from $S$, the set of all possible $M$ individual context bags, the algorithm automatically detects the subset $S_K \subseteq S$, $|S_K|=K$, of candidate individual bags that are used as the initial pool (line 1 of Alg. 1). The selection is based on some fitness (goal) function $E$. In our setup, $E(R)$ is Spearman’s $\rho$ correlation with human judgment scores obtained on the development set after training the word representation model with the configuration $R$. The selection step relies on a simple threshold: we use a threshold of $\rho\geq 0.2$ without any fine-tuning in all experiments with all word classes. We find this step to facilitate efficiency at a minor cost for accuracy. For example, since `amod` denotes an adjectival modifier of a noun, an efficient search procedure may safely remove this bag from the pool of candidate bags for verbs. The search algorithm then starts from the full $K$-set $R^{Pool}$ configuration (line 3) and tests $K$ $(K-1)$-set configurations where exactly one individual bag $r_i$ is removed to generate each such configuration (line 10). It then retains only the set of configurations that score higher than the origin $K$-set configuration (lines 11-12, see Fig. \[fig:alggraph\]). Using this principle, it continues searching only over lower-level $(l-1)$-set configurations that further improve performance over their $l$-set origin configuration. It stops if it reaches the lowest level or if it cannot improve the goal function any more (line 15). The best scoring configuration is returned (n.b., not guaranteed to be the global optimum). In our experiments with this heuristic, the search for the optimal configuration for verbs is performed only over 13 $1$-set configurations plus 26 other configurations (39 out of 133 possible configurations).[^3] For nouns, the advantage of the heuristic is even more dramatic: only 104 out of 1026 possible configurations were considered during the search.[^4] Experimental Setup {#s:exp} ================== Implementation Details ---------------------- #### Word Representation Model We experiment with SGNS [@Mikolov:2013nips], the standard and very robust choice in vector space modeling [@Levy:2015tacl]. In all experiments we use `word2vecf`, a reimplementation of `word2vec` able to learn from arbitrary $(word, context)$ pairs.[^5] For details concerning the implementation, we refer the reader to [@Goldberg:2014arxiv; @Levy:2014acl]. The SGNS preprocessing scheme was replicated from [@Levy:2014acl; @Levy:2015tacl]. After lowercasing, all words and contexts that appeared less than 100 times were filtered. When considering all dependency types, the vocabulary spans approximately 185K word types.[^6] Further, all representations were trained with $d=300$ (very similar trends are observed with $d=100,500$). The same setup was used in prior work [@Schwartz:2016naacl; @Vulic:2016acluniversal]. Keeping the representation model fixed across experiments and varying only the context type allows us to attribute any differences in results to a sole factor: the context type. We plan to experiment with other representation models in future work. #### Universal Dependencies as Labels The adopted UD scheme leans on the universal Stanford dependencies [@Marneffe:2014lrec] complemented with the universal POS tagset [@Petrov:2012lrec]. It is straightforward to “translate” previous annotation schemes to UD [@Marneffe:2014lrec]. Providing a consistently annotated inventory of categories for similar syntactic constructions across languages, the UD scheme facilitates representation learning in languages other than English, as shown in [@Vulic:2016acluniversal; @Vulic:2017eacl]. #### Individual Context Bags Standard post-parsing steps are performed in order to obtain an initial list of individual context bags for our algorithm: (1) Prepositional arcs are collapsed ([@Levy:2014acl; @Vulic:2016acluniversal], see Fig. \[fig:rex\]). Following this procedure, all pairs where the relation $r$ has the form `prep:X` (where `X` is a preposition) are subsumed to a context bag labelled `prep`; (2) Similar labels are merged into a single label (e.g., direct (`dobj`) and indirect objects (`iobj`) are merged into `obj`); (3) Pairs with infrequent and uninformative labels are removed (e.g., `punct`, `goeswith`, `cc`). Coordination-based contexts are extracted as in prior work [@Schwartz:2016naacl], distinguishing between left and right contexts extracted from the `conj` relation; the label for this bag is `conjlr`. We also utilise the variant that does not make the distinction, labeled `conjll`. If both are used, the label is simply `conj=conjlr+conjll`.[^7] Consequently, the individual context bags we use in all experiments are: `subj`, `obj`, `comp`, `nummod`, `appos`, `nmod`, `acl`, `amod`, `prep`, `adv`, `compound`, `conjlr`, `conjll`. Training and Evaluation ----------------------- We run the algorithm for context configuration selection only once, with the SGNS training setup described below. Our main evaluation setup is presented below, but the learned configurations are tested in additional setups, detailed in Sect. \[s:results\]. #### Training Data Our training corpus is the cleaned and tokenised English Polyglot Wikipedia data [@AlRfou:2013conll],[^8] consisting of approximately 75M sentences and 1.7B word tokens. The Wikipedia data were POS-tagged with universal POS (UPOS) tags [@Petrov:2012lrec] using the state-of-the art TurboTagger [@Martins:2013acl].[^9] The parser was trained using default settings (SVM MIRA with 20 iterations, no further parameter tuning) on the [train+dev]{} portion of the UD treebank annotated with UPOS tags. The data were then parsed with UD using the graph-based Mate parser v3.61 [@Bohnet:2010coling][^10] with standard settings on [train+dev]{} of the UD treebank. #### Evaluation We experiment with the verb pair (222 pairs), adjective pair (111 pairs), and noun pair (666 pairs) portions of SimLex-999. We report Spearman’s $\rho$ correlation between the ranks derived from the scores of the evaluated models and the human scores. Our evaluation setup is borrowed from : we perform 2-fold cross-validation, where the context configurations are optimised on a development set, separate from the unseen test data. Unless stated otherwise, the reported scores are always the averages of the 2 runs, computed in the standard fashion by applying the cosine similarity to the vectors of words participating in a pair. Baselines --------- #### Baseline Context Types We compare the context configurations found by Alg. 1 against baseline contexts from prior work:\ [**- BOW**]{}: Standard bag-of-words contexts.\ [**- POSIT**]{}: Positional contexts [@Schutze:1993acl; @Levy:2014conll; @Ling:2015naacl], which enrich BOW with information on the sequential position of each context word. Given the example from Fig. \[fig:rex\], POSIT with the window size $2$ extracts the following contexts for [*discovers*]{}: [*Australian\_-2*]{}, [*scientist\_-1*]{}, [*stars\_+2*]{}, [*with\_+1*]{}.\ [**- DEPS-All**]{}: All dependency links without any context selection, extracted from dependency-parsed data with prepositional arc collapsing.\ [**- COORD**]{}: Coordination-based contexts are used as fast lightweight contexts for improved representations of adjectives and verbs [@Schwartz:2016naacl]. This is in fact the `conjlr` context bag, a subset of DEPS-All.\ [**- SP**]{}: Contexts based on symmetric patterns (SPs, [@Davidov:2006acl; @Schwartz:2015conll]). For example, if the word X and the word Y appear in the lexico-syntactic symmetric pattern “X or Y” in the SGNS training corpus, then Y is an SP context instance for X, and vice versa. The development set was used to tune the window size for BOW and POSIT (to 2) and the parameters of the SP extraction algorithm.[^11] #### Baseline Greedy Search Algorithm We also compare our search algorithm to its greedy variant: at each iteration of lines 8-12 in Alg. 1, $R_n$ now keeps only the best configuration of size $l-1$ that perform better than the initial configuration of size $l$, instead of all such configurations. [1.0]{}[l XXX]{} & [Adj]{} & [Verb]{} & [Noun]{}\ (lr)[1-1]{} (lr)[2-4]{} [`conjlr` (A+N+V)]{} & & &\ [`obj` (N+V)]{} & [-0.028]{} & &\ [`prep` (N+V)]{} & [0.188]{} & &\ [`amod` (A+N)]{} & & [0.058]{} &\ [`compound` (N)]{} & [-0.124]{} & [-0.019]{} &\ [`adv` (V)]{} & [0.197]{} & & [0.104]{}\ [`nummod` (-)]{} & [-0.142]{} & [-0.065]{} & [0.029]{}\ [1.0]{}[smb]{} & [Verbs]{} & [Nouns]{}\ `amod, conjlr, conjll` & `prep, acl, obj, comp, adv, conjlr, conjll` & `amod, prep, compound, subj, obj, appos, acl, nmod, conjlr, conjll`\ [cc]{} [0.42]{}[l X]{} & [**(Verbs)**]{}\ (lr)[2-2]{} (lr)[1-1]{} [BOW (`win`=2)]{} & [0.336]{}\ [POSIT (`win`=2)]{} & [0.345]{}\ [COORD (`conjlr`)]{} & [0.283]{}\ [SP]{} & [0.349]{}\ [DEPS-All]{} & [0.344]{}\ &\ (lr)[1-1]{} [`POOL-ALL`]{} &\ [`prep+acl+obj+adv+conj`]{} &\ [`prep+acl+obj+comp+conj`]{} & [0.344]{}\ [`prep+obj+comp+adv+conj`]{} & $^\dagger$\ [`prep+acl+adv+conj`]{} (BEST) &\ [`prep+acl+obj+adv`]{} &\ [`prep+acl+adv`]{} &\ [`prep+acl+conj`]{} &\ [`acl+obj+adv+conj`]{} & [0.345]{}\ [`acl+obj+adv`]{} &\ & [0.52]{}[l X]{} & [[**(Nouns)**]{}]{}\ (lr)[2-2]{} (lr)[1-1]{} [BOW (`win`=2)]{} & [0.435]{}\ [POSIT (`win`=2)]{} & [0.437]{}\ [COORD (`conjlr`)]{} & [0.392]{}\ [SP]{} & [0.372]{}\ [DEPS-All]{} & [0.441]{}\ &\ (lr)[1-1]{} [`POOL-ALL`]{} &\ [`amod+subj+obj+appos+compound+nmod+conj`]{} &\ [`amod+subj+obj+appos+compound+conj`]{} &\ [`amod+subj+obj+appos+compound+conjlr`]{} & $^\dagger$\ [`amod+subj+obj+compound+conj`]{} (BEST) &\ [`amod+subj+obj+appos+conj`]{} &\ [`subj+obj+compound+conj`]{} &\ [`amod+subj+compound+conj`]{} &\ [`amod+subj+obj+compound`]{} &\ [`amod+obj+compound+conj`]{} &\ \[tab:verbsnouns\] Results and Discussion {#s:results} ====================== Main Evaluation Setup {#ss:main-eval} --------------------- #### Not All Context Bags are Created Equal First, we test the performance of [*individual*]{} context bags across SimLex-999 adjective, verb, and noun subsets. Besides providing insight on the intuition behind context selection, these findings are important for the automatic selection of class-specific pools (line 1 of Alg. 1). The results are shown in Tab. \[tab:single\]. The experiment supports our intuition (see Sect. \[ss:pools\]): some context bags are definitely not useful for some classes and may be safely removed when performing the class-specific SGNS training. For instance, the `amod` bag is indeed important for adjective and noun similarity, and at the same time it does not encode any useful information regarding verb similarity. `compound` is, as expected, useful only for nouns. Tab. \[tab:single\] also suggests that some context bags (e.g., `nummod`) do not encode any informative contextual evidence regarding similarity, therefore they can be discarded. The initial results with individual context bags help to reduce the pool of candidate bags (line 1 in Alg. 1), see Tab. \[tab:pool\]. [1.0]{}[l X]{} & [**(Adjectives)**]{}\ (lr)[2-2]{} (lr)[1-1]{} [BOW (`win`=2)]{} & [0.489]{}\ [POSIT (`win`=2)]{} & [0.460]{}\ [COORD (`conjlr`)]{} & [0.407]{}\ [SP]{} & [0.395]{}\ [DEPS-All]{} & [0.360]{}\ &\ (lr)[1-1]{} [`POOL-ALL: amod+conj`]{} (BEST) & $^\dagger$\ [`amod+conjlr`]{} &\ [`amod+conjll`]{} &\ [`conj`]{} & [0.470]{}\ #### Searching for Improved Configurations Next, we test if we can improve class-specific representations by selecting class-specific configurations. Results are summarised in Tables \[tab:verbsnouns\] and \[tab:adj\]. Indeed, class-specific configurations yield better representations, as is evident from the scores: the improvements with the best class-specific configurations found by Alg. 1 are approximately 6 $\rho$ points for adjectives, 6 points for verbs, and 5 points for nouns over the best baseline for each class. The improvements are visible even with configurations that simply pool all candidate individual bags (POOL-ALL), without running Alg. 1 beyond line 1. However, further careful context selection, i.e., traversing the configuration space using Alg. 1 leads to additional improvements for V and N (gains of 3 and 2.2 $\rho$ points). Very similar improved scores are achieved with a variety of configurations (see Tab. \[tab:verbsnouns\]), especially in the neighbourhood of the best configuration found by Alg. 1. This indicates that the method is quite robust: even sub-optimal[^12] solutions result in improved class-specific representations. Furthermore, our algorithm is able to find better configurations for verbs and nouns compared to its greedy variant. Finally, our algorithm generalises well: the best scoring configuration on the dev set is always the best one on the test set. #### Training: Fast and/or Accurate? Carefully selected configurations are also likely to reduce SGNS training times. Indeed, the configuration-based model trains on only 14% (A), 26.2% (V), and 33.6% (N) of all dependency-based contexts. The training times and statistics for each context type are displayed in Tab. \[tab:speed\]. All models were trained using parallel training on 10 Intel(R) Xeon(R) E5-2667 2.90GHz processors. The results indicate that class-specific configurations are not as lightweight and fast as SP or COORD contexts [@Schwartz:2016naacl]. However, they also suggest that such configurations provide a good balance between accuracy and speed: they reach peak performances for each class, outscoring all baseline context types (including SP and COORD), while training is still much faster than with “heavyweight” context types such as BOW, POSIT or DEPS-All. Now that we verified the decrease in training time our algorithm provides for the final training, it makes sense to ask whether the configurations it finds are valuable *in other setups*. This will make the fast training of practical importance. [1.0]{}[l XX]{} & [Training Time]{} & [\# Pairs]{}\ (lr)[1-1]{} (lr)[2-2]{} (lr)[3-3]{} [BOW (`win`=2)]{} & [179mins 27s]{} & [5.974G]{}\ [POSIT (`win`=2)]{} & [190mins 12s]{} & [5.974G]{}\ [COORD (`conjlr`)]{} & [4mins 11s]{} & [129.69M]{}\ [SP]{} & [1mins 29s]{} & [46.37M]{}\ [DEPS-All]{} & [103mins 35s]{} & [3.165G]{}\ [BEST-ADJ]{} & [14mins 5s]{} & [447.4M]{}\ [BEST-VERBS]{} & [29mins 48s]{} & [828.55M]{}\ [BEST-NOUNS]{} & [41mins 14s]{} & [1.063G]{}\ Generalisation: Configuration Transfer -------------------------------------- #### Another Training Setup We first test whether the context configurations learned in Sect. \[ss:main-eval\] are useful when SGNS is trained in another English setup [@Schwartz:2016naacl], with more training data and other annotation and parser choices, while evaluation is still performed on SimLex-999. [1.0]{}[l XXXX]{} & [Adj]{} & [Verbs]{} & [Nouns]{} & [All]{}\ (lr)[1-1]{} (lr)[2-2]{} (lr)[3-3]{} (lr)[4-4]{} (lr)[5-5]{} [BOW (`win`=2)]{} & [0.604]{} & [0.307]{} & [0.501]{} & [0.464]{}\ [POSIT (`win`=2)]{} & [0.585]{} & [0.400]{} & [0.471]{} & [0.469]{}\ [COORD (`conjlr`)]{} & [0.629]{} & [0.413]{} & [0.428]{} & [0.430]{}\ [SP]{} & [0.649]{} & [**0.458**]{} & [0.414]{} & [0.444]{}\ [DEPS-All]{} & [0.574]{} & [0.389]{} & [0.492]{} & [0.464]{}\ [BEST-ADJ]{} & [**0.671**]{} & [0.348]{} & [0.504]{} & [0.449]{}\ [BEST-VERBS]{} & [0.392]{} & [0.455]{} & [0.478]{} & [0.448]{}\ [BEST-NOUNS]{} & [0.581]{} & [0.327]{} & [**0.535**]{} & [0.489]{}\ & [0.616]{} & [0.402]{} & [0.519]{} & [**0.506**]{}\ In this setup the training corpus is the 8B words corpus generated by the `word2vec` script.[^13] A preprocessing step now merges common word pairs and triplets to expression tokens (e.g., [*Bilbo\_Baggins*]{}). The corpus is parsed with labelled Stanford dependencies [@Marneffe:2008sd] using the Stanford POS Tagger [@Toutanova:2003naacl] and the stack version of the MALT parser [@Goldberg:2012coling]. SGNS preprocessing and parameters are also replicated; we now train $500$-dim embeddings as in prior work.[^14] Results are presented in Tab. \[tab:roy\]. The imported class-specific configurations, computed using a much smaller corpus (Sect. \[ss:main-eval\]), again outperform competitive baseline context types for adjectives and nouns. The BEST-VERBS configuration is outscored by SP, but the margin is negligible. We also evaluate another configuration found using Alg. 1 in Sect. \[ss:main-eval\], which targets the overall improved performance without any finer-grained division to classes (BEST-ALL). This configuration ([`amod+subj+obj+compound+prep+adv+conj`]{}) outperforms all baseline models on the entire benchmark. Interestingly, the non-specific BEST-ALL configuration falls short of A/V/N-specific configurations for each class. This unambiguously implies that the “trade-off” configuration targeting all three classes at the same time differs from specialised class-specific configurations. [1.0]{}[l XXX]{} & [Adj-Q]{} & [Verb-Q]{} & [Noun-Q]{}\ (lr)[1-1]{} (lr)[2-2]{} (lr)[3-3]{} (lr)[4-4]{} [BOW (`win`=2)]{} & [31/41]{} & [14/19]{} & [16/19]{}\ [POSIT (`win`=2)]{} & [**32/41**]{} & [13/19]{} & [15/19]{}\ [COORD (`conjlr`)]{} & [26/41]{} & [11/19]{} & [8/19]{}\ [SP]{} & [26/41]{} & [11/19]{} & [12/19]{}\ [DEPS-All]{} & [31/41]{} & [14/19]{} & [16/19]{}\ [BEST-ADJ]{} & [**32/41**]{} & [12/19]{} & [15/19]{}\ [BEST-VERBS]{} & [24/41]{} & [**15/19**]{} & [16/19]{}\ [BEST-NOUNS]{} & [30/41]{} & [14/19]{} & [**17/19**]{}\ #### Experiments on Other Languages We next test whether the optimal context configurations computed in Sect. \[ss:main-eval\] with English training data are also useful for other languages. For this, we train SGNS models on the Italian (IT) and German (DE) Polyglot Wikipedia corpora with those configurations, and evaluate on the IT and DE multilingual SimLex-999 [@Leviant:2015arxiv].[^15] Our results demonstrate similar patterns as for English, and indicate that our framework can be easily applied to other languages. For instance, the BEST-ADJ configuration (the same configuration as in Tab. 4 and Tab. 7) yields an improvement of 8 $\rho$ points and 4 $\rho$ points over the strongest adjectives baseline in IT and DE, respectively. We get similar improvements for nouns (IT: 3 $\rho$ points, DE: 2 $\rho$ points), and verbs (IT: 2, DE: 4). #### TOEFL Evaluation We also verify that the selection of class-specific configurations (Sect. \[ss:main-eval\]) is useful beyond the core SimLex evaluation. For this aim, we evaluate on the A, V, and N TOEFL questions [@Landauer:1997pr]. The results are summarised in Tab. \[tab:toefl\]. Despite the limited size of the TOEFL dataset, we observe positive trends in the reported results (e.g., V-specific configurations yield a small gain on verb questions), showcasing the potential of class-specific training in this task. Conclusion and Future Work {#s:conclusion} ========================== We have presented a novel framework for selecting class-specific context configurations which yield improved representations for prominent word classes: adjectives, verbs, and nouns. Its design and dependence on the Universal Dependencies annotation scheme makes it applicable in different languages. We have proposed an algorithm that is able to find a suitable class-specific configuration while making the search over the large space of possible context configurations computationally feasible. Each word class requires a different class-specific configuration to produce improved results on the class-specific subset of SimLex-999 in English, Italian, and German. We also show that the selection of context configurations is robust as once learned configuration may be effectively transferred to other data setups, tasks, and languages without additional retraining or fine-tuning. In future work, we plan to test the framework with finer-grained contexts, investigating beyond POS-based word classes and dependency links. Exploring more sophisticated algorithms that can efficiently search richer configuration spaces is also an intriguing direction. Another research avenue is application of the context selection idea to other representation models beyond SGNS tested in this work, and experimenting with assigning weights to context subsets. Finally, we plan to test the portability of our approach to more languages. Acknowledgments {#acknowledgments .unnumbered} =============== This work is supported by the ERC Consolidator Grant LEXICAL: Lexical Acquisition Across Languages (no 648909). Roy Schwartz was supported by the Intel Collaborative Research Institute for Computational Intelligence (ICRI-CI). The authors are grateful to the anonymous reviewers for their helpful and constructive suggestions. [^1]: http://universaldependencies.org/ (V1.4 used) [^2]: A note on the nomenclature and notation: Each context configuration may be seen as a set of context bags, as it does not allow for repetition of its constituent context bags. For simplicity and clarity of presentation, we use dependency relation types (e.g., $r_i$ = `amod`, $r_j$ = `acl`) as labels for context bags. The reader has to be aware that a configuration $R=\{r_i,r_j,r_k\}$ is not by any means a set of relation types/names, but is in fact a multiset of all $(word, context)$ pairs belonging to the corresponding context bags labelled with $r_i$, $r_j$, $r_k$. [^3]: The total is 133 as we have to include 6 additional $1$-set configurations that have to be tested (line 1 of Alg. 1) but are not included in the initial pool for verbs (line 2). [^4]: We also experimented with a less conservative variant which does not stop when lower-level configurations do not improve $E$; it instead follows the path of the best-scoring lower-level configuration even if its score is lower than that of its origin. As we do not observe any significant improvement with this variant, we opt for the faster and simpler one. [^5]: https://bitbucket.org/yoavgo/word2vecf [^6]: SGNS for all models was trained using stochastic gradient descent and standard settings: $15$ negative samples, global learning rate: $0.025$, subsampling rate: $1e-4$, $15$ epochs. [^7]: Given the coordination structure [*boys and girls*]{}, `conjlr` training pairs are [*(boys, girls\_conj), (girls, $boys\_conj^{-1})$*]{}, while `conjll` pairs are [*(boys, girls\_conj), (girls, $boys\_conj)$*]{}. [^8]: https://sites.google.com/site/rmyeid/projects/polyglot [^9]: http://www.cs.cmu.edu/\~ark/TurboParser/ [^10]: https://code.google.com/archive/p/mate-tools/ [^11]: The SP extraction algorithm is available online:\ homes.cs.washington.edu/$\sim$roysch/software/dr06/dr06.html [^12]: The term *optimal* here and later in the text refers to the best configuration returned by our algorithm. [^13]: code.google.com/p/word2vec/source/browse/trunk/ [^14]: The “translation” from labelled Stanford dependencies into UD is performed using the mapping from , e.g., `nn` is mapped into `compound`, and `rcmod`, `partmod`, `infmod` are all mapped into one bag: `acl`. [^15]: http://leviants.com/ira.leviant/MultilingualVSMdata.html

ArXiv

--- abstract: 'This paper describes our approach presented for the eHealth-KD 2019 challenge. Our participation was aimed at testing how far we could go using generic tools for Text-Processing but, at the same time, using common optimization techniques in the field of Data Mining. The architecture proposed for both tasks of the challenge is a standard stacked 2-layer bi-LSTM. The main particularities of our approach are: (a) The use of a surrogate function of *F1* as loss function to close the gap between the minimization function and the evaluation metric, and (b) The generation of an ensemble of models for generating predictions by majority vote. Our system ranked second with an *F1* score of 62.18% in the main task by a narrow margin with the winner that scored 63.94%.' author: - Neus Català - Mario Martin bibliography: - 'article.bib' subtitle: Voting LSTMs for Key Phrases and Semantic Relation Identification Applied to Spanish eHealth Texts title: 'Coin\_flipper at eHealth-KD Challenge 2019' --- Introduction ============ This article describes the model presented by the *coin-flipper* team for solving the shared task presented in eHealth-KD 2019 challenge [@ehealthkd19_overview]. The main goal of the challenge is to encourage the development of software systems to automatically extract information from electronic health documents written in Spanish. The challenge involves two subtasks: 1) Identification and classification of key phrases (four different types of units) and 2) Detection of semantic relations among them (thirteen semantic relations). The tasks proposed are similar to previous competitions such as Semeval-2017 Task 10: ScienceIE [@augenstein-etal-2017-semeval] and TASS-2018-Task 3 eHealth Knowledge Discovery [@DBLP:conf/sepln/2018tass], but the types of entities and the amount of semantic relations to be identified have been changing over time. Also, we find changes in the systems approaches evolving from rule based methods to Deep Learning models some of them incorporating domain-specific knowledge. As the organizers of the challenge state, although this challenge is oriented to the health domain, the kind of knowledge to be extracted is general-purpose. With this idea in mind, we propose a model that avoids the use of knowledge or tools specific to the domain, easing the portability of the model to new domains and tasks. Tasks and data ============== The eHealth-KD 2019 challenge involves two subtasks: subtask A) Identification and classification of key phrases (four different types of units), and subtask B) Detection of semantic relations among them (thirteen semantic relations). Each subtask has been evaluated individually and also as components of a pipeline. For this purpose, three evaluation scenarios have been proposed. Scenario 1 requires subtasks A and B to be executed sequentially in a pipeline. Scenario 2 only evaluates subtask A and Scenario 3 only evaluates subtask B. A training set that contains a total of 600 sentences was provided to be used for the learning step. For the validation step, an additional set of 100 sentences was available. A detailed explanation of the corpus data, subtasks and evaluation metrics can be found in the overview description paper [@ehealthkd19_overview]. Our approach ============ Our aim was to see how far we can get in tasks described above using generic NLP tools (non-specific for the domain) and standard Data Mining strategies. About the model to be learned, given the sequential nature of the data (sentences), we proposed a stacked 2-layer bi-LSTM architecture. LSTMs have been proved a successful approach in several Natural Language Processing tasks, such as Named Entity Recognition (f.i. [@huang2015bidirectional]) and Semantic Role Labeling (f.i. [@zhou-xu-2015-end]). About the data mining strategies, we applied two following strategies: 1. Building an ensemble of models for generating predictions: It is well known in Machine Learning that ensemble models have higher performance than single models [@Zhou:2012:EMF:2381019], specially when models have a high variance which is the case of neural networks. So, for each task, we build an ensemble of models instead of a single model. There are several ways to generate an ensemble of neural networks. Usually, differences in hyperparameters or different architectures are enough to ensure diversity of models necessary for making ensemble methods profitable. In our case we achieve diversity by using random initialization of models before training [@Goodfellow2015]. There also exist a lot of different techniques to combine the output of different models into a single prediction [@Zhou:2012:EMF:2381019]. We will apply the simplest method of majority vote to combine the outputs of the models in a ensemble. 2. The use of a surrogate function of *F1* metric as the loss function: Usually, accuracy is the metric to be optimized and Cross-entropy (CE) is the continuous function used as loss function. However, in this challenge, *F1* score is used for evaluation. The use of CE is not optimal in this case because we would like to minimize the error according to the metric used for evaluation. In [@Eban2017], authors propose to solve this problem by defining continuous surrogate upper and lower bounds for true positive and false positive cases and, from them, *F1’* surrogate. Then they use $1/{F1'}$ as the loss function to minimize. We have done it differently: in our case, getting the predictions $\hat{y}$ from our Neural Network (with *sigmoid* function activation in last layer), we choose to use a soft version of *F1* metric. We define soft-true positive cases ($tp_s$) and soft-false positives ($fp_s$), where labels $y$ are in {0,1} and predictions $\hat{y}$ in \[0..1\], as: $$\begin{aligned} tp_s & = \sum_{i \in Y^+} \hat{y}_i = \sum_{i \in Y} y_i \hat{y}_i \\ fp_s & = \sum_{i \in Y^-} \hat{y}_i = \sum_{i \in Y} (1-y_i)\hat{y}_i\end{aligned}$$ Notice that when predicted labels $\hat{y}$ are hard (0 or 1), these values are exactly true and false positive cases, respectively. Replacing soft versions of true and false positives cases in *F1* definition (in the type I and type II error form), we have our surrogate $F1_s$ function to maximize. $$F1_s \; = \; \frac{2 \; tp_s}{\left|Y^+\right|+tp_s+fp_s} \; = \; %\frac{2 \left(\sum_{Y} y\hat{y}\right)}{\left|Y^+\right| + \sum_{Y} y \hat{y} + \sum_{Y} (1-y) \hat{y} } = \frac{2\; \sum_{i\in Y^+} \hat{y}_i}{\left|Y^+\right| + \sum_{i\in Y^+} \hat{y}_i+ \sum_{i\in Y^-} \hat{y}_i}$$ Note that maximize $F1_s$ implies increasing $\hat{y}$ of positive cases while decreasing $\hat{y}$ of negative cases in a proportional way to *F1* function. Loss function to minimize will be $1-F1_s$. In the next sections we describe further details about how subtasks A and B are solved. Subtask A --------- Subtask A consists in detecting key phrases in texts (from now on we will abbreviate key phrase to KPhr) and their kind or class. For this task we set as input only the sequence of words of each sentence and their *Part-of-Speech* (PoS). Words are represented using *FastText* embeddings [@bojanowski-etal-2017-enriching] trained on the *Spanish Billion Word Corpus* [@cardellinoSBWCE]. Among the advantages of *FastText* over other approaches are that, being *FastText* trained on n-grams of words, embeddings carry implicitly morphological information and also are able to output a reasonable embedding for a word not seen during training. PoS for sentences are generated using *spaCy*[@spacy2] tagger for Spanish. Probably better results would be obtained using more advanced taggers like *FreeLing* PoS tagger [@padro12]. In the definition of the output of the architecture we were inspired on image segmentation work in which, given an image, a mask over the image is generated that encodes for each pixel the class it belongs to. Similarly, for each sentence, we define as output a mask vector where for each word we have codified the kind of KPhr the word belongs to (or a label denoting that the word does not belong to any KPhr). The architecture of the model to be learned consists of two stacked layers of bi-LSTM feed with the word embedding and the one-hot encoding of the PoS for each word of the sentence. We apply sequence Padding (with length parameter equal to the length of the larger sentence in data set) and Masking. Initially, output was defined as a vector with the one-hot codification of the corresponding KPhr label for each word. However, we found empirically that learning a model to predict *all* kinds of KPhr was not very successful. So, we decided to split the task into learning models specialized in predictions for each kind of KPhr. With this modification, now the output of the architecture is a 0/1 vector with 1 denoting if a word belongs to the kind of KPhr the model is trained for, and 0 otherwise. Notice that with this modification, the use of *F1* loss become even more important because the induced sparsity of labels. The learning of each model is done by using surrogate $1-F1_s$ loss (as explained in previous section) with early stopping when $F1$ score does not improve in 100 epochs on the development dataset, and returning the model that reaches the best $F1$ score on development dataset. Default learning rate $0.001$ with *adam* optimization is used. Batch size is 32 sentences. Neither dropout nor data augmentation is used. First bi-LSTM layer has 150 units in each direction. Second bi-LSTM layer has 32 units in each direction. The same architecture is used for all KPhr detection models without fine-tuning of parameters. See a graphical representation of the model in Figure \[figura1\]. As stated in the previous section, we create an ensemble of 15 models for each target KPhr that is used later to generate the final prediction by majority voting. Diversity of models in the ensemble is obtained only from different initialization of weights. To avoid the vote of specially bad models, each model is applied to the development dataset and those models with an *F1* score far below the average of the ensemble are removed from it. Our approach has two obvious problems. The first one is that, because the way we define segmentation, we don’t know when a KPhr starts nor ends. A naive solution would be to join into the same KPhr consecutive words that are tagged with the same label. However this heuristic solution does not always work. We solve this problem by inspecting the PoS of consecutive words tagged with the same label. For each combination of PoS found, we compute how many cases covered are in fact cases where words belong to the same KPhr. From the statistics of these cases we create a set of rules that decide if two consecutive words with the same label really belong to the same KPhr or not. The second problem is that having different models, one for each kind of KPhr, words could be labeled as belonging to different kinds of KPhr at the same time. So, in case of conflict among different ensembles, in order to decide the final kind of KPhr, we count how many positive votes each ensemble has emitted and multiply it by a weighting parameter. The ensemble with the highest product is the winner and the word is labeled with the kind of KPhr the ensemble represents. Weights for each ensemble are found by applying the ensembles on the development dataset and grid-searching for the combination of weights that returns highest global *F1*. Subtask B --------- Subtask B consists in detecting semantic relations (thirteen types of relations) between KPhr pairs identified in the previous task. For this task, given a relation *rel* and a KPhr considered as *source* for the relation, we try to identify another KPhr in the sentence with the role *target* for this relation *rel*. In order to learn models able to solve this task, we define an architecture that, given a relation *rel* and a *source KPhr*, has the following inputs: 1. Sequence of words in the sentence where *source KPhr* is found. 2. Part of Speech tags for each word in the sentence. 3. Mask denoting for each word the *kind of KPhr it belongs to* (codifying in 1-4 the kind of KPhr and 0 if it does not belong to any KPhr). 4. 0/1 Mask vector for each word with 1 denoting if the word belongs to the *source KPhr*, and 0 otherwise. Words are represented as *FastText* vector embeddings and all other inputs are encoded using one-hot encoding. All vectors codifying the information of one word are concatenated. See Figure \[figura2\] for a representation of the architecture. For each relation we build a Neural Network that consists of the embedding layers explained above plus 2 stacked bi-LSTM layers. As in subtask A, we use padding and masking. The output consists of a mask vector 0/1 with 1 only for those words who belong to a *target* KPhr of the relation *rel* of current KPhr as source. When no *target* KPhr exists for the current relation, output is all zeros. When the model identifies at least one word of a target KPhr, we consider the whole KPhr detected. The output representation selected allows that in the same sentence one KPhr as *source* has several KPhr as *target* for the same relation. All training parameters and procedures are the same as those used in subtask A. Again, the learning of each model is done by using surrogate $1-F1_s$ loss with early stopping when $F1$ score does not improve in 100 epochs on the development dataset, and returning the model that reaches the best $F1$ score on development dataset. Default learning rate $0.001$ with *adam* optimization is used. Batch size is 32. Neither dropout nor data augmentation is used. First bi-LSTM layer has 150 units in each direction. Second bi-LSTM layer has 32 units in each direction. The same architecture is used for all relations without fine-tuning of parameters. For each relation, an ensemble of 15 models is built from which majority voting is implemented. In some cases, ensembles of different relations identify the same pair of KPhr as positive cases. When this happens, we assign to the pair of KPhr the relation of the ensemble with highest *precision* on the development dataset. Results ======= Table \[table2\] shows *F1* scores for the three scenarios achieved by the baseline system provided by organizers, the best method and our method. Our results were very competitive. We were second in the main evaluation (scenario 1) by a narrow margin - only 1.76 of *F1* score over 100. In scenario 2 and, specially, scenario 3 we were at a larger distance to the best method (see Table \[table2\]) in positions 5th and 4th of the ranking, respectively. This is surprising because scenario 1 is the composition of the other 2 scenarios. The main difference in evaluation among scenarios was that for scenario 1 a subset of 100 sentences out of a 8.800 sentences dataset was randomly selected for evaluation, while in scenarios 2 and 3 a dataset of 100 sentences was given to each one for evaluation. For scores shown in Table \[table2\] we did not submit results for different parameters except for subtask B. In subtask B, we noticed that not all relations had the same difficulty in being detected. We found that a first set of 6 relations (*in-context*, *subject*, *target*, *domain*, *arg* and *is-a*) were successfully identified with high recall and precision values when tested on development dataset (above 75%). Relations *in-place*, *has-property*, *same-as* and *in-time* were (in this order) harder to identify and, finally, for *part-of*, *causes* and *entails* relations, results were so bad that models for those relations were discarded. In order to obtain the highest score, we ranked relations by their *F1* scores on development dataset and submitted different sets of results considering only relations that were in 6, 7, 8, 9 and 10 top positions. Best results on testing dataset were obtained considering top 7 relations, that is, the ones in the first set (*in-context*, *subject*, *target*, *domain*, *arg* and *is-a*) plus *in-place*. Considering only a subset of possible relations had an obvious impact on final performance that achieved good precision but low recall. We think that difficulty in detecting relations *part-of*, *same-as*, *in-time*, *entails* and *has-property* comes from the low number of instances (range 50-100) in the training set. So, for these cases, some techniques like data augmentation should be used to obtain better results. An exception is relation *causes* that had 230 instances and was very difficult to identify with our approach. In this case most complex architectures should be used. \[t!\] [width=]{} [l | c c c | c c c | c c c]{} ------------------------------------------------------------------------ & &&\ & F1 & P & R & F1 & $\;\;$P & R & F1 & P & R\ ------------------------------------------------------------------------ *scenario 1*&43.09&52.04&36.77& 63.94 & $\;\;$65.06 & 62.86 & 62.18 & 74.54 & 53.34\ *scenario 2*&54.66&51.29&58.51& 82.03 & $\;\;$80.73 & 83.36 & 78.73 & 79.86 & 77.63\ *scenario 3*&12.31&48.78&7.04& 62.69 & $\;\;$66.67 & 59.15 & 49.31 & 71.33 & 37.68\ Conclusions =========== In this paper we have presented a simple approach to solve the tasks proposed in this competition. The aim was to measure how far we can get with standard tools and simple algorithms but using common optimization techniques in the Data Mining field. In particular, the architecture proposed for both tasks was a two level bi-LSTM to modelize each of the classes we wanted to capture. We didn’t use state of the art language models like BERT, ELMO or GPT, neither specialized NER techniques, knowledge bases or taggers specific for the domain (so results probably will be similar in other domains). The main particularities of our approach are: (a) The use of a surrogate loss function of *F1* metric to close the gap between the minimization function and the evaluation metric and (b) the generation of an ensemble of models for generating predictions by majority vote. From results, we conclude that it is worth using the evaluation metric (or a surrogate function of it) as loss function and also building ensemble models for generating predictions. We did not do an exhaustive study of the improvement in performance due to the use of a surrogate function of *F1* as loss, but in initial experiments we found a clear advantage of the surrogate function usage instead of standard loss functions. We neither did an exhaustive study of the impact in performance of learning ensemble models instead of learning single models, however, we noticed that usually *F1* scores obtained by the majority vote of the ensemble was about 1-2 points higher than the best model of the ensemble and 3-4 points higher than the average of the models of the ensemble. Note that combination of outputs in the ensemble is made through simple majority voting but more complex ensemble techniques could improve the performance of the system even more. For a review of ensemble methods see [@Zhou:2012:EMF:2381019]. The results show that our model achieves good precision scores for the three scenarios but sometimes at the expense of lower recall levels than other systems, specially notorious in scenario 3. The difficulties our model had in identifying some semantic relations may be due to the low number of examples of these relations in the training dataset. To deal with this problem, we are planning to use data augmentation techniques in order to increase the number of instances of the less frequent relations. One of the techniques we want to explore is backtranslation, used in text classification [@Shleifer2019]. Acknowledgments {#acknowledgments .unnumbered} =============== **Funding**: Neus Català’s work has been partially funded by the Spanish Government and by the European Union through GRAPH-MED project (TIN2016-77820-C3-3-R) and AEI/FEDER, UE. Mario Martin’s work has been partially supported by the Joint Study Agreement no. W156463 under the IBM/BSC Deep Learning Center agreement.

ArXiv

--- author: - 'Philippe Sosoe\' title: 'A remark on Gibbs-type measures for Hamiltonian PDE' --- Introduction ============ In this note, we determine the optimal threshold of normalizability for certain Gibbs-type measures based on variants of Brownian motion. These have been studied in the PDE literature because they provide invariant measures for some Hamiltonian partial differential equations [@bourgain; @LRS; @tzvetkov1; @tzvetkov2]. In particular, we resolve an issue in the proof of Theorem 3.1 of the seminal paper [@LRS] by J. Lebowitz, H. Rose and E. Speer, which was pointed out by E. Carlen, J. Lebowitz, J. Fröhlich [@carlenlebowitzfroehlich]. The same method allows us to answer the question asked by J. Bourgain and A. Bulut in [@bourgainbulut Remark 6.2], concerning the same threshold for a measure on radial functions on the disc in dimension 2, introduced by N. Tzvetkov in [@tzvetkov1]. The following result is stated in [@LRS]: \[thm: LRS\] Consider the zero mean Brownian loop $u(x)$, defined by the random series $$\label{eqn: bloop} u(x)=\sum_{|n|\neq 0}\frac{g_n}{n}e^{2\pi inx},$$ where $g_n$ are independent, standard complex Gaussian random variables, and the partition function: $$\label{eqn: Z} Z_{p,K}=\mathbf{E}[e^{\frac{1}{p}\int_0^1 |u|^p\,\mathrm{d}x}\mathbf{1}_{\{\|u\|_{L^2([0,1])}\le K\}}].$$ 1. If $p<6$, then $Z_{p,K}<\infty$ for all $p$ and $K>0$. 2. If $p=6$, then $Z_{p,k}<\infty$ if and only if $K<\|\varphi||_{L^2}$, where $\varphi$ is the (unique) optimizer for the Gagliardo-Nirenberg-Sobolev (GNS) inequality. The threshold value $p=6$ and the relevance of the GNS inequality can be understood at an intuitive level by formally rewriting as a functional integral with respect to the (periodic) Gaussian free field: $$\label{eqn: functional} Z_{p,K} ``=" \int_{\|u\|_{L^2}\le K} e^{-\frac{1}{2}\int_0^1 |u'(x)|^2\,\mathrm{d}x+\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}\,\mathcal{D}u(x).$$ Applying the GNS inequality (see equation hereafter), this quantity is bounded by $$\int_{\|u\|_{L^2}\le K} e^{-\frac{1}{2}\int_0^1 |u'(x)|^2\,\mathrm{d}x+\frac{C_{\mathrm{GNS}}}{p}K^{(p+2)/2}\left(\int_0^1 |u'(x)|^2\,\mathrm{d}x\right)^{(p-2)/2}}\,\mathcal{D}u(x),$$ so when $p<6$, $p=6$ and $K$ is small enough, we expect the Gaussian part of the measure to dominate, and the partition function to be finite. A pleasing probabilistic proof of Theorem 1 based on this idea was given in [@LRS], using the explicit joint density of the times that the Brownian path hits certain levels on a grid. Unfortunately, the proof in [@LRS] seems to apply only to the case where the expectation in the definition of $Z_{p,K}$ is taken with respect to a standard (“free”) Brownian motion started at 0, rather than the random periodic function . This was noted in [@carlenlebowitzfroehlich]; see the remark at the beginning of Section 3.2 there. A more analytic proof due to J. Bourgain appears in [@bourgain]. See also [@bourgainbulut] for the case of a Gibbs measure on the 2D ball. His argument combines basic estimates for Gaussian vectors with the Sobolev embedding to identify the tail behavior of the random variable $\int |u|^p\,\mathrm{d}x$, subject to the condition $\|u\|_{L^2}\le K$. It applies also to the case $p=6$, but shows only that $Z_{p,K}<\infty$ for sufficiently small $K$. Here we obtain the optimal threshold. Our method also applies to a Gibbs measure on radial functions first constructed, along with corresponding invariant dynamics, by N. Tzvetkov [@tzvetkov1; @tzvetkov2]. His analysis was complemented in [@bourgainbulut] by a study of the boundary case $p=6$, in the focusing situation, which is of most interest to to us here: Let $D_1=\{(x,y)\in \mathbb{R}^2: x^2+y^2<1\}$ be the unit disc. Let $J_0(r)$ be the zero order Bessel function, defined by $$J_0(x)=\sum_{j=0}^\infty \frac{(-1)^j}{(j!)^2}\left(\frac{x}{2}\right)^{2j},$$ and $z_n$, $n\ge 1$, be its successive, positive zeros. Define $$e_n(r):=J_0(z_n r), \quad 0\le r\le 1$$ and consider the random series $$\label{eqn: bloop2} v(r)=\frac{1}{\sqrt{\pi}}\sum_{n=1}^\infty \frac{g_n}{z_n} J_0(z_n r), \quad r^2=x^2+y^2.$$ As previously, the $g_n$ are independent standard Gaussian random variables. The partition function is defined as $$Z_{p,K}'= \mathbf{E}[e^{\frac{1}{p}\int_{D_1}|v|^p\,\mathrm{d}x}\mathbf{1}_{\|v\|_{L^2(D_1)}\le K}].$$ 1. If $p<4$, then $Z_{p,K}<\infty$ for any $K>0$. 2. If $p=4$, then $Z_{p,K}<\infty$ if and only if $K<\|\varphi\|_{L^2(D_1)}$, where $\varphi$ is the optimizer in the Gagliardo-Nirenberg inequality in dimension 2. Our proof is closer in spirit to Bourgain’s, since it uses the series representations and of the Brownian loops, as opposed to the path space approach taken in [@LRS]. In the next section, we review Bourgain’s argument and point out that in this approach, closing the gap between small $K$ and the optimalthreshold seems difficult. We then present our proofs of the direct implication of Theorems 1 and 2 in the subsequent sections. As in [@LRS], the idea is to make rigorous the computation suggested by by finite dimensional approximation. Bourgain’s proof ================ We reproduce Bourgain’s argument for the direct part of Theorem 1, part 1. Rewriting as $$\label{eqn: layer} Z_{p,K}=1+\int_0^\infty \lambda^{p-1}e^{\frac{1}{p}\lambda^p}\mathbf{P}(\|u\|_{L^p([0,1])}>\lambda, \|u\|_{L^2([0,1])\le K})\,\mathrm{d}\lambda,$$ we see that it suffices to show $$\mathbf{P}(\|u\|_{L^p([0,1])}>\lambda, \|u\|_{L^2([0,1])}\le K)\le e^{-c\lambda^p}.$$ For $u\in L^2([0,1])$, denote by $$u_k=P_{2^k} u:= \sum_{2^{k-1}\le |n|< 2^k} \widehat{u}(n)e^{2\pi i nx}$$ the Littlewood-Paley projection of $u$ on frequencies of order $2^k$. Similarly $$\begin{aligned} u_{\le k}&=P_{\le 2^k}u:= \sum_{|n|\le 2^k} \widehat{u}(n)e^{2\pi i nx},\\ u_{\ge k}&=P_{\ge 2^k}u:= \sum_{|n|\ge 2^k} \widehat{u}(n)e^{2\pi i nx}.\end{aligned}$$ First, by the union bound, we have for any $k$: $$\label{eqn: sacco} \mathbf{P}(\|u_{\ge k-1}\|_{L^p}>\lambda) \le \sum_{j=k-1}^\infty \mathbf{P}( \|u_{j}\|_{L^p}> \lambda_j),$$ where $$\label{eqn: grisvold} \sum_{j=k}^\infty\lambda_j = \lambda.$$ Then using the Sobolev embedding in the form of Bernstein’s inequality, we have $$\label{eqn: 1stbernstein} \|u_{j}\|_{L^p}\le C2^{j(\frac{1}{2}-\frac{1}{p})}\|u_j\|_{L^2}.$$ Thus the probability in is bounded by $$\begin{aligned} \mathbf{P}( \|u_{j}\|_{L^2}> \frac{\lambda_j}{C}2^{j(\frac{1}{p}-\frac{1}{2})} ) &=\mathbf{P}( \sum_{2^{j-1}\le |k|< 2^j}\frac{|g_k|^2}{|j|^2}> \frac{\lambda_j^2}{C^2}2^{2j(\frac{1}{p}-\frac{1}{2})}) \nonumber\\ &\le \mathbf{P}( \sum_{2^{j-1}\le |k|< 2^j}|g_k|^2> \frac{\lambda_j^2}{C^2}2^{(1+\frac{2}{p})j}). \label{eqn: rebecca}\end{aligned}$$ The next lemma follows from a simple calculation involving moment generating functions of gaussians: Let $X_n$ $n\ge 1$ be real valued, independent standard Gaussian random variables. Then, if $R\ge 3\cdot M^{1/2}$, $$\label{eqn: Rbound} \mathbf{P}\big(\sum_{n=1}^M X_n^2 \ge R^2\big)\le e^{-\frac{R^2}{4}}.$$ The estimate shows that the probability is bounded by $$\label{eqn: eurus} \exp(-\frac{\lambda_j^2}{4C^2}2^{(1+\frac{2}{p})j})$$ provided $$\label{eqn: tomwaitts} \frac{\lambda_j}{C}2^{(\frac{1}{2}+\frac{1}{p})j}\ge \frac{3}{2}\cdot 2^{j/2}.$$ Choosing $$\lambda_j = \lambda (1-2^{-r})2^{kr}2^{-jr}$$ for $0<r<1/p$, both conditions and are satisfied for all large $k$. For such $k$, the probability in is then bounded by $$\label{eqn: marcus} \exp(-\frac{\lambda^2(1-2^{-r})^2}{4C^2}2^{2kr}2^{(1+\frac{2}{p}-2r)j}).$$ Summing over $j$ in , we find that $$\label{eqn: vanguard} \mathbf{P}(\|u_{\ge k}\|_{L^p}>\lambda)\le C_{r,\lambda}\exp(-\frac{\lambda^2(1-2^{-r})^2}{4C^2}2^{(1+\frac{2}{p})k}).$$ Using Bernstein again, we have, if $\|u\|_{L^2}\le K$, $$\label{eqn: bertrand} \|u_{\le k-1}\|_{L^p}\le C2^{k(\frac{1}{2}-\frac{1}{p})}\|u_{\le k-1}\|_{L^2}\le C2^{k(\frac{1}{2}-\frac{1}{p})}K .$$ Setting $$k=\log_2 \left(\frac{\lambda}{2CK}\right)^{\frac{2p}{p-2}},$$ implies $$\|u_{\le k-1}\|_{L^p}\le \frac{\lambda}{2},$$ and so by , we have for large $\lambda$: $$\begin{split} &\mathbf{P}(\|u\|_{L^p([0,1])}>\lambda, \|u\|_{L^2}\le K)\\ \le&\mathbf{P}(\|u_{\le k-1}\|_{L^p([0,1])}>\frac{\lambda}{2}, \|u\|_{L^2}\le K)+ \mathbf{P}(\|u_{\ge k}\|_{L^p([0,1])}>\frac{\lambda}{2})\\ \le & C_{r,\lambda}\exp\big(-\frac{(1-2^{-r})^2}{16C^2\cdot (2CK)^{\frac{p+2}{p-2}}}\lambda^{\frac{4p}{p-2}}\big). \end{split}$$ The exponent $4p/(p-2)$ beats the exponent $p$ in if $p<6$ or $p=6$ and $K$ is sufficiently small. Determining the optimal threshold for $K$ would presumably require a delicate optimization in $\lambda_j$ in , an exact Gaussian tail bound to replace the appraisal , and an optimal inequality to replace the applications of Sobolev and to determine the precise tail behavior of $\|u\|_{L^p}$ given $\|u\|_{L^2}\le K$. We did not attempt this calculation. Even if it is possible to carry out, such an approach likely leads to a less transparent argument than the one we propose now. Gagliardo-Nirenberg-Sobolev =========================== The following result is due to B. V. Sz.-Nagy [@nagy] for $n=1$ and M. Weinstein [@weinstein] for $n\ge 2$. Consider the functional $$J^{p,n}(f)=\frac{\|\nabla f\|^{\frac{n(p-2)}{2}}_{L^2(\mathbb{R}^n)}\|f\|^{2+\frac{p-2}{2}(2-n)}_{L^2(\mathbb{R}^n)}}{\|f\|_{L^p(\mathbb{R}^n)}^{p}}$$ on $H^1(\mathbb{R}^n)$. Then $J^{p,n}(f)$ attains its minimum $$C_{GNS}(n,p):= \inf_{f\in H^1} J^{p,n}(f)$$ for some $\varphi\in H^1(\mathbb{R}^n)$ with $\|\varphi\|_{\dot{H}^1(\mathbb{R}^n)}=\|\varphi\|_{L^2(\mathbb{R}^n)}=1$. We have $$\label{eqn: isolde} C_{GNS}(n,p)= \frac{p}{2}\|\varphi\|_{L^2}^{2-p}$$ Moreover, $\varphi$ is a positive, radial solution of the semilinear elliptic equation $$\label{eqn: elliptic} (p-2)\Delta \varphi -(p+2)\varphi + \varphi^{p-1}=0$$ in $\mathbb{R}^n$. It is clear from the definition of $C_{GNS}(n,p)$ in the Theorem that it is the optimal constant in the *Gagliardo-Nirenberg-Sobolev* interpolation inequality: $$\label{eqn: GNSdisp} \|u\|_{L^p(\mathbb{R}^n)}^p \le C_{GNS}(n,p)\|\nabla f\|^{\frac{n(p-2)}{2}}_{L^2(\mathbb{R}^n)}\|f\|^{2+\frac{p-2}{2}(2-n)}_{L^2(\mathbb{R}^n)}.$$ See [@frank] for a pleasant exposition, including a proof of the uniqueness of positive solutions of , following [@kwong]. The scale invariance of the minimization problem implies that the inequalities hold also on the finite domains $[0,1]$ and $D_1$, with the same optimal constants. We have the following result, adapted from [@LRS Lemma 4.1]. 1. For each $m>0$, there is a constant $C=C(m)$ such that if $2<p\le 6$ and $u\in H^1([0,1])$ is periodic, $$\label{eqn: periodicGNS} \|u\|_{L^p([0,1])}^p \le (C_{GNS}(1,p)+m)\|u'\|_{L^2([0,1])}^{\frac{p-2}{2}}\|u\|_{L^2([0,1])}^{\frac{p+2}{2}}+C(m)\|u\|_{L^2([0,1])}^p.$$ 2. If $2<p\le 4$ and $u\in H^1(D_1)$ vanishes on $\partial D_1$ (that is, $u\in H_0^1(D_1)$), then $$\label{eqn: 2dGNS} \|u\|_{L^p(D_1)}^p \le C_{GNS}(2,p)\|\nabla u\|_{L^2(D_1)}^{\frac{p-2}{2}}\|u\|_{L^2(D_1)}^2.$$ The first part is proved in [@LRS]. For the second part, extend $u$ by zero to $\bar{u}\in H^1(\mathbb{R}^2)$ and apply . Proof in the 1D case ==================== We prove the direct implication. See [@LRS Theorem 2.2 (b)] for the converse. Let $\lambda>0$ and write $$\label{eqn: martin} \begin{split} \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u\|_{L^2}\le K]&= \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u_{\ge 0}\|_{L^p}\le \lambda, \|u\|_{L^2}\le K]\\ &+ \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u_{\ge 0}\|_{L^p}> \lambda, \|u\|_{L^2}\le K]. \end{split}$$ The last expectation in in turn equals $$\begin{aligned} &\mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u_{\ge 0}\|_{L^p}> \lambda, \|u_{\ge 1}\|_{L^p}\le \lambda, \|u\|_{L^2}\le K]\\ +~& \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u_{\ge 0}\|_{L^p}> \lambda, \|u_{\ge 1}\|_{L^p}> \lambda, \|u\|_{L^2}\le K].\end{aligned}$$ Continuing this way, we write $$\label{eqn: livius} \begin{split} \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u\|_{L^2}\le K]&= \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u\|_{L^p}\le\lambda, \,\|u\|_{L^2}\le K]\\ &+\sum_{k=1}^\infty \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, E_k, \,\|u\|_{L^2}\le K]. \end{split}$$ In we have set $$\begin{aligned} E_k&=\{\|u\|_{L^p}>\lambda, \ldots, \|u_{\ge k-1}\|_{L^p}>\lambda, \|u_{\ge k}\|_{L^p}\le \lambda \}\\ &\subset \{\|u_{\ge k-1}\|_{L^p}>\lambda, \|u_{\ge k}\|_{L^p}\le \lambda \}.\end{aligned}$$ Since $\sum_{k=1}^\infty \mathbf{1}_{E_k}\uparrow \mathbf{1}_{\|u\|_{L^p}>\lambda}$, by monotone convergence it suffices to show that $$\label{eqn: andre} \mathbf{E}[e^{\frac{1}{p}\int_0^1 |u(x)|^p\,\mathrm{d}x}, \|u_{\ge k-1}\|_{L^p}>\lambda, \|u_{\ge k}\|_{L^p}\le \lambda, \,\|u\|_{L^2}\le K]$$ is summable in $k$. Assuming $p$ is an integer, we have $$\begin{aligned} u&=u_{\le k-1}+u_{\ge k },\\ |u_{\le k-1}+u_{\ge k}|^p&\le \sum_{l=0}^p \binom{p}{l} |u_{\le k-1}|^{p-l}|u_{\ge k}|^l.\end{aligned}$$ Integrating and using Hölder’s inequality we have, if $\|u\|_{L^p}\le \lambda$, $$\label{eqn: ezekiel} \int_0^1|u|^p\,\mathrm{d}x \le \int_0^1|u_{\le k-1}|^p\,\mathrm{d}x+\sum_{l=1}^p \binom{p}{l} \|u_{\le k-1}\|_{L^p}^{p-l} \lambda^{l}.$$ Next, by Young’s inequality, we have for any $1\le l\le p$ and $\epsilon>0$: $$\|u_{\le k-1}\|_{L^p}^{p-l} \lambda^l \le \frac{p-l}{p}\epsilon \|u_{\le k-1}\|_{L^p}^p+ \frac{l}{p} \epsilon^{- \frac{p-l}{l}}\lambda^p,$$ so the right side of becomes $$\int_0^1|u_{\le k-1}|^p\,\mathrm{d}x \cdot \left(1+ 2^p (p-1)\epsilon\right) + p \lambda^p \epsilon^{-p+1}.$$ Thus the quantity in is $$\label{eqn: titus} \sum_{k=2}^\infty e^{p\lambda^p \epsilon^{-p+1}}\mathbf{E}[e^{\frac{(1+\delta)}{p}\int_0^1 |u_{\le k-1}(x)|^p\, \mathrm{d}x}, \|u_{\ge k-1}\|_{L^p}>\lambda ,\|u\|_{L^2}\le K],$$ with $$\delta =\delta(p) = 2^p(p-1)\epsilon.$$ Letting $\lambda=1$ and using and with $p=6$, for any $m>0$, there is a constant $C(m)$ such that the summands in are now bounded by: $$\label{eqn: lena} e^{C(m)K^p}\mathbf{E}[e^{\frac{(C_{GNS}(1,6)+m)K^4(1+\delta)}{6}\int_0^1 |u_{\le k-1}'(x)|^2\,\mathrm{d}x},\|u_{\ge k-1}\|_{L^6}>1 ].$$ Using the Hölder inequality, we have $$\label{eqn: haddock} \mathbf{E}[e^{\frac{(C_{GNS}(1,6)+m)K^4(1+\eta)(1+\delta)}{6}\int_0^1 |u_{\le k-1}'(x)|^2\,\mathrm{d}x}]^{1/(1+\eta)}\cdot \mathbf{P}(\|u_{\ge k-1}\|_{L^6}> 1)^{\eta/(1+\eta)}.$$ Applying to the probability in , we obtain the bound: $$\mathbf{P}(\|u_{\ge k-1}\|_{L^6}> 1) \le C^{\frac{\eta}{1+\eta}}\exp\left(-C\frac{\eta}{1+\eta}2^{(1+2/6)k}\right).$$ Replacing $p$ by its value $p=6$, the expectation is bounded by $$\begin{aligned} \mathbf{E}[e^{\frac{(C_{GNS}(1,6)+m)K^4(1+\eta)(1+\delta)}{6}\int_0^1 |u_{\le k-1}'(x)|^2\,\mathrm{d}x}]&= \prod_{1\le j\le 2^{k-1}} \mathbf{E}[ e^{ \frac{(C_{GNS}(1,6)+m)K^4(1+\eta)(1+\delta)}{6}|g_j|^2}]\\ &\le \left(1-2\frac{(C_{GNS}(1,6)+m) K^4 (1+\eta)(1+\delta)}{6}\right)^{-2^{k}}\\ &= \left(1- (K/\|\varphi\|_{L^2(\mathbb{R})})^4 (1+\eta)(1+\delta)\right)^{-2^{k}}.\end{aligned}$$ If $\|\varphi\|_{L^2(\mathbb{R})}> K$, then we can choose $m$, $\epsilon$ and $\eta$ such that $$\frac{K^4}{\|\varphi\|_{L^2(\mathbb{R})}^4}\left(1+\frac{m}{3}\|\varphi\|_{L^2(\mathbb{R})}^4\right)(1+\delta)(1+\eta)< c<1.$$ Then becomes $$C^{\eta/(1+\eta)}\exp(\frac{2^k}{1+\eta}\log\frac{1}{1-c})\exp\left(-C\frac{\eta}{1+\eta}2^{\frac{4}{3}k}\right).$$ It follows that the sum is finite and, in turn, the partition function is finite for any $K<\|\varphi\|_{L^2(\mathbb{R})}$. 2D case ======= We will use the following simple corollary of (the proof of) Fernique’s theorem [@fernique]: There exists a universal constant $c$ such that if $X$ is a Gaussian process with values in a Banach space $B$ with $\mathbf{E}[\|X\|_{B}]<\infty$, then $$\int e^{ c \frac{\|X\|_B^2}{\mathbf{E}[\|X\|_B]^2}}\,\mathrm{d}\mathbf{P} <\infty.$$ In particular, $$\label{eqn: jared} \mathbf{P}(\|X\|_B \ge t\mathbf{E}[\|X\|_B])\le e^{-ct^2}$$ for $t>1$. An explicit computation [@bourgainbulut Eqn. (66)] gives $$\mathbf{E}\big\|\sum_{ 2^{j-1}\le n \le 2^j} \frac{g_n}{z_n}e_n\big\|_{L^4(D_1)}\le C2^{-j/2}.$$ Applying this to $v$ in with suitable $\epsilon_j$ such that $\sum_{j\ge k}\epsilon_j = 1$, we have $$\label{eqn: jessica} \begin{split} \mathbf{P}( \|v_{\ge k}\|_{L^4(D_1)}\ge \lambda) &\le \sum_{j\ge k}\mathbf{P}(\|v_j\|_{L^4(D_1)} \ge \epsilon_j \lambda)\\ &\le \sum_{j\ge k} e^{-c' \epsilon_j^2 \lambda^2 2^j}\\ &\le Ce^{-c'' \lambda^2 2^k}. \end{split}$$ Starting from $Z_{4,K}'$, we reproduce the computations in , , with $p=4$ and $u$ replaced by $v$ defined by , and the integrals over $[0,1]$ replaced by integrals over $D_1$. We find $$\label{eqn: baptiste} Z_{p,K}'=\sum_{k=1}^\infty e^{4\lambda^4 \epsilon^{-3}}\mathbf{E}[e^{\frac{(1+\delta)}{4}\int_{D_1} |v_{\le k-1}(x)|^4\, \mathrm{d}x}, \|v_{\ge k-1}\|_{L^4(D_1)}>\lambda ,\|v\|_{L^2(D_1)}\le K],$$ where $$\delta=\delta(4)=48\epsilon.$$ We apply to find, for $v\in H^1_0(D_1)$, $$\label{eqn: GNS2d} \|v\|^4_{L^4(D_1)}\le C_{GNS}(2,4)\|v\|_{L^2(D_1)}^2\|\nabla v\|_{L^2(D_1)}^2,$$ with $$C_{GNS}(2,4)=2\|\varphi\|^{-2}_{L^2(\mathbb{R}^2)},$$ for the optimal function $\varphi$ in . The expectation in the summands of the right side of are bounded by $$\label{eqn: cassandra} \begin{split} &\mathbf{E}[e^{\frac{C_{GNS(2,4)}(1+\delta)K^2}{4}\int_{D_1}| \nabla v_{\le k-1}(x)|^2\,\mathrm{d}x},\|v_{\ge k-1}\|_{L^4}>\lambda ]\\ \le & \mathbf{E}[e^{\frac{C_{GNS}(2,4)(1+\delta)K^2}{4}\sum_{|j|\le 2^{k-1}}g_j^2} , \|v_{\ge k-1}\|_{L^4}\ge \lambda]. \end{split}$$ We have used: $$\begin{aligned} \int_{D_1} |\nabla v_{k-1}(x)|^2\,\mathrm{d}x &= -\int_{D_1} v_{k-1}(x) \Delta v_{k-1}(x)\,\mathrm{d}x\\ &= \sum_{|j|\le 2^{k-1}} \frac{1}{z_j^2} \int_0^1 e_j(r)\Delta_r e_j(r)r\,\mathrm{d}r\\ &=\sum_{|j|\le 2^{k-1}} g_j^2.\end{aligned}$$ Applying Hölder’s inequality in , we have the bound $$\label{eqn: tara} \mathbf{E}[e^{\frac{C_{GNS}(2,4)(1+\delta)(1+\eta)K^2}{4}\sum_{|j|\le 2^{k-1}}g_j^2}]^{\frac{1}{1+\eta}}\mathbf{P}( \|v_{\ge k-1}\|_{L^4}\ge \lambda)^{\frac{\eta}{1+\eta}}.$$ The first factor can be computed exactly: provided that $$\label{eqn: mathaeus} C(K,\delta,\eta):=C_{GNS}(2,4) \frac{K^2}{2}(1+\delta)(1+\eta) <1,$$ it is finite and equals $$\left(1-\frac{K^2}{\|\varphi\|_{L^2(\mathbb{R}^2)}^2}(1+\delta)(1+\eta)\right)^{-2^{k-1}}.$$ Invoking , we find that is bounded by $$\label{eqn: james} \exp\left(-\frac{2^{k-1}}{1+\eta} \log (1-C(K,\delta,\eta))\right)\exp\left(-\frac{c''\eta}{1+\eta}\lambda^2 2^k\right).$$ For a fixed $K<\|\varphi\|_{L^2(\mathbb{R}^2)}$ we choose $\delta$ and $\eta$ small enough to satisfy . The parameter $\lambda$ remains at our disposal. We choose it large enough that decays super-exponentially for large $k$. The series is then convergent. #### Acknowledgements I would like to thank Tadahiro Oh for pointing out this problem to me, and Nikolay Tzvetkov for some interesting discussions. [10]{} J. Bourgain, *Periodic nonlinear Schrödinger equation and invariant measures*, Comm. Math. Phys., **166**, 1994. E. Carlen, J. Fröhlich, J. Lebowitz, *Exponential Relaxation to Equilibrium for a One-Dimensional Focusing Non-Linear Schrödinger Equation with Noise*, Comm. Math. Phys., **342**, 2016. X. Fernique. *Intégrabilité des vecteurs gaussiens*, C.R. Acad. Sci. Paris, Sér. A-B, **270**, 1970. R. Frank, *Grounds states of semilinear equations*, Lecture notes from *Current topics in Mathematical Physics*, Luminy, 2013. M. K. Kwong, *Uniqueness of positive solutions to $\Delta u-u+u^p=0$ in $\mathbb{R}^n$*, Arch. Ration. Mech. Anal., **105**, 1989. J. Lebowitz, H. Rose and E. Speer, *Statistical Mechanics of the nonlinear Schrödinger equation*, J. Stat. Phys. **50**, 3, 1988. J. Bourgain and A. Bulut, *Almost sure global well-posedness for the radial nonlinear Schödinger equation on the unit ball I: The 2D Case*, Ann. Inst. H. Poincaré Anal. Non Linéaire, **31**, 6, 2014. B. V. Sz. Nagy, *Über Integralgleichungen zwischen einer Funktion und ihrer Ableitung*. Acta Sci. Math. (Szeged) **10**, 1941. N. Tzvetkov, *Invariant measures for the nonlinear Schrödinger equation on the disc*, Dyn. Partial Differ. Equ., **3**, 2006. N. Tzvetkov, *Invariant measures for the defocusing nonlinear Schrödinger equation*, Ann. Inst. Fourier, 2008. M. Weinstein, *Nonlinear Schrödinger equations and sharp interpolation inequalities*, Comm. Math. Phys. , **87**, 1983.

ArXiv

--- abstract: 'High dimensional correlated binary data arise in many areas, such as observed genetic variations in biomedical research. Data simulation can help researchers evaluate efficiency and explore properties of different computational and statistical methods. Also, some statistical methods, such as Monte-Carlo methods, rely on data simulation. Lunn and Davies (1998) proposed linear time complexity methods to generate correlated binary variables with three common correlation structures. However, it is infeasible to specify unequal probabilities in their methods. In this manuscript, we introduce several computationally efficient algorithms that generate high-dimensional binary data with specified correlation structures and unequal probabilities. Our algorithms have linear time complexity with respect to the dimension for three commonly studied correlation structures, namely exchangeable, decaying-product and $K$-dependent correlation structures. In addition, we extend our algorithms to generate binary data of general non-negative correlation matrices with quadratic time complexity. We provide an R package, CorBin, to implement our simulation methods. Compared to the existing packages for binary data generation, the time cost to generate a $100$-dimensional binary vector with the common correlation structures and general correlation matrices can be reduced up to $10^5$ folds and $10^3$ folds, respectively, and the efficiency can be further improved with the increase of dimensions. The R package CorBin is available on CRAN at https://cran.r-project.org/.' author: - | Wei Jiang[^1]\ Department of Biostatistics, School of Public Health, Yale University, New Haven, CT, US\ and\ Shuang Song\ Center for Statistical Science, Tsinghua University, Beijing 100084, China\ Department of Industrial Engineering, Tsinghua University, Beijing 100084, China\ and\ Lin Hou\ Center for Statistical Science, Tsinghua University, Beijing 100084, China\ Department of Industrial Engineering, Tsinghua University, Beijing 100084, China\ and\ Hongyu Zhao[^2]\ Department of Biostatistics, School of Public Health, Yale University, New Haven, CT, US bibliography: - 'refs.bib' title: '**A set of efficient methods to generate high-dimensional binary data with specified correlation structures**' --- \#1 0 [0]{} 1 [0]{} [**A set of efficient methods to generate high-dimensional binary data with specified correlation structures**]{} [*Keywords:*]{} high-dimensional correlated binary data; simulation; computational efficiency; exchangeable; decaying-product; stationary dependent Introduction {#sec:intro} ============ Binary data are commonly observed in many research areas, such as social survey responses, marketing data concerning specific issues with “yes/no” questions, responses to treatments in clinical trials, and measurements of genetic or epigenetic variations among individuals. Usually, multiple variables are collected from an individual in these studies, and correlations among these variables are ubiquitous. Instead of considering each variable individually, it is essential to model the correlated variables together as a multi-dimensional vector [@cox1972analysis; @carey1993modelling]. With the improvements of technologies over the past decade, more and more variables are collected and constitute to high-dimensional data. One example is the genomic data from biomedical research. Through high-throughput genotyping [@kennedy2003large] and sequencing technologies [@metzker2010sequencing], millions of genetic variants of an individual can be collected simultaneously. It is well known that nearby genetic variants are often highly correlated, which is known as linkage disequilibrium [@pritchard2001linkage]. Hence we cannot simply regard each variant as an independent variable, and the variants constitute a high-dimensional correlated binary vector. Other examples of high-dimensional data include community DNA fingerprints data [@wilbur2002variable], binary questionnaire data [@fieuws2006high], consumer financial history data [@diwakar2009data] and consumer behavior data [@naik2008challenges]. Data simulation can help researchers evaluate the performance of a specific method when real data are difficult to access and/or there is no ground truth about the underlying model/mechanism. We can explore the properties of the method through analyzing simulated data, such as exploring the small-sample properties of generalized estimating equations (GEE) [@hardin2002generalized]. In addition, many statistical estimation methods rely on the generation of random numbers, such as Monte-Carlo methods. With an ever increasing dimension in real world problems, it is essential to develop efficient simulation methods to generate high-dimensional data. In this article, we focus on how to efficiently generate high-dimensional binary data with specified marginal probabilities and correlation structures. To generate random numbers, a direct method is to express the probability moment function and generate random samples with probabilities equal to the function values. Bahadur established a parametric model expressing the joint mass function of binary variables with high-order correlations, but the distribution becomes computationally infeasible when the dimension is high [@bahadur1959representation]. During the 1990s, a number of more efficient methods were proposed to simulate multi-dimensional binary data. Emrich and Piedmonte first identified a multivariate normal distribution whose pairwise copulas are linked with the specified correlation matrix [@emrich1991method]. The binary variables satisfying marginal and correlation conditions can be subsequently obtained by dichotomizing the identified distribution. This algorithm has been realized in an R package, bindata [@leisch1998generation]. One drawback of this method is that the multivariate normal distribution for a given correlation structure may not exist. In addition, a large number of non-linear equations with numerical integrations need to be solved in the algorithm. Lee proposed two algorithms based on linear programming to generate binary data with the exchangeable correlation structure [@lee1993generating]. However, even with the structure assumption, a large number of non-linear equations are still needed. Gange proposed to use iterative proportional fitting approach to generate multi-dimensional categorical variables, which is more general than binary variables [@gange1995generating]. The extra iterative procedure makes the method computationally inefficient. This method has been implemented in the R package MultiOrd [@demirtas2006method]. Similar with Emrich and Piedmonte’s idea, Park et al. proposed to generate correlated binary data by dichotomizing correlated Poisson variables [@park1996simple]. The Poisson variables can be generated efficiently by summing independent Poisson variables, which may be shared among dimensions. Park et al.’s method is very efficient when the dimension is low. Nevertheless, when the dimension grows high, it performs slower than the Emrich and Piedomonte’s method. All the methods above are computationally inefficient when the dimension gets high. Instead of generating binary data with correlations specified by an arbitrary positive semi-definite matrix, Lunn and Davies proposed algorithms to generate binary data with three common correlation structures, including exchangeable, decaying-product and $1$-dependent [@lunn1998note]. The algorithms have linear time complexity with respect to the dimension. However, these algorithms only work when all the marginal probabilities are equal, which limits the applicability of these algorithms. Efficient simulating binary variables with unequal probabilities is critical in many applications. For example, let us assume we want to simulate single nucleotide polymorphism (SNP) data, most of which are biallelic [@sachidanandam2001map], in order to study some statistical behavior of a certain genomic method under different correlation structures. Since the allele frequencies (marginal probabilities) of different SNPs are naturally unequal, it will be unrealistic for us to model them with equal allele frequencies. Therefore, we may need to generate correlated binary data with a certain correlation structure and unequal marginal probabilities. The binary variables with unequal probabilities and specified correlation structures are also common in longitudinal study. A study on comparing two treatments for a common toenail infection is provided in Section 1.5.3 and Section 7.1 of [@shults2014quasi], where 294 patients were randomized to one of two treatments, and measured at various time points (baseline, 1, 2, 3, 6, 9, and 12 months post baseline) to determine the presence or absence of a severe toenail infection. Hence, the outcome variables are dichotomous. They assumed an exchangeable correlation structure in each group of same treatment, and the probabilities of each time points are obviously unequal. In addition, this kind of data are commonly used in the GEE method, where the working correlation matrix are often specified as the forms we mentioned in the paper, and the probabilities are not necessarily equal. In this paper, we generalize Lunn and Davies’ algorithms to generate high-dimensional correlated binary data with varied marginal probabilities. In line with their work, we first focus on three commonly used correlation structures including exchangeable, decaying-product, as well as $K$-dependent correlations. Our algorithms have linear time complexity with respect to the dimension. Besides, we generalize the method on $K$-dependent structure and extend the applicability to general non-negative correlation matrices. Although the time complexity has been augmented to quadratic, the algorithm is still extremely efficient compared to existing approaches. We combine and implement these algorithms in an R package CorBin, which is publicly available on the Comprehensive R Archive Network (CRAN). Compared with existing binary data generation packages, our package spends around $\frac{1}{700,000}$ of the time used in bindata and $\frac{1}{320,000}$ of MultiOrd when generating a 100-dimensional binary data with exchangeable structure. The ratios reach around $\frac{1}{4,500,000}$ and $\frac{1}{2,060,000}$ when the dimension increases to 500. Similar speed up is also observed for the other two correlation structures. For generating data with general correlation matrices, the ratios become around $\frac{1}{680}$ and $\frac{1}{320}$ of bindata and MultiOrd, respectively. The package is easy to use, and a pdf document ([`CorBin-manual.pdf`]{}, Supplementary Material) is provided to illustrate the usage for readers. Models and algorithms {#sec:meth} ===================== In this section, we propose algorithms for generating an $m$-dimensional random binary vector $\bm{X} = (X_1, X_2, \cdots, X_m)'$ with several correlation structures, where $X_i$ follows the Bernoulli distribution with marginal probability $p_i$, $i=1,2,\cdots,m.$ Compared with algorithms proposed in Lunn and Davies (1998), here we do not require all marginal probabilities to be equal, which extends the applicability of the algorithms. In the following, we denote the correlation between $X_i$ and $X_j$ as $r_{ij}$ ($1\leq i\leq j \leq n$). We assume the correlations are non-negative, i.e. $r_{ij}\geq 0$, and use $\bm{R}$ to represent the correlation matrix constituted from $r_{ij}$. We also assumed all of the Bernoulli random variables used in each algorithm are generated independently. Natural restrictions for correlated binary variables {#sec:restiction} ---------------------------------------------------- In this subsection, we present the natural restrictions for any correlated binary variables must satisfy. First, the correlation matrix $\bm{R}$ should be positive definite. This imposes the restrictions on the correlation coefficients within $\bm{R}$ matrices. Taking the $\bm{R}$ matrix presented in equation (Section 2.4) as an example, $\rho_1$ must take values in $(-\frac{1}{c_m},\frac{1}{c_m})$ to satisfy the positive definite restriction, where $c_m=2sin(\frac{\pi(m-1)}{2(m+1)})$ [@shults2014quasi]. When $m \to +\infty$, the constraints will approach $(-0.5,0.5)$. Second, natural constraints of $\bm{R}$ are also imposed by the marginal expectations. Let’s consider a simple bivariate example, in which $p_1, p_2=0.1, 0.4$, and $\rho=0.9$. The corresponding correlation matrix is obviously positive definite. However, $P(Y_1=1,Y_2=1)=0.1\times 0.4+0.9\sqrt{0.1\times 0.4\times (1-0.1)\times (1-0.4) }=0.172$; $P(Y_1=1,Y_2=0)=P(Y_1=1)-P(Y_1=1,Y_2=1)=0.1-0.172=-0.072<0$, which leads to an invalid probability. Prentice formulated the constraints imposed for marginal expectations and correlation coefficients, which are known as the Prentice constraints [@prentice1988correlated]. Specifically, under multivariate binary distribution, the correlation between any two dimensions $i, j$ should satisfy $$\max\left\{-\sqrt{\frac{p_i(1-p_j)}{p_j(1-p_i)}}, -\sqrt{\frac{p_j(1-p_i)}{p_i(1-p_j)}}\right\}\leq r_{ij}\leq \min\left\{\sqrt{\frac{p_i(1-p_j)}{p_j(1-p_i)}}, \sqrt{\frac{p_j(1-p_i)}{p_i(1-p_j)}}\right\}.$$ For the non-negative correlation structure, each element of the correlation matrix should satisfy $r_{ij}\in \left[0,\min\left\{\sqrt{\frac{p_i(1-p_j)}{p_j(1-p_i)}}, \sqrt{\frac{p_j(1-p_i)}{p_i(1-p_j)}}\right\}\right]$, based on the Prentice constraints. In our R package, we will first check whether the input parameters satisfy the Prentice constraints. If not, we will print out a warning message with exact values of the Prentice constraints for users. Please note that the Prentice constraints and the positive definiteness are the necessary but insufficient conditions to guarantee the existence of the multivariate binary distributions with the specified correlation structures. Actually, as we later presented in Section 2.2 and 2.3, for exchangeable and decaying-product correlation structure, the Prentice constraints are sufficient to guarantee the existence of the distributions. However, it is not true for all correlation structures. For example, [@chaganty2006range] presented a simple example with a 1-dependent correlation structure (detailed in Section 2.4), in which the Prentice constraints were satisfied but the corresponding distribution is not valid. Further, Shults and Hilbe provided a brief review on additional constraints for correlated binary data [@shults2014quasi]. In following subsections, we provide implementation details and related properties of algorithms to generate binary data with different correlation structures. Exchangeable correlation structure {#exchange} ---------------------------------- The exchangeable correlation structure is one of the most commonly used structures in data simulation, and is regarded as the default setting in most binary data generation packages. In this case, every pair of observations on a specific unit has the same correlation, i.e., $$r_{ij}=\left\{ \begin{aligned} & \rho , \quad i \neq j \\ & 1 , \quad i = j\\ \end{aligned} \right..$$ The correlation matrix of the exchangeable correlation structures is: $$\label{corrmatrixEX} \bm{R}={ \left( \begin{array}{ccccc} 1 & \rho & \rho &\cdots&\rho\\ \rho & 1 & \rho &\cdots&\rho\\ \rho & \rho & 1&\cdots&\rho\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ \rho&\rho&\rho&\cdots&1 \end{array} \right )}.$$ We use $p_{min}$ and $p_{max}$ to denote the minimal and maximal values in the desired marginal probabilities, i.e., $$p_{min}=\min\limits_{k\in\{1,\cdots,m\}}{p_k},\quad p_{max}=\max\limits_{k\in\{1,\cdots,m\}}{p_k}.$$ An intuitive thinking to generate binary data with exchangeable structure is to make each variable $X_i$ taking the linear combination form $(1-U_i)Y_i+U_iZ$, where $U_i\sim Bern(\alpha_i)$, $Y_i\sim Bern(\beta_i)$, $Z\sim Bern(\gamma)$, and all of the random variables are mutually independent. A careful selection of $\alpha_i$, $\beta_i$ and $\gamma$ is needed to make the constructed variables having specified marginal probabilities and correlations. We describe the constructions in detail in Algorithm \[alg1\]. In the following, we first show the justification of the algorithm, i.e., if $\alpha_i$, $\beta_i$ and $\gamma$ lies in $\left[0,1\right]$, the binary data generated from Algorithm \[alg1\] have specified marginal expectations and exchangeable correlation structure. After that, we prove that if the data to be generated satisfy the necessary condition of their existence, namely the Prentice constraints, the construction of Algorithm \[alg1\] can guarantee that the parameters lie in $\left[0,1\right]$. The expected values of the Bernoulli random variables $p_1,p_2,\cdots,p_m$; and the correlation coefficient $\rho$\ The correlated binary variables $X_1,X_2,\cdots,X_m$ Check whether the input satisfies the Prentice constraints. $\gamma=\frac{\sqrt{p_{min}p_{max}}}{\sqrt{p_{min}p_{max}}+\sqrt{(1-p_{min})(1-p_{max})}}$ Generate $Z\sim Bern(\gamma)$ $\alpha_i=\sqrt{\frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}}$ $\beta_i=\frac{p_i-\alpha_i \gamma}{1-\alpha_i}$ Generate $U_i\sim Bern(\alpha_i)$ Generate $Y_i\sim Bern(\beta_i)$ $X_i=(1-U_i)Y_i+U_iZ$ $X_1,X_2,\cdots,X_m$ \[exTHM\_cr\] If intermediate variables $\alpha_1$, $\cdots$, $\alpha_m$, $\beta_1$, $\cdots$, $\beta_m$, $\gamma \in \left[0,1\right]$, Algorithm \[alg1\] returns the binary data $X_1, \cdots,X_m$ with marginal probabilities $p_1,\cdots,p_m$ and common correlation $\rho$. $X_i$ is a binary variable with the support {0,1}. For any $i\in \{1,2,\cdots,m\}$, $$\label{a1} \begin{aligned} EX_i&=E((1-U_i)Y_i+U_iZ)=E(1-U_i)EY_i+EU_iEZ\\ &=(1-\alpha_i)\beta_i+\alpha_i\gamma=(1-\alpha_i)\frac{p_i-\alpha_i \gamma}{1-\alpha_i}+\alpha_i \gamma=p_i, \end{aligned}$$ For any $i\neq j \in \{1,2,\cdots,m\}$ $$\label{a2} \begin{aligned} r_{ij}&=\frac{cov(X_i,X_j)}{\sqrt{Var(X_i)Var(X_j)}} =\frac{cov((1-U_i)Y_i+U_iZ,(1-U_j)Y_j+U_jZ)}{\sqrt{p_ip_j(1-p_i)(1-p_j)}} \\ &=\frac{cov(U_iZ,U_jZ)}{\sqrt{p_ip_j(1-p_i)(1-p_j)}} =\frac{\alpha_i\alpha_j\gamma(1-\gamma)}{\sqrt{p_ip_j(1-p_i)(1-p_j)}}=\rho. \end{aligned}$$ Thus, the algorithm returns the variables satisfying the required marginal and correlation conditions we provide. In the following theorem, we show that if the specified binary data with non-negative exchangeable structure satisfies the Prentice constraints, the probability parameters $\alpha_i$, $\beta_i$ and $\gamma$ will automatically lie in the range of $\left[0,1\right]$. \[exTHM\] If the specified binary data with non-negative exchangeable correlation structure satisfy the Prentice constraints, the constructions in Algorithm \[alg1\] guarantee that $\alpha_1$, $\cdots$, $\alpha_m$, $\beta_1$, $\cdots$, $\beta_m$, $\gamma$ lie in the range of $\left[0, 1\right]$. By definition, $$\label{eq:thm22_1} \gamma=\frac{\sqrt{p_{min}p_{max}}}{\sqrt{p_{min}p_{max}}+\sqrt{(1-p_{min})(1-p_{max})}}.$$ Since $0 \leq \sqrt{p_{min}p_{max}} \leq \sqrt{p_{min}p_{max}}+\sqrt{(1-p_{min})(1-p_{max})}$, we have $0< \gamma< 1$. Meanwhile, $$\label{thm2.2_2} \gamma(1-\gamma)=\frac{\sqrt{p_{min}p_{max}(1-p_{min})(1-p_{max})}}{(\sqrt{p_{min}p_{max}}+\sqrt{(1-p_{min})(1-p_{max})})^2}.$$ From $Cauchy$-$Schwarz$ inequality, $$\label{cs} (\sqrt{p_{min}p_{max}}+\sqrt{(1-p_{min})(1-p_{max})})^2\leq (p_{min}+1-p_{min})(p_{max}+1-p_{max})=1.$$ According to and , we have $\gamma(1-\gamma)\geq\sqrt{p_{min}p_{max}(1-p_{min})(1-p_{max})}.$ For any $i \in \{1,2,\cdots,m\}$, $\alpha_i$ is obviously non-negative. With the Prentice constraints, we have $$\label{pren.ex} \rho\leq \sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}.$$ Therefore, $$\begin{aligned} \rho p_i(1-p_i) &\leq \rho p_{max}(1-p_{min})\leq \sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}p_{max}(1-p_{min})\\ &=\sqrt{p_{min}p_{max}(1-p_{min})(1-p_{max})}\leq \gamma(1-\gamma). \end{aligned}$$ Hence, we have $$\label{eq:thm22_2} \alpha_i=\sqrt{\frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}}\leq 1.$$ On the other hand, due to the fact that $p_{min}\leq p_i\leq p_{max}$, $$\begin{aligned} &\frac{1-p_{max}}{p_{max}}\leq\frac{1-p_i}{p_i} \iff \sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}\leq \frac{1-p_i}{p_i}\frac{\sqrt{p_{min}p_{max}}}{\sqrt{(1-p_{min})(1-p_{max})}}=\frac{1-p_i}{p_i}\frac{\gamma}{1-\gamma},\\ \text{and}\\ &\frac{p_{min}}{1-p_{min}}\leq\frac{p_i}{1-p_i} \iff \sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}\leq \frac{p_i}{1-p_i}\frac{\sqrt{(1-p_{min})(1-p_{max})}}{\sqrt{p_{min}p_{max}}}=\frac{p_i}{1-p_i}\frac{1-\gamma}{\gamma}, \end{aligned}$$ i.e., $$\sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}\leq \min\{\frac{1-p_i}{p_i}\frac{\gamma}{1-\gamma},\frac{p_i}{1-p_i}\frac{1-\gamma}{\gamma}\}.$$ Combined with inequality , $$\label{eq:thmpr1} \begin{aligned} \rho\leq \sqrt{\frac{p_{min}(1-p_{max})}{p_{max}(1-p_{min})}}\leq \min\{\frac{1-p_i}{p_i}\frac{\gamma}{1-\gamma},\frac{p_i}{1-p_i}\frac{1-\gamma}{\gamma}\}. \end{aligned}$$ Then we can derive that $$\label{eq:thmpr2} \begin{aligned} \rho\leq\frac{1-p_i}{p_i}\frac{\gamma}{1-\gamma} \iff& \frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}\leq \frac{(1-p_i)^2}{(1-\gamma)^2} \iff \alpha_i=\sqrt{\frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}}\leq\frac{1-p_i}{1-\gamma}\\ \iff& p_i-\alpha_i\gamma\leq 1-\alpha_i \iff \beta_i=\frac{p_i-\alpha_i\gamma}{1-\alpha_i}\leq1. \end{aligned}$$ Similarly, $$\label{eq:thmpr3} \begin{aligned} \rho\leq\frac{p_i}{1-p_i}\frac{1-\gamma}{\gamma} \iff& \frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}\leq\frac{p_i^2}{\gamma^2} \iff \alpha_i=\sqrt{\frac{\rho p_i(1-p_i)}{\gamma(1-\gamma)}}\leq\frac{p_i}{\gamma}\\ \iff& p_i-\alpha_i\gamma\geq 0 \iff \beta_i\geq 0. \end{aligned}$$ Combining , , and , we finish the proof. Decaying-product correlation structure {#auto} -------------------------------------- The commonly used first order autoregressive (AR(1)) correlation structure is a special case of the decaying-product correlation structure, where the correlations are highest for adjacent variables and decrease in the power of the distance between dimension indices. The correlation matrix of AR(1) is as follows: $$\bm{R}={ \left( \begin{array}{ccccc} 1 & \rho & \rho^2 &\cdots&\rho^{n-1}\\ \rho & 1 & \rho &\cdots&\rho^{n-2}\\ \rho^2 & \rho & 1&\cdots&\rho^{n-3}\\ \vdots&\vdots&\vdots&\ddots&\vdots\\ \rho^{n-1}&\rho^{n-2}&\rho^{n-3}&\cdots&1 \end{array} \right )}.$$ Here we consider the more general decaying-product correlation structure. We allow the marginal probabilities to be freely specified. Given the elements on the minor diagonal of the correlation matrix $\bm{\rho}=(\rho_1,\cdots,\rho_{n-1})$, the correlation between any two variables under this structure can be expressed as: $$r_{j,k}=\prod_{l=j}^{k-1}\rho_{l},\quad j<k,$$ i.e., $$\bm{R}={ \left( \begin{array}{cccccc} 1 & \rho_1 & \rho_1\rho_2 &\rho_1\rho_2\rho_3 &\cdots&\prod_{l=1}^{n-1}\rho_{l}\\ \rho_1 & 1 & \rho_2 &\rho_2\rho_3 &\cdots&\prod_{l=2}^{n-1}\rho_{l}\\ \rho_1\rho_2 & \rho_2 & 1&\rho_3&\cdots&\prod_{l=3}^{n-1}\rho_{l}\\ \rho_1\rho_2\rho_3 & \rho_2\rho_3 &\rho_3& 1&\cdots&\prod_{l=4}^{n-1}\rho_{l}\\ \vdots&\vdots&\vdots&\vdots&\ddots&\vdots\\ \prod_{l=1}^{n-1}\rho_{l}&\prod_{l=2}^{n-1}\rho_{l}&\prod_{l=3}^{n-1}\rho_{l}&\prod_{l=4}^{n-1}\rho_{l}&\cdots&1 \end{array} \right )}.$$ Similar with the construction method in the previous subsection, we also assume the variable, $X_i$ takes the linear combination form $(1-U_i)Y_i+U_iX_{i-1}$, where $U_i\sim Bern(\alpha_i)$, $Y_i\sim Bern(\beta_i)$, and all of the random variables are mutually independent. Here the construction of $X_i$ uses the information of $X_{i-1}$, bringing the correlation between adjacent elements. Similar to the exchangeable correlation structure, we make a subtle construction of $\alpha_i$ and $\beta_i$ in order that the constructed variables have specified marginal probability and correlations. We describe the constructions in detail in Algorithm \[alg2\]. In the following, we will first show the justification of the algorithm in Theorem \[arTHM\_cr\] and prove that the Prentice constraints are enough to guarantee the intermediate parameters in the interval $\left[0,1\right]$ in Theorem \[arTHM\]. The expected values of the Bernoulli random variables $p_1,p_2,\cdots,p_m$; and the off-diagonal correlation vector $\bm{\rho}$\ The correlated binary variables $X_1,X_2,\cdots,X_m$ Check whether the input satisfies the Prentice constraints. Generate $X_1\sim Bern(p_1)$ $\alpha_i=\rho_{i-1}\sqrt{\frac{p_i(1-p_i)}{p_{i-1}(1-p_{i-1})}}$ $\beta_i=\frac{p_i-\alpha_i p_{i-1}}{1-\alpha_i}$ Generate $U_i\sim Bern(\alpha_i)$ Generate $Y_i\sim Bern(\beta_i)$ $X_i = (1-U_i)Y_i+U_iX_{i-1}$ $X_1,X_2,\cdots,X_m$ \[arTHM\_cr\] If intermediate variables $\alpha_2$, $\cdots$, $\alpha_m$, $\beta_2$, $\cdots$, $\beta_m \in \left[0,1\right]$, Algorithm \[alg2\] returns the binary data with the specified marginal probabilities and decaying-product correlation. In Algorithm \[alg2\], $EX_1=p_1$ is naturally satisfied. Now we prove $EX_i=p_i$ ($i=2,\cdots,m$) by induction. For any $i \in \{2,\cdots,m \}$, assuming that $EX_{i-1}=p_{i-1}$, then $$EX_i=E((1-U_i)Y_i+U_iX_{i-1})=(1-\alpha_i)\beta_i +\alpha_iEX_{i-1} =p_i-\alpha_ip_{i-1}+\alpha_ip_{i-1}=p_i.$$ For any $k \in \{1,2,\cdots,m-1\}$ and $i \in \{1,2,\cdots,m-k\}$, $$\begin{aligned} r_{i,i+k}&=\frac{cov(X_i,X_{i+k})}{\sqrt{Var(X_i)Var(X_{i+k})}} =\frac{cov(X_i,U_{i+k}U_{i+k-1}\cdots U_{i+1}X_{i})}{\sqrt{p_ip_{i+k}(1-p_i)(1-p_{i+k})}}\\ &=\alpha_{i+k}\cdots\alpha_{i+1}\sqrt{\frac{p_i(1-p_i)}{p_{i+k}(1-p_{i+k})}}\\ &=\rho_{i}\rho_{i+1}\cdots\rho_{i+k-1}. \end{aligned}$$ Thus the decaying-product correlation structure holds. \[arTHM\] If the specified binary data with non-negative decaying-product correlation structure satisfy the Prentice constraints, the constructions in Algorithm \[alg2\] guarantee that $\alpha_2$, $\cdots$, $\alpha_m$, $\beta_2$, $\cdots$, $\beta_m$ lie in the range of $\left[0,1\right]$. For each $i \in \{2,\cdots,m\}$, we show $0\leq \alpha_i,\beta_i\leq 1$ in two complementary cases: When $p_{i-1}\leq p_i$, according to the Prentice constraints, we have $$0\leq \alpha_i=\rho_{i-1}\sqrt{\frac{p_i(1-p_i)}{p_{i-1}(1-p_{i-1})}}\leq \sqrt{\frac{p_i(1-p_i)}{p_{i-1}(1-p_{i-1})}}\sqrt{\frac{p_{i-1}(1-p_{i})}{p_{i}(1-p_{i-1})}}=\frac{1-p_i}{1-p_{i-1}}\leq1.$$ Thus, $$\begin{aligned} (1-p_{i-1})\alpha_i\leq1-p_i\iff p_i-\alpha_i p_{i-1}\leq1-\alpha_i\iff \beta_i=\frac{p_i-\alpha_i p_{i-1}}{1-\alpha_i}\leq1. \end{aligned}$$ At the same time, $$\beta_i=\frac{p_i-\alpha_i p_{i-1}}{1-\alpha_i}\geq\frac{p_i-\alpha_i p_i}{1-\alpha_i}=p_i\geq0.$$ Then we have $0\leq\beta_i\leq1$. In the other case, i.e., $p_{i-1}> p_i$, according to the Prentice constraints, $$0\leq \alpha_i=\rho_{i-1}\sqrt{\frac{p_i(1-p_i)}{p_{i-1}(1-p_{i-1})}}\leq \sqrt{\frac{p_i(1-p_i)}{p_{i-1}(1-p_{i-1})}}\sqrt{\frac{p_{i}(1-p_{i-1})}{p_{i-1}(1-p_{i})}}=\frac{p_i}{p_{i-1}}\leq1.$$ Then we have $$\begin{aligned} p_i\geq\alpha_i p_{i-1}\iff \beta_i=\frac{p_i-\alpha_i p_{i-1}}{1-\alpha_i}\geq 0. \end{aligned}$$ Meanwhile, $$\beta_i=\frac{p_i-\alpha_i p_{i-1}}{1-\alpha_i}\leq\frac{p_i-\alpha_i p_{i}}{1-\alpha_i}=p_i\leq1.$$ At this point, we have proved that, with the Prentice constraints satisfied, $0\leq \alpha_i,\beta_i\leq 1$, $(i = 2,\cdots,m)$ hold for any binary data with non-negative decaying-product correlation structure, indicating they can be generated by Algorithm \[alg2\]. 1-dependent correlation structure {#statm} --------------------------------- Under the stationary $K$-dependent structure, there is a band of stationary correlations, such that each of the correlation is truncated to zero after the $K$-th order band [@hardin2002generalized]. The correlation coefficient is: $$r_{ij}=\left\{ \begin{aligned} &\rho_{\left|i-j\right|},& \quad &\text{if} \ i\neq j \ \text{and}\ \left|i-j\right|\leq K \\ &0,& \quad &\text{if}\ \left|i-j\right|> K\\ & 1 ,& \quad &\text{if}\ i = j\\ \end{aligned} \right..$$ As the most common case, the stationary $1$-dependent correlation matrix can be expressed as: $$\label{1dep} \bm{R}={ \left( \begin{array}{cccccc} 1 & \rho_{1} & 0 &\cdots&0&0\\ \rho_{1} & 1 & \rho_{1} &\cdots&0& 0\\ 0 & \rho_{1} & 1&\cdots&0& 0\\ \vdots&\vdots&\vdots&\ddots&\vdots&\vdots\\ 0&0&0&\cdots&1&\rho_{1}\\ 0 & 0& 0 &\cdots&\rho_{1}&1 \end{array} \right )},$$ where the elements on the minor diagonal are equal. To expand the applicability of our method, we design the algorithms on a more general 1-dependent case, allowing the elements on the minor diagonal to vary, i.e., $\bm{\rho_1}=(\rho_{11},\rho_{12},\cdots,\rho_{1,m-1})$. For simplicity, we use $\bm{\rho}$ to represent $\bm{\rho_1}$ in this section. In the main context, we focus on the generation of binary data with $1$-dependent correlation structure. We introduce two algorithms with different applicable conditions. Both algorithms allow marginal probabilities to vary. We will describe the details of applicable conditions separately after introducing the corresponding algorithms. Intuitively, we intend to construct the variables that have correlation between adjacent elements, but independent with those not adjacent, in order to satisfy the $1$-dependent correlation structure. Our Algorithm \[alg3\] utilizes the form of $X_i=U_iY_iY_{i-1}$ to guarantee the independence of $X_j$ and $X_k$ when $|j-k|\geq 2$. In addition, the intermediate parameters are also important to make sure that the correlation coefficients, and the marginal probabilities are satisfied. We present the details in Algorithm \[alg3\]. The expected values of the Bernoulli random variables $p_1,p_2,\cdots,p_m$; and the correlation coefficient vector $\bm{\rho}=(\rho_1,\rho_2,\cdots,\rho_{m-1})$\ The correlated binary variables $X_1,X_2,\cdots,X_m$ Check whether the input satisfies the Prentice constraints. $\beta_0=1$ $Y_0 \sim Bern(1)$ $\beta_i=\frac{\sqrt{p_ip_{i+1}}}{\sqrt{p_ip_{i+1}}+\rho_i\sqrt{(1-p_i)(1-p_{i+1})}}$ $\alpha_i=\frac{p_i}{\beta_i\beta_{i-1}}$ **if** $\alpha_i \textgreater 1$: Prompt the unavailability; **break** Generate $U_i\sim Bern(\alpha_i)$ Generate $Y_i\sim Bern(\beta_i)$ $X_i = U_iY_iY_{i-1}$ $\alpha_m=\sqrt{\frac{p_m}{\beta_{m-1}}}$ $\beta_m=\sqrt{\frac{p_m}{\beta_{m-1}}}$ Generate $U_m\sim Bern(\alpha_m)$ Generate $Y_m\sim Bern(\beta_m)$ $X_m=U_mY_mY_{m-1}$ $X_1,X_2,\cdots,X_m$ \[kdepTHM\_cr\] If intermediate variables $\alpha_1$, $\cdots$, $\alpha_m$, $\beta_1$, $\cdots$, $\beta_m \in \left[0,1\right]$, Algorithm \[alg3\] returns the binary data with given marginal expectation and 1-dependent correlation structure. For each $i \in \{1,2,\cdots,m-1\}$, $$EX_i=EU_iY_iY_{i-1}=\alpha_i\beta_i\beta_{i-1}=\frac{p_i}{\beta_i\beta_{i-1}}\beta_i\beta_{i-1}=p_i.$$ Meanwhile, $$EX_m=EU_mY_mY_{m-1}=\alpha_m\beta_m\beta_{m-1}=\sqrt{\frac{p_m}{\beta_{m-1}}}\sqrt{\frac{p_m}{\beta_{m-1}}}\beta_{m-1}=p_m.$$ From the generation process, it is obvious that $r_{ij}=0$ when $ \left|j-i\right|>1$. Then for $i \in \{1,2,\cdots, m-1\}$, $$\begin{aligned} r_{i,i+1}=&\frac{cov(X_i,X_{i+1})}{\sqrt{Var(X_i)Var(X_{i+1})}}=\frac{cov(U_iY_iY_{i-1},U_{i+1}Y_{i+1}Y_i)}{\sqrt{p_ip_{i+1}(1-p_i)(1-p_{i+1})}}\\ =&\frac{E(U_iU_{i+1}Y_{i-1}Y_i^2Y_{i+1})-E(U_iY_iY_{i-1})E(U_{i+1}Y_{i+1}Y_i)}{\sqrt{p_ip_{i+1}(1-p_i)(1-p_{i+1})}}\\ =&\frac{(\alpha_{i}\beta_{i-1}\beta_i)(\alpha_{i+1}\beta_i\beta_{i+1})(1-\beta_i)}{\beta_i\sqrt{p_ip_{i+1}(1-p_i)(1-p_{i+1})}} \\=&\frac{(1-\beta_i)\sqrt{p_ip_{i+1}}}{\beta_i\sqrt{(1-p_i)(1-p_{i+1})}}=\rho_i. \end{aligned}$$ Similar to the previous algorithms, the data generated by Algorithm \[alg3\] require intermediate parameters $\alpha_i$ and $\beta_i$ to be in the range of $[0,1]$. Otherwise, intermediate variables $U_i$ and $Y_i$ cannot be generated, nor can $X_i$. In the following theorem, we show the restriction for $\pmb{\rho}$ such that the intermediate parameter requirement is satisfied. \[alg3\_property\] Only binary data with non-negative 1-dependent structure satisfying the following inequalities can be generated from Algorithm \[alg3\]: $$A_i\rho_{i-1}\rho_i+B_{1i}\rho_{i-1}+B_{2i}\rho_i\leq C_i,\quad i=2,\cdots,m-1, \label{alg_inequal}$$ where $A_i=\sqrt{(1-p_{i-1})(1-p_i)(1-p_{i+1})}$, $B_{1i}=\sqrt{(1-p_{i-1})p_ip_{i+1}}$, $B_{2i}=\sqrt{p_{i-1}p_i(1-p_{i+1})}$, and $C_i=\sqrt{p_{i-1}(1-p_i)p_{i+1}}$. When $\rho_1=\rho_2=\dots=\rho_m=\rho$, we denote $A=A_i$, $B=B_{1i}+B_{2i}$ and $C=C_i$, then the inequality can be simplified to $$0\leq\rho_i\leq \frac{-B+\sqrt{B^2+4AC}}{2A}.$$ Further, in the special case when $p_1=p_2=\dots=p_m=p$, the applicable condition is $0\leq\rho\leq\frac{\sqrt{p}}{1+\sqrt{p}}$. Since $\rho_i\geq0$, it is obvious that $\alpha_i\geq 0$ and $0\leq \beta_i\leq 1$ based on the definitions of $\alpha_i$ and $\beta_i$. Now we study the applicable condition for $\rho_i$ such that the intermediate parameter $\alpha_i\leq 1$. For each $i\in \{1,2,\cdots,m-1\}$, we have $$\begin{aligned} &\alpha_i=\frac{p_i}{\beta_i\beta_{i-1}}=\frac{(\sqrt{p_ip_{i+1}}+\rho_{i}\sqrt{(1-p_i)(1-p_{i+1})})(\sqrt{p_{i-1}p_{i}}+\rho_{i-1}\sqrt{(1-p_{i-1})(1-p_{i})})}{\sqrt{p_{i-1}p_{i+1}}}\leq 1, \end{aligned}$$ which is equivalent to inequality . When $\rho_1=\rho_2=\dots=\rho_m=\rho$, denote $A$, $B$ and $C$ as above, and we get $$\begin{aligned} \alpha_i\leq1 \iff A\rho^2+B\rho-C\leq0 \Rightarrow \rho\leq\frac{-B+\sqrt{B^2+4AC}}{2A}. \end{aligned}$$ Thus, given the marginal probabilities of a variable as well as its neighbors’, we can derive the constraint for the corresponding correlation in the Algorithm \[alg3\]. The constraint is not relevant with other correlation coefficients. In the special case when $p_1=p_2=\dots=p_m=p$, the above constraint can be reduced to $\rho\leq\frac{\sqrt{p}}{1+\sqrt{p}}$, which is not related with the dimension $m$ and only related with the marginal probability $p$. Figure \[fig:mdep1\] shows the relationship between the maximal allowed correlation $\rho_{max}$ and the marginal probability $p$. The maximal allowed correlation is monotonically increase with $p$ and converge to 0.5, which is consistent with the natural restrictions for positive definiteness discussed in Section 2.1. ![Relationship between the maximal allowed correlation ($\rho_{max}$) and given marginal probability ($p$) in Algorithm \[alg3\], when $p_1=p_2=\cdots=p_m=p$ and $\rho_1=\rho_2=\cdots=\rho_m=\rho$. The maximal correlation will monotonically increase as $p$ increases, and converge to 0.5.[]{data-label="fig:mdep1"}](Fig_1.eps){width="80.00000%"} For Algorithm \[alg4\], we used the form $W_i=(1-U_i)Y_i+U_iY_{i-1}$ to generate variables with 1-dependent correlation structure, and then multiply $W_i$ with an $A_i$ to adjust the probability. Actually, the idea of multiplying an $A_i$ has been discussed in Lunn and Davies’ paper. As they discussed in the paper, this strategy can only be applied in $1$-dependent correlation structure. Nevertheless, we provide the details of this strategy for generating binary data with $1$-dependent correlation structure in Algorithm \[alg4\]. In the following, we provide the proof of related properties for Algorithm 4. The expected values of the Bernoulli random variables $p_1,p_2,\cdots,p_m$; and the correlation coefficient vector $\bm{\rho}=(\rho_1,\rho_2,\cdots,\rho_{m-1})$\ The correlated binary variables $X_1,X_2,\cdots,X_m$ Check whether the input satisfies the Prentice constraints. $p_{max}=\max\{p_1,\cdots,p_m\}$ Denote $\alpha_i=\frac{p_i}{p_{max}}$, $ i=1,\cdots,m$ Denote $\rho_i'=\rho_i\frac{\sqrt{(1-p_i)(1-p_{i+1})}}{\sqrt{\alpha_i\alpha_{i+1}}(1-p_{max})}$, $i=1,\cdots,m-1$ $Y_1\sim Bern(p_{max})$ $W_1=Y_1$ $r_1=0$ $r_i=\frac{\rho_{i-1}'}{1-r_{i-1}}$ **if** $r_i \textgreater 1$: Prompt the unavailability; **break** Generate $U_i\sim Bern(r_i)$ Generate $Y_i\sim Bern(p_{max})$ $W_i=(1-U_i)Y_i+U_iY_{i-1}$ Generate $A_i\sim Bern(\alpha_i)$ $X_i=A_iW_i$ $X_1,X_2,\cdots,X_m$ If intermediate variables $\alpha_2$, $\cdots$, $\alpha_m$, $r_2$, $\cdots$, $r_m \in \left[0,1\right]$, Algorithm \[alg4\] returns the binary data with given marginal expectation and 1-dependent correlation structure. From the definition, $EX_1=p_1$. For $i$ in $\{2,\cdots,m\}$, we have $$EX_i=EA_iW_i=\alpha_i((1-r_i)p_{max}+r_ip_{max}) =\frac{p_i}{p_{max}}p_{max}=p_i.$$ Then for each $i \in \{1,2,\cdots,m-1\}$, we can obtain that $$\begin{aligned} r_{i,i+1}=&\frac{cov(X_i,X_{i+1})}{\sqrt{Var(X_i)Var(X_{i+1})}}= \frac{\alpha_i \alpha_{i+1}cov((1-U_i)Y_i,U_{i+1}Y_{i})}{\sqrt{p_ip_{i+1}(1-p_i)(1-p_{i+1})}}\\ =&\frac{\alpha_i \alpha_{i+1}(E((1-U_i)U_{i+1}Y_i^2)-E((1-U_i)Y_i)E(U_{i+1}Y_i))}{\sqrt{p_ip_{i+1}(1-p_i)(1-p_{i+1})}}\\ =&\frac{\rho_i'(1-p_{max})\sqrt{p_ip_{i+1}}}{p_{max}\sqrt{(1-p_i)(1-p_{i+1})}} =\frac{\rho_i\sqrt{p_ip_{i+1}}}{p_{max}\sqrt{\alpha_i\alpha_{i+1}}} =\frac{\rho_i\sqrt{p_ip_{i+1}}}{p_{max}\sqrt{\frac{p_ip_{i+1}}{p_{max}^2}}}=\rho_i. \end{aligned}$$ Meanwhile, it is obvious that $r_{i,j}=0$, $|i-j|\geq 2$. We obtain the correctness of the algorithm. The data generated by Algorithm \[alg4\] need the intermediate parameters $\gamma_i$ to locate in $[0,1]$. In the following theorem, we show the restriction for $\pmb{\rho}$ such that the intermediate parameter requirement is satisfied. \[thm2.8\] Only binary data with non-negative 1-dependent correlation structure satisfying the following inequalities can be generated from Algorithm \[alg4\]: $$0\leq\rho_i \leq \sqrt{\frac{p_ip_{i+1}}{(1-p_i)(1-p_{i+1})}}\frac{1-p_{max}}{p_{max}}(1-r_{i}),$$ where $r_i=\frac{\rho_{i-1}}{1-r_{i-1}}\sqrt{\frac{(1-p_{i-1})(1-p_i)}{p_{i-1}p_i}}\frac{p_{max}}{1-p_{max}}$ is a function of previous correlation $\rho_{i-1}$. In the special case when $p_1=p_2=\dots=p_m=p$ and $\rho_1=\rho_2=\dots=\rho_m=\rho\geq0$, the applicable condition is $\frac{1}{\sqrt{1-4\rho}}\left[(1+\sqrt{1-4\rho})^{m+1}-(1-\sqrt{1-4\rho})^{m+1}\right]\geq 0$. In Algorithm \[alg4\], $A_i$ and $Y_i$ can always be generated since their corresponding probabilities $\alpha_i$ and $p_{max}$ are in the range of $[0,1]$. Hence, we only need to consider the generation of $U_i$. Based on the definition of $U_i$’s marginal probability $r_i$, it is obvious that $r_i\geq 0$ if $r_{i-1} \leq 1$ holds. Therefore, the only requirement of Algorithm \[alg4\] is $r_i\leq 1$ ($i=1,\dots,m$). Thus, we have the applicable condition: $$r_i= \frac{\rho_{i-1}}{1-r_{i-1}}\sqrt{\frac{(1-p_{i-1})(1-p_i)}{p_{i-1}p_i}}\frac{p_{max}}{1-p_{max}} \leq1 \iff \rho_i \leq \sqrt{\frac{p_ip_{i+1}}{(1-p_i)(1-p_{i+1})}}\frac{1-p_{max}}{p_{max}}(1-r_{i}).$$ In practice, we can obtain the condition for each $\rho_i$ by iteration. In the special case when $p_1=\cdots=p_m=p$, $\rho_1=\cdots=\rho_m=\rho$, the requirement becomes the following series of inequalities: $$\label{rho4} \begin{aligned} &r_1=0\leq1\\ &r_2=\frac{\rho}{1-r_1}=\rho\leq1\\ &r_3=\frac{\rho}{1-r_2}=\frac{\rho}{1-\rho}\leq1 \Rightarrow \rho\leq\frac{1}{2}\\ &r_4=\frac{\rho}{1-r_3}=\frac{\rho}{1-\frac{\rho}{1-\rho}}=\frac{\rho(1-\rho)}{1-2\rho}\leq1 \Rightarrow \rho\leq \frac{3-\sqrt{5}}{2}\\ &\cdots \end{aligned}$$ The general term formula of $r_i$ is $r_i=2\rho\left[\frac{(1+\sqrt{1-4\rho})^{i-1}-(1-\sqrt{1-4\rho})^{i-1}}{(1+\sqrt{1-4\rho})^i-(1-\sqrt{1-4\rho})^i}\right]$. The series of inequalities can be reduced to $$\frac{1}{\sqrt{1-4\rho}}\left[(1+\sqrt{1-4\rho})^{i+1}-(1-\sqrt{1-4\rho})^{i+1}\right]\geq 0,$$ where $i=1,\dots, m$. From the inequalities above we see that as $i$ increases, the restriction of $\rho$ becomes more and more strict. Hence, we only need to satisfy the last inequality in Algorithm \[alg4\], i.e., $$\frac{1}{\sqrt{1-4\rho}}\left[(1+\sqrt{1-4\rho})^{m+1}-(1-\sqrt{1-4\rho})^{m+1}\right]\geq 0.$$ Compared to the applicable condition of Algorithm \[alg3\], the restriction of Algorithm \[alg4\] is only related with the dimension $m$ and not the marginal probability $p$. Figure \[fig:mdep0\] describes the relationship between the maximal allowed correlation $\rho_{max}$ and the dimension $m$. The $\rho_{max}$ monotonically decreases as $m$ increases, and converges to $0.25$. Therefore, if $\rho\leq 0.25$, this algorithm is suitable to generate binary data with an arbitrary dimension $m$. ![Relationship between the maximal allowed correlation $\rho_{max}$ and the dimension $m$ in Algorithm \[alg4\], when $p_1=p_2=\dots=p_m=p$ and $\rho_1=\rho_2=\dots=\rho_m=\rho$. As $m$ increases, the maximum of $\rho$ will decrease, and gradually convergent to $0.25$.[]{data-label="fig:mdep0"}](Fig_2.eps){width="80.00000%"} To conclude, both algorithms have their own specialities and applicable conditions for generating binary data with $1$-dependent correlation structure. In Algorithm \[alg3\], the limitation for each entry of correlation vector is only related with nearby marginal probabilities. In Algorithm \[alg4\], the limitation is also related with previous entries of correlation vector. It is difficult to distinguish which algorithm has more general applicable conditions. We give a detailed analysis when $p_1=p_2=\dots=p_m=p$ and $\rho_1=\rho_2=\dots=\rho_m=\rho$. In this situation, Algorithm \[alg3\] has no limitation on the dimension, meaning that the dimension can be arbitrarily large with feasible marginal probabilities. Thus, this algorithm is perfect for the situation when high dimension is required. On the other hand, Algorithm \[alg4\] is more flexible when the dimension is not too high, since it is not restricted by the marginal probabilities. In order to incorporate the two methods, we provide a function [cBern1dep]{} in our R package [CorBin]{}, which can automatically choose the suitable algorithm based on the given $\bm{p}$ and $\bm{\rho}$. In our function, we will first derive $r_i$ ($i=1,2,\cdots,m$) in Algorithm \[alg4\]. If all the $r_i$ lie in the interval $\left[0,1\right]$, we will use Algorithm \[alg4\] to generate the binary data. If not, the function will automatically call function [rhoMax1dep]{} to calculate the largest $\bm{\rho}$ allowed in Algorithm \[alg3\]. If the given $\bm{\rho}$ lies in the interval, the binary data will be generated using Algorithm \[alg3\]. $K$-dependent correlation structure and general correlation matrices {#general} -------------------------------------------------------------------- In Section \[statm\] we provide two algorithms to generate binary data with 1-dependent correlation structure. Here we discuss the generation of the binary data with $K$-dependent ($K> 1$) correlation structure by extending Algorithm \[alg3\]. Specifically, if we set $K=m-1$, we can obtain binary data with the general non-negative correlation matrices. Based on the intuition of Algorithm \[alg3\], we provide the details of the binary data generation algorithm under the $K$-dependent correlation structure (and also a general correlation matrix) in Algorithm \[alg5\]. We first denote $\bm{Y}=(Y_1,\cdots,Y_m)$ as a $K \times m$ matrix: $$\bm{Y}={ \left( \begin{array}{cccc} Y_{11} & Y_{12} & \cdots&Y_{1m}\\ Y_{21} & Y_{22} &\cdots&Y_{2m}\\ \vdots&\vdots&\ddots&\vdots\\ Y_{K1}&Y_{K2}&\cdots&Y_{Km} \end{array} \right )}.$$ In the algorithm, we use $\bm{\rho_i}$ to denote the $m-i$ elements on the $i$-th diagonal of the correlation matrix, i.e., $$\begin{aligned} \bm{\rho_i}=(\rho_{i1},\rho_{i2},\cdots,\rho_{i(m-i)})=(r_{1(i+1)},r_{2(i+2)},\cdots,r_{(m-i)m}). \end{aligned}$$ The expected values of the Bernoulli random variables $p_1,p_2,\cdots,p_m$; and the correlation coefficient vector $\bm{\rho_1}, \bm{\rho_2},\cdots,\bm{\rho_K}$\ The correlated binary variables $X_1,X_2,\cdots,X_m$ Check whether the input satisfies the Prentice constraints. Let $p_{m+1}=p_{m+2}=\cdots=p_{m+K}=p_m$ Let $\rho_{i,(m-(i-1))}=\rho_{i,(m-(i-1)+1)}=\cdots=\rho_{i,m}=0$ $\beta_{ij}=\frac{p_jp_{i+j}}{p_jp_{i+j}+\rho_{ij}\sqrt{p_jp_{i+j}(1-p_j)(1-p_{i+j})}}$ Generate $Y_{ij}\sim Bern(\beta_{ij})$ $\alpha_1=\frac{p_1}{\prod_{l=1}^{K}\beta_{li}}$ **if** $\alpha_1 \textgreater 1$: Prompt the unavailability. Generate $U_1\sim Bern(\alpha_1)$ $X_1=U_1\prod_{l=1}^{K}Y_{li}$ Denote $K_i'=\min \{i-1,K\} $ $\alpha_i=\frac{p_i}{\prod_{l=1}^{K}\beta_{li}\prod_{l=1}^{K_i'}\beta_{l(i-l)}}$ **if** $\alpha_i \textgreater 1$: Prompt the unavailability; **break** Generate $U_i\sim Bern(\alpha_i)$ $X_i=U_i\prod_{l=1}^{K}Y_{li}\prod_{l=1}^{K_i'}Y_{l(i-l)}$ $X_1,X_2,\cdots,X_m$ In the following, we show the justification of Algorithm \[alg5\] in Theorem \[thm:kdep\]. \[thm:kdep\] If intermediate variables $\alpha_1$, $\cdots$, $\alpha_m$, $\beta_{11}$, $\cdots$, $\beta_{Km} \in \left[0,1\right]$, Algorithm \[alg5\] returns the corresponding binary data with given marginal probabilities and $K$-dependent correlation structure. For $i=1$, we have $$EX_1=EU_1\prod_{l=1}^{K}Y_{li}=\alpha_1\prod_{l=1}^{K}\beta_{li}=p_1.$$ For each $i \in \{2,\cdots,m\}$, $$EX_i=EU_i\prod_{l=1}^{K}Y_{li}\prod_{l=1}^{K_i'}Y_{l(i-l)}=\alpha_i\prod_{l=1}^{K}\beta_{li}\prod_{l=1}^{K_i'}\beta_{l(i-l)}=\frac{p_i}{\prod_{l=1}^{K}\beta_{li}\prod_{l=1}^{K_i'}\beta_{l(i-l)}}\prod_{l=1}^{K}\beta_{li}\prod_{l=1}^{K_i'}\beta_{l(i-l)}=p_i.$$ From the generation process, it is obvious that $r_{ij}=0$ when $ \left|j-i\right|>K$. Considering $i=1$, $j \in \{1, 2, \cdots, m-1\}$, $$\begin{aligned} r_{1,1+j}=&\frac{cov(X_1,X_{1+j})}{\sqrt{Var(X_1)Var(X_{1+j})}}=\frac{cov(U_1\prod_{l=1}^{K_1}Y_{l1}, U_{1+j}\prod_{l=1}^{K}Y_{l(1+j)}\prod_{l=1}^{K_{1+j}'}Y_{l(1+j-l)}}{\sqrt{p_1p_{1+j}(1-p_1)(1-p_{1+j})}}\\ =&\frac{E(U_1U_{1+j}\prod_{l=1}^{K}Y_{l1}\prod_{l=1}^{K}Y_{l(1+j)}\prod_{l=1}^{K_{1+j}'}Y_{l(1+j-l)})}{\sqrt{p_1p_{1+j}(1-p_1)(1-p_{1+j})}}\\ &-\frac{E(U_1\prod_{l=1}^{K}Y_{l1})E(U_{1+j}\prod_{l=1}^{K}Y_{l(1+j)}\prod_{l=1}^{K_{1+j}'}Y_{l(1+j-l)})}{\sqrt{p_1p_{1+j}(1-p_1)(1-p_{1+j})}}\\ =&\frac{p_{1}p_{1+j}(1-\beta_{j1})}{\beta_{j1}\sqrt{p_1p_{1+j}(1-p_1)(1-p_{1+j})}}=\rho_{1j}. \end{aligned}$$ For $i \in \{2, \cdots, m-1\}$, $j \in \{1, 2, \cdots, m-i\}$, we have $$\begin{aligned} r_{i,i+j}=&\frac{cov(X_i,X_{i+j})}{\sqrt{Var(X_i)Var(X_{i+j})}}=\frac{cov(U_i\prod_{l=1}^{K}Y_{li}\prod_{l=1}^{K_i'}Y_{l(i-l)}, U_{i+j}\prod_{l=1}^{K}Y_{l(i+j)}\prod_{l=1}^{K_{i+j}'}Y_{l(i+j-l)})}{\sqrt{p_ip_{i+j}(1-p_i)(1-p_{i+j})}}\\ =&\frac{E(U_iU_{i+j}\prod_{l=1}^{K}Y_{li}\prod_{l=1}^{K_i'}Y_{l(i-l)}\prod_{l=1}^{K}Y_{l(i+j)}\prod_{l=1}^{K_{i+j}'}Y_{l(i+j-l)})}{\sqrt{p_ip_{i+j}(1-p_i)(1-p_{i+j})}}\\ &-\frac{E(U_i\prod_{l=1}^{K}Y_{li}\prod_{l=1}^{K_i'}Y_{l(i-l)})E(U_{i+j}\prod_{l=1}^{K}Y_{l(i+j)}\prod_{l=1}^{K_{i+j}'}Y_{l(i+j-l)})}{\sqrt{p_ip_{i+j}(1-p_i)(1-p_{i+j})}}\\ =&\frac{p_{i}p_{i+j}(1-\beta_{ji})}{\beta_{ji}\sqrt{p_ip_{i+j}(1-p_i)(1-p_{i+j})}}=\rho_{ij}. \end{aligned}$$ Here we have finished the proof. Due to the increased model complexity, it is difficult to derive the applicable condition of Algorithm \[alg5\] theoretically. However, given marginal probabilities and a general correlation matrix, we can still check whether the binary data can be generated using the algorithm by examining whether all intermediate parameters $\alpha_1,\cdots,\alpha_m$, $\beta_{11},\cdots,\beta_{Km}$ lie in the range of $[0,1]$. Performance {#3} =========== We implemented and integrated the above mentioned algorithms in an R package [CorBin]{}. In this section, we mainly demonstrate the effectiveness and computational efficiency of our package. If a data set is generated from the desired distribution, the sample mean should converge to the specified marginal probabilities when sample size increases. Meanwhile, the sample correlation matrix should also converge to the specified correlation matrix. Here, we demonstrate the effectiveness of our package by checking the consistency of sample mean and correlation matrix from the generated data. After that, we demonstrate the computational efficiency of our package by calculating the time needed for generating large-scale high-dimensional datasets. We further compare computational time with two commonly used binary data generation packages: [bindata]{} (Leisch et al. 1998) and [MultiOrd]{} (Demirtas 2006). Effectiveness ------------- In order to check the consistency of sample mean and correlation matrix, we generate datasets with different sample sizes in which the dimension $m$ is fixed to 100. For sample mean, we use the $l_2$ norm of the difference between the sample mean and the specified marginal probabilities as the error. For correlation matrix, we calculate the Frobenius norm of the residual matrix between sample and desired correlation matrix as the error. We randomly sampled the marginal expectations from a uniform distribution $U(0.5,0.8)$. The upper bound of correlation coefficients based on the Prentice constraints is $\sqrt{\frac{0.5\times0.2}{0.5\times0.8}}=0.5$. Then we generated a correlation coefficient from a uniform distribution $U\left(0,0.5\right)$ for exchangeable and AR(1) correlation structures. As the constraints for 1-dependent correlation structures are more stringent, we simulated the correlation coefficient from $U\left(0,0.2\right)$. Although Algorithm 5 can be applied to general cases, randomly generating the correlation coefficients cannot always satisfy the natural restrictions. Thus, without loss of generality, we fixed the structures to AR(1), and the settings were the same with simulations of Algorithm 2. We ran 10 times of simulations and calculated the average of errors for each distribution and verify the effectiveness of the algorithms. Figure \[fig:precision\] shows that under the four correlation structures we have considered and a specified general correlation matrix, both errors gradually approached to $0$ as sample size increased, indicating that the sample mean and correlation matrix of the generated data converged to the true settings we specified. These results demonstrate the effectiveness of our methods. ![As sample size increases, the sample mean and correlation matrix converge to the specified marginal probabilities and correlation matrix.[]{data-label="fig:precision"}](Fig_3.eps){width="\linewidth"} In addition, we provide five simple examples to illustrate the constructions of the algorithms and the choices for the parameters for better demonstration. The specified marginal probabilities and the specified correlation for each algorithms are summarized in Table \[tab:ex\]. The details of the data generation process for each example are attached in Supplementary Material (Example1-5.csv). Besides, we also provide their reproducing code in Supplementary Material (Example-code.R). \[tab:ex\] ----------------------------------------------------------------------------------------------------------- Algorithm Structure $m$ $\bm{p}$ $\bm{R}$ Supplementary File ----------- --------------- ----- -------------------- ------------------------------- -------------------- \[alg1\] Exchangeable 3 $(0.1,0.2,0.3)$ ${ Example1.csv \left( \begin{array}{ccc} 1 & 0.3 & 0.3\\ 0.3 & 1 & 0.3 \\ 0.3 & 0.3 & 1 \end{array} \right )}$ \[alg2\] AR(1) 3 $(0.1,0.2,0.3)$ ${ Example2.csv \left( \begin{array}{ccc} 1 & 0.2 & 0.1\\ 0.2 & 1 & 0.5 \\ 0.1 & 0.5 & 1 \end{array} \right )}$ \[alg3\] $1$-dependent 3 $(0.80,0.82,0.83)$ ${ Example3.csv \left( \begin{array}{ccc} 1 & 0.3 & 0\\ 0.3 & 1 & 0.5 \\ 0 & 0.5 & 1 \end{array} \right )}$ \[alg4\] $1$-dependent 3 $(0.80,0.82,0.83)$ ${ Example4.csv \left( \begin{array}{ccc} 1 & 0.3 & 0\\ 0.3 & 1 & 0.5 \\ 0 & 0.5 & 1 \end{array} \right )}$ \[alg5\] Generalized 3 $(0.6,0.7,0.8)$ ${ Example5.csv \left( \begin{array}{ccc} 1 & 0.3 & 0.1\\ 0.3 & 1 & 0.2 \\ 0.1 & 0.2 & 1 \end{array} \right )}$ ----------------------------------------------------------------------------------------------------------- : ****[Examples for illustration of the algorithms]{}**** The table summarized specified marginal probabilities and correlation matrices for each example, The details of data generation process for each example were presented in the corresponding file. Computational efficiency ------------------------ In this section, we demonstrate the superiority of our package in computational efficiency. All experiments performed here were based on a single processor of an Intel(R) Core(TM) 2.20GHz PC. For comparison, we also considered two commonly used packages [bindata]{} and [MultiOrd]{} to generate high dimensional binary data in the experiments. It is easy to find that all algorithms presented in Section 2.2, 2.3 and 2.4 involve only one layer of iterative process. Hence, the time complexity of our algorithms generating binary data with exchangeable, decaying-product and 1-dependent correlation structures is linear with respect to dimension $m$ theoretically. In Section 2.5, although there are two iteration layers in Algorithm \[alg5\], the time complexity is still linear with respect to $m$ if $K$ is irrelevant with $m$, which is normal in ordinary $K$-dependent correlation structure. However, when we want to generate binary data with a general correlation matrix, $K$ will be specified to $m-1$ and the time complexity will become quadratic with respect to $m$. Figure \[fig:cbtime\] presents the average time for generating binary data with different dimensions using [CorBin]{}, which further validates the linear time complexities of Algorithm 1-4 and quadratic time complexity of Algorithm 5. Here we use Algorithm 5 to generate binary data with autoregressive structure in simulation experiments. Please refer to Supplementary Table S1 for the numeric details of average time with a more general range of $m$ ($m=10^2\sim 10^6$). ![The relationship between the computational time (in seconds) and the dimension when simulating binary data using our package. The time is the average of 10 runs.[]{data-label="fig:cbtime"}](Fig_4.eps){width="\textwidth"} The calculation efficiency is impressive when generating the high-dimensional binary data. It takes only 0.2, 4.0 and 2.0 seconds to generate a $10^6$-dimensional binary data with exchangeable correlation structure, decaying-product and 1-dependent correlation structures, respectively. This dimension scale is too high for other data generation packages, such as [bindata]{} and [MultiOrd]{}. Table \[tab:time\] shows the time of different packages for generating binary data in a relative small scale ($m=100\sim 500$). The time recorded is based on experiments of 10 runs. For data generation with general correlation matrix, our algorithm is still very efficient compared to other packages. \[tab:time\] Structure $m$ CorBin bindata MultiOrd ----------- ----- ---------- -------------- -------------- 100 $(6e-6)$ 14.19 (0.07) 6.517 (0.06) 200 $(5e-6)$ 57.13 (0.38) 26.24 (0.63) 500 $(7e-6)$ 360.3 (2.46) 164.9 (0.47) 100 $(4e-5)$ 14.36 (0.06) 5.783 (0.05) 200 $(7e-5)$ 56.27 (0.93) 18.18 (0.21) 500 $(3e-4)$ 358.3 (2.24) 123.5 (0.31) 100 $(2e-5)$ 14.30 (0.05) 5.977 (0.06) 200 $(1e-4)$ 56.83 (0.28) 18.81 (0.33) 500 $(9e-5)$ 366.2 (3.27) 112.3 (0.13) : ****[Computational times (in seconds) needed for three algorithms under different correlation structures. ]{}**** The least time under each condition is highlighted in boldface. Standard deviations are in the bracket. It can be seen from Table \[tab:time\] and Figure 4, our method [CorBin]{} scales linearly with the dimension under three correlation structures and scales quadratically with general correlation matrix. As a comparison, the computational time increases rapidly with the growth of dimensions for the other two packages. Taking the exchangeable correlation structure as an example, it takes [bindata]{} around 14.2 seconds to generate a 100-dimensional data and 360.3 seconds to generate a 500-dimensional data, which is around 25 times of the former. The results are similar for [MultiOrd]{}. Moreover, regardless of the increasing rate with the dimension, our method has significant superiority over the other methods. When the dimension is 100, the time [CorBin]{} used is around $\frac{1}{700,000}$ of [bindata]{} and $\frac{1}{320,000}$ of [MultiOrd]{} in exchangeable correlation structure, and also far less than the other two in AR(1) and 1-dependent structure. When the dimension grows to 500, the advantage is even more obvious, with around $\frac{1}{4,500,000}$ of [bindata]{} and $\frac{1}{2,060,000}$ of [MultiOrd]{} in exchangeable correlation structure. For generating data with general correlation matrices, the ratios become around $\frac{1}{680}$ of [bindata]{} and $\frac{1}{320}$ of [MultiOrd]{}, respectively. Figure 5 shows the comparison among [CorBin]{}, [bindata]{} and [MultiOrd]{} in terms of computational time for general cases. ![The relationship between the computational time (in seconds) and the dimension when simulating binary data with general correlation structure using [CorBin]{}, [bindata]{} and [MultiOrd]{}. The time is the average of 10 runs.[]{data-label="fig:method_compare"}](Fig_5.eps){width="80.00000%"} Discussion and conclusions {#6} ========================== In this article, we have proposed several efficient algorithms to generate high-dimensional correlated binary data with varied marginal expectations and correlation structures. We first focus on three common correlation structures including exchangeable, decaying-product as well as $K$-dependent correlation structures, and then generalize the method on $K$-dependent structure and extend the applicability to any non-negative correlation matrices. An R package [CorBin]{} is also built based on these algorithms and uploaded on CRAN for readers to use. Compared with two state-of-the-art binary data generation packages [bindata]{} and [MultiOrd]{} [@leisch1998generation; @demirtas2006method], our algorithms require no complicated numerical procedures such as equation-solving or numerical integration and have linear time complexity with respect to the dimension when generating binary data with common correlation structures, leading to significant improvement in computational efficiency. In our simulations, [CorBin]{} needs less than 0.002 seconds to generate a $10\times1000$-dimensional binary data with exchangeable correlation structure, while generating such data takes more than 14,000 seconds and 6,400 seconds for [bindata]{} and [MultiOrd]{}, respectively. Compared with Lunn and Davies’ method, we generalize the algorithms so that the unequal probability settings can be satisfied. Concretely speaking, Lunn and Davies actually generated clusters of binary variables, and specified fixed the marginal probability and correlation coefficient in each independent cluster. Thus, it is not feasible to specify unequal probabilities in their method. Specifically, for exchangeable correlation structures, we generated each variable by $X_i=\left(1-U_i\right)Y_i+U_iZ$. Lunn and Davies first set a fixed probability $p$ and generate independent $Z\sim Bern\left(p\right)$, $Y_i\sim Bern\left(p\right)$ and $U_i\sim Bern\left(\sqrt\rho\right)$. In order to generate variables with unequal probabilities, an intuitive way is to simply fix a $\gamma$ and generate $Z\sim Bern\left(\gamma\right)$, and adjust the expectation of $Y_i$ so that the probability of $X_i$ is $p_i$. It is infeasible because there is no way to guarantee the expectation of $Y_i$ is exactly lies in $\left[0,1\right]$. A subtle construction of $\gamma$, $\alpha_i$ and $\beta_i$ is important in our algorithm, which can obtain proper $\alpha_i$s $\in\left[0,1\right]$, and derive the desired result (as proved in Theorem 2.1 and 2.2). For AR(1) correlation structures, we generated each variable by $X_i=\left(1-U_i\right)Y_i+U_iX_{i-1}$, while Lunn and Davies generated $Y_i\sim Bern\left(p\right)$ and $U_i\sim Bern\left(r_i\right)$. Thus, the expectation of $X_i$ is dependent on the expectation of $X_{i-1}$, $U_i$, and $Y_i$, and making the construction of parameters untrivial. We provide a recursive method to generate the probability of those mediating variables, guaranteeing the feasible of the algorithm (Theorem 2.3 and 2.4). For $1$-dependent correlation structures, we provide two algorithms. Algorithm 4 generalized Lunn and Davies’ method and Algorithm 3 was unrelated with Lunn and Davies’ method. We thoroughly studied two algorithms and derived their usage scopes in the manuscript. As discussed in Lunn and Davies’ paper, Algorithm 4 cannot be applied to $K>1$ situation. Our Algorithm 3 made up for this drawback. Notably, Algorithm 3 can be generalized to general cases with unequal probabilities and unequal correlation coefficients (Section 2.5, Algorithm 5). There are still some limitations with our methods. First, our package is applicable only when the correlations are non-negative, because in our algorithms we need to generate some variables following Bernoulli distribution with marginal probabilities related to the correlations. Negative correlations will lead to negative marginal probabilities, making it infeasible to generate corresponding binary vectors. Although in most situations, the capacity to generate binary data with positive correlations will suffice [@preisser2014comparison], negative correlations may still arise in some special situations. For these situations, [@guerra2014note] proposed an alternative way to simulate discrete random vectors with decaying product structure, in which negative correlations are allowed. We further derived applicable conditions for Algorithm 1-5, and surprisingly found that if the Prentice constraints are satisfied, our algorithms will be able to generate any specified binary data with non-negative exchangeable and decaying-product structures. But this is not the case for the $K$-dependent stationary structure. Therefore, we proposed two algorithms with different applicable conditions to generate binary data for $K=1$ and an algorithm with $K>1$. Our package will automatically select the suitable algorithm according to the input parameters, but an algorithm with more general applicable conditions is still needed for $K$-dependent structure. Appendix A {#appendix-a .unnumbered} ========== \[tab:addm\] $m$ $1e+2$ $1e+3$ $1e+4$ $1e+5$ $1e+6$ ----- ---------- ---------- ---------- ---------- ---------- $2e-5$ $2e-4$ $2e-3$ $2e-2$ $2e-1$ ($5e-6$) ($5e-5$) ($3e-4$) ($5e-3$) ($1e-2$) $5e-4$ $3e-3$ $4e-2$ $4e-1$ $4e+0$ ($4e-5$) ($7e-4$) ($8e-3$) ($2e-2$) ($7e-2$) $2e-4$ $2e-3$ $2e-2$ $2e-1$ $2e+0$ ($2e-5$) ($1e-4$) ($3e-3$) ($2e-2$) ($4e-2$) : ****[The computational time (in seconds) consumed in CorBin for generating binary data with different correlation structures.]{}**** The time is the average of 10 runs. Standard deviations are in the brackets. [**SUPPLEMENTARY MATERIAL**]{} CorBin: : R-package CorBin containing code to implement the algorithms described in the article. (GNU zipped tar file) CorBin-manual: : User manual for R package CorBin. (.pdf file). Examples: : The demonstration data contain five CSV files (Example1-5.csv), corresponding to five examples in described in Section 3.1 (Table 1), which illustrate the constructions of the algorithms and the choices for the parameters. (.rar file) Example-code: : The reproducing code for demonstration data. (.R file). Acknowledgements {#acknowledgements .unnumbered} ================ We thank the anonymous reviewer and the editor for their highly constructive and detailed feedback that helped us improve our manuscript substantially. Funding {#funding .unnumbered} ======= This research was supported in part by the NSF grant DMS 1713120. [^1]: These authors contributed equally to this work. [^2]: Corresponding author: [email protected]

ArXiv

--- abstract: | Tile $\mathbb{R}^2$ into disjoint unit squares $\{S_k\}_{k \geq 0}$ with the origin being the centre of $S_0$ and say that $S_i$ and $S_j$ are star-adjacent if they share a corner and plus-adjacent if they share an edge. Every square is either vacant or occupied. If the occupied plus-connected component $C^+(0)$ containing the origin is finite, it is known that the outermost boundary $\partial^+_0$ of $C^+(0)$ is a unique cycle surrounding the origin. For the finite occupied star-connected component $C(0)$ containing the origin, we prove in this paper that the outermost boundary $\partial_0$ is a unique connected graph consisting of a union of cycles $\cup_{1 \leq i \leq n} C_i$ with mutually disjoint interiors. Moreover, we have that each pair of cycles in $\partial_0$ share at most one vertex in common and we provide an inductive procedure to obtain a circuit containing all the edges of $\cup_{1 \leq i \leq n} C_i.$ This has applications for contour analysis of star-connected components in percolation. **Key words:** Star connected components, outermost boundary, union of cycles. **AMS 2000 Subject Classification:** Primary: 60J10, 60K35; Secondary: 60C05, 62E10, 90B15, 91D30. author: - | **Ghurumuruhan Ganesan** [^1]\  \ NISER, Bhubaneshwar, India title: 'Outermost boundaries for star-connected components in percolation' --- Introduction {#intro} ============ Tile $\mathbb{R}^2$ into disjoint unit squares $\{S_k\}_{k \geq 0}$ with origin being the centre of $S_0.$ We say $S_1$ and $S_2$ are *adjacent* or *star-adjacent* if they share a corner between them. We say that squares $S_1$ and $S_2$ are *plus-adjacent*, if they share an edge between them. Here we follow the notation of Penrose (2003). Suppose every square is assigned one of the two states: occupied or vacant. In many applications like for example, percolation, it is of interest to determine the outermost boundary of the plus-connected or star-connected components containing the origin. We make formal definitions below. The case of plus-connected components is well studied (Bollobas and Riordan (2006), Penrose (2003)) and in this case, the outermost boundary is simply a cycle containing the origin. Our main result is that the outermost boundary for the star-connected component is a connected union of cycles with disjoint interiors. Let $C(0)$ denote the star-connected occupied component containing the origin and throughout we assume that $C(0)$ is finite. Thus if $S_0$ is vacant then $C(0) = \emptyset.$ Else $S_0 \in C(0)$ and if $S_1, S_2 \in C(0)$ there exists a sequence of distinct occupied squares $(Y_1,Y_2,...,Y_t)$ all belonging to $C(0),$ such that $Y_i$ is adjacent to $Y_{i+1}$ for all $i$ and $Y_1 = S_1$ and $Y_t = S_2.$ Let $G_C$ be the graph with vertex set being the set of all corners of the squares $\{S_k\}_k$ in $C(0)$ and edge set consisting of the edges of the squares $\{S_k\}_k$ in $C(0).$ Two vertices $u$ and $v$ are said to be adjacent in $G_C$ if they share an edge between them. We say that an edge $e$ in $G_C$ is adjacent to square $S_k$ if it is one of the edges of $S_k.$ We say that $e$ is a *boundary edge* if it is adjacent to a vacant square and is also adjacent to an occupied square. A *path* $P$ in $G_C$ is a sequence of distinct vertices $(u_0,u_1,...,u_t)$ such that $u_i$ and $u_{i+1}$ are adjacent for every $i.$ A *cycle* $C$ in $G_C$ is a sequence of distinct vertices $(v_0,v_1,...,v_m,v_0)$ starting and ending at the same point such that $v_i$ is adjacent to $v_{i+1}$ for all $0 \leq i \leq m-1$ and $v_m$ is adjacent to $v_0.$ A *circuit* $C'$ in $G_C$ is a sequence of vertices $(w_0,w_1,...,w_r,w_0)$ starting and ending at the same point such that $w_i$ is adjacent to $w_{i+1}$ for all $0 \leq i \leq r-1,$ $w_r$ is adjacent to $w_0$ and no edge is repeated in $C'.$ Thus vertices may be repeated in circuits and for more related definitions, we refer to Chapter 1, Bollobas (2001). Any cycle $C$ divides the plane $\mathbb{R}^2$ into two disjoint connected regions. As in Bollobas and Riordan (2006), we denote the bounded region to be the *interior* of $C$ and the unbounded region to be the *exterior* of $C.$ We have the following definition. We say that edge $e$ in $G_C$ is an *outermost boundary* edge of the component $C(0)$ if the following holds true for every cycle $C$ in $G_C:$ either $e$ is an edge in $C$ or $e$ belongs to the exterior of $C.$ We define the outermost boundary $\partial _0$ of $C(0)$ to be the set of all outermost boundary edges of $G_C.$ Thus outermost boundary edges cannot be contained in the interior of any cycle in $G_C.$ Our main result is the following. \[thm3\] Suppose $C(0)$ is finite. The outermost boundary $\partial_0$ of $C(0)$ is a unique set of cycles $C_1,C_2,...,C_n$ in $G_C$ with the following properties:\ (i) The graph $\cup_{1 \leq i \leq n}C_i$ is a connected subgraph of $G_C.$\ (ii) If $i \neq j,$ the cycles $C_i$ and $C_j$ have disjoint interiors and share at most one vertex.\ (iii) Every square $S_k \in C(0)$ is contained in the interior of some cycle $C_{j}.$\ (iv) If $e \in C_{j}$ for some $j,$ then $e$ is a boundary edge of $C(0)$ adjacent to an occupied square of $C(0)$ in the interior of $C_j$ and also adjacent to a vacant square in the exterior. Moreover, there exists a circuit $C_{out}$ containing every edge of $\cup_{1 \leq i \leq n} C_i.$ The outermost boundary $\partial_0$ is therefore also an Eulerian graph with $C_{out}$ denoting the corresponding Eulerian circuit (for definitions, we refer to Chapter 1, Bollobas (2001)). We remark that the above result also provides a more detailed justification of the statement made about the outermost boundary and the corresponding circuit in the proof of Lemma 3 of Ganesan (2013). Using the above result, we also obtain the outermost circuit that is used to construct the top-down crossing in oriented percolation in a rectangle in Ganesan (2015). The proof of the above result also obtains the outermost boundary cycle in the case of plus-connected components. We recall that $S_1$ and $S_2$ are *plus-adjacent* if they share an edge between them. Analogous to the star-connected case, we define $C^+(0)$ to be the plus-connected component containing the origin and define the graph $G^+_C$ consisting of edges and corners of squares in $C^+(0).$ We have the following. \[thm2\] Suppose $C^+(0)$ is finite. The outermost boundary $\partial^+_0$ of $C^+(0)$ is unique cycle $C^+_{out}$ in $G^+_C$ with the following property:\ (i) All squares of $C^+(0)$ are contained in the interior of $C^+_{out}.$\ (ii) Every edge in $C^+_{out}$ is a boundary edge adjacent to an occupied square of $C^+(0)$ in the interior of $C^+_{out}$ and a vacant square in the exterior. This is in contrast to star-connected components which may contain multiple cycles in the outermost boundary. To prove Theorem \[thm3\], we use the following intuitive result about merging cycles. Analogous to $G_C,$ let $G$ be the graph with vertex set being the corners of the squares $\{S_k\}_k$ and edge set being the edges of the squares $\{S_k\}_k.$ \[thm1\] Let $C_1$ and $C_2$ be cycles in $G$ that have more than one vertex in common. There exists a unique cycle $C_3$ consisting only of edges of $C_1$ and $C_2$ with the following properties:\ (i) the interior of $C_3$ contains the interior of both $C_1$ and $C_2,$\ (ii) if an edge $e$ belongs to $C_1$ or $C_2,$ then either $e$ belongs to $C_3$ or is contained in its interior. Moreover, if $C_2$ contains at least one edge in the exterior of $C_1,$ then the cycle $C_3$ also contains an edge of $C_2$ that lies in the exterior of $C_1.$ The above result essentially says that if two cycles intersect at more than one point, there is a innermost cycle containing both of them in its interior. We provide an iterative construction for obtaining the cycle $C_3,$ analogous to Kesten (1980) for crossings, in Section \[pf1\]. The paper is organized as follows: In Section \[pf2\], we prove Theorem \[thm3\] and in Section \[pf1\], we prove Theorem \[thm2\] and Theorem \[thm1\]. Proof of Theorem \[thm3\] {#pf2} ========================= *Proof of Theorem \[thm3\]*: The first step is to obtain large cycles surrounding each occupied square in $C(0).$ We have the following Lemma. \[outer\] For every $S_k \in C(0),$ there exists a unique cycle $D_k$ satisfying the following properties:\ (a) $S_k$ is contained in the interior of $D_k,$\ (b) every edge in the cycle $D_k$ is a boundary edge adjacent to one occupied square of $C(0)$ in the interior and one vacant square in the exterior and\ (c) if $C$ is any cycle in $G_C$ that contains $S_k$ in the interior, then every edge in $C$ either belongs to $D_k$ or is contained in the interior. We denote $D_k$ to be the outermost boundary cycle containing the\ square $S_k.$ We prove all statements at the end. We claim that the set of distinct cycles in the set ${\cal D} := \cup_{S_k \in C(0)} \{D_k\}$ is the desired outermost boundary $\partial_0$ and satisfies the conditions (i)-(iv) mentioned in the statement of the theorem. By construction, we have that (iii) and (iv) are satisfied. To see that (ii) holds, we suppose that $D_{k_1} \neq D_{k_2}$ and that $D_{k_1}$ and $D_{k_2}$ meet at more than one vertex. We know that $D_{k_2}$ is not completely contained in $D_{k_1}.$ Thus $D_{k_2}$ contains at least one edge in the exterior of $D_{k_1}.$ From Theorem \[thm1\], we obtain a cycle $D'_{12}$ containing both $D_{k_1}$ and $D_{k_2}$ in the interior and containing an edge $e$ present in $D_{k_2}$ but not in $D_{k_1}$ or its interior. The cycle $D'_{12}$ satisfies condition (a) in Lemma \[outer\] above and thus contradicts the assumption that $D_{k_1}$ satisfies (c). Thus $D_{k_1}$ and $D_{k_2}$ cannot meet at more than one vertex. Also (i) holds, because of the following reason. First we note that by construction $G_C$ is connected; let $u_1$ and $u_2$ be vertices in $G_C.$ Each $u_i, i = 1,2$ is a corner of an occupied square $S_i \in C(0)$ and by definition, $S_1$ and $S_2$ are star-connected via squares in $C(0).$ Thus there exists a path in $G_C$ from $u_1$ to $u_2.$ To see that ${\cal D}$ is a connected subgraph of $G_C,$ we let $v_1$ and $v_2$ be vertices in ${\cal D}$ that belong to cycles $D_{r_1}$ and $D_{r_2},$ respectively, for some $r_1$ and $r_2.$ If $r_1 =$ $r_2,$ then $v_1$ and $v_2$ are connected by a path in $D_{r_1}= (z_1 = v_1,z_2,...,z_n,z_1).$ If $r_1 \neq r_2,$ let $P_{12} = (w_1 = v_1,w_2,...,w_{t-1},w_{t} = v_2)$ be a path from $v_1$ to $v_2$ in $G_C.$ We iteratively construct a path $P'_{12}$ from $P_{12}$ using only edges of cycles in ${\cal D}.$ We first note that since (iii) holds, every edge in $P_{12}$ either belongs to a cycle in ${\cal D}$ or is contained in the interior of some cycle in ${\cal D}.$ Let $i_1$ be the first time $P_{12}$ leaves $D_{r_1};$ i.e., let $i_1 = \min\{i \geq 1 : w_{i+1} \text{ belongs to exterior of } D_{r_1}\}.$ The edge formed by the vertices $w_{i_1}$ and $w_{i_1+1}$ belongs to some cycle $D_{s_1} = (x_1 = w_{i_1},x_2,...,x_r,x_1)$ or is contained in its interior. Since the cycles $D_{r_1}$ and $D_{s_1}$ have disjoint interiors, this necessarily means $D_{s_1}$ and $D_{r_1}$ meet at $w_{i_1}.$ Defining $T_1 = (z_1 = v_1,z_2,...,z_{j_1} = w_{i_1}),$ we note that $T_1$ is a path consisting only of edges in the cycle $D_{r_1}$ and containing the vertex $z_1 = v_1.$ Repeating the same procedure above, we obtain another path $T_2 = (w_{i_1} = x_1,x_2,...,x_{j_2} = w_{i_2})$ contained in $D_{s_1},$ where, as before, $i_2 = \min\{i \geq i_1+1 : w_{i+1} \text{ belongs to exterior of } D_{s_1}\}$ denotes the first time $P_{12}$ leaves $D_{s_1}.$ We continue this procedure for a finite number of steps $m,$ until we reach $v_2.$ By construction, the path $T_i$ obtained at step $i, 2 \leq i \leq m$ is connected to $\cup_{1 \leq j \leq i-1} T_j.$ The final union of paths $\cup_{1 \leq i \leq m}T_{i}$ is therefore a connected graph containing only edges in ${\cal D}$ and contains $v_1$ and $v_2.$ It remains to see that an edge $e$ belongs to the outermost boundary if and only if it belongs to some cycle in ${\cal D}.$ If $e$ is an edge in a cycle $D_k \in {\cal D}$ we have that $e$ is adjacent to an occupied square $S_e$ contained in the interior of $D_k$ and a vacant square $S'_e$ in the exterior. If there exists a cycle $C$ in $G_C$ that contains $e$ in the interior, we then have that both $S_e$ and $S'_e$ are contained in the interior of $C.$ Since $S'_e$ is exterior to $D_k,$ the cycle $C$ contains at least one edge in the exterior of $D_k.$ But if $D_e$ denotes the outermost cycle containing $S_e,$ then by the discussion in the first paragraph, we must have that $D_e = D_k.$ And thus every edge of $C$ either belongs to $D_e$ or is contained in the interior of $D_e$ which leads to a contradiction. We also see that no other edge apart from edges of cycles in ${\cal D}$ can belong to the outermost boundary since if $e_1 \notin {\cal D},$ then $e_1$ is necessarily contained in the interior of some cycle $D_r \in {\cal D}.$ Finally, to obtain the circuit we compute the cycle graph $H_{cyc}$ as follows: let $E_1,E_2,...,E_n$ be the distinct outermost boundary cycles in ${\cal D}.$ Represent $E_i$ by a vertex $i$ in $H_{cyc}.$ If $E_i$ and $E_j$ share a corner, we draw an edge between $i$ and $j.$ We have the following lemma.  \[hcyc\] We have that the graph $H_{cyc}$ described above is a tree. We provide the proof of the above at the end. We then obtain the circuit via induction on the number of vertices $n$ of $H_{cyc}.$ For $n = 1,$ it is a single cycle. Suppose we obtain the circuit of all cycle graphs containing at most $k$ vertices and let $H_{cyc}$ be a cycle graph containing $k+1$ vertices. To obtain the circuit for $H_{cyc},$ we pick a leaf $q$ of $H_{cyc}$ and apply induction assumption on the cycle graph $H'_{cyc} = H_{cyc}\setminus q.$ To fix a procedure, we choose $q$ such that the corresponding boundary cycle $E_q$ contains a square $S_j$ of least index $j$ in its interior. We have that $H'_{cyc}$ is connected and has $k$ vertices and thus has a circuit $C_{k} = (c_1,c_2,...,c_r,c_1)$ containing all edges of every cycle in $H'_{cyc}.$ Let $C_k$ meet the cycle $E_q = (d_1,d_2,...,d_i,d_1)$ at $d_t = c_1.$ We then form the new circuit $C_{k+1} = (d_1,d_2,...,d_t = c_1,c_2,...,c_r,c_1 = d_t,d_{t+1},...,d_i,d_1),$ which contains all edges of every cycle in $H_{cyc}.$ ${\hfill{\ \ \rule{2mm}{2mm}} \vspace{0.2in}}$ *Proof of Lemma \[outer\]*: We note that if there exists such a $D_k,$ then it is unique by definition. Let ${\cal E}$ be the set of all cycles in $G_C$ satisfying condition (a); i.e., if $C$ is a cycle containing $S_k$ in its interior then $C \in {\cal E}.$ The set ${\cal E}$ is not empty since the cycle formed by the four edges of $S_k$ belongs to ${\cal E}.$ We merge cycles in ${\cal E}$ two by two using Theorem \[thm1\] to obtain the desired cycle $D_k.$ We first pick a cycle $F_1$ in ${\cal E}$ using a fixed procedure; for example, using an analogous iterative procedure as described in Section 1 of Ganesan (2014) for choosing paths. We again use the same procedure to pick a cycle $F_2$ in ${\cal E} \setminus F_1$ and from Theorem \[thm1\], obtain a cycle $F'_1$ consisting of only edges of $F_1$ and $F_2$ and containing both $F_1$ and $F_2$ in its interior. The cycle $F'_1$ also satisfies (a) and thus belongs to ${\cal E}.$ Therefore, if ${\cal E}$ has $t$ cycles, then ${\cal E}_1 := ({\cal E} \setminus \{F_1,F_2\}) \cup F'_1$ has at most $t-1$ cycles; if $F_1$ contains an edge in the exterior of $F_2$ and the cycle $F_2$ also contains an edge in the exterior of $F_1,$ then ${\cal E}_1$ has $t-2$ cycles. Else $F'_1$ is either $F_1$ or $F_2$ and the set ${\cal E}_1$ therefore contains $t-1$ cycles. By construction, every cycle in ${\cal E}$ is either a cycle in ${\cal E}_1$ or is contained in the interior of a cycle in ${\cal E}_1.$ Therefore, if ${\cal E}_1$ contains one cycle, it is the desired outermost boundary cycle $D_k.$ Else we repeat the above procedure with ${\cal E}_1$ and obtain another set ${\cal E}_2$ containing at most $t-2$ cycles and again with the property that every cycle in ${\cal E}$ is either a cycle in ${\cal E}_2$ or is contained in the interior of a cycle in ${\cal E}_2.$ Continuing this process, we are finally left with a single cycle $C_{fin}.$ By construction it satisfies (a) and (c). It only remains to see that (b) is true. Suppose there exists an edge $e$ of $C_{fin}$ that is not a boundary edge. Since $e$ is an edge of $G_C,$ we then have that $e$ is adjacent to two occupied squares $S_1$ and $S_2,$ with one of the squares, say $S_1,$ contained in the interior of $C_{fin}$ and the other square $S_2,$ contained in the exterior. The cycle $C_2$ containing the four edges of the square $S_2$ and the cycle $C_{fin}$ have the edge $e$ in common and thus more than one vertex in common. Since $C_2$ contains at least one edge in the exterior of $C_{fin},$ we use Theorem \[thm1\] to obtain a larger cycle $C'_2$ containing both $C_{fin}$ and $C_2$ in the interior. The cycle $C'_2$ contains at least one edge not in $C_{fin}.$ But since $C_{fin}$ satisfies (c), this is a contradiction. Thus every edge $e$ of $C_{fin}$ is a boundary edge. By the same argument above, we also see that the edge $e$ cannot be adjacent to an occupied square in the exterior of $C_{fin}.$ Thus $e$ is adjacent to an occupied square in the interior and a vacant square in the exterior. ${\hfill{\ \ \rule{2mm}{2mm}} \vspace{0.2in}}$ *Proof of Lemma \[hcyc\]*: We already have that $H_{cyc}$ is connected. It is enough to see that it is acyclic. Before we prove that, we make the following observation. Consider a path $P = (i_1,i_2,...,i_m)$ in $H_{cyc}.$ We see that any vertex in $E_{i_1}$ and any vertex in $E_{i_m}$ is connected by a path consisting only of edges of the cycles $\{E_{i_k}\}_{1 \leq k \leq m}.$ Suppose $H_{cyc}$ contains a cycle $C = (r_1, r_2,...,r_s, r_1).$ Let the boundary cycle $E_{r_1} = (u_1,u_2,...,u_m,u_1)$ meet $E_{r_2}$ at $u_{1}$ and $E_{r_s}$ at $u_{j}.$ We have that $j\neq 1$ since three boundary cycles cannot meet at a point. This is illustrated in Figure \[fig\_second\_case\]. The occupied square $S_1$ belongs to $E_{r_2}$ and the occupied square $S_2$ belongs to $E_{r_s}.$ It is necessary that the squares $S_3$ and $S_4$ are vacant and thus cannot be on the boundary of any other cycle. Let $P_1$ and $P'_1$ be the two segments of $E_{r_1}$ starting at $u_{1}$ and ending at $u_{j}.$ Since $u_{1} \in E_{r_2}$ and $u_{j} \in E_{r_s},$ we have by the observation made in the first paragraph that there exists a path $P_2$ from $u_1$ to $u_j,$ consisting only of edges in $\{E_{r_i}\}_{2 \leq i \leq s}.$ This path necessarily lies in the exterior of $E_{r_1}$ and is illustrated in Figure \[fig\_first\_case\]. Here $ABCDA$ represents the cycle $E_{r_1},$ the path $P_1$ is the segment $ADC$ and the path $P'_1$ is the segment $ABC.$ The path $P_2$ is denoted by the exterior segment $AEC.$ Thus it is necessary that either the cycle $C_{12}$ formed by $P_1 \cup P_2$ contains $P'_1$ in the interior or the cycle $C'_{12}$ formed by $P'_1 \cup P_2$ contains $P_1$ in the interior. Suppose the former holds and let $S_{a}$ be any occupied square in the interior of $E_{r_1}.$ We know that $E_{r_1} = D_a$ is the outermost boundary cycle containing $S_a$ and satisfies conditions (a), (b) and (c) mentioned in Lemma \[outer\]. The cycle $C_{12}$ also contains $S_{a}$ in the interior and thus satisfies condition (a). Moreover, it contains at least one edge in the exterior of $E_{r_1}$ contradicting the fact that $E_{r_1}$ satisfies (c). Thus $H_{cyc}$ is acyclic. ${\hfill{\ \ \rule{2mm}{2mm}} \vspace{0.2in}}$ Proofs of Theorem \[thm2\] and Theorem \[thm1\] {#pf1} =============================================== *Proof of Theorem \[thm2\]*: Let $D_0$ be the outermost boundary cycle containing the square $S_0$ as in Lemma \[outer\]. It satisfies the conditions (i) and (ii) in the statement of the theorem and is unique and thus $C^+_{out} = D_0.$ ${\hfill{\ \ \rule{2mm}{2mm}} \vspace{0.2in}}$ *Proof of Theorem \[thm1\]*: If every edge of $C_1$ is either on $C_2$ or contained in the interior of $C_2,$ then the desired cycle $C_3 = C_2.$ If similarly, $C_2$ is completely contained in $C_1,$ we set $C_3 = C_1.$ So we suppose that $C_1$ contains at least one edge in the exterior of $C_2$ and $C_2$ also contains at least one edge in the exterior of $C_1.$ We start with cycle $C_1$ and in the first step, identify a path of $C_2$ contained in the exterior of $C_1.$ Set $C_{1,0} := C_1 = (u_0,u_1,...,u_{t-1},u_0)$ and $C_2 = (v_0,v_1,...,v_{m-1},v_0).$ For later notation, we define $u_{k} = u_{k \mod t}$ if $k \leq 0$ or $k \geq t$ and $v_{k} = v_{k \mod m}$ if $k \leq 0$ or $k \geq m.$ Start from some vertex of $C_{1,0},$ say $u_0,$ and look for the first intersection point that contains an exterior edge of $C_2;$ i.e., an edge of $C_2$ that lies in the exterior of $C_1.$ Let $$j_1 = \min\{j \geq 0 : u_j \in C_{1,0} \text{ and } u_j \text{ is an endvertex of an exterior edge of } C_2 \}$$ and let $v_{i_1} = u_{j_1}.$ We suppose that the edge of $C_2$ with endvertices $v_{i_1}$ and $v_{i_1+1}$ lies in the exterior of $C_1.$ Let $r_1 = \min\{i \geq i_1+1 : v_i \in C_{1,0}\}$ be the next time the cycles meet and define $P_1 = (v_{i_1},v_{i_1+1},...,v_{r_1}).$ We note that none of the vertices $v_{j}, i_1 +1 \leq j \leq r_1-1$ belong to $C_{1,0}.$ If $v_{i_1} = v_{r_1},$ then $P_1$ is a cycle containing the edges of $C_2$ and thus $P_1 = C_2.$ Since $C_1$ and $C_2$ contain more than one vertex in common, this cannot happen. Thus $P_1$ is a path and all edges of $P_1$ are in the exterior of $C_{1,0}.$ We then construct an outermost cycle from $C_{1,0}$ and $P_1$ as follows. Split $C_{1,0}$ into two segments based on intersection with $P_1.$ Suppose $P_1$ meets $C_{1,0}$ at $u_{a_1}$ and $u_{b_1}.$ We let $C'_{1,0} = (u_{a_1},u_{a_1+1},...,u_{b_1})$ and $C^{''}_{1,0} = (u_{a_1}, u_{a_1-1},...,u_{b_1}).$ If the interior of $C'_{1,0} \cup P_1$ contains the interior of $C''_{1,0} \cup P_1$ as in Figure \[fig\_first\_case\], we set $C_{1,1} = C'_{1,0} \cup P_1$ to be the cycle obtained in the first iteration by the concatenation of the paths $C'_{1,0}$ and $P_1.$ Here $C''_{1,0}$ is the segment $ADC,$ the path $C'_{1,0}$ is the segment $ABC$ and the path $P_1$ is denoted $AEC.$ Else necessarily we have that the interior of $C^{''}_{1,0} \cup P_1$ contains the interior of $C'_{1,0} \cup P_1$ and we set $C_{1,1} = C^{''}_{1,0} \cup P_1.$ Since $P_1 \neq \emptyset,$ we have that $C_{1,1}$ contains at least one exterior edge. We then perform the same procedure as above on the cycle $C_{1,1}$ and continue this process for a finite number of steps to obtain the final cycle $C_{1,n}.$ For each $j, 1 \leq j \leq n,$ we have that the cycle $C_{1,j}$ satisfies the following properties:\ (1) the cycle $C_{1,j}$ contains only edges from $C_1$ and $C_2,$\ (2) every edge of $C_1$ either belongs to $C_{1,j}$ or is contained in the interior of $C_{1,j},$\ (3) the cycle $C_{1,j}$ contains at least one exterior edge of $C_2$ and\ (4) the interior of $C_1$ is contained in $C_{1,j}.$\ In particular, the above properties hold true for the final cycle $C_{1,n}.$ If there exists an edge $e$ of $C_2$ in the exterior of $C_{1,n},$ then the edge $e$ belongs to a path $P_e$ of $C_2$ containing edges exterior to $C_1.$ The path $P_e$ must meet $C_1$ and thus there exists an edge of $C_2$ that lies in the exterior of $C_1$ and contains an endvertex of $C_1.$ But then the above procedure would not have terminated and thus we also have:\ (5) every edge of $C_2$ either belongs to $C_{1,n}$ or is contained in the interior of $C_{1,n}.$ Thus property (ii) stated in the result holds true and we need to see that (i) holds. For that we first prove uniqueness of the cycle $C_{1,n}$ obtained above. Suppose there exists another cycle $D'$ satisfying properties (1), (2) and (5) above. If $D'$ contains an edge $e'$ (which must necessarily belong to $C_1$ or $C_2$) in the exterior of $C_{1,n},$ it contradicts the fact that $C_{1,n}$ satisfies (2) and (5). If $D'$ is completely contained in the interior of $C_{1,n}$ and is not equal to $C_{1,n},$ then there is at least one edge of $C_{1,n}$ (which belongs to $C_1$ or $C_2$) that lies in the exterior of $D',$ contradicting the assumption that $D'$ satisfies (2) and (5). Thus any cycle satisfying properties (1), (2) and (5) is unique. We recall that $C_1$ also contains an edge in the exterior of $C_2.$ Suppose now we start from $C_{2,0} := C_2$ and identify segments of $C_1$ lying in the exterior of $C_2$ and perform the same iterative procedure as above to obtain a final cycle $C_{2,m}.$ This cycle must also satisfy (1), (2) and (5) and hence $C_{2,m} = C_{1,n}.$ Moreover, $C_{2,m}$ satisfies:\ (3$'$) the cycle $C_{2,m}$ contains at least one exterior edge of $C_1$ and\ (4$'$) the interior of $C_2$ is contained in $C_{2,m}.$ Thus the cycle $C_{1,n}$ is unique and satisfies properties (i) and (ii) stated in the result. ${\hfill{\ \ \rule{2mm}{2mm}} \vspace{0.2in}}$ Acknowledgement {#acknowledgement .unnumbered} --------------- I thank Professor Rahul Roy for crucial comments and NISER for my fellowship. [10]{} B. Bollobas. (2001). . . B. Bollobas and O. Riordan. (2006). . . G. Ganesan. (2013). . , **49**, 1130–1140. G. Ganesan. (2014). . G. Ganesan. (2015). . , **47**, 164–181. H. Kesten. (1980). . , **74**, 41–59. M. Penrose. (2003). . . [^1]: E-Mail: `[email protected]`

ArXiv

--- author: - 'Xin Zhang[^1]   and  Xun Li[^2]' title: '**Open-Loop and Closed-Loop Solvabilities for Stochastic Linear Quadratic Optimal Control Problems of Markov Regime-Switching System[^3]**' --- [**Abstract:**]{} This paper investigates the stochastic linear quadratic (LQ, for short) optimal control problem of Markov regime switching system. The representation of the cost functional for the stochastic LQ optimal control problem of Markov regime switching system is derived using the technique of It[ô]{}’s formula. For the stochastic LQ optimal control problem of Markov regime switching system, we establish the equivalence between the open-loop (closed-loop) solvability and the existence of an adapted solution to the corresponding forward-backward stochastic differential equation with constraint (the existence of a regular solution to the Riccati equation). Also, we analyze the interrelationship between the strongly regular solvability of the Riccati equation and the uniform convexity of the cost functional. [**Keywords:**]{} linear quadratic optimal control, Markov regime switching, Riccati equation, open-loop solvability, closed-loop solvability. **AMS Mathematics Subject Classification. 49N10, 49N35, 93E20.** Introduction ============ Linear-quadratic (LQ) optimal control problem plays important role in control theory. It is a classical and fundamental problem in the fields of control theory. In the past few decades, both the deterministic and stochastic linear quadratic (LQ) control problems are widely studied. Stochastic LQ optimal control problem was first carried out by Kushner [@Kushner1962] with dynamic programming method. Later, Wonham [@Wonham1968] studied the generalized version of the matrix Riccati equation arose in the problems of stochastic control and filtering. Using functional analysis techniques, Bismut [@Bismut1976] proved the existence of the Riccati equation and derived the existence of the optimal control in a random feedback form for stochastic LQ optimal control with random coefficients. Tang [@Tang2003] studied the existence and uniqueness of the associated stochastic Riccati equation for a general stochastic LQ optimal control problems with random coefficients and state control dependent noise via the method of stochastic flow, which solves Bismut and Peng’s long-standing open problems. Moreover, Tang provided a rigorous derivation of the interrelationship between the Riccati equation and the stochastic Hamilton system as two different but equivalent tools for the stochastic LQ problem. For more details on the progress of stochastic Riccati equation, interest readers may refer to [@Kohlmann2003mbsr; @Kohlmann2003mrlq; @Kohlmann2002; @Kohlmann2001ndbsre; @Tang2015]. Under some mild conditions on the weighting coefficients in the cost functional, such as positive definite of the quadratic weighting control martix, and so on, the stochastic LQ optimal control problems can be solved elegantly via the Riccati equation approach, see [@yong1999sch Chapter 6]. Chen et al. [@Chen1998] was the first to start the pioneer work of stochastic LQ optimal control problems with indefinite of the quadratic weighting control matrix, which turns out to be useful in solving the continuous time mean-variance portfolio selection problems. Since then, there has been an increasing interest in the so-called indefinite stochastic LQ optimal control, see, for example, Chen and Yong [@Chen2001], Li and Zhou [@lizhou2002islq], Li et al. [@xunli2001islqj; @xunli2003islq], and so on. Another extension to stochastic LQ optimal control problems is to involve random jumps in the state systems, such as Poisson jumps or the regime switching jumps. Wu and Wang [@wu2003fbsde] was the first to consider the stochastic LQ optimal control problems with Poisson jumps and obtain the existence and uniqueness of the deterministic Riccati equation. Using the technique of completing squares, Hu and Oksendal [@Hu2008] discussed the stochastic LQ optimal control problem with Poisson jumps and partial information. Existence and uniqueness of the stochastic Riccati equation with jumps and connections between the stochastic Riccati equation with jumps and the associated Hamilton systems of stochastic LQ optimal control problem were also presented. Yu [@Yu2017ihjd] investigated a kind of infinite horizon backward stochastic LQ optimal control problems and differential game problems under the jump-diffusion model state system. Li et al. [@Li2018] solved the indefinite stochastic LQ optimal control problem with Poisson jumps. The stochastic control problems involving regime switching jumps are of interest and of practical importance in various fields such as science, engineering, financial management and economics. The regime-switching models and related topics have been extensively studied in the areas of applied probability and stochastic controls. More recently, there has been dramatically increasing interest in studying this family of stochastic control problems as well as their financial applications, see, for examples, [@Zhou2003mmvp; @xunli2001islqj; @Yin2004mmvps; @lizhou2002islq; @xunli2003islq; @Zhang2018gsmp; @Zhang2011mrsm; @Zhang2012smp; @Zhang2010psem; @mei2017equilibrium]. Ji and Chizeck [@Ji1992jlqgc; @Ji1990csctm] formulated a class of continuous-time LQ optimal controls with Markovian jumps. Zhang and Yin [@QingZhang1999noch] developed hybrid controls of a class of LQ systems modulated by a finite-state Markov chain. Li and Zhou [@lizhou2002islq], Li et al. [@xunli2001islqj; @xunli2003islq] introduced indefinite stochastic LQ optimal controls with regime switching jumps. Liu et al. [@Liu2005nocrs] considered near-optimal controls of regime-switching LQ problems with indefinite control weight costs. Recently, Sun and Yong [@sun2014linear] investigated the two-person zero-sum stochastic LQ differential games. It was shown in [@sun2014linear] that the open-loop solvability is equivalence to the existence of an adapted solution to an forward-backward stochastic differential equation (FBSDE, for short) with constraint and closed loop solvability is equivalent to the existence of a regular solution to the Riccati equation. As a continuation work of [@sun2014linear], Sun et al. [@Sun2016olcls] studied the open-loop and closed-loop solvabilities for stochastic LQ optimal control problems. Moreover, the equivalence between the strongly regular solvability of the Riccati equation and the uniform convexity of the cost functional is established. The aim of this paper is to extend the results of Sun et al. [@Sun2016olcls] to the case of stochastic LQ optimal control problems with regime switching jumps. We will establish the above equivalences of Sun et al. [@Sun2016olcls] for the stochastic LQ optimal control problem with regime switching jumps. The first main contribution of our paper is to provide a method for obtaining the representation of the cost functional for the stochastic LQ optimal control problem with regime switching jumps. In Sun et al. [@Sun2016olcls], the representation of the cost functional, which is the summary results of Yong and Zhou [@yong1999sch], is fundamental to prove the above equivalences. Unlike the techniques of function analysis used in Yong and Zhou [@yong1999sch] or Sun et al. [@Sun2016olcls], our method for deriving the representation of the cost functional is mainly based on the technique of It[ô]{}’s formula only. The second main contribution of our paper is to use the stochastic flow theory for proving the equivalence between the closed-loop solvability and the existence of regular solution to the Riccati equation. Due to the incorporate of the regime switching jumps, the method used in Sun et al. [@Sun2016olcls] for proving the equivalence between the closed-loop solvability and the existence of regular solution to the Riccati equation does not work for the stochastic LQ optimal control problem with regime switching jumps. The rest of the paper is organized as follows. Section 2 will introduce some useful notations and collect some preliminary results and state the stochastic LQ optimal control problem with regime switching jumps. Section 3 is devoted to deriving the representation of the cost functional by using the technique of Itô formula. In section 4 and 5, we will prove the equivalence between the open-loop (closed-loop) solvability and the existence of an adapted solution to the corresponding FBSDE with constraint (the existence of a regular solution to the Riccati equation) for the stochastic LQ optimal control problem of Markov regime switching system. The equivalence between the strongly regular solvability of the Riccati equation and the uniform convexity of the cost functional is established in section 6. Preliminaries and Model Formulation =================================== Let $(\O,\cF,\dbF,\dbP)$ be a complete filtered probability space on which a standard one-dimensional Brownian motion $W=\{W(t); 0\les t < \i \}$ and a continuous time, finite-state, Markov chain $\a=\{\a(t); 0\les t< \i \}$ are defined, where $\dbF=\{\cF_t\}_{t\ges0}$ is the natural filtration of $W$ and $\a$ augmented by all the $\dbP$-null sets in $\cF$. In the rest of our paper, we will use the following notation. $$\begin{aligned} \begin{array}{ll} \mathbb{N}: & \mbox{the set of natural numbers};\\ \dbR_+, \cl{\dbR}_+: &\mbox{the sets } [0,\infty) \mbox{ and } [0,+\infty] \mbox{ respectively};\\ \mathbb{R}^n: & \mbox{the } n\mbox{-dimensional Euclidean space};\\ M^\top: & \mbox{the transpose of any vector or matrix } M;\\ \tr[M]: & \mbox{the trace of a square matrix } M;\\ \cR(M): & \mbox{the range of the matrix } M;\\ \langle \cd\,,\cd\rangle:& \mbox{the inner products in possibly different Hilbert spaces};\\ M^\dag: & \mbox{the Moore-Penrose pseudo-inverse of the matrix } M ({\rm see, \cite{penrose1955generalized}});\\ \mathbb{R}^{n\times m}: & \mbox{the space of all } n\times m \mbox{ matrices endowed with the inner product }\\ & \langle M, N\rangle \mapsto\tr[M^\top N] \mbox{ and the norm } |M|=\sqrt{\tr[M^\top M]};\\ \dbS^n: & \mbox{the set of all }n\times n \mbox{ symmetric matrices};\\ \cl{\dbS^n_+}: &\mbox{the set of all }n\times n \mbox{ positive semi-definite matrices};\\ \dbS^n_+: &\mbox{the set of all }n\times n \mbox{ positive-definite matrices}.\\ \end{array}\end{aligned}$$ Next, let $T>0$ be a fixed time horizon. For any $t\in[0,T)$ and Euclidean space $\dbH$, let $$\ba{ll} C([t,T];\dbH)=\Big\{\f:[t,T]\to\dbH\bigm|\f(\cd)\hb{ is continuous }\1n\Big\},\\ \ns\ds L^p(t,T;\dbH)=\left\{\f:[t,T]\to\dbH\biggm|\int_t^T|\f(s)|^pds<\i\right\},\q1\les p<\i,\\ \ns\ds L^\infty(t,T;\dbH)=\left\{\f:[t,T]\to\dbH\biggm|\esssup_{s\in[t,T]}|\f(s)|<\i\right\}.\ea$$ We denote $$\ba{ll} \ns\ds L^2_{\cF_T}(\O;\dbH)=\Big\{\xi:\O\to\dbH\bigm|\xi\hb{ is $\cF_T$-measurable, }\dbE|\xi|^2<\i\Big\},\\ \ns\ds L_\dbF^2(t,T;\dbH)=\left\{\f:[t,T]\times\O\to\dbH\bigm|\f(\cd)\hb{ is $\dbF$-progressively measurable},\dbE\int^T_t|\f(s)|^2ds<\i\right\},\\ \ns\ds L_\dbF^2(\O;C([t,T];\dbH))=\left\{\f:[t,T]\times\O\to\dbH\bigm|\f(\cd)\hb{ is $\dbF$-adapted, continuous, }\dbE\left[\sup_{s\in[t,T]}|\f(s)|^2\right]<\i\right\},\\ \ns\ds L^2_\dbF(\O;L^1(t,T;\dbH))=\left\{\f:[t,T]\times \O\to\dbH\bigm|\f(\cd)\hb{ is $\dbF$-progressively measurable}, \dbE\left(\int_t^T|\f(s)|ds\right)^2<\i\right\}.\ea$$ For an $\dbS^n$-valued function $F(\cd)$ on $[t,T]$, we use the notation $F(\cd)\gg0$ to indicate that $F(\cd)$ is uniformly positive definite on $[t,T]$, i.e., there exists a constant $\delta>0$ such that $$F(s)\ges\delta I,\qq\ae~s\in[t,T].$$ Now we start to formulate our system. We identify the state space of the chain $\a$ with a finite set $S :=\{1, 2 \dots, D\}$, where $D\in \mathbb{N}$ and suppose that the chain is homogeneous and irreducible. To specify statistical or probabilistic properties of the chain $\a$, we define the generator $\l(t) := [\l_{ij}(t)]_{i, j = 1, 2, \dots, D}$ of the chain under $\mathbb{P}$. This is also called the rate matrix, or the $Q$-matrix. Here, for each $i, j = 1, 2, \dots, D$, $\l_{ij}(t)$ is the constant transition intensity of the chain from state $i$ to state $j$ at time $t$. Note that $\l_{ij}(t) \ge 0$, for $i \neq j$ and $\sum^{D}_{j = 1} \l_{ij}(t) = 0$, so $\l_{ii}(t) \le 0$. In what follows for each $i, j = 1, 2, \dots, D$ with $i \neq j$, we suppose that $\l_{ij}(t) > 0$, so $\l_{ii}(t) < 0$. For each fixed $j = 1, 2, \cdots, D$, let $N_j(t)$ be the number of jumps into state $j$ up to time $t$ and set $$\l_j (t) := \int_0^t\l_{\a(s-)\, j}I_{\{\a (s-)\neq j\}}ds=\sum^{D}_{i = 1, i \neq j}\int^{t}_{0}\l_{ij}(s)I_{\{\a(s-)=i\}} d s.$$ Following Elliott et al. [@elliott1994hmm], we have that for each $j=1,2,\cdots, D$, $$\begin{aligned} \label{eq:N} \widetilde{N}_j (t):=N_j(t)-\l_j(t)\end{aligned}$$ is an $(\dbF, \dbP)$-martingale. Consider the following controlled Markov regime switching linear stochastic differential equation (SDE, for short) on a finite horizon $[t,T]$: $$\label{state} \left\{ \begin{aligned} dX^u(s;t,x,i)&=\big[A(s,\a(s))X^u(s;t,x,i)+B(s,\a(s))u(s)+b(s,\a(s))\big]ds\\ &\qq +\big[C(s,\a(s))X^u(s;t,x,i)+D(s,\a(s))u(s)+\si(s,\a(s))\big]dW(s), \qq s\in[t,T], \\ X^u(t;t,x,i)&=x,\q \a(t)=i, \end{aligned} \right.$$ where $A(\cd,\cd), B(\cd,\cd), C(\cd,\cd), D(\cd,\cd )$ are given deterministic matrix-valued functions of proper dimensions, and $b(\cd,\cd), \si(\cd,\cd)$ are vector-valued $\dbF$-progressively measurable processes. In the above, $X^u(\cd\,;t,x,i)$, valued in $\dbR^n$, is the [*state process*]{}, and $u(\cd)$, valued in $\dbR^m$, is the [*control process*]{}. Any $u(\cd)$ is called an [*admissible control*]{} on $[t,T]$, if it belongs to the following Hilbert space: $$\cU[t,T]=\left\{u:[t,T]\times\O\to\dbR^m\bigm|u(\cd)\hb{ is $\dbF$-progressively measurable, }\dbE\int_t^T|u(s)|^2ds<\i\right\}.$$ For any admissible control $u(\cd)$, we consider the following general quadratic cost functional: $$\label{cost} {\small \begin{aligned} J(t,x,i;u(\cd))\deq&\dbE\Bigg\{\Blan G(T,\a(T))X^u(T;t,x,i)+2g(T,\a(T)),X^u(T;t,x,i)\Bran\\ &\qq +\int_t^T\bigg[\Blan Q(s,\a(s))X^u(s;t,x,i)+2q(s,\a(s)), X^u(s;t,x,i)\Bran\\ &\qq\qq\qq+2\Blan\1nS(s,\a(s))X^u(s;t,x,i),u(s)\Bran +\Blan R(s,\a(s))u(s)+2\rho(s,\a(s)),u(s)\Bran\bigg] ds\Bigg\}, \end{aligned} }$$ where $G(T,i)$ is a symmetric matrix, $Q(\cd,i)$, $S(\cd,i)$, $R(\cd,i), i=1,\cdots,D$ are deterministic matrix-valued functions of proper dimensions with $Q(\cd,i)^\top=Q(\cd,i)$, $R(\cd,i)^\top=R(\cd,i)$; $g(T,\cd)$ is allowed to be an $\cF_T$-measurable random variable and $q(\cd,\cd), \rho(\cd,\cd)$ are allowed to be vector-valued $\dbF$-progressively measurable processes. The following standard assumptions will be in force throughout this paper. [**(H1)**]{} The coefficients of the state equation satisfy the following: for each $i\in \cS$, $$\left\{\2n\ba{ll} \ns\ds A(\cd,i)\in L^1(0,T;\dbR^{n\times n}),\q B(\cd,i)\in L^2(0,T;\dbR^{n\times m}), \q b(\cd,i)\in L^2_\dbF(\O;L^1(0,T;\dbR^n)),\\ \ns\ds C(\cd,i)\in L^2(0,T;\dbR^{n\times n}),\q D(\cd,i)\in L^\i(0,T;\dbR^{n\times m}), \q\si(\cd,i)\in L_\dbF^2(0,T;\dbR^n).\ea\right.$$ [**(H2)**]{} The weighting coefficients in the cost functional satisfy the following: for each $i\in \cS$ $$\left\{\2n\ba{ll} \ns\ds G(T,i)\in\dbS^n,\q Q(\cd, i)\in L^1(0,T;\dbS^n),\q S(\cd, i)\in L^2(0,T;\dbR^{m\times n}),\q R(\cd, i)\in L^\infty(0,T;\dbS^m),\\ \ns\ds g(T,i)\in L^2_{\cF_T}(\O;\dbR^n),\q q(\cd, i)\in L^2_\dbF(\O;L^1(0,T;\dbR^n)),\q\rho(\cd, i)\in L_\dbF^2(0,T;\dbR^m).\ea\right.$$ Now we sate the stochastic LQ optimal control problem for the Markov regime switching system as follows. **(M-SLQ) For any given initial pair $(t,x,i)\in[0,T)\times\dbR^n\times \cS$, find a $u^*(\cd)\in\cU[t,T]$, such that $$\label{optim} J(t,x,i;u^*(\cd))=\inf_{u(\cd)\in\cU[t,T]}J(t,x,i;u(\cd))\deq V(t,x,i).$$** Any $u^*(\cd)\in\cU[t,T]$ satisfying is called an [*optimal control*]{} of Problem (M-SLQ) for the initial pair $(t,x,i)$, and the corresponding path $X^*(\cd)\equiv X^{u^*}(\cd\,; t,x,i)$ is called an [*optimal state process*]{}; the pair $(X^*(\cd),u^*(\cd))$ is called an [*optimal pair*]{}. The function $V(\cd\,,\cd\, , \cd)$ is called the [*value function*]{} of Problem (M-SLQ). When $b(\cd,\cd), \si(\cd,\cd), g(T,\cd), q(\cd,\cd), \rho(\cd,\cd)=0$, we denote the corresponding Problem (M-SLQ) by Problem $\hb{(M-SLQ)}^0$. The corresponding cost functional and value function are denoted by $J^0(t,x,i;u(\cd))$ and $V^0(t,x,i)$, respectively. Similar to Sun et al. [@Sun2016olcls], we introduce the following definitions of open-loop (closed-loop) optimal control. \[sec:defnofopen-closeloop\] (i) An element $u^*(\cd)\in\cU[t,T]$ is called an [*open-loop optimal control*]{} of Problem (M-SLQ) for the initial pair $(t,x,i)\in[0,T]\times\dbR^n\times\cS$ if $$\begin{aligned} \label{open-opti} J(t,x,i;u^*(\cd))\les J(t,x,i;u(\cd)),\qq\forall u(\cd)\in\cU[t,T]. \end{aligned}$$ (ii) A pair $(\Th^*(\cd),v^*(\cd))\in L^2(t,T;\dbR^{m\times n})\times\cU[t,T]$ is called a [*closed-loop optimal strategy*]{} of Problem (M-SLQ) on $[t,T]$ if $$\begin{aligned} \label{closed-opti}\ba{ll} \ns\ds J(t,x,i;\Th^*(\cd)X^*(\cd)+v^*(\cd))\les J(t,x,i;u(\cd)),\qq\forall (x,i)\in\dbR^n\times\cS,\q u(\cd)\in\cU[t,T],\ea \end{aligned}$$ where $X^*(\cd)$ is the strong solution to the following closed-loop system: $$\begin{aligned} \label{closed-loop0}\left\{\2n\ba{ll} dX^*(s)=\Big\{\big[A(s,\a(s))+B(s,\a(s))\Th^*(s)\big]X^*(s)+B(s,\a(s))v^*(s)+b(s,\a(s))\Big\}ds\\ \qq\qq+\Big\{\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]X^*(s)+D(s,\a(s))v^*(s)+\si(s,\a(s))\Big\}dW(s), \\ X^*(t)=x.\ea\right.\hspace{-2cm} \end{aligned}$$ We emphasize that in the definition of closed-loop optimal strategy, (\[closed-opti\]) has to be true for all $(x,i)\in\dbR^n\times \cS$. One sees that if $(\Th^*(\cd),v^*(\cd))$ is a closed-loop optimal strategy of problem (M-SLQ) on $[t,T]$, then the outcome $u^*(\cd)\equiv\Th^*(\cd)X^*(\cd)+v^*(\cd)$ is an open-loop optimal control of Problem (M-SLQ) for the initial pair $(t,X^*(t),\a(t))$. Hence, the existence of closed-loop optimal strategies implies the existence of open-loop optimal controls. But, the existence of open-loop optimal controls does not necessarily imply the existence of a closed-loop optimal strategy. To simply notation of our further analysis, we introduce the following forward-backward stochastic differential equation (FBSDE for short) on a finite horizon $[t,T]$: $$\label{generalstate} \left\{ \begin{aligned} dX^u(s;t,x,i)=&\big[A(s,\a(s))X^u(s;t,x,i)+B(s,\a(s))u(s)+b(s,\a(s))\big]ds\\ &+\big[C(s,\a(s))X^u(s;t,x,i)+D(s,\a(s))u(s)+\si(s,\a(s))\big]dW(s),\\ dY^u(s;t,x,i)=&-\big[A(s,\a(s))^\top Y^u(s;t,x,i)+C(s,\a(s))^\top Z^u(s;t,x,i)\\ &+Q(s,\a(s))X^u(s;t,x,i)+S(s,\a(s))^\top u(s)+q(s,\a(s))\big]ds\\ &+Z^u(s;t,x,i)dW(s)+\sum_{k=1}^D\G_k^u(s;t,x,i)d\widetilde{N}_k(s)\qq s\in[t,T], \\ X^u(t;t,x,i)=&x, \q \a(t)=i, \q Y^u(T;t,x,i)=G(T,\a(T))X^u(T;t,x,i)+g(T,\a(T)). \end{aligned}\right.$$ The solution of the above FBSDE system is denoted by $(X^u(\cd\,;t,x,i), Y^u(\cd\,;t,x,i), Z^u(\cd\,;t,x,i),\G^u(\cd\,;t,x,i))$, where $\G^u(\cd\,;t,x,i):=(\G_1^u(\cd\,;t,x,i),\cdots, \G_D^u(\cd\,;t,x,i))$. If the control $u(\cd)$ is chose as $\Th(\cd)X(\cd)+v(\cd)$, we will use the notation $$(X^{\Th,v}(\cd\,;t,x,i), Y^{\Th,v}(\cd\,;t,x,i), Z^{\Th,v}(\cd\,;t,x,i),\G^{\Th,v}(\cd\,;t,x,i))$$ denoting by the solution of the above FBSDE. If $b(\cd,\cd)=\si(\cd,\cd)=q(\cd,\cd)=g(\cd,\cd)=0$, the solution of the above FBSDE is denoted by $$(X_0^u(\cd\,;t,x,i), Y_0^u(\cd\, ;t,x,i), Z_0^u(\cd\, ;t,x,i),\G_0^u(\cd\, ;t,x,i)).$$ Representation of the Cost Functional ===================================== In this section, we will present a representation of the cost functional for Problem (M-SLQ), which plays a crucial role in the study of open-loop/closed-loop solvability of Problem (M-SLQ). Unlike the method used in Yong and Zhou [@yong1999sch], we derive the representation of the cost functional using the technique of It[ô]{}’s formula. \[RP-cost\] *Let [(H1)–(H2)]{} hold and $(X^u(\cd\,;t,x,i), Y^u(\cd\,;t,x,i), Z^u(\cd\,;t,x,i),\G^u(\cd\,;t,x,i))$ is the solution of . Then for $(x,i,u(\cd))\in\dbR^n\times\cS\times\cU[t,T]$, $$\begin{aligned} \label{J-rep1} \begin{aligned} J^0(t,x,i;u(\cd))&=\langle M_2(t,i)u,u\rangle+2\langle M_1(t,i)x,u\rangle+\langle M_0(t,i)x,x\rangle,\\ J(t,x,i;u(\cd))&=\langle M_2(t,i)u,u\rangle+2\langle M_1(t,i)x,u\rangle+\langle M_0(t,i)x,x\rangle +2\langle \nu_t, u\rangle+2\langle y_t, x\rangle+c_t, \end{aligned} \end{aligned}$$ where $$\begin{aligned} M_0(t,i)x=&\dbE[Y_0^0(t;t,x,i)],\\ (M_1(t,i)x)(s)=&B(s,\a(s))^\top Y_0^0(s;t,x,i)+D(s,\a(s))^\top Z_0^0(s;t,x,i)\\ &+S(s,\a(s))X_0^0(s;t,x,i), \qq s\in[t,T],\\ (M_2(t,i)u(\cd))(s)=&B(s,\a(s))^\top Y_0^u(s;t,0,i)+D(s,\a(s))^\top Z_0^u(s;t,0,i)\\ &+S(s,\a(s))X_0^u(s;t,0,i)+R(s,\a(s))u(s), \qq s\in[t,T], \end{aligned}$$ and $$ \begin{aligned} y_t=&\dbE[Y^0(t;t,0,i)],\\ v_t(s)=&\big[B(s,\a(s))]^\top Y^0(s;t,0,i)+D(s,\a(s))^\top Z^0(s;t,0,i)\\ &\q+S(s,\a(s))X^0(s;t,0,i)+\rho(s,\a(s)), \qq s\in[t,T],\\ c_t=&\dbE\bigg[\llan G(T,\a(T))X^0(T;t,0,i)+2g(T,\a(T)),X^0(T;t,0,i)\rran\\ &\q+\int_t^T\llan Q(s,\a(s))X^0(s;t,0,i)+2q(s,\a(s)),X^0(s;t,0,i)\rran ds\bigg]. \end{aligned}$$* Let $$\begin{aligned} I_1:=&\dbE\bigg[\Blan G(T,\a(T))X_0^u(T;t,x,i),X_0^u(T;t,x,i)\Bran\bigg],\\ I_2:=&\dbE\bigg\{\int_t^T\bigg[\Blan Q(s,\a(s))X_0^u(s;t,x,i),X_0^u(s;t,x,i)\Bran\\ &\qq+2\Blan S(s,\a(s))X_0^u(s;t,x,i), u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\bigg]ds\bigg\}, \end{aligned}$$ and we have $$\begin{aligned} J^0(t,x,i;u(\cd))=I_1+I_2. \end{aligned}$$ Observing that $$\begin{aligned} X_0^u(\cd\,;t,x,i)=X_0^u(\cd\,;t,0,i)+X_0^0(\cd\,;t,x,i),\end{aligned}$$ and therefore $$\begin{aligned} I_1=&\dbE\bigg[\Blan G(T,\a(T))X_0^u(T;t,0,i),X_0^u(T;t,0,i)\Bran,\\ &\qq\q+2\Blan G(T,\a(T))X_0^0(T;t,x,i),X_0^u(T;t,0,i)\Bran+\Blan G(T,\a(T))X_0^0(T;t,x,i),X_0^0(T;t,x,i)\Bran\bigg],\\ I_2=&\dbE\bigg\{\int_t^T\bigg[\Blan Q(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\Bran+\Blan Q(s,\a(s))X_0^0(s;t,x,i), X_0^0(s;t,x,i)\Bran\\ &\qq\q+2\Blan Q(s,\a(s))X_0^0(s;t,x,i), X_0^u(s;t,0,i)\Bran+2\Blan S(s,\a(s))X_0^u(s;t,0,i), u(s)\Bran\\ &\qq\q+2\Blan S(s,\a(s))X_0^0(s;t,x,i),u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\bigg]ds\bigg\}. \end{aligned}$$ Applying Itô’s formula to $\langle Y_0^u(s;t,0,i),X_0^u(s;t,0,i)\rangle, \langle Y_0^0(s;t,x,i),X_0^u(s;t,0,i)\rangle$ and $\langle Y_0^0(s;t,x,i),X_0^0(s;t,x,i)\rangle$, we have $$\begin{aligned} J^0(t,x,i;u(\cd))&=&I_1+I_2\\ &=&\dbE\int_t^T\lan (M_2(t,i)u(\cd))(s), u(s)\ran ds+2\dbE\int_t^T\lan (M_1(t,i)x)(s), u(s)\ran ds +\lan\dbE[Y_0^0(t;t,x,i)],x\ran\\ &=&\langle M_2(t,i)u,u\rangle+2\langle M_1(t,i)x,u\rangle+\langle M_0(t,i)x,x\rangle.\end{aligned}$$ Let $$\begin{aligned} I_3&:=&\dbE\bigg[\Blan G(T,\a(T))X^u(T;t,x,i)+2g(T,\a(T)),X^u(T;t,x,i)\Bran\bigg],\\ I_4&:=&\dbE\bigg\{\int_t^T\bigg[\Blan Q(s,\a(s))X^u(s;t,x,i)+2q(s,\a(s)),X^u(s;t,x,i)\Bran\\ &&\qq\qq\q+2\Blan S(s,\a(s))X^u(s;t,x,i), u(s)\Bran+\Blan R(s,\a(s))u(s)+2\rho(s,\a(s)),u(s)\Bran\bigg]ds\bigg\},\end{aligned}$$ and we have $$\begin{aligned} J(t,x,i;u(\cd))=I_3+I_4.\end{aligned}$$ Observing that $$\begin{aligned} X^u(\cd\,;t,x,i)=X_0^u(\cd\,;t,x,i)+X^0(\cd\,;t,0,i),\end{aligned}$$ and therefore $$\begin{aligned} I_3=I_{31}+I_{32}+I_{33},\qq I_4=I_{41}+I_{42}+I_{43},\end{aligned}$$ where $$\begin{aligned} &&I_{31}:=\dbE\Blan G(T,\a(T))X_0^u(T;t,x,i),X_0^u(T;t,x,i)\Bran,\\ &&I_{32}:=2\dbE \Blan G(T,\a(T))X^0(T;t,0,i)+g(T,\a(T)),X_0^u(T;t,x,i)\Bran,\\ &&I_{33}:=\dbE\Blan G(T,\a(T))X^0(T;t,0,i)+2g(T,\a(T)),X^0(T;t,0,i),\end{aligned}$$ and $$\begin{aligned} I_{41}&:=\dbE\int_t^T\bigg[\Blan Q(s,\a(s))X_0^u(s;t,x,i),X_0^u(s;t,x,i)\Bran\\ &\qq\qq+2\Blan S(s,\a(s))X_0^u(s;t,x,i),u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\bigg]ds,\\ I_{42}&:=2\dbE\int_t^T\bigg[\Blan Q(s,\a(s))X^0(s;t,0,i)+q(s,\a(s)), X_0^u(s;t,x,i)\Bran\\ &\qq\qq+2\Blan S(s,\a(s))X^0(s;t,0,i)+\rho(s,\a(s)),u(s)\Bran\bigg]ds,\\ I_{43}&:=\dbE\int_t^T\bigg[\Blan Q(s,\a(s))X^0(s;t,0,i)+2q(s,\a(s)), X^0(s;t,0,i)\Bran\bigg]ds. \end{aligned}$$ Applying Itô’s formula to $\lan Y^0(s;t,0,i), X_0^u(s;t,x,i)\ran$ yields $$\begin{aligned} I_{32}+I_{42}=2\lan \dbE Y^0(t;t,0,i),x\ran+2\dbE\int_t^T\lan v_t(s), u(s)\ran ds=2\lan y_t, x\ran+2\lan \nu_t, u\ran.\end{aligned}$$ Noting that $$\begin{aligned} J^0(t,x,i;u(\cd))=I_{31}+I_{41},\qq c_t=I_{33}+I_{43}\end{aligned}$$ and therefore, $$\begin{aligned} J(t,x,i;u(\cd))&= I_3+I_4=(I_{31}+I_{41})+(I_{32}+I_{42})+(I_{33}+I_{43})\\ & = \langle M_2(t,i)u,u\rangle+2\langle M_1(t,i)x,u\rangle+\langle M_0(t,i)x,x\rangle +2\langle \nu_t, u\rangle+2\langle y_t, x\rangle+c_t. \end{aligned}$$ Next we shall show that the above characterizes of operators $M_0(t,i)$ and $M_2(t,i)$ is equivalent to the results obtained by using the technique of function analysis. \[sec:F-K\] *$M_0(\cd,i)$ defined in \[RP-cost\] admits the following Feynman-Kac representation: $$\label{L_0} M_0(t,i)=\dbE\bigg[\F(T;t,i)^\top G(T,\a(T))\F(T;t,i)+\int_t^T\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)ds\bigg],$$ where $\F(\cd\,;t,i)$ is the solution to the following SDE for $\dbR^{n\times n}$-valued process: $$\label{F} \left\{\begin{aligned} d\F(s;t,i)&=A(s,\a(s))\F(s;t,i)ds+C(s,\a(s))\F(s;t,i)dW(s),\qq s\in[t,T],\\ \F(t;t,i)&=I, \q \a(t)=i. \end{aligned} \right.$$ Furthermore, $M_0(t,i)$ also solves the following ordinary differential equations $$\label{3.8} \left\{\begin{aligned} \dot M_0(t,i)&+M_0(t,i)A(t,i)+A(t,i)^\top M_0(t,i)\\ &+C(t,i)^\top M_0(t,i)C(t,i)+Q(t,i)+\sum_{k=1}^D\lambda_{ik}(t)M_0(t,k)=0,\q (t,i)\in[0,T]\times\cS,\\ M_0(T,i)&=G(T,i), \q i\in\cS. \end{aligned} \right.$$* Let $\F(\cd;t,i)$ be the solution to (\[F\]). Then it is easy to verify that $$X_0^0(s;t,x,i)=\F(s;t,i)x.$$ Applying Itô’s formula to $\lan Y_0^0(s;t,x,i),X_0^0(s;t,x,i)\ran$, we can easily obtain $$\begin{aligned} &\dbE\big[\lan G(T,\a(T))X_0^0(T;t,x,i), X_0^0(T;t,x,i)\ran\big]\\ &=\lan \dbE [Y_0^0(t;t,x,i)], x\ran-\dbE\bigg[\int_t^TX_0^0(s;t,x,i)^\top Q(s,\a(s))X_0^0(s;t,x,i)ds\bigg].\end{aligned}$$ Therefore, $$\begin{aligned} \lan \dbE [Y_0^0(t;t,x,i)], x\ran&=\dbE\big[\lan G(T,\a(T))X_0^0(T;t,x,i), X_0^0(T;t,x,i)\ran\big]\\ &\qq+\dbE\bigg[\int_t^TX_0^0(s;t,x,i)^\top Q(s,\a(s))X_0^0(s;t,x,i)ds\bigg]\\ &=\dbE\big[\lan G(T,\a(T))\F(T;t,i)x, \F(T;t,i)x\ran\big]\\ &\qq+\dbE\bigg[\int_t^Tx^\top\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)xds\bigg]\\ &=\dbE\big[\lan \F(T;t,i)^\top G(T,\a(T))\F(T;t,i)x, x\ran\big]\\ &\qq+\dbE\bigg[\int_t^T\lan\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)x,x\ran ds\bigg]\\ &=\Blan\dbE\big[\F(T;t,i)^\top G(T,\a(T))\F(T;t,i)+\int_t^T\lan\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)ds\big]x, x\Bran.\\ \end{aligned}$$ Thus observing that $M_0(t,i)x=\dbE\big[Y_0^0(t;t,x,i)\big]$, we have $$\begin{aligned} M_0(t,i)=\dbE\bigg[\F(T;t,i)^\top G(T,\a(T))\F(T;t,i)+\int_t^T\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)ds\bigg].\end{aligned}$$ Suppose $\wt M(\cd,i)$ satisfy the ODE . Next we shall prove that $\wt M(\cd,i)=M_0(\cd,i)$. Observing that $$\begin{aligned} d\wt M(s,\a(s))=\dot{\wt{M}}(s,\a(s))ds+\sum_{k=1}^D\big[\wt M(s,k)-\wt M(s,\a(s-))\big]d\l_k(s)+\sum_{k=1}^D\big[\wt M(s,k)-\wt M(s,\a(s-))\big]d\wt N_k(s).\end{aligned}$$ Thus applying the Itô’s formula to $\F(s;t,i)^\top \wt M(s,\a(s))\F(s;t,i)$ leads to $$\begin{aligned} \begin{array}{rl} \wt M(t,i)=\5n & \dbE\bigg[\F(T;t,i)^\top G(T,\a(T))\F(T;t,i)+\int_t^T\F(s;t,i)^\top Q(s,\a(s))\F(s;t,i)ds\bigg]\\ =\5n & M_0(t,i). \end{array}\end{aligned}$$ Thus we complete our proof. *The operator $M_2(\cd,i)$ defined in \[RP-cost\] admits the following representation: $$\begin{aligned} M_2(t,i)=\h L_t^*G(T,\a(T))\h L_t+ L_t^*Q(\cd,\a(\cd))L_t + S(\cd,\a(\cd))L_t+ L_t^*S(\cd,\a(\cd))^\top +R(\cd,\a(\cd)), \end{aligned}$$ where the operators $$\begin{aligned} L_t:\cU[t,T]\rightarrow L_\dbF^2(t,T;\dbR^n), \qq \h L_t:\cU[t,T]\rightarrow L_{\dbF_T}^2(\O;\dbR^n) \end{aligned}$$ are defined as follows: $$\begin{aligned} (L_tu)(\cd)&=\F(\cd\,;t,i)\bigg\{\int_t^\cdot \F(r;t,i)^{-1}\big[B(r,\a(r))-C(r,\a(r))D(r,\a(r))\big]u(r)dr\\ &\qq\q\qq\qq+\int_t^\cdot \F(r;t,i)^{-1}D(r,\a(r))u(r)dW(r)\bigg\},\nonumber\\ \h L_tu&=(L_tu)(T),\end{aligned}$$ and $L_t^*$ and $\h L_t^*$ are the adjoint operators of $L_t$ and $\h L_t$, respectively.* Noting that the solution $X_0^u(\cd;t,0,i)$ of can be written as follows: $$\begin{aligned} X_0^u(s;t,0,i)&\5n=\5n&\F(s;t,i)\bigg\{\int_t^s\F(r;t,i)^{-1}\big[B(r,\a(r))-C(r,\a(r))D(r,\a(r))\big]u(r)dr\\ &&\qq\qq+\int_t^s\F(r;t,i)^{-1}D(r,\a(r))u(r)dW(r)\bigg\}\nonumber\\ &\5n=\5n&(L_tu)(s).\nonumber \end{aligned}$$ Applying Itô’s formula to $\lan Y_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran$ yields $$\begin{aligned} \lan (M_2(t,i))u,u\ran &\5n=\5n&\dbE\bigg\{\lan G(T,\a(T))X_0^u(T;t,0,i), X_0^u(T;t,0,i)\ran\\ &&\qq+\int_t^T\bigg[\lan Q(s,\a(s))X_0^u(s;t,0,i), X_0^u(s;t,0,i)\ran+\lan S(s,\a(s))X_0^u(s;t,0,i),u(s)\ran\\ &&\qq\qq\qq+\lan S(s,\a(s))^\top u(s), X_0^u(s;t,0,i)\ran+\lan R(s,\a(s))u(s),u(s)\ran \bigg]ds\bigg\}\\ &\5n=\5n&\dbE\big[\lan G(T,\a(T))\h L_t u, \h L_t u\ran\big]+\lan Q(\cd,\a(\cd))L_tu, L_tu\ran\\ &&\qq+\lan S(\cd,\a(\cd))L_tu,u\ran+\lan S(\cd,\a(\cd))^\top u, L_tu\ran+\lan R(\cd,\a(\cd))u,u\ran\\ &\5n=\5n&\Blan \big[\h L_t^*G(T,\a(T))\h L_t+ L_t^*Q(\cd,\a(\cd))L_t + S(\cd,\a(\cd))L_t+ L_t^*S(\cd,\a(\cd))^\top +R(\cd,\a(\cd))\big]u,u\Bran.\end{aligned}$$ Thus we complete the proof. From the representation of the cost functional, we have the following simple corollary. \[sec:frechetdifferential\] *Let [(H1)–(H2)]{} hold and $t\in[0,T)$ be given. For any $x\in\dbR^n, \eps\in\dbR$ and $u(\cd), v(\cd)\in\cU[t,T]$, the following holds: J(t,x,i;u()+v())=J(t,x,i;u())+\^2J\^0(t,0,i;v())+2\_t\^T|[M]{}(t,i)(x,u)(s),v(s)ds, where $$\begin{aligned} \label{eq:barM} \begin{aligned} \bar{M}(t,i)(x,u)(s):=\,&B(s,\a(s))^\top Y^u(s;t,x,i)\1n+\1nD(s,\a(s))^\top Z^u(s;t,x,i)\\ &+S(s,\a(s))X^u(s;t,x,i)\1n +\1nR(s,\a(s))u(s)\1n+\1n\rho(s,\a(s)), \q s\in[t, T]. \end{aligned} \end{aligned}$$ Consequently, the map $u(\cd)\mapsto J(t,x,i;u(\cd))$ is Fréchet differentiable with the Fréchet derivative given by $$\begin{aligned} \label{DJ} \cD J(t,x,i;u(\cd))(s)=2\bar{M}(t,i)(x,u)(s),\qq s\in[t,T], \end{aligned}$$ and (\[u+v-1\]) can also be written as $$\begin{aligned} \label{u+v-1*}J(t,x,i;u(\cd)+\eps v(\cd))=J(t,x,i;u(\cd))+\eps^2J^0(t,0,i;v(\cd))+\eps\dbE\int_t^T\lan\cD J(t,x,i;u(\cd))(s),v(s)\ran ds.\hspace{-0.8cm} \end{aligned}$$* From Proposition \[RP-cost\], we have $$\ba{ll} \ns\ds J(t,x,i;u(\cd)+\eps v(\cd))\\ \ns\ds=\lan M_2(t,i)(u+\eps v),u+\eps v\ran+2\lan M_1(t,i)x,u+\eps v\ran+\lan M_0(t,i)x,x\ran +2\lan \nu_t, u+\eps v\ran+2\lan y_t, x\ran+c_t\\ \ns\ds=\lan M_2(t,i)u,u\ran+2\eps\lan M_2(t,i)u,v\ran+\eps^2\lan M_2(t,i)v,v\ran+2\lan M_1(t,i)x,u\ran +2\eps\lan M_1(t,i)x,v\ran+\lan M_0(t,i)x,x\ran\\ \ns\ds\q~+2\lan \nu_t, u\ran+2\eps\lan \nu_t, v\ran+2\lan y_t, x\ran+c_t\\ \ns\ds=J(t,x,i;u(\cd))+\eps^2J^0(t,0;v(\cd))+2\eps\lan M_2(t,i)u+M_1(t,i)x+\nu_t,v\ran.\ea$$ From the representation of $M_1(t,i)$, $M_2(t,i)$ and $\nu_t$ in Proposition \[RP-cost\] and the fact $$X^u(\cd\,;t,x,i)=X_0^u(\cd\,;t,x,i)+X^0(\cd\,;t,0,i),$$ we see that $$\begin{aligned} (M_2(t,i)u)(s)+(M_1(t,i)x)(s)+\nu_t(s)&\5n=\5n&B(s,\a(s))^\top Y^u(s;t,x,i)+D(s,\a(s))^\top Z^u(s;t,x,i)\\ &&+S(s,\a(s))X^u(s;t,x,i)+R(s,\a(s))u(s)+\rho(s,\a(s))\\ &\5n=\5n&\bar{M}(t,i)(x,u)(s),\q s\in[t,T]. \end{aligned}$$ Open-loop Solvabilities ======================= We first present the equivalence between the open-loop solvability and the corresponding forward-backward differential equation system. *Let [(H1)–(H2)]{} hold and $(t,x,i)\in [t,T]\times \dbR^n\times \cS$ be given. An element $u(\cd)\in\cU[t,T]$ is an open-loop optimal control of Problem [(M-SLQ)]{} if and only if $J^0(t,0,i;v(\cd))\ge 0, \forall v(\cd)\in \cU[t,T]$ and the following stationary condition hold: $$\label{J_u=0} \begin{aligned} &B(s,\a(s))^\top Y^{u}(s;t,x,i)+D(s,\a(s))^\top Z^{u}(s;t,x,i)\\ &+S(s,\a(s))X^{u}(s;t,x,i)+R(s,\a(s))u(s)+\rho(s,\a(s))=0,\qq s\in[t,T], \end{aligned}$$ where $(X^{u}(\cd\,;t,x,i), Y^{u}(\cd\,;t,x,i), Z^{u}(\cd\,;t,x,i))$ is the adapted solution to the FBSDE .* By definition, $u(\cd)$ is an open-loop optimal control if and only if the following hold: $$\label{eq:oloopopticondi} J(t,x,i; u(\cd)+\eps v(\cd))-J(t,x,i;u)\ges 0,\q \forall v(\cd)\in \cU[t,T].$$ While from Corollary \[sec:frechetdifferential\], we have $$\begin{aligned} J(t,x,i; u(\cd)+\eps v(\cd))- J(t,x,i;u)=\eps^2J^0(t,0,i;v(\cd))+2\eps\dbE\int_t^T\lan \bar{M}(t,i)(x,u)(s),v(s)\ran ds. \end{aligned}$$ Therefore, holds if and only if $J^0(t,0,i;v(\cd))\ges 0, \forall v(\cd)\in \cU[t,T]$ and $\bar{M}(t,i)(x,u)(s)=0, s\in[t,T]$. Note the definition of $\bar{M}$ in and so the proof is completed. Note that if $u(\cd)$ happens to be an open-loop optimal control of Problem (M-SLQ), then the [*stationarity condition*]{} holds, which brings a coupling into the FBSDE . We call , together with the stationarity condition (\[J\_u=0\]), the [*optimality system*]{} for the open-loop optimal control of Problem (M-SLQ). Next we shall investigate the relationships between open-loop solvability and uniform convexity of the cost functional. We first introduce the definition of uniform convexity, which is from Zalinescu [@Zalinescu2002cagvs page 203] or [@Zalinescu1983ucf]. For a general normed space $(\dbH, \lVert\cd\rVert )$, the function $f:(\dbH, \lVert\cd\rVert)\mapsto \cl{\dbR}$ is said to be uniformly convex if there exists $h:\dbR_+\mapsto \cl{\dbR}_+$ with $h(t)>0$ for $t>0$ and $h(0)=0$ such that $$\begin{aligned} f(\eps x+(1-\l)y)\les \eps f(x)+(1-\eps)f(y)-\eps(1-\eps)h(\rVert x-y\rVert),\ \forall x, y\in \mbox{\rm dom} f, \, \eps\in [0,1]. \end{aligned}$$ *The cost functional $J(t,x,i;u(\cd))$ is uniformly convex if and only if $M_2(t,i)\ges\eps I$ for some $\eps>0$, which is also equivalent to $$\begin{aligned} \label{J>l} J^0(t,0,i;u(\cd)) \ges \eps \dbE\int_t^T|u(s)|^2ds,\qq\forall u(\cd)\in\cU[t,T], \end{aligned}$$ for some $\eps>0$.* From Proposition \[RP-cost\], we can see that for any $u(\cd), v(\cd)\in \cU[t,T]$ and $\eps\in[0,1]$, $$\begin{aligned} &J(t,x,i;\eps u(\cd)+(1-\eps)v(\cd))\\ &=\langle M_2(t,i)(\eps u+(1-\eps)v,\eps u+(1-\eps)v\rangle+2\langle M_1(t,i)x,\eps u+(1-\eps)v\rangle\\ &\q+\langle M_0(t,i)x,x\rangle +2\langle \nu_t, \eps u+(1-\eps)v\rangle+2\langle y_t, x\rangle+c_t\\ &=\eps \big[\langle M_2(t,i)u,u\rangle+2\langle M_1(t,i)x,u\rangle+\langle M_0(t,i)x,x\rangle +2\langle \nu_t, u\rangle+2\langle y_t, x\rangle+c_t\big]\\ &\q+(1-\eps)\big[\langle M_2(t,i)v,v\rangle+2\langle M_1(t,i)x,v\rangle+\langle M_0(t,i)x,x\rangle +2\langle \nu_t, v\rangle+2\langle y_t, x\rangle+c_t\big]\\ &\q-\eps(1-\eps)\lan M_2(t,i)(u-v),u-v\ran. \end{aligned}$$ Thus from the definition of uniformly convex, the cost functional $J(t,x,i;u(\cd))$ is uniformly convex if and only if there exists $h:\dbR_+\mapsto \cl{\dbR}_+$ with $h(t)>0$ for $t>0$ and $h(0)=0$ such that $$\lan M_2(t,i)(u-v),u-v\ran\ges h(\lVert u-v\rVert),$$ which equivalent to $M_2(t,i)\ges \eps I$ for some $\eps>0$. From Proposition \[RP-cost\], we have $$\begin{aligned} J^0(t,0,i;u(\cd))&=\langle M_2(t,i)u,u\rangle. \end{aligned}$$ Therefore, $M_2(t,i)>\eps I$ for some $\eps>0$ if and only if $$\begin{aligned} J^0(t,0,i;u(\cd)) \ges \eps \dbE\int_t^T|u(s)|^2ds,\qq\forall u(\cd)\in\cU[t,T].\end{aligned}$$ Thus the proof is completed. From the definition of uniform convexity, one can easily verify that $J^0(t,x,i;u(\cd))$ is uniformly convex if and only if is satisfied. So the uniform convexity of $J(t,x,i;u(\cd))$ is equivalent to the uniform convexity of $J^0(t,x,i;u(\cd))$. It is obvious that if the following standard conditions $$\begin{aligned} \label{classical}G(T,i)\ges0,\q R(s,i)\ges\d I,\q Q(s,i)-S(s,i)^\top R(s,i)^{-1}S(s,i)\ges0,\q i\in \cS, \q \ae~s\in[0,T], \end{aligned}$$ hold for some $\d>0$, then $$\begin{aligned} M_2(t,i)&=&\h L_t^*G(T,\a(T))\h L_t+L_t^*\big[Q(\cd,\a(\cd))-S(\cd,\a(\cd))^\top R(\cd,\a(\cd))^{-1}S(\cd,\a(\cd))\big]L_t\\ &&+\big[L_t^*S(\cd,\a(\cd))^\top R(\cd,\a(\cd))^{-{1\over2}}+R(\cd,\a(\cd))^{1\over2}\big]\big[R(\cd,\a(\cd))^{-{1\over2}}S(\cd,\a(\cd)L_t+R(\cd,\a(\cd))^{1\over2}\big]\\ &\ges& 0, \end{aligned}$$ which means that the functional $u(\cd)\mapsto J^0(t,0,i;u(\cd))$ is convex. In fact, one actually has the uniform convexity of the cost functional $J^0(t,0,i;u(\cd))$ under standard conditions (\[classical\]). We first present a lemma for proving the uniform convexity of $J^0(t,x,i;u(\cd))$. \[uniformconvex\] *For any $u(\cd)\in\cU[t,T]$, let $X_0^u(\cd\,;t,0,i)$ be the solution of with $x=0, b(\cd,\cd)=\si(\cd,\cd)=0.$ Then for any $\Th(\cd,i)\in L^2(t,T;\dbR^{m\times n}), i\in\cS$, there exists a constant $\gamma>0$ such that \_t\^T|u(s)-(s) X\_0\^u(s;t,0,i)|\^2ds \_t\^T|u(s)|\^2ds,u().* The proof is similar to Lemma 2.3 of Sun et al. [@Sun2016olcls] and so we omit it here. *Let [(H1)–(H2)]{} and [(\[classical\])]{} hold. Then for any $(t,i)\in[0,T)\times\cS$, the map $u(\cd)\mapsto J^0(t,0,i;u(\cd))$ is uniformly convex.* By Lemma \[uniformconvex\] (taking $\Th(\cd)=-R(\cd,\cd)^{-1}S(\cd,\cd)$), we have $$\begin{aligned} J^0(t,0,i;u(\cd)) &=&\dbE\bigg\{\langle G(T,\a(T))X_0^u(T;t,0,i),X_0^u(T;t,0,i)\rangle\\ &&\qq+\int_t^T\[\Blan Q(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\Bran \\ &&\qq\qq\qq+2\Blan S(s,\a(s))X_0^u(s;t,0,i),u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\]ds\bigg\}\\ &\ges&\dbE\int_t^T\[\Blan Q(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\Bran\\ &&\qq\qq+2\Blan S(s,\a(s))X_0^u(s;t,0,i),u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\]ds\\ &=&\dbE\int_t^T\[\Blan\big[Q(s,\a(s))\1n-\1nS(s,\a(s))^\top R(s,\a(s))^{-1}S(s,\a(s))\big]X_0^u(s;t,0,i),X_0^u(s;t,0,i)\Bran\1n\\ &&\qq\qq+\Blan R(s,\a(s))\big[u(s)\1n+\1nR(s,\a(s))^{-1}S(s,\a(s))X_0^u(s;t,0,i)\big],\1n\\ &&\qq\qq\qq\qq\qq\qq~u(s)+\1nR(s,\a(s))^{-1}S(s,\a(s))X_0^u(s;t,0,i)\Bran\]ds\\ &\ges&\d\dbE\int_t^T\big|u(s)+R(s,\a(s))^{-1}S(s,\a(s))X_0^u(s;t,0,i)\big|^2 ds\\ &\ges&\d\g\dbE\int_t^T|u(s)|^2 ds,\q \forall u(\cd)\in\cU[t,T], \end{aligned}$$ for some $\g>0$. This completes the proof. Next, we shall show that the uniform convexity of $J^0(t,x,i;u(\cd)$ implies the open-loop solvability of Problem (M-SLQ). \[sec:valueuniformconvex\] *Let [(H1)–(H2)]{} hold. Suppose the map $u(\cd)\mapsto J^0(t,0,i;u(\cd))$ is uniformly convex. Then Problem [(M-SLQ)]{} is uniquely open-loop solvable, and there exists a constant $\g\in\dbR$ such that V\^0(t,x,i)|x|\^2, (t,x)\^n. Note that in the above, the constant $\g$ does not have to be nonnegative.* First of all, by the uniform convexity of $u(\cd)\mapsto J^0(t,0,i;u(\cd))$, we may assume that $$J^0(t,0,i;u(\cd))\ges\l\,\dbE\1n\int_t^T|u(s)|^2ds,\qq\forall u(\cd)\in\cU[0,T],$$ for some $\l>0$. Thus, $u(\cd)\mapsto J^0(t,x,i;u(\cd))$ is uniformly convex for any given $(t,x)\in[0,T)\times\dbR^n$. By Corollary \[sec:frechetdifferential\], we have J(t,x,i;u())=J(t,x,i;0)+J\^0(t,0,i;u())+\_t\^TJ(t,x,i;0)(s),u(s)ds\ J(t,x,i;0)+J\^0(t,0,i;u())-[ł2]{}\_t\^T|u(s)|\^2ds-[12ł]{}\_t\^T|J(t,x,i;0)(s)|\^2ds\ \_t\^T|u(s)|\^2ds+J(t,x,i;0)-[12ł]{}\_t\^T| J(t,x,i;0)(s)|\^2ds,u(). Consequently, by a standard argument involving minimizing sequence and locally weak compactness of Hilbert spaces, we see that for any given initial pair $(t,x,i)\in [0,T)\times\dbR^n\times\cS$, Problem (M-SLQ) admits a unique open-loop optimal control. Moreover, when $b(\cd), \si(\cd), g, q(\cd), \rho(\cd)=0$, (\[uni-convex-prop1\]) implies that V\^0(t,x,i)J\^0(t,x,i;0)-[12ł]{}\_t\^T|J\^0(t,x,i;0)(s)|\^2ds. Note that the functions on the right-hand side of (\[uni-convex-prop2\]) are quadratic in $x$ and continuous in $t$. (\[uni-convex-prop0\]) follows immediately. Closed-loop Solvabilities ========================= In this section, we shall establish the equivalence between the closed-loop solvability and the existence of a regular solution to the Riccati equation. In the following, we first introduce some notation and the Riccati equation. Let $$\label{eq:hatsr} \begin{aligned} \hat S(s,i)&:=B(s,i)^\top P(s,i)+ D(s,i)^\top P(s,i)C(s,i)+S(s,i),\\ \hat R(s,i)&:=R(s,i)+D(s,i)^\top\1n P(s,i)D(s,i). \end{aligned}$$ The Riccati equation associated with Problem (M-SLQ) is $$\label{Riccati} \left\{ \begin{aligned} \dot P(s,i)&+P(s,i)A(s,i)+A(s,i)^\top P(s,i)+C(s,i)^\top P(s,i)C(s,i)\\ &-\hat{S}(s,i)^\top\hat{R}(s,i)^\dag\hat{S}(s,i)+Q(s,i)+\sum_{k=1}^D\l_{ik}(s)P(s,k)=0,\q \ae~s\in[0,T],\\ P(T,i)&=G(T,i). \end{aligned} \right.$$ A solution $P(\cd,\cd)\in C([0,T]\times \cS;\dbS^n)$ of (\[Riccati\]) is said to be [*regular*]{} if $$\label{eq:regular} \begin{aligned} \cR\big(\hat{S}(s,i)\big)&\subseteq\cR\big(\hat{R}(s,i)\big),\q \ae~s\in[0,T],\\ \hat{R}(\cd,\cd)^\dag\hat{S}(\cd,\cd)&\in L^2(0,T;\dbR^{m\times n}),\\ \hat{R}(s,i)&\ges0,\qq\ae~s\in[0,T]. \end{aligned}$$ A solution $P(\cd,\cd)$ of (\[Riccati\]) is said to be [*strongly regular*]{} if $$\begin{aligned} \label{strong-regular}\hat{R}(s,i)\ges \l I,\qq\ae~s\in[0,T], \end{aligned}$$ for some $\l>0$. The Riccati equation (\[Riccati\]) is said to be ([*strongly*]{}) [*regularly solvable*]{}, if it admits a (strongly) regular solution. Clearly, condition (\[strong-regular\]) implies (\[eq:regular\]). Thus, a strongly regular solution $P(\cd)$ must be regular. Moreover, if a regular solution of (\[Riccati\]) exists, it must be unique. *\[sec:closedloop-regusolu\] Let [(H1)–(H2)]{} hold. Problem [(M-SLQ)]{} is closed-loop solvable on $[0,T]$ if and only if the Riccati equation [(\[Riccati\])]{} admits a regular solution $P(\cd,\cd)\in C([0,T]\times\cS;\dbS^n)$ and the solution $(\eta(\cd),\z(\cd), \xi_1(\cd),\cds,\xi_D(\cd))$ of the following BSDE: $$\begin{aligned} \label{eta-zeta-xi}\left\{\2n\ba{ll} d\eta(s)=-\Big\{\big[A(s,\a(s))^\top\2n-\hat S(s,\a(s))^\top \hat R(s,\a(s))^\dag B(s,\a(s))^\top\big]\eta(s)\\ \qq\qq\q+\big[C(s,\a(s))^\top\2n-\hat S(s,\a(s))^\top \hat R(s,\a(s))^\dag D(s,\a(s))^\top\big]\z(s)\\ \ns\ds\qq\qq\q+\big[C(s,\a(s))^\top\2n-\hat S(s,\a(s))^\top \hat R(s,\a(s))^\dag D(s,\a(s))^\top\big]P(s,\a(s))\si(s,\a(s))\\ \ns\ds\qq\qq\q-\hat S(s,\a(s))^\top \hat R(s,\a(s))^\dag \rho(s,\a(s))+P(s,\a(s))b(s,\a(s))+q(s,\a(s))\Big\}ds\\ \ns\ds\qq\qq\q+\z(s) dW(s)+\sum_{k=1}^D\xi_k(s)d\wt{N}_k(s),\q s\in[0,T],\\ \ns\ds\eta(T)=g(T,i),\ea\right. \end{aligned}$$ satisfies $$\begin{aligned} \label{eta-zeta-regularity}\left\{\ba{ll} \hat \rho(s,i)\in\cR(\hat R(s,i)), \qq \ae~\as\\ \ns\ds \hat R(s,i)^\dag\hat \rho(s,i)\in L_\dbF^2(0,T;\dbR^m),\ea\right. \end{aligned}$$ with $$\begin{aligned} \label{eq:hatrrho} \hat \rho(s,i)&=B(s,i)^\top\eta(s)+D(s,i)^\top\z(s)+D(s,i)^\top P(s,i)\si(s,i)+\rho(s,i). \end{aligned}$$ In this case, Problem [(M-SLQ)]{} is closed-loop solvable on any $[t,T]$, and the closed-loop optimal strategy $(\Th^*(\cd),v^*(\cd))$ admits the following representation: $$\begin{aligned} \label{Th-v-rep}\left\{\2n\ba{ll} \Th^*(s)=-\hat R(s,\a(s))^\dag\hat S(s,\a(s)) +\big[I-\hat R(s,\a(s))^\dag\hat R(s,\a(s))\big]\Pi,\\ \ns\ds v^*(s)=-\hat R(s,\a(s))^\dag\hat\rho(s,\a(s))+\big[I-\hat R(s,\a(s))^\dag\hat R(s,\a(s))\big]\n(s),\ea\right. \end{aligned}$$ for some $\Pi(\cd)\in L^2(t,T;\dbR^{m\times n})$ and $\n(\cd)\in L_\dbF^2(t,T;\dbR^m)$, and the value function is given by $$\begin{aligned} \label{Value} V(t,x,i)=\dbE\bigg\{&\langle P(t,i)x,x\rangle+2\langle\eta(t),x\rangle+\int_t^T\[\hat P(s,\a(s))-\lan\hat R(s,\a(s))^\dag\hat\rho(s,\a(s)),\hat\rho(s,\a(s)) \ran\]ds\bigg\}, \end{aligned}$$ where $$\begin{aligned} \hat P(s,i):=\langle P(s,i)\si(s,i)+2\z(s),\si(s,i)\rangle+2\langle\eta(s),b(s,i)\rangle. \end{aligned}$$* [*Necessity.*]{} Let $(\Th^*(\cd),v^*(\cd))$ be a closed-loop optimal strategy of Problem (M-SLQ) over $[t,T]$ and set $$\begin{aligned} (X^*(\cd),Y^*(\cd),Z^*(\cd),\G^*(\cd)):=(X^{\Th^*,v^*}(\cd\,;t,x,i),Y^{\Th^*,v^*}(\cd\,;t,x,i),Z^{\Th^*,v^*}(\cd\,;t,x,i),\G^{\Th^*,v^*}(\cd\,;t,x,i)). \end{aligned}$$ Then the following stationary condition hold: $$\label{eq:statcondorig} \begin{aligned} B(s,\a(s))^\top Y^*(s)&+D(s,\a(s))^\top Z^*(s)+\big[S(s,\a(s))+R(s,\a(s))\Th^*(s)\big]X^*(s)\\ &+R(s,\a(s))v^*(s)+\rho(s,\a(s))=0\quad \mbox{a.e. a.s.} \end{aligned}$$ Since the above admits a solution for each $x\in\dbR^n$, and $(\Theta^*(\cd,\cd),v^*(\cd))$ is independent of $x$, by subtracting soulutions corresponding to $x$ and $0$, the later from the former, we see that for any $x\in \dbR^n$, as long as $(X(\cd),Y(\cd),Z(\cd),\G(\cd))$ is the adapted solution to the FBSDE $$\left\{ \begin{aligned} dX(s)=&\big[A(s,\a(s))+B(s,\a(s))\Th^*(s)\big] X(s)ds+\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]X(s)dW(s), \\ dY(s)=&-\Big[A(s,\a(s))^\top Y(s)+C(s,\a(s))^\top Z(s)+\big[Q(s,\a(s))+S(s,\a(s))^\top\Th^*(s)\big]X(s)\Big]ds\\ &+Z(s)dW(s)+\sum_{k=1}^D\G_k(s)d\widetilde{N}_k(s), \q s\in[t,T], \\ X(t)=& x, \qq Y(T)=G(T,\a(T))X(T), \end{aligned}\right.$$ one must have the following stationary condition: $$\label{eq:Newstationary} \begin{aligned} B(s,\a(s))^\top Y(s;t,x,i)&+D(s,\a(s))^\top Z(s;t,x,i)\\ &+\big[S(s,\a(s))+R(s,\a(s))\Th^*(s)\big]X(s;t,x,i)=0\quad \mbox{a.e. a.s.}, \end{aligned}$$ where $$\begin{aligned} &(X(\cd\,;t,x,i),Y(\cd\,;t,x,i),Z(\cd\,;t,x,i),\G(\cd\,;t,x,i))\\ &:=(X_0^{\Th^*,0}(\cd\,;t,x,i),Y_0^{\Th^*,0}(\cd\,;t,x,i),Z_0^{\Th^*,0}(\cd\,;t,x,i),\G_0^{\Th^*,0}(\cd\,;t,x,i)). \end{aligned}$$ Let $e_i$ denote the unit vector of $\dbR^n$ whose $i$-th component is one. Define, for $t\leq s\leq T$, $$\begin{aligned} X(s;t,i):=&(X(s;t,e_1,i),\cdots,X(s;t,e_n,i))\\ Y(s;t,i):=&(Y(s;t,e_1,i),\cdots,Y(s;t,e_n,i))\\ Z(s;t,i):=&(Z(s;t,e_1,i),\cdots,Z(s;t,e_n,i))\\ \G_{k}(s;t,i):=&(\G_{k}(s;t,e_1,i),\cdots,\G_{k}(s;t,e_n,i)). \end{aligned}$$ It is easy to verify that $$\label{eq:RelationXYZ} \begin{aligned} & X(s;t,x,i)=X(s;t,i)x, \q Y(s;t,x,i)=Y(s;t,i)x,\\ &Z(s;t,x,i)=Z(s;t,i)x, \q \G_{k}(s;t,x,i)=\G_{k}(s;t,i)x. \end{aligned}$$ In particular, if we set $P(t,i):=Y(t;t,i)$, then $$\begin{aligned} Y(t;t,x,i)=Y(t;t,i)x=P(t,i)x. \end{aligned}$$ Therefore, $$\begin{aligned} Y(s;t,i)x&=Y(s;t,x,i)=Y(s;s,X(s;t,x,i),\a(s))=Y(s;s,\a(s))X(s;t,x,i)\\ &=P(s,\a(s))X(s;t,i)x,\qq \mbox{for any } x\in\dbR^n, \end{aligned}$$ which leads to $$\label{eq:YP} Y(s;t,i)=P(s,\a(s))X(s;t,i).$$ Applying the It[ô]{}’s formula to $P(s,\a(s))X(s;t,i)$ yields $$\label{eq:PX} \begin{aligned} d[P(s,\a(s))X(s;t,i)]=&\bigg[\dot P(s,\a(s))+P(s,\a(s))\big[A(s,\a(s))+B(s,\a(s))\Th^*(s)\big]\\ &\q+\sum_{k=1}^D\l_{\a(s-)k}(s)\big[P(s,k)-P(s,\a(s-))\big]\bigg]X(s;t,i)ds\\ &+P(s,\a(s))\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]X(s;t,i)dW(s)\\ &+\sum_{k=1}^D\big[P(s,k)-P(s,\a(s-))\big]X(s;t,i)d\wt{N}_k(s) \end{aligned}$$ Observing that $Y(s;t,i)$ satisfied the following SDE $$\label{eq:Y} \left\{ \begin{aligned} dY(s;t,i)=&-\Big[A(s,\a(s))^\top Y(s;t,i)+C(s,\a(s))^\top Z(s;t,i)\\ &\q+\big[Q(s,\a(s))+S(s,\a(s))\Th^*(s)\big]X(s;t,i)\Big]ds\\ &+Z(s;t,i)dW(s)+\sum_{k=1}^D\G_k(s;t,i)d\widetilde{N}_k(s), \q s\in[0,T], \\ Y(T;0,i)=&G(T,\a(T))X(T;0,i). \end{aligned}\right.$$ Comparing the coefficients of and , we must have $$\label{eq:ZGP} \begin{aligned} Z(s;t,i)=&P(s,\a(s))\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]X(s;t,i),\\ \G_k(s;t,i)=&\big[P(s,k)-P(s,\a(s-))\big]X(s;t,i), \end{aligned}$$ and $$\label{eq:7} \begin{aligned} &\bigg\{\dot P(s,\a(s))+A(s,\a(s))^\top P(s,\a(s))+P(s,\a(s))A(s,\a(s))+C(s,\a(s))^\top P(s,\a(s))C(s,\a(s))\\ &\q+\bigg[P(s,\a(s))B(s,\a(s))+C(s,\a(s))^\top P(s,\a(s))D(s,\a(s))+S(s,\a(s))^\top\bigg]\Th^*(s)+Q(s,\a(s))\\ &\q+\sum_{k=1}^D\l_{\a(s-)k}(s)\big[P(s,k)-P(s,\a(s-))\big]\bigg\}X(s;t,i)=0, \end{aligned}$$ where the last equation leads to $$\label{eq:Pdiff} \begin{aligned} &\dot P(s,\a(s))+A(s,\a(s))^\top P(s,\a(s))+P(s,\a(s))A(s,\a(s))+C(s,\a(s))^\top P(s,\a(s))C(s,\a(s))\\ &\q+\bigg[P(s,\a(s))B(s,\a(s))+C(s,\a(s))^\top P(s,\a(s))D(s,\a(s))+S(s,\a(s))^\top\bigg]\Th^*(s)+Q(s,\a(s))\\ &\q+\sum_{k=1}^D\l_{\a(s-)k}(s)\big[P(s,k)-P(s,\a(s-))\big]=0. \end{aligned}$$ From , and , and the definition of $\hat{S}(\cd,\cd)$ and $\hat{R}(\cd,\cd)$ in , the stationary condition can be rewritten as $$\big[\hat{S}(s,\a(s))+\hat{R}(s,\a(s))\Th^*(s)\big]X(s;t,i)=0\quad \mbox{a.e. a.s.},$$ which yields $$\label{eq:statcond2} \hat{S}(s,\a(s)) +\hat{R}(s,\a(s))\Th^*(s)=0, \q i\in \dbS, \quad \mbox{a.e.}.$$ This implies $$\begin{aligned} \cR\big(\hat{S}(s,i)\subseteq\cR\big(\hat{R}(s,i)\big),\q \ae~s\in[0,T]. \end{aligned}$$ Using , we can rewrite as $$\label{eq:Pdiff2} \begin{aligned} \dot P(s,\a(s))&+\big[A(s,\a(s))+B(s,\a(s))\Th^*(s)\big]^\top P(s,\a(s))\\ &+P(s,\a(s))\big[A(s,\a(s))+B(s,\a(s))\Th^*(s)\big]\\ &+\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]^\top P(s,\a(s))\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]\\ &+\Th^*(s)^\top R(s,\a(s))\Th^*(s)+S(s,\a(s))^\top \Th^*(s)+ \Th^*(s)^\top S(s,\a(s))\\ &+Q(s,\a(s))+\sum_{k=1}^D\l_{\a(s-)k}(s)\big[P(s,k)-P(s,\a(s-))\big]=0. \end{aligned}$$ Since $P(T,i)=G(T,i)\in \dbS^n$ and $Q(\cd,\cd), R(\cd,\cd)$ are symmetric, we must have $P(\cd,\cd)\in C([t,T]\times S; \dbS^n) $ due to the uniqueness of the solution of . Let $\hat{R}(\cd,\cd)^\dag$ be the pseudo inverse of $\hat{R}(\cd,\cd)$, then the solution of admits the following representation $$\label{eq:3} \Th^*(s)=-\hat{R}(s,\a(s))^\dag \hat{S}(s,\a(s))+\big(I-\hat{R}(s,\a(s))^\dag\hat{R}(s,\a(s))\big)\Pi(s,\a(s)),$$ for some $\Pi(\cd,\cd)\in L^2(t,T;\dbR^{m\times n})$. Noting that $$\label{eq:stheta} \begin{aligned} \hat{S}(s,\a(s))^\top\Th^*(s)&=-\Th^*(s)\hat{R}(s,\a(s))\Th^*(s)\\ &=-\Th^*(s)\hat{R}(s,\a(s))\big[-\hat{R}(s,\a(s))^\dag \hat{S}(s,\a(s))+\big(I-\hat{R}(s,\a(s))^\dag\hat{R}(s,\a(s))\big)\Pi(s,\a(s))\big]\\ &=-\hat{S}(s,\a(s))^\top\hat{R}(s,\a(s))^\dag\hat{S}(s,\a(s)) \end{aligned}$$ Observing $\sum_{k=1}^D\l_{ik}(s)=0$ and substituting the above equation into , we obtain $$\label{eq:Pdiff3} \begin{aligned} \dot P(s,\a(s))&+A(s,\a(s))^\top P(s,\a(s))+P(s,\a(s))A(s,\a(s))\\ &+C(s,\a(s))^\top P(s,\a(s))C(s,\a(s))-\hat{S}(s,\a(s))^\top\hat{R}(s,\a(s))^\dag\hat{S}(s,\a(s))\\ &+Q(s,\a(s))+\sum_{k=1}^D\l_{\a(s-)k}(s)P(s,k)=0, \end{aligned}$$ which is equivalent to the Riccati equation . In the next, we try to determine $v^*(\cd)$. Let $$\left\{ \begin{aligned} \eta(s)=Y^*(s)&-P(s,\a(s))X^*(s)\\ \z(s)=Z^*(s)&-P(s,\a(s))[C(s,\a(s))+D(s,\a(s))\Th^*(s)]X^*(s) \q s\in[t,T]\\ &-P(s,\a(s))D(s,\a(s))v^*(s)-P(s,\a(s))\si(s,\a(s))\\ \xi_k(s)=\G_k^*(s)&-\big[P(s,k)-P(s,\a(s-))\big]X^*(s). \end{aligned} \right.$$ Then $$\begin{aligned} \label{eq:etas} d\eta(s)&=&dY^*(s)-dP(s,\a(s))\cd X^*(s)-P(s,\a(s))dX^*(s)\nonumber\\ &=&-\bigg[A(s,\a(s))^\top Y^*(s)+C(s,\a(s))^\top Z^*(s)+\big(Q(s,\a(s))+S(s,\a(s))^\top\Th^*(s)\big)X^*(s)\nonumber\\ &&\qq+S(s,\a(s))^\top v^*(s)+q(s,\a(s))\bigg]ds+Z^*(s)dW(s)+\sum_{k=1}^D\G_k^*(s)d\widetilde{N}_k(s)\nonumber\\ &&+\bigg\{\bigg[A(s,\a(s))^\top P(s,\a(s))+P(s,\a(s))A(s,\a(s))+C(s,\a(s))^\top P(s,\a(s))C(s,\a(s))\nonumber\\ &&\qq-\hat{S}(s,\a(s))^\top\hat{R}(s,\a(s))^\dag\hat{S}(s,\a(s))+Q(s,\a(s))\bigg]X^*(s)\nonumber\\ &&\qq-P(s,\a(s))\bigg[\bigg(A(s,\a(s))+B(s,\a(s))\Th^*(s)\bigg)X^*(s)\nonumber\\ &&\qq+B(s,\a(s))v^*(s)+b(s,\a(s))\bigg]\bigg\}ds\nonumber\\ &&-P(s,\a(s))\bigg[\bigg(C(s,\a(s))+D(s,\a(s))\Th^*(s)\bigg)X^*(s)+D(s,\a(s))v^*(s)\nonumber\\ &&\qq\qq\qq\q+\si(s,\a(s))\bigg]dW(s)-\sum_{k=1}^D\big[P(s,k)-P(s,\a(s-))\big]X^*(s)d\wt{N}_k(s)\nonumber\\ &=&-\bigg[A(s,\a(s))^\top \eta(s)+C(s,\a(s))^\top \z(s)+\hat{S}(s,\a(s))^\top\big[\Th^*(s)X^*(s)+v^*(s)\big]\nonumber\\ &&\qq+C(s,\a(s))^\top P(s,\a(s))\si(s,\a(s))+P(s,\a(s))b(s,\a(s))+q(s,\a(s))\nonumber\\ &&\qq+\hat{S}(s,\a(s))^\top\hat{R}(s,\a(s))^\dag\hat{S}(s,\a(s))X^*(s)\bigg]ds+\z(s)dW(s)+\sum_{k=1}^D\xi_k(s)\wt{N}_k(s)\nonumber\\ &=&-\bigg[A(s,\a(s))^\top \eta(s)+C(s,\a(s))^\top \z(s)+\hat{S}(s,\a(s))^\top v^*(s)+C(s,\a(s))^\top P(s,\a(s))\si(s,\a(s))\\ &&\qq+P(s,\a(s))b(s,\a(s))+q(s,\a(s))\bigg]ds+\z(s)dW(s)+\sum_{k=1}^D\xi_k(s)\wt{N}_k(s), \nonumber \end{aligned}$$ where the last equality follows from the equation . According to , we have $$\begin{aligned} 0=& B(s,\a(s))^\top Y^*(s)+D(s,\a(s))^\top Z^*(s)\\ &+\big[S(s,\a(s))+R(s,\a(s))\Th^*(s)\big]X^*(s)+R(s,\a(s))v^*(s)+\rho(s,\a(s))\\ =&B(s,\a(s))^\top \big[\eta(s)+P(s,\a(s))X^*(s)\big]\\ &+D(s,\a(s))^\top \bigg\{\z(s)+P(s,\a(s))\big[C(s,\a(s))+D(s,\a(s))\Th^*(s)\big]X^*(s)\\ &-P(s,\a(s))D(s,\a(s))v^*(s)-P(s,\a(s))\si(s,\a(s))\bigg\}\\ &+\big[S(s,\a(s))+R(s,\a(s))\Th^*(s)\big]X^*(s)+R(s,\a(s))v^*(s)+\rho(s,\a(s))\\ =&\big[\hat{S}(s,\a(s))+\hat{R}(s,\a(s))\Th^*(s)]X^*(s)+\hat{\rho}(s,\a(s))+\hat{R}(s,\a(s))v^*(s)\\ =&\hat{\rho}(s,\a(s))+\hat{R}(s,\a(s))v^*(s), \end{aligned}$$ where $\h \rho(s,i)$ is defined by . Thus we have $$\hat{\rho}(s,i)\in \cR(\hat{R}(s,i)),$$ and $$v^*(s)=-\hat{R}(s,\a(s))^\dag\hat{\rho}(s,\a(s))+\big[I-\hat{R}(s,\a(s))^\dag\hat{R}(s,\a(s))]\n(s,\a(s)),$$ for some $\n(\cd,i)\in L_\dbF^2(t,T;\dbR^m)$. Consequently, $$\begin{aligned} \hat{S}(s,\a(s))^\top v^*(s)&=-\Th^*(s)^\top \hat{R}(s,\a(s))v^*(s)\\ &=\Th^*(s)^\top \hat{R}(s,\a(s)) \hat{R}(s,\a(s))^\dag \hat{\rho}(s,\a(s))\\ &=-\hat{S}(s,\a(s))^\top\hat{R}(s,\a(s))^\dag \hat{\rho}(s,\a(s)). \end{aligned}$$ Thus observing the definition of $\hat{\rho}(s,\a(s))$ and substituting the above equation into yield the desired result of equation . [*Sufficiency.*]{} Applying It[ô]{}’s formula to $s\mapsto \langle P(s,\a(s))X(s)+2\eta(s),X(s)\rangle$ yields $$\label{eq:Jtxiu} \begin{aligned} &J(t,x,i;u(\cd))\\ &=\dbE \bigg\{\lan P(t,i)x+2\eta(t),x\ran+\int_t^T\bigg[\lan P(s,\a(s))\si(s,\a(s))+2\z(s),\si(s,\a(s))\ran +2\lan\eta(s),b(s,\a(s))\ran\bigg]ds\\ &\qq\q+\int_t^T\bigg[\Blan \hat Q(s,\a(s))X(s),X(s)\Bran+\Blan \hat R (s,\a(s))u(s)+2\hat S(s,\a(s)) X(s)+2\hat{\rho}(s,\a(s)),u(s)\Bran\\ &\qq\qq\qq\q+2\Blan\hat{S}(s,\a(s))^\top \hat{R}(s,\a(s))^\dag \hat\rho(s,\a(s)),X(s)\Bran\bigg]ds\bigg\}, \end{aligned}$$ where $$\label{eq:hatQ} \begin{aligned} \h Q(s,i)&:=\dot P(s,i)+P(s,i)A(s,i)+A(s,i)^\top P(s,i)\\ &\qq+C(s,i)^\top P(s,i)C(s,i)+Q(s,i)+\sum_{k=1}^D\l_{ik}(s)P(s,k). \end{aligned}$$ Let $\Th^*(\cd)$ and $v^*(\cd)$ be defined by . It is easy to verify that $$\begin{aligned} \hat S(s,\a(s))&=-\hat R(s,\a(s))\Th^*(s),\\ \hat Q(s,\a(s))&=\Th^*(s)^\top \hat R(s,\a(s))\Th^*(s),\\ \hat \rho(s,\a(s))&=-\hat R(s,\a(s))v^*(s),\\ -\hat S(s,\a(s))^\top \hat R(s,\a(s))^\dag\hat \rho(s,\a(s))&=-\Th^*(s)^\top \hat R(s,\a(s))v^*(s). \end{aligned}$$ Substituting these equation into yields $$\begin{aligned} &J(t,x,i;u(\cd))\\ &=\dbE \bigg\{\lan P(t,i)x+2\eta(t),x\ran +\int_t^T\bigg[\lan P(s,\a(s))\si(s,\a(s))+2\z(s),\si(s,\a(s))\ran +2\lan\eta(s),b(s,\a(s))\ran\bigg]ds \\ &\qq\q+\int_t^T\bigg[\Blan \Th^*(s)^\top \hat R(s,\a(s))\Th^*(s)X(s),X(s)\Bran\\ &\qq\qq\qq\q+\Blan\hat R(s,\a(s))u(s)-2 \hat R (s,\a(s))\big[\Th^*(s)X(s)+v^*(s)\big],u(s)\Bran \\ &\qq\qq\qq\q+2\Blan\Th^*(s)^\top \hat R(s,\a(s))v^*(s),X(s)\Bran \bigg]ds\bigg\}\\ &=\dbE \bigg\{\lan P(t,i)x+2\eta(t),x\ran +\int_t^T\bigg[\lan P(s,\a(s))\si(s,\a(s))+2\z(s),\si(s,\a(s))\ran \\ &\qq\q+2\lan\eta(s),b(s,\a(s))\ran-\lan\hat R(s,\a(s))v^*(s),v^*(s)\ran\bigg]ds\\ &\qq\q+\int_t^T\Blan \hat R(s,\a(s))\big[u(s)-\Th^*(s)X(s)-v^*(s)\big],u(s)-\Th^*(s)X(s)-v^*(s)\Bran ds\Bigg\}\\ &=J(t,x,i;\Th^*(\cd)X^*(\cd)+v^*(\cd))\\ &\qq+\dbE\int_t^T\Blan \hat R(s,\a(s))\big[u(s)-\Th^*(s)X(s)-v^*(s)\big],u(s)-\Th^*(s)X(s)-v^*(s)\Bran ds.\end{aligned}$$ For any $v(\cd)\in \cU[t,T]$, let $u(\cd):=\Th^*(\cd)X(\cd)+v(\cd)$ with $X(\cd)$ being the solution to the state equation under the closed-loop strategy $(\Th^*(\cd),v(\cd))$. Then the above implies that $$\begin{aligned} J(t,x,i;\Th^*(\cd)X(\cd)+v(\cd))=&J(t,x,i;\Th^*(\cd)X^*(\cd)+v^*(\cd))\\ &+\dbE\int_t^T\lan \hat R(s,\a(s))\big[v(s)-v^*(s)], v(s)-v^*(s)\ran ds.\end{aligned}$$ Therefore, $(\Th^*(\cd),v^*(\cd))$ is a closed-loop optimal strategy if and only if $$\begin{aligned} \dbE\int_t^T\lan \hat R(s,\a(s))\big[v(s)-v^*(s)], v(s)-v^*(s)\ran ds\ge 0, \q\forall v(\cd)\in\cU[t,T],\end{aligned}$$ or equivalently, $$\begin{aligned} \hat R(s,\a(s))\ge 0, \q a.e. s\in[t,T].\end{aligned}$$ Finally, the representation of the value function follows from the identity $$\begin{aligned} \lan \hat R(s,\a(s))v^*(s),v^*(s)\ran=\lan\hat R(s,\a(s))^\dag \hat \rho(s,\a(s)),\hat\rho(s,\a(s))\ran.\end{aligned}$$ Uniform convexity of the cost functional and the strongly regular solution of the Riccati equation ================================================================================================== We first present some properties for the solution to Lyapunov equation, which play a crucial role on establishing the equivalence between uniform convexity of the cost functional and the strongly regular solution of the Riccati equation. \[sec:reprecostclosedloop\] Let [(H1)–(H2)]{} hold and $\Th(\cd)\in L^2(0,T;\dbR^{m\times n})$ for $i\in \cS$. Let $P(\cd,i)\in C([0,T];\dbS^n), i\in\cS$ be the solution to the following Lyapunov equation: $$\label{P-Th} \left\{\begin{aligned} \dot P(s,i)&+P(s,i)A(s,i)+A(s,i)^\top P(s,i)+C(s,i)^\top P(s,i)C(s,i)\\ &+\hat S(s,i)^\top\Th(s)+\Th(s)^\top \hat S(s,i)+\Th(s)^\top \hat R(s,i)\Th(s)\\ &+Q(s,i)+\sum_{k=1}^D\lambda_{ik}(s)P(s,k)=0,\qq\ae~s\in[0,T],\\ P(T,i)&=G(T,i). \end{aligned}\right.$$ Then for any $(t,x,i)\in[0,T)\times\dbR^n\times\cS$ and $u(\cd,\cd)\in\cU[t,T]$, we have $$\begin{aligned} &J^0(t,x,i;\Th(\cd)X_0^{\Th,u}(\cd\,;t,x,i)+u(\cd))=\langle P(t,i)x,x\rangle+\dbE\int_t^T\Big\{\lan T_\a^1u(s),u(s)\ran+2\lan T_\a^2X_0^{\Th,u}(s;t,x,i),u(s)\ran\Big\}ds. \end{aligned}$$ where $X_0^{\Th,u}(\cd\,;t,x,i)$ is the solution of and $$\begin{aligned} T_\a^1u(\cd) &:=&\hat R(\cd,\a(\cd))u(\cd)\\ T_\a^2X_0^{\Th,u}(\cd;t,x,i) &:=&\big[\hat S(\cd,\a(\cd))+\hat R(\cd,\a(\cd))\Th(\cd)\big]X_0^{\Th,u}(\cd\,;t,x,i). \end{aligned}$$ For any $(t,x)\in[0,T)\times\dbR^n$ and $u(\cd)\in\cU[t,T]$, let $X_0^{x,u}$ be the solution of and set $$\begin{aligned} T_\a^0X_0^{\Th,u}(\cd\,;t,x,i):= \bigg[\dot P(\cd,\a(\cd))&+P(\cd,\a(\cd))A(\cd,\a(\cd))+A(\cd,\a(\cd))^\top P(\cd,\a(\cd))+C(\cd,\a(\cd))^\top P(\cd,\a(\cd))C(\cd,\a(\cd))\\ &+\hat S(\cd,\a(\cd))^\top\Th(\cd)+\Th(\cd)^\top \hat S(\cd,\a(\cd))+\Th(\cd)^\top \hat R(\cd,\a(\cd))\Th(\cd)\\ &+Q(\cd,\a(\cd))+\sum_{k=1}^D\lambda_{\a(\cd)k}(\cd)P(\cd,k)\bigg] X_0^{\Th,u}(\cd\,;t,x,i) \end{aligned}$$ Applying Itô’s formula to $s\mapsto\langle P(s,\a(s))X(s),X(s)\rangle$, we have $$\begin{aligned} &J^0(t,x,i;\Th(\cd)X_0^{\Th,u}(\cd\,;t,x,i)+u(\cd))\\ &=\dbE\Bigg\{\Blan G(T,\a(T))X_0^{\Th,u}(T;t,x,i),X_0^{\Th,u}(T;t,x,i)\Bran+\int_t^T\bigg[\Blan Q(s,\a(s))X_0^{\Th,u}(s;t,x,i), X_0^{\Th,u}(s;t,x,i)\Bran\\ &\qq\qq+2\Blan\1nS(s,\a(s))X_0^{\Th,u}(s;t,x,i),u(s)\Bran +\Blan R(s,\a(s))u(s),u(s)\Bran\bigg] ds\Bigg\}\\ &=\langle P(t,i)x,x\rangle+\dbE\int_t^T\Big\{\lan T_\a^0X_0^{\Th,u}(s;t,x,i),X_0^{\Th,u}(s;t,x,i)\ran+\lan T_\a^1u(s),u(s)\ran+2\lan T_\a^2X_0^{\Th,u}(s;t,x,i),u(s)\ran\Big\}ds\\ &=\langle P(t,i)x,x\rangle+\dbE\int_t^T\Big\{\lan T_\a^1u(s),u(s)\ran+2\lan T_\a^2X_0^{\Th,u}(s;t,x,i),u(s)\ran\Big\}ds. \end{aligned}$$ This completes the proof. \[sec:closedlooplyapunov\] Let [(H1)–(H2)]{} and [(\[J&gt;l\])]{} hold. Then for any $\Th(\cd)\in L^2(0,T;\dbR^{m\times n})$, the solution $P(\cd,\cd)\in C([0,T];\dbS^n)$ to the Lyapunov equation (\[P-Th\]) satisfies $$\begin{aligned} \label{Convex-prop-1} \hat R(t,i)\ges\l I, \q\ae~t\in[0,T],\qq\hb{and}\qq P(t,i)\ges \g I,\q\forall t\in[0,T], \end{aligned}$$ where $\g\in\dbR$ is the constant appears in [(\[uni-convex-prop0\])]{}. Let $\Th(\cd)\in L^2(0,T;\dbR^{m\times n})$ and let $P(\cd,\cd)$ be the solution to [(\[P-Th\])]{}. By [(\[J&gt;l\])]{} and Lemma \[sec:reprecostclosedloop\], we have $$\ba{ll} \ns\ds \l\dbE\int_t^T|\Th(s)X_0^{\Th,u}(s;t,0,i)+u(s)|^2ds\les J^0(t,0,i;\Th(\cd)X_0^{\Th,u}(\cd\,;t,0,i)+u(\cd))\\ \ns\ds=\dbE\int_t^T\Big\{\lan \hat R(s,\a(s))u(s),u(s)\ran+2\lan [\hat S(s,\a(s))+\hat R(s,\a(s))\Th(s)]X_0^{\Th,u}(s;t,0,i),u(s)\ran\Big\}ds.\ea$$ Hence, for any $u(\cd)\in\cU[t,T]$, the following holds: $$\begin{aligned} \label{P>LI}\ba{ll} \ns\ds\dbE\int_t^T\Big\{2\lan [\hat S(s,\a(s))+\big(\hat R(s,\a(s))-\l I\big)\Th(s)]X_0^{\Th,u}(s;t,0,i),u(s)\ran\\ \ns\ds\qq\q~+\lan \big(\hat R(s,\a(s))-\l I\big)u(s),u(s)\ran\Big\}ds\ges \l\dbE\int_0^T|\Th(s)X_0^{\Th,u}(s;t,0,i)|^2ds\ges0.\ea \end{aligned}$$ Let $$\F^\Th(\cd\,;t,i):=(X_0^{\Th,0}(\cd\,;t,e_1,i),\cds,X_0^{\Th,0}(\cd\,;t,e_n,i)).$$ Then it is easy to verify that $\F^\Th(\cd\,;t,i)$ is the solution to the following SDE for $\dbR^{n\times n}$-valued process: $$\begin{aligned} \label{FTh}\left\{\2n\ba{ll} \ns\ds d\F^\Th(s;t,i)=\big[A(s,\a(s))+B(s,\a(s))\Th(s)\big]\F^\Th(s;t,i)ds\\ \qq\qq\q+\big[C(s,\a(s))+D(s,\a(s))\Th(s)\big]\F^\Th(s;t,i)dW(s),\qq s\ges0,\\ \ns\ds\F^\Th(t;t,i)=I,\q \a(t)=i.\ea\right. \end{aligned}$$ Thus, $X_0^{\Th,u}(\cd\,;t,0,i)$ can be written as $$\begin{aligned} X_0^{\Th,u}(s;t,0,i)&=\F^\Th(s;t,i)\bigg\{\int_t^s\F^\Th(r;t,i)^{-1}\big[B(r,\a(r))-[C(r,\a(r))+D(r,\a(r))\Th(r)]D(r,\a(r))\big]u(r)dr\\ &\qq\qq\qq\q+\int_t^s\F^\Th(r;t,i)^{-1}D(r,\a(r))u(r)dW(r)\bigg\}.\end{aligned}$$ Now, fix any $u_0\in\dbR^m$, take $u(s)=u_0{\bf 1}_{[t,t+h]}(s)$, with $ 0\les t\les t+h\les T$. Consequently, (\[P&gt;LI\]) becomes $$\begin{aligned} \ba{ll} \ns\ds\dbE\int_t^{t+h}\Big\{2\lan [\hat S(s,\a(s))+\big(\hat R(s,\a(s))-\l I\big)\Th(s)]\hat \F (s;t,i),u_0\ran+\lan \big(\hat R(s,\a(s))-\l I\big)u_0,u_0\ran\Big\}ds\ges0,\ea\end{aligned}$$ where $$\begin{aligned} \hat \F(s;t,i)&=\F^\Th(s;t,i)\bigg\{\int_t^s\F^\Th(r;t,i)^{-1}\big[B(r,\a(r))-[C(r,\a(r))+D(r,\a(r))\Th(r)]D(r,\a(r))\big]u_0dr\\ &\qq\q\qq\qq+\int_t^s\F^\Th(r;t,i)^{-1}D(r,\a(r))u_0dW(r)\bigg\}.\end{aligned}$$ Dividing both sides of the above by $h$ and letting $h\to 0$, we obtain $$\lan\big(\hat R(t,i)-\l I\big)u_0,u_0\ran\ges 0,\qq\ae~t\in[0,T],\q \forall u_0\in\dbR^m.$$ The first inequality in (\[Convex-prop-1\]) follows. To prove the second, for any $(t,x)\in[0,T)\times\dbR^n$ and $u(\cd)\in\cU[t,T]$ and by Proposition \[sec:valueuniformconvex\] and Lemma \[sec:reprecostclosedloop\], we have $$\ba{ll} \ns\ds \g|x|^2\les V^0(t,x,i)\les J^0(t,x,i;\Th(\cd)X_0^{\Th,u}(\cd\,;t,x,i)+u(\cd))\\ \ns\ds\qq~\1n=\langle P(t,i)x,x\rangle\1n+\dbE\int_t^T\Big\{\lan \hat R(s,\a(s))u(s),u(s)\ran+2\lan [\hat S(s,\a(s))+\hat R(s,\a(s))\Th(s)]X_0^{\Th,u}(s;t,0,i),u(s)\ran\Big\}ds.\ea$$ In particular, by taking $u(\cd)=0$ in the above, we obtain $$\langle P(t,i)x,x\rangle\ges\g|x|^2,\qq\forall (t,x,i)\in[0,T]\times\dbR^n\times \cS,$$ and the second inequality therefore follows. Now we are in the position to prove the equivalence between the uniform convexity of the cost functional and the strongly regular solution of the Riccati equation. \[sec:unifconv-strongregusolu\] *Let [(H1)–(H2)]{} hold. Then the following statements are equivalent:* [(i)]{} The map $u(\cd)\mapsto J^0(t,0;u(\cd))$ is uniformly convex, i.e., there exists a $\l>0$ such that [(\[J&gt;l\])]{} holds. The Riccati equation [(\[Riccati\])]{} admits a strongly regular solution $P(\cd,\cd)\in C([0,T]\times \cS;\dbS^n)$. \(i) $\Ra$ (ii). Let $P_0(\cd,\cd)$ be the solution of $$\begin{aligned} \left\{\ba{l} \dot P_0(s,i)+P_0(s,i)A(s,i)+A(s,i)^\top P_0(s,i)\\ \qq\q\hspace{0.1cm}+C(s,i)^\top P_0(s,i)C(s,i)+Q(s,i)+\sum_{k=1}^D\l_{ik}(s)P_0(s,k)=0,\qq\ae~s\in[0,T],\\ P_0(T,i)=G(T,i).\ea\right. \end{aligned}$$ Applying Proposition \[sec:closedlooplyapunov\] with $\Th(\cd)=0$, we obtain that $$\hat R(s,i)\ges\l I,\q P_0(s,i)\ges\g I,\qq\ae~s\in[0,T].$$ Next, inductively, for $n = 0,1,2, \cdots$, we set $$\begin{aligned} \label{Iteration-i} \left\{\ba{l} \Th_n(s,i)=-\hat R(s,i)^{-1}\big[B(s,i)^\top P_n(s,i)+D(s,i)^\top P_n(s,i)C(s,i)+S(s,i)\big],\\ A_n(s,i)=A(s,i)+B(s,i)\Th_n(s,i),\\ C_n(s,i)=C(s,i)+D(s,i)\Th_n(s,i),\ea\right. \end{aligned}$$ and let $P_{n+1}$ be the solution of $$\begin{aligned} \left\{\ba{l} \dot P_{n+1}(s,i)+P_{n+1}(s,i)A_n(s,i)+A_n(s,i)^\top P_{n+1}(s,i)\\ \qq\q\hspace{0.5cm}+C_n(s,i)^\top P_{n+1}(s,i)C_n(s,i)+Q_n(s,i)+\sum_{k=1}^D\l_{ik}(s)P_{n+1}(s,k)=0,\qq\ae~s\in[0,T],\\ P_{n+1}(T,i)=G(T,i).\ea\right. \end{aligned}$$ By Proposition \[sec:closedlooplyapunov\], we see that $$\begin{aligned} \label{R+Pi-lowerbound}\left\{\ba{l} R(s,i)+D(s,i)^\top P_{n+1}(s,i)D(s,i)\ges\l I,\\ P_{n+1}(s,i)\ges\g I,\q\ae~s\in[0,T],\q n=0,1,2,\cdots. \ea\right. \end{aligned}$$ We now claim that $\{P_n(s,i)\}_{n=1}^\i$ converges uniformly in $C([0,T];\dbS^n)$. To show this, let $$\D_n(s,i)\deq P_n(s,i)-P_{n+1}(s,i),\qq \L_n(s,i)\deq\Th_{n-1}(s,i)-\Th_n(s,i),\qq n\ges1.$$ Then for $n\ges1$, we have $$\begin{aligned} \label{Di-equa1} -\dot \D_n(s,i)=&\dot{P}_{n+1}(s,i)-\dot{P}_n(s,i) \nonumber\\ =&P_n(s,i)A_{n-1}(s,i)+A_{n-1}(s,i)^\top P_n(s,i)+C_{n-1}(s,i)^\top P_n(s,i)C_{n-1}(s,i)\nonumber\\ &+\Th_{n-1}(s,i)^\top R(s,i)\Th_{n-1}(s,i)+S(s,i)^\top\Th_{n-1}(s,i)+\Th_{n-1}(s,i)^\top S(s,i)\nonumber\\ &-P_{n+1}(s,i)A_n(s,i)-A_n(s,i)^\top P_{n+1}(s,i)-C_n(s,i)^\top P_{n+1}(s,i)C_n(s,i)\nonumber\\ &-\Th_n(s,i)^\top R(s,i)\Th_n(s,i)-S(s,i)^\top\Th_n(s,i)-\Th_n(s,i)^\top S(s,i)+\sum_{k=1}^D\l_{ik}(s)\D_n(s,k)\\ =&\D_n(s,i)A_n(s,i)+A_n(s,i)^\top\D_n(s,i)+C_n(s,i)^\top\D_n(s,i)C_n(s,i)\nonumber\\ &+P_n(s,i)(A_{n-1}(s,i)-A_n(s,i))+(A_{n-1}(s,i)-A_n(s,i))^\top P_n(s,i)\nonumber\\ &+C_{n-1}(s,i)^\top P_n(s,i)C_{n-1}(s,i)-C_n(s,i)^\top P_n(s,i)C_n(s,i)\nonumber\\ &+\Th_{n-1}(s,i)^\top R(s,i)\Th_{n-1}(s,i)-\Th_n(s,i)^\top R(s,i)\Th_n(s,i)\nonumber\\ &+S(s,i)^\top\L_n(s,i)+\L_n(s,i)^\top S(s,i)+\sum_{k=1}^D\l_{ik}(s)\D_n(s,k).\nonumber \end{aligned}$$ By (\[Iteration-i\]), we have the following: $$\begin{aligned} \label{Di-equa2}\left\{\2n\ba{ll} \ns\ds A_{n-1}(s,i)-A_n(s,i)=B(s,i)\L_n(s,i),\\ C_{n-1}(s,i)-C_n(s,i)=D(s,i)\L_n(s,i),\\ \ns\ds C_{n-1}(s,i)^\top P_n(s,i)C_{n-1}(s,i)-C_n(s,i)^\top P_n(s,i)C_n(s,i)\\ =\L_n(s,i)^\top D(s,i)^\top P_n(s,i)D(s,i)\L_n(s,i)+C_n(s,i)^\top P_n(s,i)D(s,i)\L_n(s,i)\\ \q+\L_n(s,i)^\top D(s,i)^\top P_n(s,i)C_n(s,i),\\ \ns\ds \Th_{n-1}(s,i)^\top R(s,i)\Th_{n-1}(s,i)-\Th_n(s,i)^\top R(s,i)\Th_n(s,i)\\ =\L_n(s,i)^\top R(s,i)\L_n(s,i)+\L_n(s,i)^\top R(s,i)\Th_n(s,i)+\Th_n(s,i)^\top R(s,i)\L_n(s,i).\ea\right. \end{aligned}$$ Note that $$\begin{aligned} &B(s,i)^\top P_n(s,i)+D(s,i)^\top P_n(s,i)C_n(s,i)+R(s,i)\Th_n(s,i)+S(s,i)\\ &=B(s,i)^\top P_n(s,i)+D(s,i)^\top P_n(s,i)C(s,i)+S(s,i)+(R(s,i)+D(s,i)^\top P_n(s,i)D(s,i))\Th_n(s,i)=0. \end{aligned}$$ Thus, plugging (\[Di-equa2\]) into (\[Di-equa1\]) yields $$\begin{aligned} \label{Di-equa3}\ba{ll} \ns\ds&-\,\big[\dot\D_n(s,i)+\Delta_n(s,i)A_n(s,i)+A_n(s,i)^\top\D_n(s,i)+C_n(s,i)^\top\D_n(s,i)C_n(s,i)+\sum_{k=1}^D\l_{ik}(s)\D_n(s,k)\big]\\ \ns\ds&=P_n(s,i)B(s,i)\L_n(s,i)+\L_n(s,i)^\top B(s,i)^\top P_n(s,i)+\L_n(s,i)^\top D(s,i)^\top P_n(s,i)D(s,i)\L_n(s,i)\\ &\q+C_n(s,i)^\top P_n(s,i)D(s,i)\L_n(s,i)+\L_n(s,i)^\top D(s,i)^\top P_n(s,i)C_n(s,i)+\L_n(s,i)^\top R(s,i)\L_n(s,i)\\ \ns\ds&\q+\L_n(s,i)^\top R(s,i)\Th_n(s,i)+\Th_n(s,i)^\top R(s,i)\L_n(s,i)+S(s,i)^\top\L_n(s,i)+\L_n(s,i)^\top S(s,i)\\ \ns\ds&=\L_n(s,i)^\top\big[R(s,i)+D(s,i)^\top P_n(s,i)D(s,i)\big]\L_n(s,i)\\ &\q+\big[P_n(s,i)B(s,i)+C_n(s,i)^\top P_n(s,i)D(s,i)+\Th_n(s,i)^\top R(s,i)+S(s,i)^\top\big]\L_n(s,i)\\ &\q+\L_n(s,i)^\top\big[B(s,i)^\top P_n(s,i)+D(s,i)^\top P_n(s,i)C_n(s,i)+R(s,i)\Th_n(s,i)+S(s,i)\big]\\ \ns\ds&=\L_n(s,i)^\top\big[R(s,i)+D(s,i)^\top P_n(s,i)D(s,i)\big]\L_n(s,i)\ges0.\ea \end{aligned}$$ Noting that $\D_n(T,i)=0$ and using Proposition \[sec:F-K\], also noting (\[R+Pi-lowerbound\]), we obtain $$P_1(s,i)\ges P_n(s,i)\ges P_{n+1}(s,i)\ges\a I,\qq\forall s\in [0,T],\q\forall n\ges1.$$ Therefore, the sequence $\{P_n(s,i)\}_{n=1}^\i$ is uniformly bounded. Consequently, there exists a constant $K>0$ such that (noting (\[R+Pi-lowerbound\])) $$\begin{aligned} \label{Di-equa4}\left\{\2n\ba{ll} \ns\ds|P_n(s,i)|,\ |R_n(s,i)|\les K,\\ \ns\ds|\Th_n(s,i)|\les K\big(|B(s,i)|+|C(s,i)|+|S(s,i)|\big),\\ \ns\ds|A_n(s,i)|\les |A(s,i)|+K|B(s,i)|\big(|B(s,i)|+|C(s,i)|+|S(s,i)|\big),\\ \ns\ds|C_n(s,i)|\les |C(s,i)|+K\big(|B(s,i)|+|C(s,i)|+|S(s,i)|\big),\ea\right. \ae s\in [0,T],\forall i\in\cS, \forall n\ges0, \end{aligned}$$ where $R_n(s,i)\deq R(s,i)+D^\top(s,i)P_n(s,i)D(s,i)$. Observe that $$\begin{aligned} \label{Di-equa5} \L_n(s,i)=&\Th_{n-1}(s,i)-\Th_n(s,i) \nonumber\\ =&R_n(s,i)^{-1}D(s,i)^\top\D_{n-1}(s,i)D(s,i)R_{n-1}(s,i)^{-1}\hat S_n(s,i)\\ &-R_{n-1}(s,i)^{-1}\big[B(s,i)^\top\D_{n-1}(s,i)+D(s,i)^\top\D_{n-1}(s,i)C(s,i)\big]. \nonumber \end{aligned}$$ where $\hat S_n(s,i):=B(s,i)^\top P_n(s,i)+D(s,i)^\top P_n(s,i)C(s,i)+S(s,i)$. Thus, noting (\[Di-equa4\]), one has $$\begin{aligned} \label{3.22}\ba{ll} \ns\ds|\L_n(s,i)^\top R_n(s,i)\L_n(s,i)|\les\(|\Th_n(s,i)|+|\Th_{n-1}(s,i)|\)\,|R_n(s,i)|\, |\Th_{n-1}(s,i)-\Th_n(s,i)|\\ \ns\ds\qq\qq\qq\qq\qq\q\2n~\les K\(|B(s,i)|+|C(s,i)|+|S(s,i)|\)^2|\D_{n-1}(s,i)|.\ea \end{aligned}$$ Equation (\[Di-equa3\]), together with $\D_n(T,i)=0$, implies that $$\begin{aligned} \D_n(s,i)=\int^T_s\big[&\D_n(r,i)A_n(r,i)+A_n(r,i)^\top\D_n(r,i)+C_n(r,i)^\top\D_n(r,i)C_n(r,i)\\ &+\L_n(r,i)^\top R_n(r,i)\L_n(r,i)+\sum_{k=1}^D\l_{ik}(r)\D_n(r,k)\big]dr. \end{aligned}$$ Making use of (\[3.22\]) and still noting (\[Di-equa4\]), we get $$\begin{aligned} \label{eq:D_nineq} |\D_n(s,i)|\les \int^T_s\f(r)\[\Big\vert\sum_{k=1}^D\D_n(r,k)\Big\vert+|\D_{n-1}(r,i)|\]dr,\qq\forall s\in[0,T],\q\forall n\ges1, \end{aligned}$$ where $\f(\cd)$ is a nonnegative integrable function independent of $\D_n(\cd, \cd)$. Let $$\begin{aligned} \Vert\D_n(s)\Vert:=\max_{k=1}^D|\D_n(s,k)|. \end{aligned}$$ Thus from (\[eq:D\_nineq\]), we have $$\begin{aligned} \Vert\D_n(s)\Vert\les \int^T_s\f(r)\[\Vert\D_n(r)\Vert+\Vert\D_{n-1}(r)\Vert\]dr,\qq\forall s\in[0,T],\q\forall n\ges1, \end{aligned}$$ By Gronwall’s inequality, $$\Vert\D_n(s)\Vert\les e^{\int_0^T\f(r)dr}\int^T_s\f(r)\Vert\D_{n-1}(r)\Vert dr\equiv c\int^T_s\f(r)\Vert\D_{n-1}(r) .$$ Set $$a\deq\max_{0\les s\les T}\Vert\D_0(s)\Vert.$$ By induction, we deduce that $$||\D_n(s)||\les a{c^n\over n!}\(\int_s^T\f(r)dr\)^n,\qq\forall s\in[0,T],$$ which implies the uniform convergence of $\{P_n(\cd,\cd)\}_{n=1}^\i$. We denote $P(\cd,\cd)$ the limit of $\{P_n(\cd,\cd)\}_{n=1}^\infty$, then (noting (\[R+Pi-lowerbound\])) $$R(s,i)+D(s,i)^\top P(s,i)D(s,i)=\lim_{n\to\i}R(s,i)+D(s,i)^\top P_n(s,i)D(s,i)\ges\eps I, \qq\ae~s\in[0,T],$$ and as $n\to\infty$, $$\left\{\2n\ba{ll} \ns\ds\Th_n(s,i)\to-\hat R(s,i)\hat S(s,i)\equiv\Th(s) & \hb{in $L^2$},\\ \ns\ds A_n(s,i)\to A(s,i)+B(s,i)\Th(s) & \hb{in $L^1$},\\ \ns\ds C_n(s,i)\to C(s,i)+D(s,i)\Th(s) & \hb{in $L^2$}.\ea\right.$$ Therefore, $P(\cd,\cd)$ satisfies the following equation: $$\left\{\2n\ba{ll} \ns\ds\dot P(s,i)+P(s,i)\big[A(s,i)+B(s,i)\Th(s)\big]+\big[A(s,i)+B(s,i)\Th(s)\big]^\top P(s,i)\\ \ns\ds \q+\big[C(s,i)+D(s,i)\Th(s)\big]^\top P(s,i)\big[C(s,i)+D(s,i)\Th(s)\big]+\Th(s)^\top R(s,i)\Th(s)\\ \q+S(s,i)^\top\Th(s)+\Th(s)^\top S(s,i)+Q(s,i)+\sum_{k=1}^D\l_{ik}(s)P(s,k)=0,\qq\ae~s\in[0,T],\\ \ns\ds P(T,i)=G(T,i),\ea\right.$$ which is equivalent to (\[Riccati\]). \(ii) $\Ra$ (i). Let $P(\cd,\cd)$ be the strongly regular solution of (\[Riccati\]). Then there exists a $\eps>0$ such that R(s,i) I,æ s. Set $$\Th(s)\deq -\hat R(s,\a(s))\hat S(s,\a(s))\in L^2(0,T;\dbR^{m\times n}).$$ For any $u(\cd)\in\cU[0,T]$, let $X_0^u(\cd\,;t,0,i)$ be the solution of $$\left\{\2n\ba{ll} \ns\ds dX_0^u(s;t,0,i)=\big[A(s,\a(s))X_0^u(s;t,0,i)+B(s,\a(s))u(s)\big]ds\\ \qq\qq\q+\big[C(s,\a(s))X_0^u(s;t,0,i)+D(s,\a(s))u(s)\big]dW(s),\qq s\in[t,T], \\ \ns\ds\hspace{0.2cm} X_0^{0,u}(t)=0.\ea\right.$$ Applying Itô’s formula to $s\mapsto\langle P(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\rangle$, we have $$\begin{aligned} &J^0(t,0;u(\cd))\\ =&\dbE\Bigg\{\Blan G(T,\a(T))X_0^u(T;t,0,i),X_0^u(T;t,0,i)\Bran + \int_t^T\bigg[\Blan Q(s,\a(s))X_0^u(s;t,0,i), X_0^u(s;t,0,i)\Bran\\ &\qq\q+2\Blan\1nS(s,\a(s))X_0^u(s;t,0,i),u(s)\Bran+\Blan R(s,\a(s))u(s),u(s)\Bran\bigg] ds\Bigg\}\\ =&\dbE\int_t^T\[\lan\dot P(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran\\ &\qq\q+\lan P(s,\a(s))\big[A(s,\a(s))X_0^u(s;t,0,i)+B(s,\a(s))u(s)\big],X_0^u(s;t,0,i)\ran\\ &\qq\q+\lan P(s,\a(s))X_0^u(s;t,0,i),A(s,\a(s))X_0^u(s;t,0,i)+B(s,\a(s))u(s)\ran\\ &\qq\q+\lan P(s,\a(s))\big[C(s,\a(s))X_0^u(s;t,0,i)+D(s,\a(s))u(s)\big],C(s,\a(s))X_0^u(s;t,0,i)+D(s,\a(s))u(s)\ran\\ &\qq\q+\lan Q(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran+2\lan S(s,\a(s))X_0^u(s;t,0,i),u(s)\ran\\ &\qq\q+\lan R(s,\a(s))u(s),u(s)\ran+\lan \sum_{k=1}^D\l_{\a(s-),k}(s)P(s,k)X_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran\]ds\\ =&\dbE\int_t^T\[\lan\h Q(s,\a(s))X_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran+2\lan\h S(s,\a(s))X_0^u(s;t,0,i),u(s)\ran+\lan\h R(s,\a(s))u(s),u(s)\ran\]ds\\ =&\dbE\int_t^T\big[\lan\Th(s)^\top\h R(s,\a(s))\Th(s) X_0^u(s;t,0,i),X_0^u(s;t,0,i)\ran \\ &\qq\q-2\lan\h R(s,\a(s))\Th(s) X_0^u(s;t,0,i),u(s)\ran+\lan\h R(s,\a(s))u(s),u(s)\ran\big]ds\\ =&\dbE\int_t^T\lan\big[\hat R(s,\a(s))\big[u(s)-\Th(s) X_0^u(s;t,0,i)\big],u(s)-\Th(s) X_0^u(s;t,0,i)\ran ds. \end{aligned}$$ Noting (\[iitoi\]) and making use of Lemma \[uniformconvex\], we obtain that $$\begin{aligned} J^0(t,0;u(\cd))=&\dbE\int_t^T\lan \hat R(s,\a(s))\big[u(s)-\Th(s) X_0^u(s;t,0,i)\big],u(s)-\Th(s) X_0^u(s;t,0,i)\ran ds\\ \ges&\l\g\dbE\int_t^T|u(s)|^2ds, \q\forall u(\cd)\in\cU[t,T], \end{aligned}$$ for some $\g>0$. Then (i) holds. From the first part of the proof of Theorem 4.6, we see that if (\[J&gt;l\]) holds, then the strongly regular solution of (\[Riccati\]) satisfies (\[strong-regular\]) with the same constant $\l>0$. Combining Theorem \[sec:closedlooplyapunov\] and Theorem \[sec:unifconv-strongregusolu\], we obtain the following corollary. \[sec:opt-open-cont-u\] Let $P(\cd,\cd)$ be the unique strongly regular solution of [(\[Riccati\])]{} with $(\eta(\cd),\z(\cd))$ being the adapted solution of [(\[eta-zeta-xi\])]{}. $\hat R(\cd,\cd)$ and $\hat \rho(\cd,\cd)$ are defined by and respectively. Suppose that [(H1)–(H2)]{} and [(\[J&gt;l\])]{} hold. Then Problem [(M-SLQ)]{} is uniquely open-loop solvable at any $(t,x,i)\in[0,T)\times\dbR^n\times \cS$ with the open-loop optimal control $u^*(\cd)$ being of a state feedback form: $$\begin{aligned} \label{opti-biaoshi} u^*(\cd)=-\hat R(\cd,\a(\cd))^{-1}\hat S(\cd,\a(\cd))X^*(\cd)-\hat R(\cd,\a(\cd))^{-1}\hat\rho(\cd,\a(\cd)) \end{aligned}$$ where $X^*(\cd)$ is the solution of the following closed-loop system: $$\begin{aligned} \label{eclosed-loop-state}\left\{\2n\ba{ll} \ns\ds dX^*(s)=\Big\{\big[A(s,\a(s))-B(s,\a(s))\hat R(s,\a(s))^{-1}\hat S(s,\a(s))\big]X^*(s)\\ \ns\ds\qq\qq\q~-B(s,\a(s))\hat R(s,\a(s))^{-1}\hat \rho(s,\a(s))+b(s,\a(s))\Big\}ds\\ \ns\ds\qq\qq~~\1n+\Big\{\big[C(s,\a(s))-D(s,\a(s))\hat R(s,\a(s))^{-1}\hat S(s,\a(s))\big]X^*\\ \ns\ds\qq\qq\qq~-D(s,\a(s))\hat R(s,\a(s))^{-1}\hat\rho(s,\a(s))+\si(s,\a(s))\Big\}dW(s),\qq s\in[t,T], \\ \ns\ds X^*(t)=x.\ea\right. \end{aligned}$$ By Theorem \[sec:unifconv-strongregusolu\], the Riccati equation (\[Riccati\]) admits a unique strongly regular solution $P(\cd,\cd)\in C([0,T]\times\cS;\dbS^n)$. Hence, the adapted solution $(\eta(\cd),\z(\cd))$ of [(\[eta-zeta-xi\])]{} satisfies (\[eta-zeta-regularity\]) automatically. Now applying Theorem \[sec:closedloop-regusolu\] and noting the remark right after Definition \[sec:defnofopen-closeloop\], we get the desired result. Under the assumptions of Corollary \[sec:opt-open-cont-u\], when $b(\cd,\cd), \si(\cd,\cd), g(\cd,\cd), q(\cd,\cd), \rho(\cd,\cd)=0$, the adapted solution of (\[eta-zeta-xi\]) is $(\eta(\cd),\z(\cd))\equiv(0,0)$. Thus, for Problem ${\rm(M-SLQ)}^0$, the unique optimal control $u^*(\cd)$ at initial pair $(t,x)\in[0,T)\times\dbR^n$ is given by $$\begin{aligned} \label{opti-biaoshi-0} u^*(\cd)=-\hat R(\cd,\a(\cd))^{-1}\hat S(\cd,\a(\cd))X^*(\cd), \end{aligned}$$ with $P(\cd,\cd)$ being the unique strongly regular solution of [(\[Riccati\])]{} and $X^*(\cd)$ being the solution of the following closed-loop system: $$\begin{aligned} \label{closed-loop-state-0}\left\{\2n\ba{ll} \ns\ds dX^*(s)=\big[A(s,\a(s))-B(s,\a(s))\hat R(s,\a(s))^{-1}\hat S(s,\a(s))\big]X^*(s)ds\\ \ns\ds\qq\qq~~\1n+\big[C(s,\a(s))-D(s,\a(s))\hat R(s,\a(s))^{-1}\hat S(s,\a(s))\big]X^*(s)dW(s),\qq s\in[t,T], \\ \ns\ds X^*(t)=x.\ea\right. \end{aligned}$$ Moreover, by (\[Value\]), the value function of Problem ${\rm(M-SLQ)}^0$ is given by V\^0(t,x,i)=P(t,i)x,x, (t,x,i)\^n. [99]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} J.M. Bismut. [Linear quadratic optimal stochastic control with random coefficients]{}. *[SIAM]{} Journal on Control and Optimization*, 140 (3): 0 419–444, 1976. S.P. Chen and J.M. Yong. [Stochastic linear quadratic optimal control problems]{}. *Applied Mathematics and Optimization*, 430 (1): 0 21–45, 2001. S.P. Chen, X.J. Li, and X.Y. Zhou. [Stochastic linear quadratic regulators with indefinite control weight costs]{}. *[SIAM]{} Journal on Control and Optimization*, 360 (5): 0 1685–1702, 1998. Y.Z. Hu and Bernt Oksendal. [Partial information linear quadratic control for jump diffusions]{}. *[SIAM]{} Journal on Control and Optimization*, 470 (4): 0 1744–1761, 2008. Y. Ji and H.J. Chizeck. [Controllability, stabilizability, and continuous-time Markovian jump linear quadratic control]{}. *[IEEE]{} Transactions on Automatic Control*, 350 (7): 0 777–788, 1990. Y. Ji and H.J. Chizeck. [Jump linear quadratic Gaussian control in continuous time]{}. *[IEEE]{} Transactions on Automatic Control*, 370 (12): 0 1884–1892, 1992. M. Kohlmann and S.J. Tang. [New developments in backward stochastic riccati equations and their applications]{}. In Michael Kohlmann and Shanjian Tang, editors, *Mathematical Finance*, pages 194–214. Birkhäuser, Basel, 2001. M. Kohlmann and S.J. Tang. [Global adapted solution of one-dimensional backward stochastic Riccati equations, with application to the meanvariance hedging]{}. *Stochastic Processes and their Applications*, 970 (2): 0 255–288, 2002. M. Kohlmann and S.J. Tang. [Multidimensional backward stochastic Riccati equations and applications]{}. *[SIAM]{} Journal on Control and Optimization*, 410 (6): 0 1696–1721, 2003. M. Kohlmann and S.J. Tang. [Minimization of risk and linear quadratic optimal control theory]{}. *[SIAM]{} Journal on Control and Optimization*, 420 (3): 0 1118–1142, 2003. H. Kushner. [Optimal stochastic control]{}. *[IRE]{} Transactions on Automatic Control*, 70 (5): 0 120–122, 1962. N. Li, Z. Wu, and Z.Y. Yu. [Indefinite stochastic linear-quadratic optimal control problems with random jumps and related stochastic Riccati equations]{}. *Science China Mathematics*, 610 (3): 0 563–576, 2018. X. Li and X.Y. Zhou. [Indefinite stochastic LQ controls with Markovian jumps in a finite time horizon]{}. *Communications in Information and Systems*, 20 (3): 0 265–282, 2002. X. Li, X.Y. Zhou, and M. Ait Rami. [Indefinite stochastic LQ control with jumps]{}. In *Proceedings of the 40th [IEEE]{} Conference on Decision and Control (Cat. No.01CH37228)*. [IEEE]{}, 2001. X. Li, X.Y. Zhou, and Mustapha Ait Rami. [Indefinite stochastic linear quadratic control with Markovian jumps in infinite time horizon]{}. *Journal of Global Optimization*, 270 (2): 0 149–175, 2003. Y.J. Liu, G. Yin, and X.Y. Zhou. [Near-optimal controls of random-switching LQ problems with indefinite control weight costs]{}. *Automatica*, 410 (6): 0 1063–1070, 2005. H.W. Mei and J.M. Yong. [Equilibrium strategies for time-inconsistent stochastic switching systems]{}. 0 arXiv:1712.09505 , 2017. R. Penrose. [A generalized inverse for matrices]{}. *Mathematical Proceedings of the Cambridge Philosophical Society*, 510 (3): 0 406–413, 1955. J.R. Sun and J.M. Yong. [Linear quadratic stochastic differential games: open-loop and closed-loop saddle points]{}. *SIAM Journal on Control and Optimization*, 520 (6): 0 4082–4121, 2014. J.R. Sun, X. Li, and J.M. Yong. [Open-loop and closed-loop solvabilities for stochastic linear quadratic optimal control problems]{}. *[SIAM]{} Journal on Control and Optimization*, 540 (5): 0 2274–2308, 2016. S.J. Tang. [General linear quadratic optimal stochastic control problems with random coefficients: Linear stochastic hamilton systems and backward stochastic riccati equations]{}. *[SIAM]{} Journal on Control and Optimization*, 420 (1) : 0 53–75, 2003. S.J. Tang. [Dynamic programming for general linear quadratic optimal stochastic control with random coefficients]{}. *[SIAM]{} Journal on Control and Optimization*, 530 (2): 0 1082–1106, 2015. W.M. Wonham. [On a matrix Riccati equation of stochastic control]{}. *[SIAM]{} Journal on Control*, 60 (4): 0 681–697, 1968. Z. Wu and X.R. Wang. [FBSDE with Poisson process and its application to linear quadratic stochastic optimal control problem with random jumps]{}. *Acta Automatica Sinica*, 290 (6): 0 821–826, 2003. G. Yin and X.Y. Zhou. [Markowitz’s mean-variance portfolio selection with regime switching: From discrete-time models to their continuous-time limits]{}. *IEEE Transactions on Automatic Control*, 490 (3): 0 349–360, 2004. J.M. Yong and X.Y. Zhou. [*Stochastic Controls: Hamiltonian Systems and HJB Equations*]{}. Springer New York, 1999. Z.Y. Yu. [Infinite horizon jump-diffusion forward-backward stochastic differential equations and their application to backward linear-quadratic problems]{}. *[ESAIM]{}: Control, Optimisation and Calculus of Variations*, 230 (4): 0 1331–1359, 2017. C. Zalinescu. [On uniformly convex functions]{}. *Journal of Mathematical Analysis and Applications*, 950 (2): 0 344–374, 1983. C. Zalinescu. [*Convex Analysis in General Vector Spaces*]{}. World Scientific, 2002. Q. Zhang and G. Yin. [On nearly optimal controls of hybrid LQG problems]{}. *IEEE Transactions on Automatic Control*, 440 (12): 0 2271–2282, 1999. X. Zhang, T.K. Siu and Q.B. Meng. [Portfolio selection in the enlarged Markovian regime-switching market]{}. *[SIAM]{} Journal on Control and Optimization*, 480 (5): 0 3368–3388, 2010. X. Zhang, R.J. Elliott, T.K. Siu and J.Y. Guo. [Markovian regime-switching market completion using additional markov jump assets]{}. *[IMA]{} Journal of Management Mathematics*, 230 (3): 0 283–305, 2011. X. Zhang, R.J. Elliott, and T.K. Siu. [A stochastic maximum principle for a Markov regime-switching jump-diffusion model and its application to finance]{}. *[SIAM]{} Journal on Control and Optimization*, 500 (2): 0 964–990, 2012. X. Zhang, Z. Sun, and J. Xiong. [A general stochastic maximum principle for a Markov regime switching jump-diffusion model of mean-field type]{}. *[SIAM]{} Journal on Control and Optimization*, 560 (4): 0 2563–2592, 2018. X.Y. Zhou and G. Yin. [Markowitz’s mean-variance portfolio selection with regime switching: A continuous-time model]{}. *[SIAM]{} Journal on Control and Optimization*, 420 (4): 0 1466–1482, 2003. [^1]: School of Mathematics, Southeast University, Nanjing, Jiangsu Province, 211189, China ([email protected]). [^2]: Department of Applied Mathematics, The Hong Kong Polytechnic University, Hong Kong, China ([email protected]). [^3]: This work is supported by the National Natural Science Foundation of China (grant nos. 11771079, 11371020), and RGC Grants 15209614 and 15255416.

ArXiv

--- abstract: 'We prove that an element of the symplectic Steinberg group is trivial if and only if its image under any maximal localisation homomorphism is trivial.' author: - Andrei Lavrenov title: 'A local–global principle for symplectic ${\mathop{\mathrm{K_2}}\nolimits}$' --- =2 Introduction {#introduction .unnumbered} ============ The main goal of the present paper is to establish a symplectic $\mathrm K_2$-analogue of Quillen’s Patching Theorem which is the key ingredient in his solution of Serre’s problem (see [@n13])[^1]. Let $R$ be a commutative ring, and $P$ be a finitely generated projective module over the polynomial ring $R[t_1,\ldots,t_n]$. Then $P$ is extended from $R$ [(]{}i.e., there exists a projective $R$-module $Q$ such that $P\cong R[t_1,\ldots,t_n]\otimes Q$[)]{} if and only if $P_{\mathfrak m}$ is extended from $R_{\mathfrak m}$ for every maximal ideal $\mathfrak m$ of $R$. Later Suslin proved his ${\mathop{\mathrm{K_1}}\nolimits}$-analogue of Serre’s problem with the use of a similar statement concerning elementary matrices (see [@n11]). Let $n\geq3$ and $g\in{\mathop{\mathrm{GL}}\nolimits}_n(R[t],\,tR[t])$. Then $g\in{{\mathrm{E}}}_n(R[t])$ if and only if $g_{\mathfrak m}\in{{\mathrm{E}}}_n(R_{\mathfrak m}[t])$ for every maximal ideal $\mathfrak m$ of $R$. Subsequently, ${\mathop{\mathrm{K_1}}\nolimits}$-analogues for Chevalley groups [@n12; @n5; @n1] were obtained. Presently, ${\mathop{\mathrm{K_1}}\nolimits}$-analogue is known in much larger generality, namely, for isotropic reductive groups [@n9] and in the framework of Stepanov’s universal localisation [@n10]. As opposed to that, ${\mathop{\mathrm{K_2}}\nolimits}$-analogues are proven only for ${\mathop{\mathrm{GL}}\nolimits}_n$ [@n14] and, recently, for Chevalley groups of type $\mathrm E_l$ [@n8]. The Main Theorem of the present paper is the following ${\mathop{\mathrm{K_2}}\nolimits}$-analogue to the local-global principle for ${\mathop{\mathrm{Sp}}\nolimits}_{2n}$. Let $R$ be an arbitrary commutative ring [(]{}with 1[)]{}, $n\geq3$, and $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])$. Then $g=1$ [(]{}corr., lies in the image of ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R[t])$[)]{} if and only if $g_{\mathfrak m}=1\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R_{\mathfrak m}[t])$ [(]{}corr., lies in the image of ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R_{\mathfrak m}[t])$[)]{} for every maximal ideal $\mathfrak m$ of $R$. The paper is organised as follows. In the first section we recall results of [@l1], where a “basis-free” presentation of the symplectic Steinberg group id given. We make an essential use of these results in the Section 4. In the next section we show that three possible definitions of the relative symplectic Steinberg group coincide. One of them is the “correct” definition, and the other two are used in our proof. In the Section 3 we establish the local-global principle modulo the main technical lemma. Finally, in the Section 4, we give a proof to this lemma, namely, construct a symplectic analogue of Tulenbaev map. Absolute Steinberg groups ========================= In the present paper $R$ always denotes an arbitrary associative commutative unital ring, $R^{2n}$ is the free right $R$-module, we number its basis as follows: $e_{-n}$, $\ldots$, $e_{-1}$, $e_1$, $\ldots$, $e_n$, $n\geq3$. [*The symplectic group*]{} ${\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ is the group of automorphisms of $R^{2n}$ preserving the standard symplectic form ${\langle}\,,\,{\rangle}$, where ${\langle}e_i,\,e_{-i}{\rangle}=1,\ i>0$. We denote [*the elementary symplectic transvections*]{} by $$T_{ij}(a)=1+e_{ij}\cdot a-e_{-j,-i}\cdot a\cdot{\mathrm{sign}(i)}\cdot{\mathrm{sign}(j)},\quad T_{i,-i}(a)=1+e_{i,-i}\cdot a\cdot{\mathrm{sign}(i)},$$ where $a\in R$, $i$, $j\in\{-n,$ $\ldots,$ $-1,$ $1,$ $\ldots,$ $n\}$, $i\not\in\{\pm j\}$, $e_{ij}$ is a matrix unit. They generate [*the elementary symplectic group*]{} ${\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)$. The [*symplectic Steinberg group*]{} ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ is generated by the formal symbols $X_{ij}(a)$ for $i\neq j$, $a\in R$ subject to the Steinberg relations $$\begin{aligned} &X_{ij}(a)=X_{-j,-i}(-a\cdot{\mathrm{sign}(i)}{\mathrm{sign}(j)}),\\ &X_{ij}(a)X_{ij}(b)=X_{ij}(a+b),\\ &[X_{ij}(a),\,X_{hk}(b)]=1,\text{ for }h\not\in\{j,-i\},\ k\not\in\{i,-j\},\\ &[X_{ij}(a),\,X_{jk}(b)]=X_{ik}(ab),\text{ for }i\not\in\{-j,-k\},\ j\neq-k,\\ &[X_{i,-i}(a),\,X_{-i,j}(b)]=X_{ij}(ab\cdot{\mathrm{sign}(i)})X_{-j,j}(-ab^2),\\ &[X_{ij}(a),\,X_{j,-i}(b)]=X_{i,-i}(2\,ab\cdot{\mathrm{sign}(i)}).\end{aligned}$$ There is a natural projection $\phi\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R){\twoheadrightarrow}{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)$ sending the generators $X_{ij}(a)$ to $T_{ij}(a)$. In other words, the Steinberg relations hold for the elementary symplectic transvections. To define the elementary symplectic group ${\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)$ instead of $T_{ij}(a)$ one can use “basis-independent” ESD-transformations $T(u,\,v,\,a)$ defined by $$w\mapsto w+u({\langle}v,\,w{\rangle}+a{\langle}u,\,w{\rangle})+v{\langle}u,\,w{\rangle}$$ as a set of generators. Here $u$, $v\in R^{2n}$, ${\langle}u,\,v{\rangle}=0$, $a\in R$. See section 1 of [@l1] for details. On the level of Steinberg groups, this idea leads to the following presentation, inspired by van der Kallen’s paper [@n3]. It is the main theorem of [@l1]. The symplectic Steinberg group ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ can be defined by the set of generators $$\big\{[u,\,v,\,a]\,\big|\ u\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)e_1,\,v\in R^{2n},\ \ {\langle}u,\,v{\rangle}=0,\ a\in R\big\}$$ and relations $$\begin{aligned} &[u,\,v_1,\,a_1][u,\,v_2,\,a_2]=[u,\,v_1+v_2,\,a_1+a_2+{\langle}v_1,\,v_2{\rangle}],\\ &[u_1,\,u_2b,\,0]=[u_2,\,u_1b,\,0]\,\text{ for any }\, b\in R,\\ &[u',\,v',\,a'][u,\,v,\,a][u',\,v',\,a']{^{-1}}=[T(u',\,v',\,a')u,\,T(u',\,v',\,a')v,\,a],\end{aligned}$$ For the usual generators of the symplectic Steinberg group the following identities hold $$X_{ij}(a)=[e_i,\,e_{-j}\cdot a\cdot{\mathrm{sign}(-j)},\,0]\ \text{for $j\neq-i$},\quad X_{i,-i}(a)=[e_i,\,0,\,a],$$ and, moreover, $\phi$ sends $[u,\,v,\,a]$ to $T(u,\,v,\,a)$. Furthermore, the following relations are automatically satisfied $$\begin{aligned} &[u,\,ua,\,0]=[u,\,0,\,2a],\\ &[ub,\,0,\,a]=[u,\,0,\,ab^2],\\ &[u+v,\,0,\,a]=[u,\,0,\,a][v,\,0,\,a][v,\,ua,\,0]\,\text{ for }{\langle}u,\,v{\rangle}=0.\end{aligned}$$ More precisely, we need elements $X(u,\,v,\,a)$ in ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ defined in [@l1] for arbitrary $u$, $v\in R^{2n}$ and $a\in R$ (see the definition of $X(u,\,0,\,a)$ in section 4, then the definition of $X(u,\,v,\,0)$ on p. 3775 and the general case on p. 3777, or an overview in section 1). For $u\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)$ these elements coincide with the generators $[u,\,v,\,a]$ above (see the last section of [@l1] where the isomorphism between two presentations is described). Below we list their properties, proven in [@l1]. \[xlist\] For all $u$, $v\in R^{2n}$, ${\langle}u,\,v{\rangle}=0$, $a$, $b\in R$, $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ one has $$\begin{aligned} &\phi\big(X(u,\,v,\,a)\big)=T(u,\,v,\,a),\\ &g\,X(u,\,v,\,a)g{^{-1}}=X(\phi(g)u,\,\phi(g)v,\,a),\\ &X(ub,\,0,\,a)=X(u,\,0,\,b^2a),\\ &X(u,\,0,\,a)X(u,\,0,\,b)=X(u,\,0,\,a+b),\\ &X(u,\,v,\,0)=X(v,\,u,\,0),\\ &X(ua,\,ub,\,0)=X(u,\,0,\,2ab),\\ &X(u+v,\,0,\,1)=X(u,\,0,\,1)X(v,\,0,\,1)X(u,\,v,\,0),\\ &X(u,\,v,\,a)=X(u,\,v,\,0)X(u,\,0,\,a).\end{aligned}$$ For X0 see section 1 of [@l1], properties X1–X4 are proven in Lemmas 26, 29, 31, 33, 34 of [@l1], X6 and X7 are actually definitions (see pp. 3775 and 3777 of [@l1]). To get X5 use X6, then X2 and X3. We also need to use a few more properties of $X(u,\,v,\,a)$ listed in Lemma \[xlist2\] below. To prove them, we pass to another type of elements $Y(u,\,v,\,a)$ defined in [@l1]. As soon as Lemma \[xlist2\] is proven, we do not need $Y$’s any more in this paper and use only $X$’s. For $u\in R^{2n}$ having a pair of zeros in symmetric positions, i.e., such that $u_i=u_{-i}=0$ for some $i$, and $v\in R^{2n}$, such that ${\langle}u,\,v{\rangle}=0$, $a\in R$ also elements $Y_{(i)}(u,\,v,\,a)$ are defined. Their definition is given in three steps (see definitions and remarks on pp. 3762, 3764 and 3766 of [@l1]). If, in addition, $u$ has another symmetric pair of zeros, i.e., $u_i=u_{-i}=u_j=u_{-j}=0$ for some $i\neq\pm j$, then $Y_{(i)}(u,\,v,\,a)=Y_{(j)}(u,\,v,\,a)$ (see Y1 below) and in this situation we omit the index and denote this element by $Y(u,\,v,\,a)$. We use the following properties of these elements, proven in [@l1]. \[ylist\] For a fixed index $\,i$ and any $j\neq\pm i$, vectors $u$, $v$, $v'$, $w$, $w'$, $q$, $q'$, $r$, $r'$ $s\in R^{2n}$, such that $u_i=u_{-i}=0$, ${\langle}u,\,v{\rangle}={\langle}u,\,v'{\rangle}=0$, ${\langle}w,\,e_i{\rangle}={\langle}w',\,e_i{\rangle}=0$, $v'_i=v'_{-i}=q_i=q_{-i}=r_i=r_{-i}=r_j=r_{-j}=0$, $q'=e_iq'_i+e_{-i}q'_{-i}$, $r'=e_jr'_j+e_{-j}r'_{-j}$, $s=e_is_i+e_{-i}s_{-i}$, and elements $a$, $a'\in R$, one has $$\begin{aligned} &\phi\big(Y_{(i)}(u,\,v,\,a)\big)=T(u,\,v,\,a),\\ &Y_{(i)}(u,\,v,\,a)=Y_{(j)}(u,\,v,\,a)\,\text{ if also }\,u_j=u_{-j}=0,\\ &Y(e_i,\,w,\,a)Y(e_i,\,w',\,a')=Y(e_i,\,w+w',\,a+a'+{\langle}w,\,w'{\rangle}),\\ &Y(e_i,\,w,\,a)=X(e_i,\,w,\,a),\\ &\!\begin{aligned} Y_{(i)}(u,\,v',\,a)=[Y(e_i,\,u,\,0)&,\,Y(e_{-i},\,v'\,{\mathrm{sign}(i)},\,a)]\cdot\\ \cdot Y&(e_i,\,ua\,{\mathrm{sign}(-i)},\,0), \end{aligned}\\ &\!\begin{aligned} Y_{(i)}(u,\,v,\,a)=Y_{(i)}(u,\,v-e_iv_i-e_{-i}v_{-i}&,\,a-v_iv_{-i}\,{\mathrm{sign}(i)})\cdot\\ \cdot&Y(e_i,\,uv_i,\,0)Y(e_{-i},\,uv_{-i},\,0), \end{aligned}\\ &X(q+q',\,0,\,a)=Y_{(i)}(q,\,0,\,a)Y(q',\,0,\,a)Y(q',\,qa,\,0),\\ &Y(r,\,s,\,0)Y(r',\,s,\,0)=Y_{(i)}(r+r',\,s,\,0).\end{aligned}$$ For Y0 see section 1 of [@l1], for Y1 and Y2 see Lemmas 20 and 9 of [@l1], for Y3 use Lemma 36 of [@l1], Y6, Lemma \[xlist\](X7) and Y2, Y4–Y6 are actually definitions (pp. 3764, 3766, 3768 and remark on p. 3770 of [@l1]), Y7 is Lemma 28 of [@l1]. Let us establish some further properties of $Y_{(i)}(u,\,v,\,a)$. For an index $\,i$ and any vectors $u$, $v$, $w\in R^{2n}$, such that $u_i=u_{-i}=0$, ${\langle}u,\,v{\rangle}={\langle}u,\,w{\rangle}=0$, elements $a$, $b\in R$, one has $$\begin{aligned} &Y_{(i)}(u,\,v,\,a+b)=Y_{(i)}(u,\,v,\,a)Y_{(i)}(u,\,0,\,b),\\ &Y_{(i)}(u,\,v,\,a)=Y_{(i)}(u,\,v-e_iv_i-e_{-i}v_{-i},\,a)Y_{(i)}(u,\,e_iv_i+e_{-i}v_{-i},\,0),\\ &X(u+v,\,0,\,a)=X(u,\,0,\,a)X(v,\,0,\,a)Y_{(i)}(u,\,va,\,0),\\ &Y_{(i)}(u,\,va,\,0)=Y_{(i)}(v,\,ua,\,0)\,\text{ if also }\,v_i=v_{-i}=0,\\ &Y_{(i)}(u,\,v,\,a)=X(u,\,v,\,a),\\ &Y_{(i)}(u,\,v,\,0)Y_{(i)}(u,\,w,\,0)=Y_{(i)}(u,\,v+w,\,{\langle}v,\,w{\rangle}).\end{aligned}$$ We prove Y8 and Y9 together. First, assume that $v_i=v_{-i}=0$. In this assumption one can get Y8 repeating the proof of Lemma 21 of [@l1]. Use Y2 and Y4 to decompose elements and Y3 and X1 to show, that some of them commute (instead of complicated arguments in the original proof). Next, use this result to obtain Y9, more precisely, repeating the proof of Lemma 22 of [@l1] use it instead of Lemma 21. Now, Y8 follows from Y9 in full generality. One needs that $Y_{(i)}(u,\,0,\,b)=X(u,\,0,\,b)$ by Y6, and commutes with $Y_{(i)}(u,\,e_iv_i+e_{-i}v_{-i},\,0)$ by X1. To obtain Y10–Y13 we also proceed in several steps. First, consider Y10 and assume that $v_i=v_{-i}=0$. Then, repeat the proof of Lemma 23 of [@l1] interchanging the roles of $u$ and $v$ and using Y3 and X1 instead of the arguments presented there. Obviously, by X1, $X(u,\,0,\,a)$ and $X(v,\,0,\,a)$ commute. Using this fact we get Y11. Then, we can get Y12 in the same assumptions on $v$. For $a=0$ it follows from Y10 and X6, for the general case use Y8 and X7. Next, consider Y13 and assume that $v_i=v_{-i}=w_i=w_{-i}=0$. By Y4, we have $$\begin{gathered} Y(u,\,v+w,\,{\langle}v,\,w{\rangle})=\\=[Y(e_i,\,u,\,0),\,Y(e_{-i},\,(v+w)\,{\mathrm{sign}(i)},\,{\langle}v,\,w{\rangle})]Y(e_{-i},\,u{\langle}v,\,w{\rangle}{\mathrm{sign}(-i)}).\end{gathered}$$ Now, we get $$\begin{gathered} Y(e_{-i},\,(v+w)\,{\mathrm{sign}(i)},\,{\langle}v,\,w{\rangle})=Y(e_{-i},\,v\,{\mathrm{sign}(i)},\,0)Y(e_{-i},\,w\,{\mathrm{sign}(i)},\,0)\end{gathered}$$ by Y2 and use an identity $[a,\,bc]=[a,\,b]\cdot[a,\,c]\cdot[[c,\,a],\,b]$ for $a=Y(e_i,\,u,\,0)$, $b=Y(e_{-i},\,v\,{\mathrm{sign}(i)},\,0)$, $c=Y(e_{-i},\,w\,{\mathrm{sign}(i)},\,0)$. As above, $[a,\,b]=Y_{(i)}(u,\,v,\,0)$ and $[a,\,c]=Y_{(i)}(u,\,w,\,0)$. One can check, that $$[[c,\,a],\,b]=Y(e_{-i},\,-u{\langle}w,\,v{\rangle}{\mathrm{sign}(i)},\,0)$$ with the use of Y3, X1 and Y2. Now, we prove Y13 for arbitrary $v$, $w$ and $u$ having two pairs of zeros, i.e., such that $u_j=u_{-j}=0$ for $j\neq\pm i$ also. Decompose $v=\tilde v+v'$, where $v'=e_iv_i+e_{-i}v_{-i}$, and similarly for $w$. Use Y9, then Y8, then Y6 and X1 to change the order of factors, then Y8 again to obtain $$Y(u,\,v+w,\,{\langle}v,\,w{\rangle})=Y(u,\,\tilde v+\tilde w,\,{\langle}\tilde v,\,\tilde w{\rangle})Y(u,\,v'+w',\,{\langle}v',\,w'{\rangle}).$$ For each factor we can use the previous steps. To reorder the factors in the result use that $Y(u,\,v',\,0)=X(u,\,v',\,0)$ (we already have Y12 for this situation) and X1. Then use Y9 again. Next, we establish Y10 in full generality. Decompose $v=\tilde v+v'$ as above and use Y6 to obtain $$X(u+v,\,0,\,a)=X(u+\tilde v,\,0,\,a)X(v',\,0,\,a)Y(v',\,(u+\tilde v)a,\,0).$$ For the first factor we can already use Y10, and for the last one we can already use Y13. Reordering factors by X1 we get $$X(u+v,\,0,\,a)=X(u,\,0,\,a)Y_{(i)}(u,\,\tilde va,\,0)Y(v',\,ua,\,0)X(v,\,0,\,a)$$ with the use of Y6. Decompose $u=\tilde u+u'$, where $u'=e_ju_j+e_{-j}u_{-j}$, then decompose $Y(v',\,ua,\,0)$ by Y9, apply Y11 to each factor and use Y7 to get $$\begin{gathered} Y(v',\,ua,\,0)=Y(v',\,\tilde ua,\,0)Y(v',\,u'a,\,0)=\\=Y_{(j)}(\tilde u,\,v'a,\,0)Y_{(k)}(u',\,v'a,\,0)=Y(u,\,v'a,\,0).\end{gathered}$$ Now, we are done by Y9. Proceeding as above, we get Y12 in full generality. Finally, consider Y13. As above, decompose $v=\tilde v+v'$ and $w=\tilde w+w'$, and get $$Y_{(i)}(u,\,v+w,\,{\langle}v,\,w{\rangle})=Y_{(i)}(u,\,\tilde v+\tilde w,\,{\langle}\tilde v,\,\tilde w{\rangle})Y_{(i)}(u,\,v'+w',\,{\langle}v',\,w'{\rangle}).$$ For the first factor Y13 holds by previous steps, for the second one use Y5, then Y8 and Y2. Then change the order of factors. To interchange positions of $Y(e_i,\,uw_i,\,0)$ and $Y(e_{-i},\,uv_{-i},\,0)$ we need to plug in an extra commutator equal to $$Y_{(i)}(-uw_i,\,-uv_{-i}\,{\mathrm{sign}(i)},\,0)=X(u,\,0,\,2w_iv_{-i}\,{\mathrm{sign}(i)})$$ by Y12 and X5. With the use of Y5 and Y8 one gets $$Y_{(i)}(u,\,v'+w',\,{\langle}v',\,w'{\rangle})=Y_{(i)}(u,\,v',\,0)Y_{(i)}(u,\,w',\,0).$$ It remains to change the order of factors and use Y9. In the following lemma we collect some new relations among X’s which we need in the sequel. \[xlist2\] Consider $u$, $v$, $w\in R^{2n}$, $a$, $b\in R$, such that ${\langle}u,\,v{\rangle}={\langle}u,\,w{\rangle}=0$. Assume either that $u_i=u_{-i}=0$ or that $v_i=v_{-i}=w_i=w_{-i}=0$. Then one has $$\begin{aligned} &X(u+vr,\,0,\,a)=X(u,\,0,\,a)X(v,\,0,\,r^2a)X(u,\,vra,\,0),\\ &X(u,\,va,\,0)=X(v,\,ua,\,0),\\ &X(u,\,v,\,a)X(u,\,w,\,b)=X(u,\,v+w,\,a+b+{\langle}v,\,w{\rangle}).\end{aligned}$$ First, assume $u_i=u_{-i}=0$. Then X8 follows from Y10, X2 and Y12. Denote $u'=e_ju_j+e_{-j}u_{-j}$ and $v'=e_iv_i+e_{-i}v_{-i}$ for some $j\neq\pm i$ and decompose $u=\tilde u+u'$, $v=\tilde v+v'$. Then $$X(u,\,va,\,0)=Y_{(i)}(u,\,\tilde va,\,0)Y(\tilde u,\,v'a,\,0)Y(u',\,v'a,\,0)$$ by Y9 and Y7. Now apply Y11 twice to each factor $$X(u,\,va,\,0)=Y_{(i)}(ua,\,\tilde v,\,0)Y(\tilde ua,\,v',\,0)Y(u'a,\,v',\,0).$$ Next, X10 for $a=b=0$ follows from Y12 and Y13, and the general case from Y8, X7 and X3. Now, consider the second case $v_i=v_{-i}=w_i=w_{-i}=0$. For X8 use the previous step (X8 and X9). It remains to prove X10. Assume $a=b=0$ (the case of arbitrary $a$ and $b$ may be treated as in the previous step). Denote $u'=e_iu_i+e_{-i}u_{-i}$ and $\tilde u=u-u'$ (previously we took $\pm j$-th components instead). Using X4, Y12 and Y9 we get $$X(u,\,v,\,0)=X(\tilde u,\,v,\,0)X(u',\,v,\,0)$$ and similarly for $w$. Using X1 and the previous case (X10) we can compute $$[X(u',\,v,\,0),\,X(\tilde u,\,w,\,0)]=X(\tilde u,\,u'{\langle}v,\,w{\rangle},\,0).$$ Thus, reordering factors and using previous case we get $$\begin{gathered} X(u,\,v,\,0)X(u,\,w,\,0)=\\=X(\tilde u,\,v+w,\,{\langle}v,\,w{\rangle})X(u',\,v+w,\,{\langle}v,\,w{\rangle})X(\tilde u,\,u'{\langle}v,\,w{\rangle},\,0).\end{gathered}$$ Now, we can finish the proof with the use of X7, X1, X4, Y12, Y9 and Y6. At the end of this section we establish one more relation automatically satisfied by the generators $[u,\,v,\,a]$ from [*Another presentation*]{}. We need it later to get a map from the relative symplectic Steinberg group to the absolute one. First, observe that for $v\in R^{2n}$ such that $v_{-1}=0$ (equiv., ${\langle}v,\,e_1{\rangle}=0$) holds $$X(e_1+vr,\,0,\,a)=X(e_1,\,0,\,a)X(v,\,0,\,r^2a)X(e_1,\,vra,\,0)$$ by X8. Next, take $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ and denote $u=\phi(g)e_1$ and $w=\phi(g) v$. Conjugate the above identity by $g$ and use X1 to get $$X(u+wr,\,0,\,a)=X(u,\,0,\,a)X(w,\,0,\,r^2a)X(u,\,wra,\,0).$$ The above identity holds for any $u\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)e_1$ and $w$ orthogonal to it. Finally, take $(u,\,w)\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)(e_1,\,e_2)$. Then $u+wr$ also lies in ${\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)e_1$ and we can replace X’s above by the generators from the [*Another presentation*]{}. For $(u,\,w)\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)(e_1,\,e_2)$ and $a$, $r\in R$ the following identity holds $$\begin{aligned} &[u+wr,\,0,\,a]=[u,\,0,\,a][w,\,0,\,ar^2][u,\,war,\,0].\end{aligned}$$ Relative Steinberg groups ========================= In the present section we give three definitions of a relative symplectic Steinberg group. For our purposes we can concentrate on splitting ideals $I\trianglelefteq R$, i.e., those ideals for which the natural projection $\rho\colon R{\twoheadrightarrow}R/I$ splits. Obviously, $tR[t]\trianglelefteq R[t]$ is a splitting ideal. We show that for splitting ideals all three definitions of relative symplectic Steinberg group coincide. The correct approach to relative Steinberg groups is described in [@n4; @n7; @n8]. But for splitting ideals we can define it in the following naive way. Afterwards, we show that (for splitting ideals) it coincides with the usual one. Let $I\trianglelefteq R$ be a [*splitting*]{} ideal. Define [*the relative symplectic Steinberg group*]{} ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R,\,I)={\mathop{\mathrm{Ker}}\nolimits}\big(\rho^*\colon {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R){\twoheadrightarrow}{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R/I)\big)$. Obviously, ${\mathop{\mathrm{Ker}}\nolimits}(\rho_*)$ coincides with the normal subgroup of ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ generated by $\{X_{ij}(a)\mid a\in I\}$. This is tantamount to saying that applying $\rho^*$ is the same as forcing an additional relation $X_{ij}(a)=1,\ a\in I$, in ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$. The next definition is a symplectic version of the Keune–Loday presentation in the linear case (see [@n4; @n7]). It is a relative version of the definition via Steinberg relations. For a group $G$ acting on a group $H$ on the left, we will denote the image of $h\in H$ under the homomorphism corresponding to the element $g\in G$ by $\!\,^gh$, the element $\!\,^gh\cdot h{^{-1}}$ by $\llbracket g,\,h]$ and the element $h\cdot\!\,^gh{^{-1}}$ by $[h,\,g\rrbracket$. Let [*the Keune–Loday relative symplectic Steinberg group*]{} ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)$ be a group with the action of the (absolute) Steinberg group ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)$ defined by the set of relative generators $Y_{ij}(a), i\neq j,\ a\in I$, subject to the relations $$\begin{aligned} &Y_{ij}(a)=Y_{-j,-i}(-a\cdot{\mathrm{sign}(i)}\cdot{\mathrm{sign}(j)}),\\ &Y_{ij}(a)Y_{ij}(b)=Y_{ij}(a+b),\\ &\llbracket X_{ij}(r),\,Y_{hk}(a)]=1,\text{ for }h\not\in\{j,-i\},\ k\not\in\{i,-j\},\\ &\llbracket X_{ij}(r),\,Y_{jk}(a)]=Y_{ik}(ra),\text{ for }i\not\in\{-j,-k\},\ j\neq-k,\\ &\llbracket X_{i,-i}(r),\,Y_{-i,j}(a)]=Y_{ij}(ra\cdot{\mathrm{sign}(i)})Y_{-j,j}(-ra^2),\\ &[Y_{i,-i}(a),\,X_{-i,j}(r)\rrbracket=Y_{ij}(ar\cdot{\mathrm{sign}(i)})Y_{-j,j}(-ar^2),\\ &\llbracket X_{ij}(r),\,Y_{j,-i}(a)]=X_{i,-i}(2\,ra\cdot{\mathrm{sign}(i)}),\\ &\!\,^{X_{ij}(a)}\Big(\!\,^{X_{hk}(r)}Y_{st}(b)\Big)=\!\,^{Y_{ij}(a)}\Big(\!\,^{X_{hk}(r)}Y_{st}(b)\Big).\end{aligned}$$ In other words, we consider a free group generated by symbols $(g,\,x)=\!\,^gx$ where $g$ is from the absolute Steinberg group and $x$ is from the set of relative generators, ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)$ naturally acts on this free group via $\!\,^f(g,\,x)=(fg,\,x)$ and then we define the relative symplectic Steinberg group as the quotient of the above free group modulo equivariant normal subgroup generated by KL0–KL7. There is an obvious equivariant mapping from ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)$ to ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)$ sending $Y_{ij}(a)$ to $X_{ij}(a)$, and its image is the normal subgroup generated by $\{X_{ij}(a)\mid a\in I\}$, i.e., coincides with ${\mathop{\mathrm{Ker}}\nolimits}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R){\twoheadrightarrow}{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R/I)\big)$. Let $I\trianglelefteq R$ be a splitting ideal. Then the natural map $$\iota\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)$$ is injective. In other words, ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)={\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R,\,I)$. The proof is actually the same as in the linear case (see [@n4; @n7] and [@n8] for the simply-laced case). Denote by $\rho\colon R{\twoheadrightarrow}R/I$ the natural projection and by $\sigma\colon R/I\rightarrow R$ its splitting. Then ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R/I)$ acts on ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)$ via $\sigma^*$ and one can consider the semi-direct product ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)\leftthreetimes{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R/I)$ which maps to ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)$ via $\iota\leftthreetimes\sigma^*$ $${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)\leftthreetimes{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R/I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R),\quad (x,\,y)\mapsto\iota(x)\cdot\sigma^*(y).$$ We construct an inverse map $$\psi\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)\leftthreetimes{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R/I),$$ sending $$X_{ij}(r)\mapsto\big(Y_{ij}(r-\sigma\rho(r)),\,X_{ij}(\rho(r))\big).$$ Obviously, the fact that $\iota\leftthreetimes\sigma^*$ is an isomorphism implies that $\iota$ is injective. To check that $\psi$ is well-defined one has to verify relations S0–S5 for the images of the generators. Consider, say, S4. We will show that the images of $$X_{i,-i}(a)X_{-i,j}(b)X_{i,-i}(-a)\text{ and }X_{ij}(ab\cdot{\mathrm{sign}(i)})X_{-j,j}(-ab^2)X_{-i,j}(b)$$ under $\psi$ coincide. Indeed, $$\begin{gathered} \begin{aligned} &\psi\Big(X_{i,-i}(a)X_{-i,j}(b)X_{i,-i}(-a)\Big)=\\ &=\Big(Y_{i,-i}(a-\sigma\rho(a))\,^{X_{i,-i}(\sigma\rho(a))}Y_{-i,j}(b-\sigma\rho(b))\cdot \end{aligned}\\ \cdot\,^{X_{i,-i}(\sigma\rho(a))X_{-i,j}(\sigma\rho(b))}Y_{i,-i}(-a+\sigma\rho(a)),\\ X_{i,-i}(\rho(a))X_{i,-j}(\rho(b))X_{i,-i}(-\rho(a))\Big).\end{gathered}$$ Rewriting $$\begin{gathered} \,^{X_{-i,j}(\sigma\rho(b))}Y_{i,-i}(-a+\sigma\rho(a))=\\ =Y_{i,-i}(-a+\sigma\rho(a))[Y_{i,-i}(a-\sigma\rho(a)),\,X_{-i,j}(\sigma\rho(b))\rrbracket\end{gathered}$$ we get with the use of KL7 $$\begin{gathered} \psi\Big(X_{i,-i}(a)X_{-i,j}(b)X_{i,-i}(-a)\Big)=\\ =\Big(\,^{X_{i,-i}(a)}Y_{-i,j}(b-\sigma\rho(b))[Y_{i,-i}(a-\sigma\rho(a)),\,X_{-i,j}(\sigma\rho(b))\rrbracket,\,\\ X_{i,-i}(\rho(a))X_{i,-j}(\rho(b))X_{i,-i}(-\rho(a))\Big),\end{gathered}$$ and finally $$\begin{gathered} \begin{aligned}&\psi\Big(X_{i,-i}(a)X_{-i,j}(b)X_{i,-i}(-a)\Big)=\\&=\Big(Y_{ij}((ab-\sigma\rho(ab))\cdot{\mathrm{sign}(i)})\cdot\end{aligned}\\ \cdot Y_{-i,j}(b-\sigma\rho(b))Y_{-j,j}(-ab^2+2\,\sigma\rho(ab)b-\sigma\rho(ab^2)),\\ X_{ij}(\sigma\rho(ab)\cdot{\mathrm{sign}(i)})X_{-j,j}(-\sigma\rho(ab^2))X_{-i,j}(\sigma\rho(b))\Big).\end{gathered}$$ On the other hand, $$\begin{gathered} \begin{aligned}&\psi\Big(X_{ij}(ab\cdot{\mathrm{sign}(i)})X_{-j,j}(-ab^2)X_{-i,j}(b)\Big)=\\&=\Big(Y_{ij}((ab-\sigma\rho(ab))\cdot{\mathrm{sign}(i)})\cdot\end{aligned}\\ \cdot Y_{-j,j}(-ab^2+\sigma\rho(ab^2))\cdot\,^{X_{ij}(\sigma\rho(ab){\mathrm{sign}(i)})}Y_{-i,j}(b-\sigma\rho(b)),\\ X_{ij}(\sigma\rho(ab)\cdot{\mathrm{sign}(i)})X_{-j,j}(-\sigma\rho(ab^2))X_{-i,j}(\sigma\rho(b))\Big).\end{gathered}$$ Other relations are similar and much less tedious. Obviously, $\iota\leftthreetimes\sigma^*\circ\psi=1$ and it only remains to show that $\psi$ is surjective. All elements of types $(1,\,X_{ij}(s))$ and $(Y_{ij}(a),\,1)$ lie in the image of $\psi$, and then elements of type $(\,^{X_{hk}(r)}Y_{ij}(a),\,1)$ lie as well. In the proof of the local–global principle we also need another presentation for the relative Steinberg group. It is inspired by the definition of the relative linear Steinberg groups given by Tulenbaev. Let [*Tulenbaev relative symplectic Steinberg group*]{} ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2n}(R,\,I)$ be a group defined by the set of generators $$\{[u,\,v,\,a,\,b]\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)e_1\times R^{2n}\times I\times I\mid {\langle}u,\,v{\rangle}=0\}$$ subject to the relations $$\begin{aligned} &[u,\,vr,\,a,\,b]=[u,\,v,\,ra,\,b]\ \ \forall\,r\in R,\\ &[u,\,v,\,a,\,b][u,\,w,\,a,\,c]=[u,\,v,\,a,\,b+c+a^2{\langle}v,\,w{\rangle}],\\ &[u,\,v,\,a,\,0][u,\,v,\,b,\,0]=[u,\,v,\,a+b,\,0],\\ &[u,\,u,\,a,\,0]=[u,\,0,\,0,\,2a],\\ &[u,\,v,\,a,\,0]=[v,\,u,\,a,\,0]\ \ \forall\,(u,\,v)\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)(e_1,\,e_2),\\ &\begin{aligned}\! [u+vr,\,0,\,0,\,a]=[u,\,0,\,0,\,a][v,\,&0,\,0,\,ar^2][u,\,v,\,ar,\,0]\\&\forall\,r\in R\ \ \forall\,(u,\,v)\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(R)(e_1,\,e_2), \end{aligned}\\ &\begin{aligned}\! [u',\,v',\,a',\,b'][u,\,v,\,a,\,b][u',\,&v',\,a',\,b']{^{-1}}=\\ &=[T(u',\,v'a',\,b')u,\,T(u',\,v'a',\,b')v,\,a,\,b]. \end{aligned}\end{aligned}$$ There is a natural map $\kappa\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2n}(R,\,I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R)$ sending $[u,\,v,\,a,\,b]$ to $[u,\,va,\,b]$ (here we need the relation K7 established in the previous section). Its image is contained in ${\mathop{\mathrm{Ker}}\nolimits}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R){\twoheadrightarrow}{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R/I)\big)$ and contains all elements of the form $\,^{g}X_{ij}(a)=[\phi(g)e_i,\,\phi(g)e_{-j}\,a\,{\mathrm{sign}(-j)},\,0]$ and $\,^{g}X_{i,-i}(a)=[\phi(g)e_i,\,0,\,a]$ for $a\in I$, and thus actually coincides with this kernel. Any triple $(u,\,v,\,a)\in V\times V\times R$ defines a homomorphism $$\alpha_{u,v,a}\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2n}(R,\,I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2n}(R,\,I)$$ sending generator $[u',\,v',\,a',\,b']$ to $[T(u,\,v,\,a)u',\,T(u,\,v,\,a)v',\,a',\,b']$. To show that $\alpha_{u,v,a}$ is well-defined we have to check that T0–T6 hold for the images of the generators, but that is straightforward. Next, there exists a well-defined homomorphism $${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R)\rightarrow\mathrm{Aut}\,({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^*_{2l}(R,\,I))$$ sending $X(u,\,v,\,a)$ to $\alpha_{u,v,a}$, i.e., the absolute Steinberg group acts on Tulenbaev group. Obviously, we need to verify that K1–K3 hold for $\alpha_{u,v,a}$, but that is also straightforward. \[allthesame\] Let $I\trianglelefteq R$ be a splitting ideal. Then $${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2l}(R,\,I)={\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R,\,I)={\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I).$$ We identify ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2l}(R,\,I)$ with ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)$ and construct a map inverse to $\kappa$. With this end define $$Y^*_{ij}(a)=[e_i,\,e_{-j},\,a\,{\mathrm{sign}(-j)},\,0]\quad\text{and}\quad Y^*_{i,-i}(a)=[e_i,\,0,\,0,\,a]$$ inside ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2l}(R,\,I)$. These elements satisfy relations KL0–KL7. KL0–KL2 and KL7 are obvious. Consider, say, KL4. $$\begin{gathered} \llbracket X_{i,-i}(r),\,Y^*_{-i,j}(b)]=\\ =[e_{-j},\,T(e_i,\,0,\,r)e_{-i},\,b{\varepsilon_{-j}},\,0][e_{-j},\,e_{-i},\,b{\varepsilon_{-j}},\,0]{^{-1}}=\\ =[e_{-j},\,e_ir{\varepsilon_{i}},\,b{\varepsilon_{-j}},\,-rb^2]=\\ =[e_{-j},\,e_ir{\varepsilon_{i}},\,b{\varepsilon_{-j}},\,0][e_{-j},\,0,\,b{\varepsilon_{-j}},\,-rb^2]=\\ =Y^*_{ij}(rb{\varepsilon_{i}})\cdot Y^*_{-j,j}(-rb^2);\end{gathered}$$ One can check other relations similarly. For KL5 use T5 and for KL6 use T3. Thus, we have a map $\theta\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm KL}_{2l}(R,\,I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}^{\rm T}_{2l}(R,\,I)$ preserving the action. Obviously, $\kappa\theta=1$ thus $\theta$ is injective. It remains to show that it is also surjective. First, observe that $[e_1,\,v,\,a,\,b]=[e_1,\,0,\,0,\,b-a^2\sum v_kv_{-k}]\prod[e_1,\,e_k,\,v_ka,\,0]$ lie in the image of $\theta$ (here $v_k$ is a $k$-th coordinate of $v$; we use that $v_{-1}=0$ and T3). Thus, all generators $[u,\,v,\,a,\,b]$ lie in the image of $\theta$, since it preserves the action. In the sequel for splitting ideals we do not distinguish relative Steinberg groups defined in this section. Local-global principle ====================== In this section we prove the Main Theorem. Fix a non-nilpotent element $a\in R$. Let $\lambda_a\colon R\rightarrow R_a$ be a principal localisation of $R$ in $a$. For any $x\in R[t]$ consider the evaluation map $\mathrm{ev}_x\colon R[t]\rightarrow R[t]$, which is the only $R$-algebra homomorphism sending $t$ to $x$. For $p\in R[t]$ denote its image under $\mathrm{ev}_x$ by $p(x)$, e.g., $p=p(t)$. Similarly, for $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t])$ denote its image under $\mathrm{ev}_{x}^*$ by $g(x)$. We claim the following. \[tul-inj\] Consider $g(t)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])$ such that $$\lambda_{a}^*(g(t))=1\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R_a[t]).$$ Then there exists an $N\in\mathbb N$ such that $g(a^Nt)=1$. Similarly, assume that $$\lambda_{a}^*(g(t))\in\mathrm{Im}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R_a[t])\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R_a[t])\big).$$ Then there exists an $N\in\mathbb N$ such that $$g(a^Nt)\in\mathrm{Im}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R[t],\,tR[t])\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])\big).$$ Now, we define the symplectic analogue of Tulenbaev map and use it to prove Lemma \[tul-inj\] Denote $ B=R\ltimes tR_a[t] $ the ring with component-wise addition and multiplication given by $$(r,\,f)\cdot(s,\,g)=(rs,\,\lambda_a(r)g+f\lambda_a(s)+fg).$$ One may think of elements of $B$ as polynomials in $t$ with the constant term from $R$ and all other coefficients from $R_a$. Consider a direct system of rings $$\xymatrix{ R[t]\ar@<-0.0ex>[r]^{\mathrm{ev}_{at}}&R[t]\ar@<-0.0ex>[r]^{\mathrm{ev}_{at}}&R[t]\ar@<-0.0ex>[r]^{\mathrm{ev}_{at}}&\cdots }$$ i.e., $(S_i,\,\psi_{ij})_{0\leq i\leq j}$, where $S_i=R[t]$ and $\psi_{ij}\colon t\mapsto a^{j-i}t$. It induces a direct system of Steinberg groups. The following facts are left to the reader. A system of maps $\varphi_i\colon S_i\rightarrow B$ sending $$p(t)\mapsto\big(p+tR[t],\,\lambda_{a}^*(p)(a^{-i}t)-\lambda_{a}^*(p)(0)\big)$$ induces 1. an isomorphism $$\varinjlim S_i\rightarrow ^{\!\!\!\!\!\!\!\sim}B;$$ 2. an isomorphism $$\varinjlim{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(S_i)\rightarrow ^{\!\!\!\!\!\!\!\sim}\ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B).$$ Indeed, Steinberg group functor commutes with directed limits. Now, we claim that the composition of $\varphi_{0}^*$ with the inclusion $$\xymatrix{ \mu\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])\ar@{^(->}[r]^{}&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t])\ar@<-0.0ex>[r]^{\varphi_{0}^*}&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B) }$$ factors through the localisation in $a$. More generally, the following statement holds. \[tul\] Let $B$ be a ring, $a\in B$, and $I\trianglelefteq B$ be an ideal such that for any $x\in I$ there exists a unique $y\in I$ such that $ya=x$ [(]{}equivalently, a localisation map $\lambda_a\colon I\rightarrow I_a=I\otimes_RR_a$ is an isomorphism[)]{}. Then, there exists a map $$\mathrm T\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I_a)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$$ making the diagram $$\xymatrix{ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B,\,I)\ar@<-0.0ex>[rr]^{}\ar@<-0.0ex>[d]_{\lambda_{a}^*}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)\ar@<-0.0ex>[d]_{\lambda_{a}^*}\\ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I_a)\ar@<-0.0ex>[rr]^{}\ar@{-->}[rru]_{\mathrm T}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B_a)\\ }$$ commutative. Moreover, for $g\in\mathrm{Im}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}^{\mathrm T}(R_a[t],tR_a[t])\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I_a)\big)$ one has $\mathrm T(g)\in\mathrm{Im}\big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(B)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)\big)$. The next section is devoted to the proof of Lemma \[tul\]. Now, we deduce Lemma \[tul-inj\] from it. Apply Lemma \[tul\] for $a\in R\subseteq B$, $B=R\ltimes tR_a[t]$ as above, $I=tR_a[t]\trianglelefteq B$. Consider the following commutative diagram. $$\xymatrix{ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])\ar@{^(->}[rr]^{}\ar@<-0.0ex>[rd]_{\varphi_0^*}\ar@<-0.0ex>[dd]_{\lambda_{a}^*}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t])\ar@<-0.0ex>[dd]_{\varphi_{0}^*}\\ &{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B,\,I)\ar@{^(->}[rd]^{}\ar@<-0.0ex>[ld]_{\lambda_a^*}&\\ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R_a[t],\,tR_a[t])\ar@<-0.0ex>[rr]^{\mathrm T}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)\\ }$$ Take $g(t)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t],\,tR[t])$ such that $\lambda_{a}^*(g(t))=1$. Then $$\varphi_0^*(g(t))=(T\circ\lambda_{a}^*)(g(t))=1$$ as well, i.e., $g(t)$ becomes trivial in the limit. But it can only happen if $\psi_{0,N}^*(g(t))=1$ for some $N$ (use the construction of direct limit as disjoint union modulo an equivalence relation). The proof of the second statement is similar. For the next lemma the proof of Lemma 16 of [@n8] works verbatim. There are two references in that proof: instead of Lemma 8 of [@n8] use Lemma \[allthesame\], and instead of Lemma 15 of [@n8] use Lemma \[tul-inj\]. The second statement is not proven in [@n8], but the proofs of both statements are the same. \[almost\] Consider $a$, $b\in R$ generating $R$ as an ideal, $Ra+Rb=R$. Assume that for $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(R[t], tR[t])$ one has $\lambda_{a}^*(g)=\lambda_{b}^*(g)=1$ . Then $g=1$. Similarly, assume that $\lambda_{a}^*(g)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R_a[t], tR_a[t])$ and $\lambda_{b}^*(g)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R_b[t], tR_b[t])$. Then $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(R[t], tR[t])$. Now, the Main Theorem also follows. For the first statement the proof of Theorem 2 of [@n8] can be repeated verbatim. The only reference in that proof is Lemma 16 of [@n8], which should be replaced by Lemma \[almost\]. For the second statement the same proof works. Tulenbaev map ============= This section is devoted to the construction of the map $$\mathrm T\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I_a)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$$ from Lemma \[tul\]. As there, let $B$ be a ring, $a\in B$ a non-nilpotent element, $I\trianglelefteq B$, such that for any $x\in I$ there exists a unique $y\in I$ such that $ya=x$. We denote such a $y$ by $\frac xa$. Elements $\frac x{a^N}$ are also well-defined. The localisation map $\lambda_a\colon I\rightarrow I_a$ is an isomorphism and we identify $I$ and $I_a$. To define the map $\mathrm T$ we need to find elements $Z(u,\,v,\,b,\,c)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$ for any $u\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(B_a)e_1$, $v\in B_a^{2n}$, ${\langle}u,\,v{\rangle}=0$, and $b$, $c\in I$ subject to relations T0–T6. We start with the following definition. For $u$, $v_1,\ldots,v_N\in B^{2n}$ such that ${\langle}u,\,v_k{\rangle}=0$ for all $k$ define $$\begin{gathered} Z(u;\,v_1,\ldots,v_N)=X(u,\,v_1,\,0)\ldots X(u,\,v_N,\,0)\cdot X(u,\,0,\,-\sum\limits_{i<j}{\langle}v_i,\,v_j{\rangle}).\end{gathered}$$ Consider $u$, $v$, $w\in B^{2n}$, ${\langle}u,\,v{\rangle}={\langle}u,\,w{\rangle}=0$. Suppose that $w$ has a pair of zero coordinates, i.e., $w_i=w_{-i}=0$ for some $i$. Then $$[X(u,\,v,\,0),\,X(u,\,w,\,0)]=X(u,\,0,\,2{\langle}v,\,w{\rangle}).$$ Use Lemma \[xlist\](X1) to compute the conjugate and decompose the result by Lemma \[xlist\](X6) $$\begin{gathered} \,^{X(u,\,v,\,0)}X(u,\,w,\,0)=\\=X(u,\,0,\,-1)X(w+u{\langle}v,\,w{\rangle},\,0,\,-1)X(w+u(1+{\langle}v,\,w{\rangle}),\,0,\,1),\end{gathered}$$ then decompose the first and the third factors by Lemma \[xlist2\](X8). Next, change the order of factors and then simplify the product using Lemmas \[xlist\] and \[xlist2\]. As a result, we have $$\,^{X(u,\,v,\,0)}X(u,\,w,\,0)=X(u,\,0,\,2{\langle}v,\,w{\rangle})X(w,\,u,\,0).$$ For the following corollary use that the symmetric group is generated by fundamental transpositions. Take $u$ and $v_1,\ldots,v_N\in B^{2n}$ such that ${\langle}u,\,v_k{\rangle}=0$ and each $v_k$ has a pair of zero coordinates. Then for any permutation $\sigma\in S_N$ we have $$Z(u;\,v_1,\ldots,v_N)=Z(u;\,v_{\sigma(1)},\ldots,v_{\sigma(N)}).$$ For $u$, $v_1,\ldots,v_N\in B^{2n}$ such that ${\langle}u,\,v_k{\rangle}=0$ and each $v_k$ has a symmetric pair of zero coordinates we denote $$Z(u;\,\{v_k\}_{1\leq k\leq N})=Z(u;\,v_1,\ldots,v_N).$$ Observe also that the following easy fact holds (use Lemma \[xlist\](X2) and Lemma \[xlist2\](X9). \[forgotten\] For $u$, $v_1,\ldots,v_N\in B^{2n}$ such that ${\langle}u,\,v_k{\rangle}=0$ and each $v_k$ has a pair of zero coordinates, $r\in B$, one has $$Z(ur;\,\{v_k\}_{1\leq k\leq N})=Z(u;\,\{rv_k\}_{1\leq k\leq N}).$$ The following result is well-known (see [@n2; @n5]). \[suslin\] For $w$, $u$, $v\in B^{2n}$ such that ${\langle}w,\,u{\rangle}=A\in B$, ${\langle}u,\,v{\rangle}=0$ denote $$v_{ij}=v_{ij}^w=(e_iu_{-j}\,{\mathrm{sign}(j)}-e_ju_{-i}\,{\mathrm{sign}(i)})(v_iw_j-v_jw_i)$$ for any distinct $\,-n\leq i,j\leq n$. Then one has $v_{ij}=v_{ji}$, ${\langle}u,\,v_{ij}{\rangle}=0$, and $$\sum_{i<j}v_{ij}=vA.$$ Compare the next result with Lemma \[xlist2\](X10): we do not need that $v_i=v_{-i}=0$, but we assume that ${\langle}v,\,w{\rangle}=0$. \[orth\] Take $u$, $v$, $w\in B^{2n}$ such that ${\langle}u,\,v{\rangle}={\langle}u,\,w{\rangle}={\langle}v,\,w{\rangle}=0$ and $w_i=w_{-i}=0$. Then $$X(u,\,v+w,\,0)=X(u,\,v,\,0)X(u,\,w,\,0).$$ By Lemma \[xlist\](X6) we have $$X(u,\,v+w,\,0)=X(u,\,0,\,-1)X(v+w,\,0,\,-1)X(u+v+w,\,0,\,1).$$ Decompose the second and the third factors by Lemma \[xlist2\](X8). We get a factor $X(w,\,u+v,\,0)$, and decompose it by Lemma \[xlist2\](X10). Vectors $u$, $v$, $w$ are orthogonal, thus all factors commute. Simplify the product by Lemma \[xlist2\](X10) and get the claim with the use of Lemma \[xlist\](X6). \[x=z\] Take $w$, $u$, $v\in B^{2n}$ such that ${\langle}w,\,u{\rangle}=A$ and ${\langle}u,\,v{\rangle}=0$. Assume in addition that $v$ has a symmetric pair of zero coordinates. Then $$X(u,\,vA,\,0)=Z(u,\,\{v_{ij}\}_{i\leq j}).$$ Say, $v_1=v_{-1}=0$. Then $v_{1,-1}=0$. Decompose $$vA=\sum_{i<j}v_{ij}=\underbrace{\sum_{i\neq\pm1}v_{-1,i}}_p+\underbrace{\sum_{i\neq\pm1}v_{1,i}}_{q}+\underbrace{\sum_{i,j\neq\pm1}v_{ij}}_r$$ with the use of Suslin’s Lemma (Lemma \[suslin\]). Obviously, $$p_{-1}=\Big(\sum_{i\neq\pm1}v_{-1,i}\Big)_{-1}=\Big(\sum_{i<j}v_{ij}\Big)_{-1}=v_{-1}A=0$$ and similarly $q_1=v_1A=0$. Also, $p_1=q_{-1}=r_{-1}=r_1=0$. Thus, by Lemma \[xlist2\](X10) we get $$\begin{gathered} X(u,\,vA,\,0)=X(u,\,p+q+r,\,0)=\\ =X(u,\,p,\,0)X(u,\,q,\,0)X(u,\,r,\,0)X(u,\,0,\,-{\langle}p,\,q{\rangle}-{\langle}p,\,r{\rangle}-{\langle}q,\,r{\rangle}).\end{gathered}$$ For $1<i\leq n$ denote $z_i=v_{-1,i}+v_{-1,-i}$. Then ${\langle}z_i,\,z_j{\rangle}=0$ for $i\neq j$, each $z_i$ has a pair of zero coordinates and $\sum_{i=2}^nz_i=p$. By Lemma \[orth\] we get $$X(u,\,p,\,0)=\prod_{i=2}^nX(u,\,z_i,\,0).$$ Next, observe that $v_{-1,i}$ and $v_{-1,-i}$ have a common symmetric pair of zero coordinates. Thus, by Lemma \[xlist2\](X10), one has $$X(u,\,z_i,\,0)=X(u,\,v_{-1,i},\,0)X(u,\,v_{1,i},\,0)X(u,\,0,\,-{\langle}v_{-1,i},\,v_{1,i}{\rangle}),$$ so that $$X(u,\,p,\,0)=Z(u;\,\{v_{-1,i}\}_{i\neq\pm1}).$$ Similarly, $X(u,\,q,\,0)=Z(u;\,\{v_{1,i}\}_{i\neq\pm1})$ and by Lemma \[xlist2\](X10), $X(u,\,r,\,0)=Z(u;\,\{v_{ij}\}_{i,j\neq\pm1})$ what finishes the proof. \[decomposition\] Take $w$, $u$, $v\in B^{2n}$, ${\langle}w,\,u{\rangle}=A$ and ${\langle}u,\,v{\rangle}=0$. Consider $x^1,\ldots x^N\in B^{2n}$ such that each $x^k$ has a pair of zero coordinates, ${\langle}u,\,x^k{\rangle}=0$, and $\sum_{k=1}^Nx^k=vA$. Then one has $$Z(u;\,\{x^kA\}_{k=1}^N)=Z(u;\,\{v_{ij}A\}_{i<j}).$$ Since ${\langle}u,\,x^k{\rangle}=0$ consider $x_{ij}^k=(x^k)_{ij}^w$ from Suslin’s Lemma (Lemma \[suslin\]) and use Lemma \[x=z\] to get $$X(u,\,x^kA,\,0)=Z(u,\,\{x_{ij}^k\}_{i<j}).$$ Then, $$Z(u;\,\{x^kA\}_{k=1}^N)=Z(u,\,\{x_{ij}^k\}_{k,\,i<j}).$$ On the other hand, for fixed $i$ and $j$ all $x_{ij}^k$ are scalar multiples of the same vector having a pair of zero coordinates and $$\sum_{k=1}^Nx_{ij}^k=(e_iu_{-j}\,{\mathrm{sign}(j)}-e_ju_{-i}\,{\mathrm{sign}(i)})\Big(\big(\sum_{k=1}^Nx_i^k\big)w_j-\big(\sum_{k=1}^Nx_j^k\big)w_i\Big)=v_{ij}A.$$ Thus, $$X(u,\,v_{ij}A,\,0)=\prod_{k=1}^NX(u,\,x_{ij}^k,\,0)$$ by Lemma \[xlist2\](X10) and $$Z(u;\,\{v_{ij}A\}_{i<j})=Z(u,\,\{x_{ij}^k\}_{k,\,i<j}).$$ Take $u$, $v\in B^{2n}$ such that ${\langle}u,\,v{\rangle}=0$ and denote by $$I(u)=\sum_{k=-n}^nBu_k$$ the ideal generated by entries of $u$. Then for an $A\in I(u)$ take any $w\in B^{2n}$ such that ${\langle}w,\,u{\rangle}=A$ and denote $$Z^A(u,\,v)=Z(u;\,\{v_{ij}^wA\}_{i<j}).$$ By the previous lemma, this element does not depend on the choice of $w$. The projection of $Z^A(u,\,v)$ to the elementary group $$\phi(Z^A(u,\,v))=T(u,\,vA^2,\,0).$$ We start to prove properties of the elements $Z^A(u,\,v)$. \[conj\] Take $u$, $v\in B^{2n}$ such that ${\langle}u,\,v{\rangle}=0$, $A\in I(u)$ and $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$. Then $$g\,Z^A(u,\,v)g{^{-1}}=Z^A(\phi(g)u,\,\phi(g)v).$$ Take $w$ such that ${\langle}w,\,u{\rangle}=A$. Then, ${\langle}\phi(g)w,\,\phi(g)u{\rangle}={\langle}w,\,u{\rangle}=A$ as well, so that $A\in I(\phi(g)u)$ and the right hand side is well-defined. We may assume that $g=X_{ij}(b)$. Obviously, $$g\,X(u,\,v_{hk}A,\,0)g{^{-1}}=X(\phi(g)u,\,\phi(g)v_{hk}A,\,0)$$ and $$-\sum{\langle}v_{hk}A,\,v_{st}A{\rangle}=-\sum{\langle}\phi(g)v_{hk}A,\,\phi(g)v_{st}A{\rangle},$$ so that, $$g\,Z^A(u,\,v)g{^{-1}}=Z(\phi(g)u;\,\{\phi(g)v_{hk}A\}_{h<k}).$$ For $n\geq4$ each $T_{ij}(b)v_{hk}$ still has at least one pair of zero coordinates what finishes the proof for this case by Lemma \[decomposition\]. Now, consider the case of $n=3$. For $j=-i$ any $T_{ij}(b)v_{hk}$ still has a pair of zero coordinates. Now, assume $j\neq\pm i$. If $h$, $k\not\in\{j,-i\}$, then $\phi(g)v_{hk}=v_{hk}$. If $\{h,k\}=\{j,-i\}$ we also get that $T_{ij}(b)v_{hk}$ has a pair of zero coordinates. Thus, we may assume that $h\in\{j,-i\}$, say, $h=-i$, and $k\not\in\{\pm i,\pm j\}$. Set $$u_{k,-i}=e_ku_{i}\,{\mathrm{sign}(-i)}-e_{-i}u_{-k}\,{\mathrm{sign}(k)},$$ then $v_{k,-i}=u_{k,-i}(v_kw_{-i}-v_{-i}w_k)$. One has $$\begin{aligned} &T_{ij}(b)u=u+e_iu_jb-e_{-j}u_{-i}\,b\,{\mathrm{sign}(ij)},\\ &T_{ij}(b)u_{k,-i}=u_{k,-i}+e_{-j}u_{-k}\,b\,{\mathrm{sign}(ijk)}.\end{aligned}$$ Direct computation shows that $${\langle}T_{ij}(b)u,\,u_{k,-i}-e_ku_j\,b\,{\mathrm{sign}(i)}{\rangle}=0.$$ Set $$q=u_{k,-i}-e_ku_j\,b\,{\mathrm{sign}(i)}\,\text{ and }\,r=e_ku_j\,b\,{\mathrm{sign}(i)}+e_{-j}u_{-k}\,b\,{\mathrm{sign}(ijk)}.$$ One has $T_{ij}(b)u_{k,-i}=q+r$, so that $r$ is also orthogonal to $T_{ij}(b)u$. Both $q$ and $r$ have a pair of zero coordinates and, moreover, they are orthogonal. Set $c=(v_kw_{-i}-v_{-i}w_k)$, then by Lemma \[orth\] $$X(T_{ij}(b)u,\,T_{ij}(b)v_{k,-i}A,\,0)=X(T_{ij}(b)u,\,qcA,\,0)X(T_{ij}(b)u,\,rcA,\,0).$$ Finally, the claim follows from Lemma \[decomposition\]. The next lemma follows from Lemma \[decomposition\]. \[add\] Take $u$, $v$, $w\in B^{2n}$ such that ${\langle}u,\,v{\rangle}={\langle}u,\,w{\rangle}=0$, $A\in I(u)$. Then $$Z^A(u,\,v)Z^A(u,\,w)=Z^A(u,\,v+w)X(u,\,0,\,{\langle}v,\,w{\rangle}\cdot A^4).$$ For $u$, $v$, ${\langle}u,\,v{\rangle}=0$, $A\in I(u)$ one has $$\begin{aligned} &Z^A(u,\,0)=1,\\ &Z^A(u,\,v){^{-1}}=Z^A(u,\,-v).\end{aligned}$$ \[symm\] Take $u$, $v\in B^{2n}$ such that ${\langle}u,\,v{\rangle}=0$, $A\in I(u)\cap I(v)$, $b\in B$. Assume that there exist $p$, $q\in B^{2n}$ such that $${\langle}u,\,p{\rangle}={\langle}u,\,q{\rangle}={\langle}v,\,p{\rangle}={\langle}v,\,q{\rangle}=0,$$ and ${\langle}p,\,q{\rangle}=A$. Then one has $$Z^A(u,\,vb\cdot A^3)=Z^A(v,\,ub\cdot A^3).$$ Denote $g=Z^A(u,\,pb)$ and $h=Z^A(v,\,q)$. Compute the commutator in two ways $$\!\,^gh\cdot h{^{-1}}=g\cdot\,^hg{^{-1}}.$$ Recall that $\phi\big(Z^A(u,\,pb)\big)=T(u,\,pbA^2,\,0)$ and use Lemmas \[conj\] and \[add\] to get $$\!\,^gh\cdot h{^{-1}}=Z^A(v,\,q+ubA^3)Z^A(v,\,-q)=Z^A(v,\,ubA^3).$$ Similarly, $g\cdot\,^hg{^{-1}}=Z^A(u,\,vbA^3)$. Consider $w$, $u\in B^{2n}$, $b\in B$, denote $A={\langle}w,\,u{\rangle}$. Assume that there exist $z$, $v\in B^{2n}$ such that ${\langle}z,\,v{\rangle}=A$ and $${\langle}u,\,v{\rangle}={\langle}u,\,z{\rangle}={\langle}w,\,v{\rangle}={\langle}w,\,z{\rangle}=0.$$ Then one has $$Z^A(u,\,ubA^3)=X(u,\,0,\,2bA^5).$$ Set $g=Z^A(u,\,zb)$, $h=Z^A(u,\,v)$ and compute $[g,\,h]$ in two ways. On the one hand, $$\!\,^gh\cdot h{^{-1}}=Z^A(u,\,v+ubA^3)Z^A(u,\,-v)=Z(u,\,ubA^3)$$ by Lemma \[conj\]. On the other hand, $$\begin{gathered} gh\cdot g{^{-1}}h{^{-1}}=\\=Z^A(u,\,zb+v)X(u,\,0,\,bA^5)Z^A(u,\,-zb-v)X(u,\,0,\,bA^5)=\\=X(u,\,0,\,2bA^5)\end{gathered}$$ by Lemma \[add\]. \[5+6\] Consider $u$, $v\in B^{2n}$, $A\in I(u)\cap I(v)$, $b$, $c\in B$. Assume that there exist $w$, $z$, $x$, $y\in B^{2n}$, such that $${\langle}w,\,u{\rangle}={\langle}z,\,v{\rangle}={\langle}x,\,y{\rangle}=A$$ and pairs $(w,\,u)$, $(z,\,v)$ and $(x,\,y)$ are mutually orthogonal. Then one has $$X(u+vb,\,0,\,cA^{11})=X(u,\,0,\,cA^{11})X(v,\,0,\,b^2cA^{11})Z^A(u,\,vbcA^9).$$ First, use Lemma \[add\] $$\begin{gathered} Z^A(u+vb,\,xA^3)Z^A(u+vb,\,ycA^3)=\\=Z^A(u+vb,\,(x+yc)A^3)X(u+vb,\,0,\,cA^{11}).\end{gathered}$$ We want to show that $$Z^A(u+vb,\,xA^3)=Z^A(x,\,(u+vb)A^3).$$ With this end, use Lemma \[symm\] with $p=z-wb$ and $q=v$. Next, decompose $$Z^A(x,\,(u+vb)A^3)=Z^A(x,\,uA^3)Z^A(x,\,vbA^3)$$ by Lemma \[add\] and use Lemma \[symm\] with $p=z$ and $q=v$ to show that $$Z^A(x,\,uA^3)=Z^A(u,\,xA^3)$$ and with $p=w$, $q=u$ to get that $$Z^A(x,\,vbA^3)=Z^A(v,\,xbA^3).$$ Similarly, one shows that $$Z^A(u+vb,\,ycA^3)=Z^A(u,\,ycA^3)Z^A(v,\,ybcA^3)$$ and $$Z^A(u+vb,\,-(x+yc)A^3)=Z^A(u,\,-(x+yc)A^3)Z^A(v,\,-(x+yc)bA^3).$$ Decompose also $$\begin{aligned} &Z^A(u,\,-(x+yc)A^3)=Z^A(u,\,-ycA^3)Z^A(u,\,-xA^3)X(u,\,0,\,cA^{11}),\\ &Z^A(v,\,-(x+yc)bA^3)=Z^A(v,\,-ybcA^3)Z^A(v,\,-xbA^3)X(v,\,0,\,b^2cA^{11})\end{aligned}$$ by Lemma \[add\]. Now, we can express $X(u+vb,\,0,\,cA^{11})$ in terms of these ten elements. Most of the factors will cancel, but we will need to interchange positions of $Z^A(u,\,xA^3)$ and $Z^A(v,\,ybcA^3)$, thus we obtain their commutator as an extra factor $$[Z^A(u,\,xA^3),\,Z^A(v,\,ybcA^3)]=Z^A(u,\,vbcA^9).$$ Now, we focus on the case $A=a^N$. Take $b\in I$ and $u$, $v\in B^{2n}$, such that $a^N\in I(u)$ for some $N\in\mathbb N$, ${\langle}u,\,v{\rangle}=0$. Then set $$Z(u,\,v,\,b)=Z^{(a^N)}\Big(u,\,v\frac b{a^{2N}}\Big).$$ This element does not depend on the choice of $N$. Take $w\in B^{2n}$ such that ${\langle}w,\,u{\rangle}=a^N$, then ${\langle}wa^M,\,u{\rangle}=a^{N+M}$ and by the very definition $$\Big(v\frac{b}{a^{2(N+M)}}\Big)_{ij}^{(wa^M)}=v_{ij}^{(wa^M)}\cdot\frac{b}{a^{2(N+M)}}=v_{ij}^w\cdot a^M\cdot\frac{b}{a^{2(N+M)}},$$ so that $$Z^{(a^{N+M})}\Big(u,\,v\frac{b}{a^{2(N+M)}}\Big)=Z\Big(u;\,\Big\{\big(v_{ij}^w\frac{b}{a^{2N+M}}\big)a^{N+M}\Big\}_{i<j}\Big)=Z^{(a^N)}\Big(u,\,v\frac b{a^{2N}}\Big).$$ Observe that $\phi\big(Z(u,\,v,\,b)\big)=T(u,\,vb,\,0)$. Below we list the properties of our new elements $Z(u,\,v,\,b)$. They follow directly from the definition and Lemmas \[conj\]–\[5+6\]. \[zlist\] Take $u$, $v$, $v'\in B^{2n}$ such that $a^N\in I(u)$ for some $N\in\mathbb N$, ${\langle}u,\,v{\rangle}={\langle}u,\,v'{\rangle}=0$, $b$, $c\in I$, $r\in B$, $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$. Then one has $$\begin{aligned} &\phi\big(Z(u,\,v,\,b)\big)=T(u,\,vb,\,0),\\ &Z(u,\,vr,\,b)=Z(u,\,v,\,rb),\\ &Z(u,\,v,\,b)Z(u,\,v',\,b)=Z(u,\,v+v',\,b)X(u,\,0,\,b^2{\langle}v,\,v'{\rangle}),\\ &Z(u,\,v,\,b)Z(u,\,v,\,c)=Z(u,\,v,\,b+c),\\ &g\,Z(u,\,v,\,b)g{^{-1}}=Z(\phi(g)u,\,\phi(g)v,\,b).\end{aligned}$$ Assume that there also exist $w$, $z\in B^{2n}$ such that holds ${\langle}w,\,u{\rangle}={\langle}z,\,v{\rangle}=a^N$ and pairs $(w,\,u)$, $(z,\,v)$ are orthogonal. Then one also has $$\begin{aligned} &Z(u,\,u,\,b)=X(u,\,0,\,2b).\end{aligned}$$ If in addition there exist $x$, $y\in B^{2n}$ such that ${\langle}x,\,y{\rangle}=a^N$ and the pair $(x,\,y)$ is orthogonal to pairs $(w,\,u)$ and $(z,\,v)$, then $$\begin{aligned} &Z(u,\,v,\,b)=Z(v,\,u,\,b),\\ &X(u+vr,\,0,\,b)=X(u,\,0,\,b)X(v,\,0,\,br^2)Z(u,\,v,\,br).\end{aligned}$$ We need yet another property of $Z(u,\,v,\,b)$. \[onemore\] For $u$, $v\in B^{2n}$, $b\in I$, $M$, $N\in\mathbb N$, such that $a^N\in I(u)$, ${\langle}u,\,v{\rangle}=0$, holds $$Z(ua^M,\,v,\,b)=Z(u,\,v,\,a^Mb).$$ Firstly, we clearify the notations. Take $w\in B^{2n}$ such that ${\langle}w,\,u{\rangle}=a^N$, then ${\langle}w,\,ua^M{\rangle}=a^{N+M}$. Denote $u_{ij}=e_iu_{-j}\,{\mathrm{sign}(j)}-e_ju_{-i}\,{\mathrm{sign}(i)}$, then $(ua^M)_{ij}=u_{ij}a^M$. Denote as usually $v_{ij}=u_{ij}(v_iw_j-v_jw_i)$. Then in the definition of $$Z^{a^{N+M}}\Big(ua^M,\,v\frac b{a^{2N+2M}}\Big)$$ we actually use $(ua^M)_{ij}\cdot(v_iw_j-v_jw_i)=v_{ij}a^M$. Thus, we have $$\begin{gathered} Z(ua^M,\,v,\,b)=Z^{a^{N+M}}\Big(ua^M,\,v\frac b{a^{2N+2M}}\Big)=\\ =Z\Big(ua^M;\,\Big\{\big(v_{ij}a^M\frac{b}{a^{2N+2M}}\big)a^{N+M}\Big\}_{i<j}\Big).\end{gathered}$$ Now use Lemma \[forgotten\] and get $$Z\Big(ua^M;\,\Big\{\big(v_{ij}\frac{b}{a^{2N}}\big)a^{N}\Big\}_{i<j}\Big)=Z\Big(u;\,\Big\{\big(v_{ij}\frac{a^Mb}{a^{2N}}\big)a^{N}\Big\}_{i<j}\Big)=Z(u,\,v,\,a^Mb).$$ Finally, introduce yet another notation. For $u$, $v\in B^{2n}$, such that $a^N\in I(u)$ for some $N\in\mathbb N$, ${\langle}u,\,v{\rangle}=0$, $b$, $c\in I$ denote $$Z(u,\,v,\,b,\,c)=Z(u,\,v,\,b)X(u,\,0,\,c).$$ At this point, we are ready to construct Tulenbaev map $$\mathrm T\colon{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B).$$ For each quadruple $$(u,\,v,\,b,\,c)\in\Big({\mathop{\mathrm{Ep}}\nolimits}_{2n}(B_a)e_1\Big)\times B_a^{2n}\times I\times I$$ we associate an element in ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$. We proceed as follows. First, if $u=Me_1$, denote $w=-Me_{-1}$, then ${\langle}w,\,u{\rangle}=1$. Next, $w$, $u$, $v\in B_a$, thus there exists an $N\in\mathbb N$ such that the entries of $wa^N$, $ua^N$, $va^N$ do not have denominators, i.e., lie in the image of the localisation homomorphism $\lambda_a\colon B\rightarrow B_a$. Then for each of these elements take its their preimages and get vectors $\tilde w$, $\tilde u$, $\tilde v\in B^{2n}$. Since ${\langle}\tilde u,\,\tilde v{\rangle}$ localises to zero and ${\langle}\tilde w,\,\tilde u{\rangle}$ localises to $a^{2N}$, there exist $M\in\mathbb N$ such that ${\langle}\tilde u,\,\tilde va^M{\rangle}=0$ and ${\langle}\tilde wa^M,\,\tilde u{\rangle}=a^{2N+M}$. Then the element $$Z\Big(\tilde u,\,\tilde va^M,\,\frac b{a^{2N+M}},\,\frac c{a^{2N}}\Big)$$ in ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$ is defined. Using Lemma \[xlist\](X2), Lemma \[zlist\](Z1) and Lemma \[onemore\] one can show that this element does not depend on the above choices. Thus, we have a well-defined set-theoretic map from the set of generators of ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I)$ to ${\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$. Next, we need to show that the images of $(u,\,v,\,b,\,c)$ under this map satisfy relations T0–T6. This is a straightforward consequence of Lemmas \[xlist\] and \[zlist\] and the fact that the above map is well-defined. The only trick one should use to get T3–T5 is the following. For $u=Me_1$, where $M\in{\mathop{\mathrm{Ep}}\nolimits}_{2n}(B_a)$, one can take $w=-Me_{-1}$, $v=Me_{-2}$, $z=Me_2$ and use their lifts to deduce T3 from Lemma \[zlist\](Z5). Similarly, one can use $(e_3,\,e_{-3})$ for T4 and T5. Now, we have to show that the diagram $$\xymatrix{ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B,\,I)\ar@<-0.0ex>[rr]^{\kappa}\ar@<-0.0ex>[d]_{\lambda_{a}^*}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)\ar@<-0.0ex>[d]_{\lambda_{a}^*}\\ {\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}^{\mathrm T}(B_a,\,I_a)\ar@<-0.0ex>[rr]_{\kappa}\ar@<-0.0ex>[rru]_{\mathrm T}&&{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B_a)\\ }$$ is commutative. Start with the upper triangle. Lemma \[zlist\](Z7) and Lemma \[xlist2\](X8) imply that for $u$ with a pair of symmetric zeros holds $Z(u,\,v,\,b,\,c)=X(u,\,vb,\,c)$. Next, take $X(u,\,v,\,b,\,c)\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B,\,I)$ and take $g\in{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)$ such that $\phi(g)u=e_1$. Then $\kappa$ sends it to $X(u,\,vb,\,c)$ and $T\circ\lambda_a^*$ to $$Z\Big(\lambda_a(u)a^N,\,\lambda_a(v)a^{N+M},\,\frac{b}{a^{2N+M}},\,\frac{c}{a^{2N}}\Big).$$ Now, conjugate both elements by $g$ and use the previous observation to show that they coincide. The lower triangle can be treated similarly. Finally, we show that $\mathrm T$ maps $$g\in\mathrm{Im}\Big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(B_a,I_a)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B_a,I_a)\Big)$$ to the element of $\mathrm{Im}\Big({\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(B)\rightarrow{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n}(B)\Big)$. For $n=2$ there exist no obvious analogue of the Tulenbaev map, so that for $n=3$ we argue as follows (for $n>3$ this argument works as well). We can assume that $g=X(u,\,v,\,b,\,c)$ for $u$ and $v$ such that $u_n=u_{-n}=v_n=v_{-n}=0$. Then, consider lifts $\tilde u$, $\tilde v$ of $ua^N$ and $va^N$. Their $\,\pm n$-th coordinates localise to zeros, thus increasing $N$ we may assume that they actually are zeros. Thus, it remains to show that for $u$ and $v$ such that $u_n=u_{-n}=v_n=v_{-n}=0$ one has $Z(u,\,v,\,b,\,c)\in\mathrm{Im}{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(B)$. As above, in this situation $Z(u,\,v,\,b,\,c)=X(u,\,vb,\,c)$. By Lemma \[xlist\](X6), we only need to consider $X(u,\,0,\,c)$ with $u_n=u_{-n}=0$. With the use of Lemma \[ylist\](Y3–Y6) we reduce it to the case $X(e_i,\,v,\,c)$ with $i\neq\pm n$, $v_n=v_{-n}=0$. To conclude the proof, decompose this element by Lemma \[xlist2\](X10) as a product of usual elementary generators of Steinberg group $$X(e_i,\,v,\,c)=X(e_i,\,0,\,c-\sum v_kv_{-k})\prod X(e_i,\,e_kv_k,\,0).$$ Al of them lie in $\mathrm{Im}{\mathop{\mathrm{St}}\nolimits}\!{\mathop{\mathrm{Sp}}\nolimits}_{2n-2}(B)$. [99]{} E. Abe, Whitehead groups of Chevalley groups over polynomial rings, [*Comm. Algebra*]{} [**11**]{} (1983) 1271–1307. A. Bak, N. Vavilov, Structure of hyperbolic unitary groups I: elementary subgroups, [*Algebra Colloq.*]{} [**7**]{} (2) (2000) 159–196. W. van der Kallen, Another presentation for Steinberg groups, [*Indag. Math*]{} [**80**]{} (1977) 304–312. F. Keune, The relativisation of $\mathrm K_2$, [*J. Algebra*]{} [**54**]{} (1) (1987) 159–177. V. Kopeiko, Stabilization of symplectic groups over rings of polynomials (Russian), [*Mat. Sb.*]{} ([*N. S.*]{}) [**106**]{}, ([**148**]{}) (1) (1978) 94–107. A. Lavrenov, Another presentation for symplectic Steinberg groups, [*J. Pure Appl. Algebra*]{} [**219**]{} (9) (2015) 3755–3780. J.-L. Loday, Cohomologie et groupes de Steinber relatifs, [*J. Algebra*]{} [**54**]{} (1) (1978) 178–202. S. Sinchuk, On centrality of $\mathrm K_2$ for Chevalley groups of type $E_l$, [*J. Pure Appl. Algebra*]{} [**220**]{} no. 2 (2016) 857–875. A. Stavrova, Homotopy invariance of non-stable $\mathrm K_1$-functors, [*J. K-theory*]{} [**13**]{} no.2 (2014) 199–248. A. Stepanov, Structure of Chevalley groups over rings via universal localization, [*J. Algebra*]{} [**450**]{} (2016) 522–548. A. Suslin, On the structure of special linear group over polynomial rings, [*Math. USSR Izv.*]{} [**11**]{} (1977) 221–238. A. Suslin, V. Kopeiko, Quadratic modules and the orthogonal group over polynomial rings, [*J. Soviet Math.*]{} [**20**]{} (1982) 2665–2691. D. Quillen, Projective modules over polynomial rings, [*Invent. Math.*]{} [**36**]{} (1976) 167–171. M. Tulenbaev, The Steinberg group of a polynomial ring, [*Math. USSR Sb.*]{} [**45**]{} (1) (1983) 131–144. [^1]: The author acknowledges financial support from Russian Science Foundation grant 14–11–00297.

ArXiv

--- abstract: 'From the early days of computing, games have been important testbeds for studying how well machines can do sophisticated decision making. In recent years, machine learning has made dramatic advances with artificial agents reaching superhuman performance in challenge domains like Go, Atari, and some variants of poker. As with their predecessors of chess, checkers, and backgammon, these game domains have driven research by providing sophisticated yet well-defined challenges for artificial intelligence practitioners. We continue this tradition by proposing the game of Hanabi as a new challenge domain with novel problems that arise from its combination of purely cooperative gameplay and imperfect information in a two to five player setting. In particular, we argue that Hanabi elevates reasoning about the beliefs and intentions of other agents to the foreground. We believe developing novel techniques capable of imbuing artificial agents with such theory of mind will not only be crucial for their success in Hanabi, but also in broader collaborative efforts, and especially those with human partners. To facilitate future research, we introduce the open-source Hanabi Learning Environment, propose an experimental framework for the research community to evaluate algorithmic advances, and assess the performance of current state-of-the-art techniques.' author: - Nolan Bard - 'Jakob N. Foerster' - Sarath Chandar - Neil Burch - Marc Lanctot - 'H. Francis Song' - Emilio Parisotto - Vincent Dumoulin - Subhodeep Moitra - Edward Hughes - Iain Dunning - Shibl Mourad - Hugo Larochelle - 'Marc G. Bellemare' - Michael Bowling bibliography: - 'hanabi-challenge.bib' title: 'The Hanabi Challenge: A New Frontier for AI Research' --- multi-agent learning ,challenge paper ,reinforcement learning ,games ,theory of mind ,communication ,imperfect information ,cooperative 00-01,99-00

ArXiv

--- address: 'Sergii Favorov, Karazin’s Kharkiv National University Svobody sq., 4, 61022, Kharkiv, Ukraine' author: - 'S.Yu. Favorov' title: Tempered distributions with discrete support and spectrum --- introduction ============ Denote by $S(\R^d)$ Schwartz space of test functions $\p\in C^\infty(\R^d)$ with finite norms $$\label{n} N_{n,m}(\p)=\sup_{\R^d}\{\max\{1,|x|^n\}\max_{\|k\|\le m} |D^k\p(x)|\},\quad n,m=0,1,2,\dots,$$ where $$k=(k_1,\dots,k_d)\in(\N\cup\{0\})^d,\ \|k\|=k_1+\dots+k_d,\ D^k=\partial^{k_1}_{x_1}\dots\partial^{k_d}_{x_d}.$$ These norms generate the topology on $S(\R^d)$. Elements of the space $S^*(\R^d)$ of continuous linear functionals on $S(\R^d)$ are called tempered distributions. For each tempered distribution $f$ there are $c<\infty$ and $n,\,m\in\N\cup\{0\}$ such that for all $\p\in S(\R^d)$ $$\label{d} |f(\p)|\le cN_{n,m}(\p).$$ Moreover, this estimate is sufficient for distribution $f$ to belong to $S^*(\R^d)$ (see [@V], Ch.3). The Fourier transform of a tempered distribution $f$ is defined by the equality $$\hat f(\p)=f(\hat\p)\quad\mbox{for all}\quad\p\in S(\R^d),$$ where $$\hat\p(y)=\int_{\R^d}\p(x)\exp\{-2\pi i\la x,y\ra\}dx$$ is the Fourier transform of the function $\p$. Note that the Fourier transform of every tempered distribution is also a tempered distribution. An element $f\in\R^d$ is called [*a quasicrystal Fourier*]{} if $f$ and $\hat f$ are discrete measures on $\R^d$. In this case the support of $\hat f$ is called [*spectrum*]{} of $f$. These notions were inspired by experimental discovery made in the middle of 80’s of non-periodic atomic structures with diffraction patterns consisting of spots. There are a lot of papers devoted to investigation of properties of quasicrystals Fourier (see, for example, collections of works [@D], [@Q], papers [@F1]-[@M], and so on). A set $A\subset\R^d$ is [*discrete*]{} if it has no finite limit points, and $A$ is [*uniformly discrete*]{} if it has a strictly positive separating constant $$\eta(A):=\inf\{|x-x'|:\,x,\,x'\in A,\,x\neq x'\}.$$ Complex Radon measure or distribution is discrete (uniformly discrete) if its support is discrete (uniformly discrete). we will call the support of Fourier transform of a tempered distribution $f$ spectrum of $f$ as well. Following [@L2], we will say that a discrete set $A\subset\R^p$ is [*a set of finite type*]{}, if the set $$A-A=\{x-x':\,x,x'\in A\}$$ is discrete. A set $L\subset\R^d$ is called [*a full-rank lattice*]{} if $L=T\Z^d$ for some nondegenerate linear operator $T$ on $\R^d$, the lattice $L^*=(T^t)^{-1}\Z^d$ is called the [*conjugate lattice*]{} for $L$. A set $A$ is a [*pure crystal*]{} with respect to a full-rank lattice $L$ if it is a finite union of cosets of $L$. We begin with the following result of N.Lev and A.Olevskii [@LO1], [@LO2] on quasicrystals. \[TO\] Let $\mu$ be a uniformly discrete positive quasicrystal Fourier on $\R^d$ with uniformly discrete spectrum. Then the support of $\mu$ is a subset of a pure crystal with respect to a full-rank lattice $L$, and $\hat\mu$ is a subset of a pure crystal with respect to the conjugate lattice $L^*$. The same assertion is valid for an arbitrary uniformly discrete quasicrystal Fourier with the spectrum of a finite type. In the dimension $d=1$ the assertion is valid for every uniformly discrete quasicrystal Fourier with uniformly discrete spectrum. On the other hand, there are quasicrystals Fourier with discrete support and spectrum such that the above assertions do not valid. Note that for dimension $d>1$ there are non-positive quasicrystals Fourier with uniformly discrete support and spectrum such that their support is not a pure crystal ([@F2]). There is another type of results. \[TF\] Let $\mu$ be a complex measure on $\R^d$ with discrete support $\L$ of a finite type such that $\inf_{x\in\L}|\mu(x)|>0$, let the Fourier transform be a measure $\hat\mu=\sum_{y\in\G} b(y)\d_y$ with the countable $\G\subset\R^d$ such that $$\sum_{|y|<r}|b(y)|=O(r^T),\quad r\to\infty,$$ with some $T<\infty$. Then $\L$ is a pure crystal. Let us now go over from measures to distributions. An analog of Theorem \[TO\] for tempered distributions was proved by V.Palamodov [@P]. Let $f\in S^*(\R^d)$ be such that its support $\L$ and spectrum $\G$ are discrete sets of a finite type and, moreover, one of the differences $\L-\L$ and $\G-\G$ is uniformly discrete. Then $\L$ is a pure crystal with respect to a lattice $L$ and $\G$ is a pure crystal with respect to the conjugate lattice $L^*$. In the present paper we obtain two analogs of Theorem \[TF\] for tempered distributions. the main results ================ By [@Ru], every distribution $f$ with discrete support $\L$ has the form $$\label{r} f=\sum_{\l\in\L} P_\l(D)\d_\l, \quad P_\l(x)=\sum_{\|k\|\le K_\l}p_{\l,k}x^k, \quad x\in\R^d,\ p_{\l,k}\in\C,\ K_\l<\infty.$$ Here $\d_y$ means, as usual, the unit mass at the point $y\in\R^d$ and $x^k=x_1^{k_1}\cdots x_d^{k_d}$. Moreover, $\operatorname{ord}f=\sup_\l \deg P_\l<\infty$ (see Proposition \[P1\] below). Therefore we will consider distributions $$\label{r1} f=\sum_{\l\in\L}\sum_{\|k\|\le m}p_{\l,k}D^k\d_\l,\quad k\in(\N\cup\{0\})^d.$$ If the Fourier transform $\hat f$ has a discrete support $\G$, we also have $$\label{r2} \hat f=\sum_{\g\in\G}\sum_{\|j\|\le m'}q_{\g,j}D^j\d_\g,\quad j\in(\N\cup\{0\})^d.$$ We will suppose that $m=\operatorname{ord}f$ and $m'=\operatorname{ord}\hat f$. Also, we will consider the case of distributions $f$ and $\hat f$ of forms (\[r1\]) and (\[r2\]) with arbitrary countable $\L$ and $\G$. If this is the case, we will also say that $\L$ is support and $\G$ is spectrum of $f$. Denote by $B(x,r)$ the ball in $\R^d$ of radius $r$ with the center in $x$, $B(r)=B(0,r)$, by $\#A$ denote a number of elements of a discrete set $A$, and put $n_A(r)=\#(A\cap B(r))$. We say that discrete $A$ is of [*bounded density*]{} if $$\sup_{x\in\R^d}\#(A\cap B(x,1))<\infty.$$ Clearly, every uniformly discrete set $A$ is of bounded density, and every set $A$ of bounded density satisfies the condition $$n_A(r)=O(r^d),\quad r\to\infty.$$ A set $A\subset\R^d$ is [*relatively dense*]{} if there is $R<\infty$ such that every ball of radius $R$ intersects with $A$. Also, for any $f$ of the form (\[r1\]) set $$\kappa_f(\l)=\sup_{\|k\|\le m}|p_{\l,k}|,\qquad \rho_f(r)=\sum_{|\l|<r}\kappa_{f}(\l).$$ Using properties of almost periodic measures and sets, we prove the following theorems in Section 4. \[T1\] Let $f_1,f_2$ be tempered distributions on $\R^d$ with discrete supports $\L_1,\,\L_2$, respectively, such that $\L_1-\L_2$ is a discrete set and $$\label{k1} \inf_{\l\in\L_j}\kappa_{f_j}(\l)>0,\quad j=1,2.$$ If $\hat f_1,\,\hat f_2$ are both measures with countable supports such that $$\rho_{\hat f_1}(r)+\rho_{\hat f_2}(r)=O(r^T),\quad r\to\infty,$$ with some $T<\infty$, then $\L_1,\L_2$ are pure crystals with respect to a unique full-rank lattice. \[C1\] Let $f$ be a tempered distribution on $\R^d$ with discrete support $\L$ of a finite type such that $\inf_{\l\in\L}\kappa_{f}(\l)>0$. If $\hat f$ is a measure with countable support such that $$\rho_{\hat f}(r)=O(r^T),\quad r\to\infty$$ with some $T<\infty$, then $\L$ is a pure crystal. \[T2\] Let $f_1,f_2$ be tempered distributions on $\R^d$ with discrete relatively dense supports $\L_1,\,\L_2$ and discrete spectrums $\G_1,\,\G_2$, let $\L_1-\L_2$ be a discrete set, and let $$n_{\G_1}(r)+n_{\G_2}(r)=O(r^T),\quad r\to\infty$$ with some $T<\infty$. If conditions (\[k1\]) and $$\label{k2} \sup_{\l\in\L_j}\kappa_{f_j}(\l)<\infty,\quad j=1,2,$$ satisfy, then $\L_1,\L_2$ are pure crystals with respect to a unique full-rank lattice. \[C2\] Let $f$ be tempered distribution on $\R^d$ with discrete support $\L$ of a finite type and discrete spectrum $\G$ such that $$n_\G(r)=O(r^T),\quad r\to\infty$$ with some $T<\infty$. If $$0<\inf_{\l\in\L}\kappa_{f}(\l)\le\sup_{\l\in\L}\kappa_{f}(\l)<\infty,$$ then $\L$ is a pure crystal. preliminary properties of distributions with discrete supports ============================================================== \[P1\] Suppose $f\in S^*(\R^d)$ has form (\[r\]) with discrete $\L$ and satisfies (\[d\]) for some $n,m$. Then $$\sup_{\l\in\L}\deg P_\l\le m,$$ and there exists $C<\infty$ such that $$\label{p} |p_{\l,k}|\le C\max\{1,|\l|^n\} \mbox{ for all $\l\in\L$ and $k$ such that }\|k\|=m.$$ If $\L$ is uniformly discrete, then for all $k$, $\|k\|\le m$, $$\label{e} |p_{\l,k}|\le C\max\{1,|\l|^n\}.$$ The second part of the Proposition was earlier proved by V.Palamodov [@P]. [**Proof of Proposition \[P1\]**]{}. Set $\l\in\L$ and $\e\in(0,1)$ such that $$\inf\{|\l-\l'|:\,\l'\in\L,\,\l'\neq\l\}>\e.$$ Let $\p$ be a function on $\R$ such that $$\p(|x|)\in C^\infty(\R^d),\quad\p(|x|)=0\mbox{ for }|x|>1/2,\quad \p(|x|)=1\mbox{ for }|x|<1/3.$$ Put $$\p_{\l,k,\e}(x)=\frac{(x-\l)^k}{k!}\p(|x-\l|/\e)\in S(\R^d),$$ where, as usual, $k!=k_1!\cdots k_d!$. It is easily shown that $$f(\p_{\l,k,\e})=(-1)^{\|k\|}p_{\l,k}.$$ On the other hand, we get for some $c(\a,\b)<\infty$, where $\a,\b\in(\N\cup\{0\})^d$, $$|f(\p_{\l,k,\e})|\le\sup_{|x-\l|<\e}\max\{1,|x|^n\}\sum_{\|\a+\b\|\le m} c(\a,\b)\left|D^\a\p\left(\frac{|x-\l|}{\e}\right)D^\b\left(\frac{(x-\l)^k}{k!}\right)\right|.$$ Note that $$|D^\a\p(|x-\l|/\e)|\le \e^{-\|\a\|}c(\a)$$ and $$D^\b(x-\l)^k=\left\{\begin{array}{l} 0\mbox{ if }k_j<\b_j\mbox{ for at least one }j,\\c(k,\b)(x-\l)^{k-\b}\mbox{ if }k_j\ge\b_j\mbox{ for all }j . \end{array}\right.$$ Since $|x-\l|<1$, we get $$\max\{1,|x|^n\}\le 2^n\max\{1,|\l|^n\}.$$ Taking into account that $$\operatorname{supp}\p(|x-\l|/\e)\subset B(\l,\e),$$ we get $$\label{p1} |p_{\l,k}|\le\sum_{\|\a+\b\|\le m,\,\b_j\le k_j\,\forall j} c(k,\a,\b)\max\{1,|\l|^n\}\e^{\|k\|-\|\a+\b\|}.$$ Note that coefficients $c(k,\a,\b)$ do not depend on $\l$ and $\e$. If $\|k\|>m$, we take $\e\to0$ and obtain $p_{\l,k}=0$, hence $\deg P_\l\le m$ for all $\l$. If $\|k\|=m$, we obtain $|p_{\l,k}|\le C\max\{1,|\l|^n\}$. If $\L$ is uniformly discrete, we pick $$\e=\e_0<\min\{1,\,\eta(\L)/2\},$$ and (\[p1\]) implies the estimate $$|p_{\l,k}|\le C\e_0^{-m}\max\{1,|\l|^n\}$$ for all $k$. [**Remark**]{}. There are tempered distributions with discrete support such that conditions (\[e\]) do not valid for $\|k\|<m$. Set $$\l_j=j,\ \l'_j=j+2^{-2j},\ \L=\{\l_j,\,\l'_j\}_{j\in\N},\ f=\sum_{j\in\N}2^j(\d_{\l'_j}-\d_{\l_j}).$$ For any $\psi\in S(\R)$ we get $$f(\psi)=\sum_{j\in\N}2^j(\psi(j+2^{-2j})-\psi(j))=\sum_{j\in\N}2^{-j}\psi'(j+\t_j2^{-2j})$$ with $\t_j\in (0,1)$. Therefore, $$|f(\psi)|\le\sup_{x\in\R}|\psi'(x)|,$$ and $f\in S^*(\R)$. Since $$p_{\l'_j,0}=-p_{\l_j,0}=2^j,$$ we see that estimate (\[e\]) does not valid for $k=0$. \[P2\] Suppose $f\in S^*(\R^d)$ satisfies (\[d\]) with some $n,\,m$, $f$ has form (\[r1\]) with countable $\L$, and $\hat f$ has form (\[r2\]) with discrete $\G$. Then $$n\ge d,\quad \operatorname{ord}\hat f\le n-d,\quad |q_{\g,j}|\le C'\max\{1,|\g|^m\}\mbox{ for }\|j\|=n-d.$$ For the case of uniformly discrete $\G$ we get $$|q_{\g,j}|\le C\max\{1,|\g|^m\}\quad\forall j,\ \|j\|\le n-d.$$ \[C3\] Suppose $f\in S^*(\R^d)$ satisfies (\[d\]) with $n=d$ and some $m$, has form (\[r1\]) with countable $\L$, and spectrum $\G$ of $f$ is discrete. Then $\hat f$ is a measure, and $$\hat f=\sum_{\g\in\G}q_\g\d_\g,\quad |q_\g|\le C'\max\{1,|\g|\}^m.$$ [**Proof of Proposition \[P2\]**]{}. Set $\g\in\G$ and pick $\e\in(0,1)$ such that $$\inf\{|\g-\g'|:\,\g'\in\G,\,\g'\neq\g\}>\e.$$ Let $\p$ be the same as in the proof of Proposition \[P1\]. Put $$\p_{\g,l,\e}(y)=\frac{(y-\g)^l}{l!}\p(|y-\g|/\e)\in S(\R^d).$$ We have $$\label{q} (-1)^{\|l\|}q_{\g,l}=\sum_{\|j\|\le m'}q_{\g,j}D^j\d_\g(\p_{\g,l,\e}(y))=(\hat f,\p_{\g,l,\e})=(f,\hat\p_{\g,l,\e}).$$ Note that $$\hat\p_{\g,l,\e}(x)=e^{-2\pi i\la x,\g\ra}(l!)^{-1}(-2\pi i)^{-\|l\|}D^l(\widehat{\p(\cdot/\e)})=c(l)e^{-2\pi i\la x,\g\ra}\e^{d+\|l\|}(D^l\hat\p)(\e x).$$ Therefore, $$D^k(\hat\p_{\g,l,\e}(x))=\e^{d+\|l\|}\sum_{\a+\b=k}c(\a,\b)D^\a\left[e^{-2\pi i\la x,\g\ra}\right]D^\b[(D^l\hat\p)(\e x)]$$ $$=\sum_{\a+\b=k}c(\a,\b)(-2\pi i)^{\|\a\|}\g^\a e^{-2\pi i\la x,\g\ra}\e^{d+\|l\|+\|\b\|}(D^{\b+l}\hat\p)(\e x).$$ Since $\hat\p\in S(\R^d)$, we get for every $k$, $\|k\|\le m$, and every $M<\infty$ $$\label{D} |D^k(\hat\p_{\g,l,\e}(x))|\le C(M)\e^{d+\|l\|}\max\{1,|\g|^m\}(\max\{1,|\e x|^M\})^{-1}.$$ Pick $M=n$. By (\[d\]), we have $$|(f,\hat\p_{\g,l,\e})|\le c(f)\sup_{\R^d}[\max\{1,|x|^n\}\max_{\|k\|\le m} |D^k(\hat\p_{\g,l,\e}(x))|].$$ Hence, $$|q_{\g,l}|=|(f,\hat\p_{\g,l,\e})|\le C'(\hat\p)\e^{d+\|l\|}\max\{1,|\g|^m\}\sup_{x\in\R^d}[\max\{1,|x|^n\} (\max\{1,|\e x|^n\})^{-1}].$$ Since $\sup$ in the right-hand side of the inequality equals $\e^{-n}$, we obtain $$|q_{\g,l}|\le C'\max\{1,|\g|^m\}\e^{\|l\|+d-n}.$$ If $\|l\|>n-d$, we take $\e\to0$ and get $q_{\g,l}=0$, hence $\operatorname{ord}\hat f\le n-d$. For $\|l\|=n-d$ we get $|q_{\g,l}|\le C'\max\{1,|\g|^m\}$. If $\G$ is uniformly discrete, we take $\e=\e_0<\eta(\G)/2$ for all $\g\in\G$ and obtain the bound $$|q_{\g,l}|\le \e_0^{d-n} C'\max\{1,|\g|^m\}\quad\forall l,\ \|l\|\le n-d.\qquad\qquad\qquad \bs$$ . Since $$N_{n,m}(\hat\p)\le C(n,m) N_{m,n}(\p)$$ for every $\p\in S(\R^d)$, we see that estimate (\[d\]) for $f\in S^*(\R^d)$ implies the estimate $|\hat f(\p)|\le c'N_{m,n}(\p)$. Therefore the direct application of Proposition \[P1\] gives only the bound $\operatorname{ord}\hat f\le n$. M.Kolountzakis, J.Lagarias proved in [@KL] that Fourier transform of every measure $\mu$ on the line $\R$ with support of bounded density, bounded masses $\mu(x)$, and discrete spectrum is also a measure $\hat\mu=\sum_{\g\in\G}q_\g\d_\g$ with uniformly bounded $q_\g$. The following proposition generalizes this result for distributions from $S^*(\R^d)$. \[P3\] Suppose $f\in S^*(\R^d)$ has form (\[r1\]) with some $m$ and countable $\L$, and discrete spectrum $\G$. If $$\rho_f(r)=O(r^{d+H}),\quad r\to\infty,\quad H\ge0,$$ then $\operatorname{ord}\hat f\le H$. If, in addition, $H$ is integer, then for $\|j\|=H$ we get $|q_{\g,j}|\le C'\max\{1,|\g|^m\}$. For the case of uniformly discrete $\G$ we get $$|q_{\g,j}|\le C\max\{1,|\g|^m\}\quad\forall j,\quad \|j\|\le H.$$ \[C4\] If $f\in S^*(\R^d)$ has form (\[r1\]) with some $m$ and countable $\L$, discrete spectrum $\G$, and $$\rho_f(r)=O(r^d)\quad r\to\infty,$$ then $\hat f$ is a measure, and $$\hat f=\sum_{\g\in\G}q_\g\d_\g,\ |q_\g|\le C'\max\{1,|\g|^m\}.$$ \[C4a\] If $f\in S^*(\R^d)$ has form (\[r1\]) with some $m$, bounded $\kappa_f(\l)$, discrete support $\L$ of bounded density, and discrete spectrum $\G$, then $\hat f$ is a measure, and $$\hat f=\sum_{\g\in\G}q_\g\d_\g,\ |q_\g|\le C'\max\{1,|\g|^m\}.$$ [**Proof of Proposition \[P3\]**]{}. Let $\p_{\g,l,\e}$ be the same as in the proof of Proposition \[P2\]. By (\[q\]) and (\[r1\]), $$(-1)^{\|l\|}q_{\g,l}=(f,\hat\p_{\g,l,\e})=\sum_{\l\in\L}\sum_{\|k\|\le m}p_{\l,k}(-1)^{\|k\|}D^k(\hat\p_{\g,l,\e}(\l)).$$ Using (\[D\]), we get $$|q_{\g,l}|\le C'(m,M)\e^{d+\|l\|}\max\{1,|\g|^m\}\sum_{\l\in\L}\kappa_f(\l)(\max\{1,|\e\l|^M\})^{-1}.$$ We have $$\sum_{\l\in\L}\kappa_f(\l)(\max\{1,|\e\l|^M\})^{-1}=\rho_f(1/\e)+\e^{-M}\int_{1/\e}^\infty t^{-M}d\rho_f(t)$$ Pick $M>d+H$. Integrating by parts and using the estimate for $\rho_f(r)$, we see that the right-hand side is equal to $$O(\e^{-d-H})+\e^{-M}M\int_{1/\e}^\infty \rho_f(t)t^{-M-1}dt=O(\e^{-d-H})\quad\hbox{ as }\e\to\infty.$$ Finally, $$|q_{\g,l}|\le C'\max\{1,|\g|^m\}\e^{\|l\|-H}.$$ If $\|l\|>H$, we take $\e\to0$ and get $q_{\g,l}=0$, hence, $m'=\operatorname{ord}\hat f\le H$. If $H$ is integer, we get $|q_{\g,l}|\le C'\max\{1,|\g|^m\}$ for $\|l\|=H$. If $\G$ is uniformly discrete, we take $\e=\e_0<\eta(\G)/2$ for all $\g\in\G$ and obtain the bound $$|q_{\g,l}|\le \e_0^{-H} C'\max\{1,|\g|^m\}\quad \forall l, \|l\|\le m'.\qquad\qquad\qquad \bs$$ almost periodic distributions and proofs of Theorems \[T1\] and \[T2\] ====================================================================== Recall that a continuous function $g$ on $\R^d$ is almost periodic if for any $\e>0$ the set of $\e$-almost periods of $g$ $$\{\tau\in\R^d:\,\sup_{x\in\R^d}|g(x+\tau)-g(x)|<\e\}$$ is a relatively dense set in $\R^d$. Almost periodic functions are uniformly bounded on $\R^d$. The class of almost periodic functions is closed with respecting to taken absolute values, linear combinations, maximum, minimum of a finite family of almost periodic functions. A limit of a uniformly in $\R^d$ convergent sequence of almost periodic functions is also almost periodic. A typical example of an almost periodic function is an absolutely convergence exponential sum $\sum c_n\exp\{2\pi i\la x,\o_n\ra\}$ with $\o_n\in\R^d$ (see, for example, [@C]). A measure $\mu$ on $\R^d$ is called almost periodic if the function $$(\psi\star\mu)(t)=\int_{\R^d}\psi(x-t)d\mu(x)$$ is almost periodic in $t\in\R^d$ for each continuous function $\psi$ on $\R^d$ with compact support. A distribution $f\in S^*(\R)$ is almost periodic if the function $(\psi\star f)(t)=f(\psi(\cdot-t))$ is almost periodic in $t\in\R^d$ for each $\psi\in S(\R^d)$ (see [@A],[@M],[@R]). Clearly, every almost periodic distribution has a relatively dense support. \[T6\] Let $f\in S^*(\R^d)$ and $\hat f$ be a measure $\sum_{\g\in\G}q_\g\d_\g$ with countable $\G$ such that $$\label{ro} \rho_{\hat f}(r)=O(r^T),\quad r\to\infty$$ for some $T<\infty$. Then $f$ is an almost periodic distribution. [**Proof of Theorem \[T6\]**]{}. Let $\psi\in S(\R^d)$. The converse Fourier transform of the function $\psi(x-t)$ is $\hat\psi(-y)e^{2\pi i\la y,t\ra}$. Therefore, $$\label{s} f(\psi(\cdot-t))=\hat f(\hat\psi(-y)e^{2\pi i\la y,t\ra})=\sum_{\g\in\G}q_\g\hat\psi(-\g)e^{2\pi i\la\g,t\ra}.$$ Taking into account that $\hat\psi(-y)\in S(\R^d)$, we see that the sum in (\[s\]) is majorized by the sum $$\sum_{\g\in\G}C(\psi,T)|q_\g|\max\{1,|\g|\}^{-T-1}=C(\psi,T)\left[\rho_{\hat f}(1)+\int_1^\infty r^{-T-1}d\rho_{\hat f}(r)\right].$$ Integrating by parts, we obtain that the integral converges, then the sum in (\[s\]) converges absolutely, and $\psi\star f$ is almost periodic. Combining this theorem with the results of the previous section, we obtain \[T7\] Suppose $f\in S^*(\R^d)$ has form (\[r1\]) with countable $\L$, has discrete spectrum $\G$ such that $n_\G(r)=O(r^T)$ as $r\to\infty$ for some $T<\infty$, and satisfies one of the following conditions i\) inequality (\[d\]) holds with $n=d$ and some $m$; ii\) $\rho_f(r)=O(r^d)$ as $r\to\infty$; iii\) $\kappa_f(\l)$ is bounded and support $\L$ is of bounded density. Then $f$ is an almost periodic distribution. [**Proof of Theorem \[T7\]**]{}. Using Corollaries \[C3\], or \[C4\], or \[C4a\], we obtain that $\hat f$ is a measure, $$\hat f=\sum_{\g\in\G}q_\g\d_\g,\qquad |q_\g|\le C'\max\{1,|\g|\}^m.$$ Therefore, integrating by parts, we get $$\rho_{\hat f}(r)\le C'\left[\rho_{\hat f}(1)+\int_1^r t^m dn_\G(t)\right]=O(r^{m+T}),\quad r\to\infty.$$ Then we apply Theorem \[T6\] completes the proof. Now we can proof the main results of our paper. [**Proof of Theorem \[T1\]**]{}. By Theorem \[T6\], $f_j,\,j=1,2,$ are almost periodic distributions. In particular, $\L_1$ is relatively dense. If there are $\l_n, \l'_n\in\L_2$ such that $\l_n\neq\l'_n$ and $\l_n-\l'_n\to0$ as $n\to\infty$, then there are $R<\infty$ and $x_n\in\L_1$ such that $\l_n, \l'_n\in B(x_n,R)$. Therefore there exists infinitely many points of the set $\L_1-\L_2$ in the ball $B(R)$, that is impossible. Hence $\eta(\L_2)>0$ and, similarly, $\eta(\L_1)>0$. By Proposition \[P1\], $\operatorname{ord}f_j<\infty,\,j=1,2$. We apply Proposition \[P3\] to the measure $\tilde\mu(y)=\hat f_1(-y)$ with $m=0$ and $H=T-d$ (we may suppose that $T\ge d$). Since the Fourier transform of $\tilde\mu(y)$ is just the tempered distribution $f_1$ with uniformly discrete support, we get representation (\[r1\]) for $f_1$ with $$\sup_{\l\in\L_1}\sup_{\|k\|\le\operatorname{ord}f_1}|p_{\l,k}|<\infty.$$ Put $s=\inf_{\l\in\L_1}\kappa_{f_1}(\l)$, and pick a number $\e\in(0,\eta(\L_1)/5)$ such that $$\sum_{\|k\|\le\operatorname{ord}f_1}|p_{\l,k}|<s/(2\e)\quad \forall \l\in\L_1.$$ Let $\p_\e(x)$ be nonnegative $C^\infty$-function such that $$\operatorname{supp}\p_\e\subset B(\e+\e^2),\quad \p_\e(x)=1 \mbox{ for }|x|<\e.$$ Put $$\p_{k,\e}(x)=\p_\e(x)x^k/k!,\quad k\in(\N\cup\{0\})^d.$$ Since $f_1$ is almost periodic, we see that the function $g_{k,\e}=f_1\star\p_{k,\e}$ is almost periodic. Fix $\l\in\L_1$. By (\[k1\]), there exists $k'$ such that $|p_{\l,k'}|\ge s$. Put $$A=\{k\in(\N\cup\{0\})^d:\,k_j\le k_j'\ \forall j=1,\dots,d\}.$$ For $x\in B(\l,\e)$ we have $$g_{k',\e}(x)=\left(\sum_{\|k\|\le\operatorname{ord}f_1}p_{\l,k}D^k\d_\l\right)\frac{(\l-x)^{k'}}{{k'}!}=\sum_{k\in A}p_{\l,k}(-1)^{\|k\|}(\l-x)^{k'-k}/(k'-k)!$$ Therefore, $$|g_{k',\e}(x)|\ge |p_{\l,k'}|-\e\sum_{k\in A,k\neq k'}|p_{\l,k}|>s/2.$$ Now set $$h_\e(x)=\min\{1,2s^{-1}\sup_{\|k\|\le\operatorname{ord}f_1}|g_{k,\e}(x)|\}.$$ Clearly, $h_\e(x)$ is an almost periodic function and $$h_\e(x)=1\mbox{ for } x\in\bigcup_{\l\in\L_1}B(\l,\e),\qquad \operatorname{supp}h_\e\subset\bigcup_{\l\in\L_1}B(\l,\e+\e^2),\qquad 0\le h_\e(x)\le1.$$ Let $\psi$ be an arbitrary continuous function on $\R^d$ with support in the ball $B(\eta(\L_1)/5)$. It is readily seen that the function $\psi\star h_\e$ is almost periodic as well. Since $\e<\eta(\L_1)/5$, we see that for each fixed $t\in\R^d$ the support of the function $\psi(\cdot-t)$ intersects with at most one ball $B(\l,\e+\e^2)$. Therefore we get $$\begin{aligned} \label{h1} \left|\frac{\psi\star h_\e(t)}{\o_d\e^d}-\frac{1}{\o_d\e^d}\int_{B(\l,\e)}h_\e(x)\psi(x-t)dx\right| \le \left|\frac{1}{\o_d\e^d}\int_{B(\l,\e+\e^2)\setminus B(\l,\e)}h_\e(x)\psi(x-t)dx\right|\nonumber \\ \le\frac{\o_d(\e+\e^2)^d-\o_d\e^d}{\o_d\e^d}\sup_{\R^d}|\psi(x)|,\qquad\qquad\qquad\qquad\qquad\qquad, \end{aligned}$$ where $\o_d$ is a volume of the unit ball in $\R^d$, and $$\label{h2} \left|\frac{1}{\o_d\e^d}\int_{B(\l,\e)}h_\e(x)\psi(x-t)dx-\psi(\l-t)\right|\le\sup_{x\in B(\l,\e)}|\psi(x-t)-\psi(\l-t)|.$$ It follows from (\[h1\]) and (\[h2\]) that the almost periodic functions $(\o_d\e^d)^{-1}(\psi\star h_\e)$ uniformly converge as $\e\to0$ to the function $\psi\star\d_{\L_1}$, where $\d_{\L_1}=\sum_{\l\in\L_1}\d_{\l}$. Hence the function $\psi\star\d_{\L_1}$ is almost periodic. It is readily seen that the same is true for every continuous function $\psi$ on $\R^d$ with compact support. The similar construction works for $\d_{\L_2}=\sum_{\l\in\L_2}\d_{\l}$. Thus $\d_{\L_1}, \d_{\L_2}$ are almost periodic measures with discrete set of differences $\L_1-\L_2$. Applying Theorem 6 from [@F2], we obtain that $\L_1,\L_2$ are pure crystals with respect to a unique full-rank lattice. . Since the sets $\L_j,\,j=1,2$, are relatively dense, we get, arguing as before, the bounds $\eta(\L_j)>0$ for $j=1,2$. Hence, both $\L_j$ are of bounded density. By Corollary \[C4a\], $\hat f_j$ are measures such that $$\kappa_{\hat f_j}(\l)=O(|\l|^m),\quad |\l|\to\infty,$$ for some $m<\infty$. Using Theorem \[T7\] iii), we obtain that both $f_j$ are almost periodic distributions, hence we can repeat the proof of Theorem \[T1\]. [**Remark**]{}. It is clear that the bound $\eta(\L)>0$ follows immediately from the conditions of Corollaries \[C1\] or \[C2\]. [8]{} J.G.de Lamadrid, L.N.Argabright, Almost Periodic Measures. Memoirs of the AMS, No.428, Providence RI, (1990), 218p. Directions in Mathematical Quasicrystals, M. Baake and R. Moody, eds., CRM Monograph series, [**13**]{}, AMS, Providence RI, 379p. C. Corduneanu, Almost Periodic Functions, Second English ed.) Chelsea, New-York, 1989 (Distributed by AMS and Oxford University Press). S.Yu. Favorov. [*Bohr and Besicovitch almost periodic discrete sets and quasicrystals*]{}, Proc. Amer. Math. Soc. 140 (2012), 1761–1767. S.Yu. Favorov. [*Fourier quasicrystals and Lagarias’ conjecture*]{}, Proc. Amer. Math. Soc. 144 (2016) , 3527-3536. S.Yu. Favorov. [*Large Fourier quasicrystals and Wiener’s Theorem*]{}, Journal of Fourier Analysis and Applications, DOI 10.1007/s00041-017-9576-0 Lev, N., Olevskii, A.: [*Measures with uniformly discrete support and spectrum*]{}. C.R.Acad.Sci.,ser.1 [**351**]{}, 599-603 (2013) Lev, N., Olevskii, A.: [*Quasicrystals and Poisson’s summation formula. Invent.Math.*]{} [**200**]{}, 585–606 (2015) Lev, N., Olevskii, A.: [*Fourier quasicrystals and discreteness of the diffraction spectrum*]{}. Advances in Mathematics, [**315**]{}, 1-26 (2017) M.N. Kolountzakis. [*On the structure of multiple translations tilings by polygonal regions*]{}, Preprint, 1999, 16p. M.N. Kolountzakis. [*Fourier Pairs of Discrete Support with Little Structure*]{} February 2015 Journal of Fourier Analysis and Applications 22(1) DOI10.1007/s00041-015-9416-z M.N. Kolountzakis, J.C. Lagarias, [*Structure of Tilings of the Line by a Function*]{}, Duke Math.Journal, 82, (1996), 653-678. J.C. Lagarias, [*Mathematical quasicrystals and the problem of diffraction*]{}, in [@D], 61-93. J.C. Lagarias, [*Geometric models for quasicrystals I.Delone set of finite type*]{}, Discr.and Comp.Geometry 1998. Meyer, Y.: Quasicrystals, Almost Periodic Patterns, Mean–periodic Functions, and Irregular Sampling. African Diaspora Journal of Mathematics, [**13**]{} (1), 1–45 (2012) V.P. Palamodov, [*A geometric characterization of a class of Poisson type distributions*]{}, Journal of Fourier Analysis and Applications (2017). Quasicrystals and Discrete Geometry. J.Patera,ed., Fields Institute Monographs, AMS, Providence RI, 289p. L.I.Ronkin [*Almost periodic distributions and divisors in tube domains,*]{} Zap. Nauchn. Sem. POMI [**247**]{} (1997). P. 210–236 (Russian). W. Rudin [*Functional Analysis*]{}, McGraw-Hill Book Company, New York, St.Louis, Sun Francisco, (1973), 443p. V.S. Vladimirov [*Equations of Mathematical Physics*]{}, Marcel Dekker, Inc.,New-York, 1971, 418p.

ArXiv

--- abstract: 'The lattice thermal conductivity of crystalline Si nanowires is calculated. The calculation uses complete phonon dispersions, and does not require any externally imposed frequency cutoffs. No adjustment to nanowire thermal conductivity measurements is required. Good agreement with experimental results for nanowires wider than 35 nm is obtained. A formulation in terms of the transmission function is given. Also, the use of a simpler, nondispersive “Callaway formula”, is discussed from the complete dispersions perspective.' author: - 'N. Mingo' title: Calculation of nanowire thermal conductivity using complete phonon dispersion relations --- Determining the thermal conductivity of semiconducting nanowires plays a crucial role in the development of a new generation of thermoelectric materials [@Mahan]. However, experimental results are still scarce [@Fon]. The first measurements of the thermal conductivity of Si nanowires in the $\sim 20-115 nm$ range have been recently reported in ref.. From a theoretical point of view, it is important to be able to quantitatively calculate lattice thermal conductivities of nanowires in a predictive fashion. In the past, several techniques have become widespread in the calculation of thermal conductivities of bulk materials [@Callaway; @Holland; @Parrott; @Berman]. These works are based on [*linearized dispersion*]{} models that involve a certain number of adjustable parameters. In principle, one might hope that a proper account of the boundary scattering in the nanowires might enable us to predict nanowire thermal conductivities using the procedures of ref.  or ref. , without any parameters other than those obtained from fitting the bulk data. Nevertheless, calculations using those two approaches did not yield reasonable results for nanowires [@APL]. A first goal of this paper is to show that, using accurate phonon dispersion relations, [*predictive*]{} calculations of the nanowire thermal conductivities can be done, which [*do not involve any adjustable parameter to fit the nanowire experimental curves*]{}. A second result in this paper shows that the use of a “modified” Callaway approach is still possible for nanowires, although this second approach is not predictive, and requires adjustment to nanowire measurements. For a nanowire suspended between two thermal reservoirs, its thermal conductance, $\sigma(T)$, is defined as the flow of heat current through the wire when a small temperature difference $\Delta T$ exists between the reservoirs, divided by $\Delta T$ and taking the limit $\Delta T \rightarrow 0$. If the phonons transmit ballistically, the conductance is $$\begin{aligned} \sigma_{b}(T)=\sum_{\alpha}\int_0^{\pi/a_z} {dk_z\over 2\pi}[\hbar \omega v_z(\alpha,k_z) {df_{B}\over dT}]\end{aligned}$$ where $\alpha$ is a set of discrete quantum numbers labeling the particular subbands in the one dimensional phonon dispersion relations. $f_B$ is the Bose distribution, and $v_z$ is the speed of the phonon in the axial direction of the nanowire, $z$. Using $v_z dk_z=d\omega$ we have [@Rego] $$\begin{aligned} \sigma_{b}(T)=\sum_{\alpha}\int_{\omega^i_{\alpha}}^{\omega^f_{\alpha}} {d\omega\over 2\pi}[\hbar \omega {df_{B}\over dT}] \equiv \int_0^{\infty} \Xi_{b}(\omega){\hbar \omega \over 2\pi}{df_{B}\over dT}d\omega \label{Eq:conductance}\end{aligned}$$ where $\omega^{i(f)}_{\alpha}$ is the lower (upper) frequency limit of subband $\alpha$. Here we have defined the ballistic transmission function, $\Xi_b(\omega)$, which simply corresponds to the number of phonon subbands crossing frequency $\omega$. However, in general, phonon scattering takes place in the wire and at the contacts, so that the transmission function is not given by the simple mode count stated above. In general, the transmission function can be obtained from the Green function of the system. This allows to study the interesting issues of ballistic versus diffusive transport [@deJong], or the effect of one dimensional localization.[@Beenakker] We shall not be concerned with those phenomena here, but shall assume that diffusive transport takes place and the Boltzmann transport equation is valid. From this standpoint, it has been shown that a mode’s transmission can be described in terms of a characteristic relaxation length in the wire’s direction, $\lambda_{\alpha}(\omega)$, such that $\sigma_{diff}(T)\simeq \sum_{\alpha}\int {\lambda_{\alpha}(k_z)\over L} {\hbar \omega_{\alpha}(k_z) \over 2\pi}{df_{B}\over dT}v_z(\alpha,k_z)d k_z$ where $L$ is the nanowire’s length [@Klemens]. The thermal conductivity is defined as $\kappa(T)={L\over s}\sigma$, where $s$ is the cross section of the nanowire. Therefore, in this diffusive regime $$\begin{aligned} \kappa(T)={1\over s}\sum_{\alpha} \int_0^{\pi/a_z} \lambda_{\alpha}(k_z) {\hbar \omega_{\alpha}(k_z) \over 2\pi}{df_{B}\over dT}v_z(\alpha,k_z)d k_z.\end{aligned}$$ To obtain the thermal conductivity we must compute the complete dispersion relations for the wire, $\omega_{\alpha}(k_z)$, which, using $v_z(\alpha,k_z)={d\omega_{\alpha}(k_z)\over dk_z}$, allows us to express the previous equation as $$\begin{aligned} \label{Eq:thcond1} \kappa(T)={1\over s}\int_0^{\infty} (\sum_{\alpha} \lambda_{\alpha}(\omega)) {\hbar \omega \over 2\pi}{df_{B}\over dT}d\omega,\\ \label{Eq:lambdas} \hbox{with }\ \lambda_{\alpha}(\omega) \equiv \left\{ \matrix{ \lambda_{\alpha}(k_z(\alpha,\omega)), & \omega_{\alpha}^i < \omega < \omega_{\alpha}^f \cr 0, & \hbox{otherwise.} \cr} \right.\end{aligned}$$ If the modes’ lifetimes $\tau(\omega)$ are independent of the subband, and only depend on the frequency, then, the relaxation lengths are given by $$\begin{aligned} \lambda_{\alpha}(\omega)=v_z(\alpha,\omega)\tau(\omega),&& \omega_{\alpha}^i < \omega < \omega_{\alpha}^f \label{Eq:lambdas2}\end{aligned}$$ Knowing $\omega_{\alpha}(k_z)$ and $\tau(\omega)$, Eqs. (\[Eq:thcond1\]-\[Eq:lambdas2\]) already allow us to compute the thermal conductivity. Nevertheless, it is useful to recast Eq. (\[Eq:thcond1\]) into an equivalent form, so as to explicitly retain the “transmission function” concept. For this, we define $N_b(\omega)$ as the number of phonon subbands crossing frequency $\omega$: [^1] $$\begin{aligned} \label{Eq:N} N_b(\omega) \equiv \sum_{\alpha}\Theta(\omega-\omega_{\alpha}^i)\Theta(\omega_{\alpha}^f-\omega),\\ \Theta(x) = \left\{ \matrix{ 1 & x \ge 0 \cr 0 & x < 0. \cr} \right.\end{aligned}$$ The average group velocity in the axial direction is defined as $$\begin{aligned} \langle v_z(\omega)\rangle \equiv \left (\sum_{\alpha} v_z(\alpha,\omega)\right )/N_b(\omega) \label{Eq:vave}\end{aligned}$$ where the sum extends to subbands $\alpha$ such that $\omega_{\alpha}^i<\omega<\omega_{\alpha}^f$. From Eqs. (\[Eq:lambdas\]-\[Eq:lambdas2\]) and (\[Eq:vave\]), it follows that $\sum_{\alpha}\lambda_{\alpha}(\omega) = \tau(\omega)\sum_{\alpha} v_z(\alpha,\omega) = \tau(\omega)N_b(\omega)\langle v_z (\omega) \rangle$. Thus Eq. (\[Eq:thcond1\]) can be recast into its completely equivalent form [^2] $$\begin{aligned} \label{Eq:thcond2} \kappa(T)={L\over s}\int_0^{\infty} \Xi(\omega) {\hbar \omega \over 2\pi}{df_{B}\over dT}d\omega\\ \hbox{with } \ \ {L\over s} \Xi(\omega)= \tilde N_b(\omega) \tau(\omega) \langle v_z(\omega)\rangle. \label{Eq:transmission}\end{aligned}$$ Here, the diffusive transmission function $\Xi$ has been defined, and $$\begin{aligned} \tilde N_b(\omega)\equiv N_b(\omega)/s.\end{aligned}$$ Computation of the transmission function requires the obtention of the full phonon dispersions for the nanowire. We now proceed to separately study each of the three factors in Eq. (\[Eq:transmission\]): $\tilde N_b(\omega)$, $\tau(\omega)$, and $\langle v_z(\omega)\rangle$. $\tilde N_b(\omega)$ is obtained from the complete dispersion relations for the wire. To compute them, we have used the interatomic potential proposed by ref. , in which the energy of the system is given as the sum of two and three body terms. The two body terms are defined as $$\delta E_0 (i,j)= \hbox{$ {1\over 2} $ } C_0 {(d_{i,j} - d_0)^2 / d_0^2}$$ for every pair of nearest neighbors $i$ and $j$, where $d_{i,j}$ is the distance between the atoms and $d_0$ is the lattice equilibrium distance. The three body terms are defined as $$\delta E_1(i,j,k)=\hbox{$ {1\over 2} $ } C_1 \delta \Theta_{i,j,k}^2$$ for every pair of bonds joining atoms $i$, $j$ and $k$, where $\delta\Theta$ is the deviation with respect to the equilibrium angle between the two bonds in the lattice. The constants are $C_0=49.1eV$ and $C_1=1.07eV$ [@Harrison]. With this potential the dynamical matrix of the system is constructed, and the dispersion relations for the wire are calculated. The inset of fig. \[Fig:trans\] shows one example. The bulk dispersions obtained with the Harrison potential are close to those experimentally measured. Although a more sophisticated calculation of the interatomic potential is possible using ab initio techniques, this would only result in minor differences in the calculated $\tilde N_b$ and the conductivities computed from it. The dispersions in general depend on the cross sectional shape of the wire, and on whether its surface is reconstructed or not, clean or coated by an overlayer, etc. As the surface/volume ratio decreases, these differences become smaller, and for wide enough wires of cross section $s$, the function $\tilde N_b(\omega)\equiv N_b(\omega)/s$ approaches a limiting form, $\tilde N_b^{\infty}(\omega)$, independent of the surface features. To ascertain whether it is valid to approximate $\tilde N_b(\omega)$ by $\tilde N_b^{\infty}(\omega)$ for nanowires wider than 35 nm, the former was calculated for nanowires of increasing width, with frozen boundary conditions. (As compared to “free boundary”, frozen boundary conditions are more similar to the experimental situation, where 1 nm of $SiO_2$ coats the wire.) For a romboidal cross section wire of 17 nm side, $\tilde N_b$ deviates from the limiting value $\tilde N_b^{\infty}$ by less than $0.1 \%$. [^3] Therefore, for the widths considered here ($l>35 nm$), $\tilde N_b=\tilde N_b^{\infty}$ is a valid assumption. In the thermal conductivity calculations presented here, $\tilde N_b^{\infty}$ for the 110 direction was used (fig. \[Fig:trans\]). ![[*Solid line:*]{} function $\tilde N_b^{\infty}(\omega)$ for the (110) direction, calculated using complete phonon dispersions. [*Dotted line:*]{} parabolic approximation to $\tilde N_b^{\infty}(\omega)$, for a non dispersive medium. [*Inset:*]{} dispersion relations for a 2.2 nm wide wire with frozen boundary. []{data-label="Fig:trans"}](fig1mingo_rev.eps){width="8."} The phonon lifetime is commonly given by Matheissen’s rule, expressing the total inverse lifetime as the sum of the inverse lifetimes corresponding to each scattering mechanism. For bulk or a macroscopically thick whisker, the expressions used for boundary, anharmonic and impurity scattering are [@Asen-Palmer]: $$\begin{aligned} \tau^{-1}_b =& (\langle 1/c \rangle_b lF)^{-1}\nonumber\\ \label{Eq:times} \tau_a^{-1} =& B T \omega^2 e^{-C/T} \\ \tau^{-1}_i =& A \omega^4\nonumber\end{aligned}$$ From here on, $\langle \rangle_b$ denotes averaging in the three acoustic branches, i.e. $\langle 1/c \rangle_b \equiv {1\over 3}\sum_{i=1}^3 1/c_i$, where $c_i$ is the speed of sound for each branch. (Do not confuse with averages over subbands, denoted by $\langle\rangle$.) A, B and C are numerical constants specified later. $l$ is the wire’s lateral dimension, and $F$ relates the boundary scattering rate to the shape and specularity of the sample’s boundary. Eqs. (\[Eq:times\]) are always valid for wire thicknesses larger than a certain threshold, $L_{3D}$. On the other hand, if the width of the whisker is very small, such that $L<<L_{3D}$, the frequency difference between consecutive phonon subbands gets large. This confinement modifies the inter-subband scattering, and the validity of Eqs. (\[Eq:times\]) breaks down. [^4] A criterion to estimate $L_{3D}$ is: confinement is important if the energy spacings between consecutive subbands, $\hbar \Delta\omega$, are larger or comparable to the thermal energy $k_B T$. The largest intersubband spacings are of the order of $\hbar\Delta\omega\sim {2\pi\over l}\hbar c$. Thus, for Eqs. \[Eq:times\] to be valid at temperatures $40^oK$ or higher, we need $l>10 nm$. This is well fulfilled for nanowires in the 37-115 nm range, as the ones experimentally measured [@Li]. Therefore, for these widths, we use Eqs. \[Eq:times\] without modification. The good agreement with experimental results corroborates their validity. For sub 20 nm widths, confinement effects may be expected. [^5] Anharmonic scattering should include the effects of umklapp as well as normal processes. In some works normal processes have been omitted in favor of umklapp scattering, without extensive discussion [@Fon; @Balandin]. In general, if resistive processes dominate, the normal relaxation time can be disregarded [@Klemens]. For the higher frequencies, umklapp scattering is proportional to $\omega^2$, and it thus dominates over normal processes, which depend linearly on frequency [@HanKlemens]. At the lower frequencies, in the case of nanowires, boundary scattering dominates. This justifies the form of $\tau_a$ in Eq. \[Eq:times\], which is the functional dependence for umklapp scattering at the higher frequencies. Constants B and C were adjusted to reproduce the [*bulk material*]{}’s experimental thermal conductivity curve, yielding $B=1.73\times 10^{-19} s/K$ and $C=137.3K$. [^6] Parameter $A=1.32\times 10^{-45}s^3$ is analytically determined from the isotope concentration, and it should not be adjusted [@Asen-Palmer], so we maintain the value given in \[\]. The third term in Eq. (\[Eq:transmission\]), $\langle v_z(\omega)\rangle$, is also calculated from the dispersion relations. In a frozen boundary wire, it is a matter of simple geometry to prove that at the lower part of the spectrum,[^7] $$\begin{aligned} \langle v_z \rangle \simeq {2\over 3}\langle 1/c \rangle_b / \langle 1/c^2 \rangle_b \label{Eq:linspeed}\end{aligned}$$ where $\langle 1/c^2 \rangle_b \equiv {1\over 3} \sum_{i=1}^3 1/c^2_i$. The relaxation times, Eq. (\[Eq:times\]) are quite rough, since they only depend on the lower spectrum speeds of sound. Thus, no additional approximation is incurred if Eq. (\[Eq:linspeed\]) is used for the third term in Eq. (\[Eq:transmission\]). Fig. \[Fig:thcond\] shows the results of the calculation using the [*complete dispersions*]{} transmission function into Eq. (\[Eq:thcond2\]). Calculations for nanowires of diameters $F l = 38.85$, $72.8$ and $132.25$ nm are shown, together with experimental results from ref. . Although the TEM measured diameters of the experimental nanowires were 37, 56 and 115 nm, variations in cross sectional shape, as well as thickness of the nanowire’s oxide surrounding layer are possible, so that the cross sectional area is not accurately known. Assuming $l$ equal to the TEM diameters, values can be assigned to $F$, which in the current case are $F\simeq 1.05$, $1.3$ and $1.15$. These values suggest that the boundary scattering is very diffusive, rather than specular. ![Thermal conductivities versus temperature calculated using the complete dispersions transmission function, for $F l=1.05\times 37 nm$ (solid), $1.3\times 56 nm$ (dotted) and $1.15\times 115 nm$ (dashed). Dots: experimental results from ref. . [*Inset:*]{} bulk. []{data-label="Fig:thcond"}](fig2mingo_revNEW.eps){width="6.5"} Let us see what would happen if, instead of the atomistically calculated complete dispersions, a nondispersive medium were used. In such a case, for frozen boundary, function $\tilde N_b^{\infty}(\omega)$ is analytical, having a parabolic dependence: $$\begin{aligned} \hbox{(non dispersive)}\ \ \tilde N_b^{\infty}(\omega) \simeq &\sum_{i=1}^3\int_0^{\omega/c_i} 2\pi k_{\perp} {dk_{\perp}\over (2\pi)^2} \nonumber \\ =&{3\over 4\pi}\omega^2 \langle 1/c^2 \rangle_b \label{Eq:NCallaway}\end{aligned}$$ Substituting Eq. (\[Eq:NCallaway\]) into Eqs. (\[Eq:thcond2\]) and (\[Eq:transmission\]), the Callaway formula [@Callaway; @Holland] is obtained. Since now the subbands extend to infinite frequency, an upper cutoff has to be imposed, which was not necessary when using the complete dispersions, because for these ones the upper bands’ limits are well defined. Traditionally, the Debye frequency, $\omega_D$ (86 THz for Si), has been used as cutoff [@Callaway]. If we do this, the calculated thermal conductivities for nanowires are in large disagreement with experimental measurements. As fig. \[Fig:Callaway\](a) shows, no calculated curve agrees in shape with any experimental curve: the theoretical inflection points are always too high. Comparison of the parabolic approximation to $\tilde N_b^{\infty}$ (Eq. \[Eq:NCallaway\]) with the correct $\tilde N_b^{\infty}$, in fig. \[Fig:trans\], shows that the former grossly overestimates the transmission of the high frequency phonons, if its cutoff is set at $\omega_D$. Nevertheless, if a lower cutoff $\omega_C$ is chosen, the “Callaway” and “correct” $\tilde N_b^{\infty}(\omega)$ curves are more similar. Using $\omega_C$, the shapes of the theoretical thermal conductivity curves agree well with the experimental ones (fig. \[Fig:Callaway\](b)). The values of the lifetime parameters B and C are those that yield the correct bulk limit: $3.9\times 10^{-19}s/K$ and $223.1^oK$ for fig. \[Fig:Callaway\](a), and $1.7\times 10^{-19}s/K$ and $151.8^oK$ for fig. \[Fig:Callaway\](b). For bulk, it is possible to obtain a good fit of experimental data using either $\omega_C$ or $\omega_D$ (see inset of fig. \[Fig:Callaway\]), because the dominant umklapp scattering makes the role of high frequencies less important. For nanowires, boundary scattering dominates, and high frequencies play a more important role. Therefore, using the experimental Debye frequency as a cutoff for the Callaway formula does not yield correct results for nanowires, although good results can be obtained with a lower, adjusted cutoff. This one must be determined by comparison with nanowire experimental results ([*non predictive*]{} approach). On the other hand, if the atomistically calculated, [*complete dispersion*]{} relations are used, good agreement with experimental results is obtained (fig. \[Fig:thcond\]), with no need for prior knowledge of any experimental measurements for nanowires. ![Thermal conductivity normalized by its value at $320^oK$, for nanowires of different widths, calculated by the Callaway formula with cutoff at $\omega_D$ (86 THz) (a), and at $\omega_C$ (42 THz) (b). (Experimental results from ref. .) [*Inset:*]{} bulk thermal conductivity for the two cases (overlapping almost completely) and experimental curve from ref.  (dots). []{data-label="Fig:Callaway"}](fig3mingo_revNEW.eps){width="8.5"} To conclude, it was shown that, by using complete, atomistically computed phonon dispersions, it is possible to [*predictively*]{} calculate lattice thermal conductivity curves for nanowires, in good agreement with experiments. On the other hand, it was not possible to obtain correct results with the approximated Callaway formula. Still, good results could be obtained with a modified Callaway formula, although non predictively. The results of this paper are only expected to apply to “nanowhiskers” for which phonon confinement effects are unimportant. Si nanowires wider than $\sim 35 nm$ are within this category. I am indebted to D. Li, A. Majumdar and Ph. Kim for communication of their experimental data, and discussions, and to A. Balandin and Liu Yang for discussions. [1-100]{} G. Mahan, B. Sales and J. Sharp, Physics Today, March 1997, p.42. W. Fon, K. C. Schwab, J. M. Worlock and M. L. Roukes, Phys. Rev. B [**66**]{}, 045302 (2002). D. Li, Y. Wu, P. Kim, L. Shi, P. Yang and A. Majumdar, Appl. Phys. Lett., in press. J. Callaway, Phys. Rev. [**113**]{}, 1046 (1959). M. G. Holland, Phys. Rev. [**132**]{}, 2461 (1963). J. E. Parrott and A. D. Stuckes, Thermal conductivity of solids, Pion Ltd., London (1975). R. Berman, Thermal conductivity of solids, Clarendon Press, Oxford (1976). N. Mingo, L. Yang, D. Li, and A. Majumdar, to be published. For a nondispersive medium see L. G. C. Rego and G. Kirczenow, Phys. Rev. Lett. [**81**]{}, 232 (1998), or D. H. Santamore and M. C. Cross, Phys. Rev. B [**63**]{}, 184306 (2001). M. J. M. de Jong, Phys. Rev. B 49, 7778 (1994). C. W. J. Beenakker, Rev. Mod. Phys. 69, 731 (1997). P. G. Klemens, in [*Solid State Physics*]{}, edited by F. Seitz and D. Turnbull (Academic, New York, 1958), Vol. 7, p.1. W. A. Harrison, Electronic Structure and the Properties of Solids, Dover (1989). J. Zou and A. Balandin, J. Appl. Phys. 89, 2932 (2001). M. Asen-Palmer, K. Bartkowski, E. Gmelin, M. Cardona, A. P. Zhernov, A. V. Inyushkin, A. Taldenkov, V. I. Ozhogin, K. M. Itoh and E. E. Haller, Phys. Rev. B [**56**]{}, 9431 (1997). A. Balandin and K. Wang, Phys. Rev. B [**58**]{}, 1544 (1998). Y. -J. Han and P. G. Klemens, Phys. Rev. B 48, 6033 (1993). Yong-Jin Han, Phys. Rev. B 54, 8977 (1996). [^1]: $N_b$ is equivalent to the “ballistic” transmission function. [^2]: Note the similarity with Eq. (\[Eq:conductance\]). [^3]: $\tilde N_b$ was also calculated for a “free boundary” $11\times 11$ unit cells cross section wire, verifying that it is already quite similar to $\tilde N_b^{\infty}$, and very similar results for $\kappa$ were yielded when $\tilde N_b$ was replaced by $\tilde N_b^{\infty}$, keeping the remaining parameters unchanged. [^4]: The case of strong confinement can be dealt with by using modified expressions for $\tau$ involving the subbands’ group velocities rather than the speed of sound, as done in ref. . [^5]: Measurements of narrow nanowires around $20 nm$ wide (not shown here) have yielded thermal conductivities much smaller than those predicted, suggesting differences in the transport process at those small diameters. [^6]: An estimation of the order of magnitude of B can be made as [@Zou] $B\sim 2\gamma^2 {k_B \over \mu V_0 \omega_D}\sim 10^{-19}$, which agrees with our obtained value ($\gamma=$ Gruneisen parameter, $\mu=$ shear modulus, $V_0=$ volume per atom, and $\omega_D=$ Debye frequency.) Constant $C$ has the order of magnitude of the transverse phonons’ frequency near the zone edge [@Han; @Balandin]. [^7]: For this, use $\omega_{\alpha}(k_z)\simeq c_i\sqrt{k_z^2+k_{\perp}^2}$, $\alpha \equiv \{i,\vec k_{\perp}\}$, and $\langle v_z \rangle \simeq {\sum_{i=1}^3\int_0^{\omega/c_i} c_i {\sqrt{(\omega/c_i)^2-k_{\perp}^2} \over \omega/c_i}2\pi k_{\perp} d k_{\perp}\over \sum_{i=1}^3\int_0^{\omega/c_i} 2\pi k_{\perp} d k_{\perp}}$.

ArXiv

--- abstract: 'Evolution of periodic matter waves in one-dimensional Bose-Einstein condensates with time dependent scattering length is described. It is shown that variation of the effective nonlinearity is a powerful tool for controlled generation of bright and dark solitons starting with periodic waves.' author: - 'F. Kh. Abdullaev$^1$' - 'A.M. Kamchatnov$^{2,3}$' - 'V. V. Konotop$^3$' - 'V. A. Brazhnyi$^3$' title: ' Adiabatic dynamics of periodic waves in Bose-Einstein condensate with time dependent atomic scattering length' --- Observation of Bose-Einstein condensate (BEC) in gases of weakly interacting alkali metals have stimulated intensive studies of the nonlinear matter waves. A new area of physics—nonlinear matter waves and nonlinear atomic optics—was originated. Generation and dynamics of solitary wave pulses in BEC’s is one of the most important related problems. Experimental observations of dark [@dark; @Den] and bright [@Khayk; @Hulet] solitons have recently been reported. Theoretically, several methods of creating solitary waves have been proposed. First of all, it is a modulational instability [@Konotop], which is a universal phenomenon of the nonlinear physics (especially intensively explored in nonlinear optics, see [@AbdRev]). This method, however, cannot predict exactly the parameters of generated solitons. Another method, which is controllable in the above sense, is the so-called phase engineering [@Den], which consists of imposing an initial phase on a BEC and is appropriate for generating dark solitons. The phase imprinting, however, affects the whole background condensate which acquires nonzero initial velocity and starts to oscillate in a trap potential. The problem becomes even more complicated when one is interested in generating trains (or lattices) of solitons in BEC’s. In the present Letter we show that a powerful tool for generating and managing matter soliton trains can be provided by variation of the effective nonlinearity, which in practical terms can be achieved by means of variation of the $s$-wave scattering length $a_{s}$ due to the Feshbach resonance [@Moer]: $$\label{SL} a_{s}(t) = a(1 + \Delta/(B_{0} - B(t)).$$ Here $a$ is the asymptotic value of the scattering length far from resonance, $B(t)$ is the time-dependent external magnetic field, $\Delta$ is the width of the resonance and $B_{0}$ is the resonant value of the magnetic field. Feshbach resonances have been observed in Na at 853 and 907 G [@Inouye], in $^7$Li at 725 G [@Hulet], and in $^{85}$Rb at 164 G with $\Delta = 11$G [@Court]. Also, rapid variation in time of $a_{s}$ has been recently used for generation of bright solitons in BEC [@Khayk; @Hulet]. Here we want to indicate that in quasi-one-dimensional geometry an initially weak modulation of the condensate can be amplified by means of proper variation of the scattering length. As a result, the condensate evolves into either a sequence of bright solitons for $a_s<0$, or “domains” separated by dark solitons for $a_s>0$. In the case of bright solitons the attractive forces between atoms exactly compensate the wave-packet dispersion in the longitudinal direction, so that the confining trap potential in this direction becomes unnecessary. Then the motion of bright solitons in the longitudinal direction can be controlled by means of application of external forces. Actually, oscillations of bright solitons in the trap observed in [@Khayk; @Hulet] give a simple example of such controllable motion. Thus, quasi-one-dimensional bright BEC solitons behave as separate entities and their investigation seems to be a quite promising field of research. Dark solitons may be considered as moving “domain walls” which separate regions of a condensate with different values of the order parameter. Investigation of dark solitons is also useful for understanding the properties of BEC. In general, the problem of the controllable soliton generation is important for a number of BEC applications, like atomic interferometry [@Wright], and different kinds of the quantum phase transitions [@Ueda], as well as in the context of the nonlinear physics, including nonlinear optics and hydrodynamics. Our approach is based on the well established concept that the BEC dynamics at low enough temperature is well described by the three-dimensional (3D) Gross-Pitaevskii (GP) equation. In some physically important cases it admits a self-consistent reduction to the 1D nonlinear Schrödinger (NLS) equation $$\label{NLS} iu_{t} + u_{xx} - \frac 12 \nu^2x^2 u - 2\sigma g|u|^{2}u= 0.$$ In particular this is the case of a cigar-shaped BEC of low density when $\epsilon=\frac{{\cal N}|a|}{a_\bot}\ll 1$ and $\frac{a_\bot^2}{a_0^2}=\frac{\epsilon^2\nu}{\sqrt{2}}$, where ${\cal N}$ is a total number of atoms, $a_\bot=(\frac{\hbar}{m_a\omega_\bot})^{1/2}$ and $a_0=(\frac{\hbar}{m_a\omega_0})^{1/2}$ are linear oscillator lengths in the transverse and in cigar-axis directions, respectively (in the small amplitude limit they are of order of effective sizes of the condensate), $\omega_\bot$ and $\omega_0$ being respective harmonic oscillator frequencies, $\nu\lesssim 1$ is a positive parameter , and $m_a$ is the atomic mass. In (\[NLS\]) time $t$ and coordinate $x$ are measured in units $2/(\epsilon^2\omega_\bot)$ and $a_\bot/\epsilon$ respectively. The order parameter in the leading order is searched in the form $\psi({\bf r}, t)=\frac{\epsilon}{\sqrt{2\pi |a|}a_\bot}\exp\left(-i\omega_\bot t-\frac{y^2+ z^2}{2a_\bot^2}\right)u\left(\frac{\epsilon x}{a_\bot}, \frac{\epsilon^2\omega_\bot t}{2}\right)$, where $\sigma={\rm sign}(a_s)$, and $g(t)\equiv a_s(t)/a_s(0)$. It will be assumed that $a_s(t)$ does not change its sign and thus $g(t)$ is a positive-valued function. We notice that smallness of the density rules out a possibility of collapse phenomenon (if $a<0$). We start with analytical estimates supposing that the initial wave function $u(x,0)$ is modulated along the cigar axis with the wavelength $L$ of modulation much less than the longitudinal dimension of the condensate, i.e. of the $l$: $L\ll l$. Therefore, at this stage we neglect the smooth trap potential and impose cyclic boundary conditions. Then the initial wave function can be approximated well enough by exact periodic solutions of Eq. (\[NLS\]) at $\nu=0$. For example, if at $t=0$ we take into account only one space harmonic of the initial wave function, $u(x,0)=u_0+u_1\cos(x/L)$, then this distribution can be approximated by well-known elliptic function solutions of Eq. (\[NLS\]) with a small parameter $m$ (see below). We are interested in evolution of such solutions due to slow change of $g(t)$ with time. At the same time we suppose that the total time of adiabatically slow change of $g(t)$ is much less than the period $\sim2\pi/\nu$ of soliton oscillations in the trap potential, so that we can neglect the influence of the trap potential on the motion of solitons during the formation of soliton trains and put $\nu=0$ in Eq. (\[NLS\]). This means that we shall consider analytically relatively small segments of the modulated wave much greater than the wavelength $L$ and much smaller than the size $l$ of the whole condensate in the trap. To solve the problem of the condensate evolution, we note that substitution $$\label{u-v} u(x,t)=v(x,t)/\sqrt{g(t)}$$ transforms Eq. (\[NLS\]) with $\nu=0$ into $$\label{modNLS} iv_t+v_{xx}-2\sigma |v|^2v=i{\varepsilon}v$$ with $ {\varepsilon}(t)=g'(t)/2g(t). $ Thus, for slowly varying $g(t)$ the right hand side of Eq. (\[modNLS\]) can be considered as a small perturbation: $|\varepsilon(t)|\ll 1$. As it follows from Eq. (\[u-v\]), for the initial distribution one has $v(x,0)=u(x,0)$. For our purposes it is enough to consider typical particular solutions of the unperturbed NLS equation which are parameterized by two parameters ${\lambda}_{1,2}$ [@com]. Under influence of the perturbation, these parameters in the adiabatic approximation become slow functions of time $t$. Equations which govern their evolution can be derived by the following simple method. First, the initial values of ${\lambda}_{1,2}$, as well as the coefficients in Eq. (\[modNLS\]) are supposed to be independent of $x$, hence the wavelength $L$ of the nonlinear wave evolving according to Eq. (\[modNLS\]) is constant, $$\label{wl} {dL({\lambda}_1(t),{\lambda}_2(t))}/{dt}=0.$$ Second, we find that the variable $ N({\lambda}_1(t),{\lambda}_2(t))=\int_0^L|v|^2dx $ changes with time according to $$\label{Neq} {dN({\lambda}_1(t),{\lambda}_2(t))}/{dt}=2{\varepsilon}N({\lambda}_1(t),{\lambda}_2(t)).$$ Then, if the expressions for $L$ and $N$ in terms of ${\lambda}_{1,2}$ are known, Eqs. (\[wl\]) and (\[Neq\]) reduce to two equations linear with respect to derivatives $d{\lambda}_{1,2}/dt$, which yield the desired system of differential equations for ${\lambda}_{1,2}$. The form of this system depends, of course, on the choice of the parameters ${\lambda}_{1,2}$. It is well known from the theory of modulations of nonlinear periodic waves that for completely integrable equations (as the NLS equation) the most convenient choice is provided by the so-called “finite-gap integration method” of obtaining periodic solutions. Therefore we shall use the parametrization of the periodic solutions of the NLS equation obtained by this method (see, e.g. [@Kamch2000]), and consider three most typical cases. [*Case 1*]{}: cn-[*wave in a BEC with a negative scattering length*]{}. In the case of a BEC with negative scattering length, $\sigma=-1$, there are two simple two-parametric periodic solutions of unperturbed Eq. (\[modNLS\]). One of them has the form $$\label{cn} v(x,t)=2{\lambda}_1e^{-4i({\lambda}_1^2-{\lambda}_2^2)t}{{\mathrm{cn}} }[2\sqrt{{\lambda}_1^2+{\lambda}_2^2}\, x,m],$$ where the parameter of elliptic function is given by $m={{\lambda}_1^2}/({{\lambda}_1^2+{\lambda}_2^2})$. Then straightforward calculations give $$\label{L-cn} L=\frac{2{{\mathrm{K}} }(m)}{\sqrt{{\lambda}_1^2+{\lambda}_2^2}}, \quad N=8\sqrt{{\lambda}_1^2+{\lambda}_2^2}\,{{\mathrm{E}} }(m)-4{\lambda}_2^2 L$$ where ${{\mathrm{K}} }(m)$ and ${{\mathrm{E}} }(m)$ are complete elliptic integrals of the first and the second kind, respectively. Substitution of these expressions into Eqs. (\[wl\]) and (\[Neq\]) yields the system $$\label{cn-syst} \begin{array}{l}\displaystyle{ \frac{d{\lambda}_1}{dz}=\frac{(({\lambda}_1^2+{\lambda}_2^2){{\mathrm{E}} }(m)-{\lambda}_2^2{{\mathrm{K}} }(m)){{\mathrm{E}} }(m){\lambda}_1} {{\lambda}_1^2{{\mathrm{E}} }^2(m)+{\lambda}_2^2({{\mathrm{K}} }(m)-{{\mathrm{E}} }(m))^2},}\\ \displaystyle{ \frac{d{\lambda}_2}{dz}=\frac{({\lambda}_2^2{{\mathrm{K}} }(m)-({\lambda}_1^2+{\lambda}_2^2){{\mathrm{E}} }(m))({{\mathrm{K}} }(m)-{{\mathrm{E}} }(m)){\lambda}_2} {{\lambda}_1^2{{\mathrm{E}} }^2(m)+{\lambda}_2^2({{\mathrm{K}} }(m)-{{\mathrm{E}} }(m))^2}} \end{array}$$ where $$\label{z} z=z(t)=2\int_0^t{\varepsilon}(t') dt'=\ln g(t), \qquad z(0)=0.$$ If dependence of ${\lambda}_1$ and ${\lambda}_2$ on $z$ is found from (\[cn-syst\]), then Eq. (\[u-v\]) gives evolution of the periodic wave $u(x,t)$ with slow change of the parameter $z$ connected with time $t$ by Eq. (\[z\]). In particular, the density of particles in the condensate is given by $$\label{cn-dens} |u|^2=4e^{-z}{\lambda}_1^2(z)\,{{\mathrm{cn}} }^2[2\sqrt{{\lambda}_1^2(z)+{\lambda}_2^2(z)}\, x,m],$$ where transformation to the time variable should be performed with the use of Eq. (\[z\]). In Fig. \[figone\]a we present an example of the evolution of the respective density distribution in the presence of a harmonic trap potential where the parabolic parameter as well as experimentally feasible parameters are used. The figure shows that in the case of a negative scattering length given by (\[SL\]) increase of the magnetic field $B(t)$ within the region $B(0)<B(t)<B_0$ results in compression of the atomic density and formation of a lattice of matter solitons. ![Numerical solution of Eq.(\[NLS\]) with $g(t)=e^{t/\tau}$. Initial conditions are chosen in the form $u(x,0)e^{-\nu x^2/2^{3/2}}$ where $u(x,0)$ is given by (a) Eq. (\[cn\]) with $\lambda_1(0)=1$, $\lambda_2(0)=.2$, $\tau\approx 2$ $\nu=.02$; (b) Eq. (\[dn\]) with $\lambda_1(0)=1.5$, $\lambda_2(0)=.2$, $\tau\approx 1.5$ and $\nu=.01$; (c) Eq. (\[sn\]) with $\lambda_1(0)=3$, $\lambda_2(0)=.3$, $\tau\approx 1$ and $\nu=.01$. In the boxes we show density distributions at initial (thin lines) and final (thick lines) moments of time. []{data-label="figone"}](3dd.eps){width="5cm" height="8cm"} [*Case 2*]{}: dn-[*wave in a BEC with a negative scattering length*]{}. Another simple solution of the NLS equation (\[modNLS\]) with $\sigma=-1$ is given by $$\label{dn} v(x,t)=({\lambda}_1+{\lambda}_2)e^{-2i({\lambda}_1^2+{\lambda}_2^2)t}{{\mathrm{dn}} }[({\lambda}_1+{\lambda}_2)x,m],$$ where $ m={4{\lambda}_1{\lambda}_2}/{({\lambda}_1+{\lambda}_2)^2}$. By analogy with (\[cn-syst\]) we derive the following equations for ${\lambda}_1$ and ${\lambda}_2$: $$\label{dn-syst} \begin{array}{l}\displaystyle{ \frac{d{\lambda}_1}{dz}=\frac{{\lambda}_1({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)} {({\lambda}_1-{\lambda}_2){{\mathrm{K}} }(m)+({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)},}\\ \displaystyle{ \frac{d{\lambda}_2}{dz}=-\frac{{\lambda}_2({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)} {({\lambda}_1-{\lambda}_2){{\mathrm{K}} }(m)-({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)},} \end{array}$$ where it is supposed that ${\lambda}_1>{\lambda}_2$ and $z$ is defined by Eq. (\[z\]). Now the density of particles is given by $$\label{dn-dens} |u|^2=e^{-z}({\lambda}_1(z)+{\lambda}_2(z))^2{{\mathrm{dn}} }^2[({\lambda}_1(z)+{\lambda}_2(z)) x,m].$$ An example of the respective evolution in the presence of the potential is given in Fig. \[figone\]b. One again observes formation of a lattice of matter solitons starting with a weakly modulated periodic wave. [*Case 3*]{}: sn-[*wave in a BEC with a positive scattering length*]{}. In the case of (\[modNLS\]) with $\sigma=1$ there exists simple periodic solution $$\label{sn} v(x,t)=({\lambda}_1-{\lambda}_2)e^{2i({\lambda}_1^2+{\lambda}_2^2)t}{{\mathrm{sn}} }[({\lambda}_1+{\lambda}_2)x,m],$$ where $ m=\left(\frac{{\lambda}_1-{\lambda}_2}{{\lambda}_1+{\lambda}_2}\right)^2 $ and it is supposed that ${\lambda}_1>{\lambda}_2$. Now we obtain $$\label{sn-syst} \begin{array}{l}\displaystyle{ \frac{d{\lambda}_1}{dz}=\frac{{\lambda}_1({\lambda}_1+{\lambda}_2)({{\mathrm{K}} }(m)-{{\mathrm{E}} }(m))} {2{\lambda}_1{{\mathrm{K}} }(m)-({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)},}\\ \displaystyle{ \frac{d{\lambda}_2}{dz}=\frac{{\lambda}_2({\lambda}_1+{\lambda}_2)({{\mathrm{K}} }(m)-{{\mathrm{E}} }(m))} {2{\lambda}_2{{\mathrm{K}} }(m)-({\lambda}_1+{\lambda}_2){{\mathrm{E}} }(m)},} \end{array}$$ with $z$ defined again by Eq. (\[z\]). The density of particles in the condensate is given by $$\label{sn-dens} |u|^2=e^{-z}({\lambda}_1(z)-{\lambda}_2(z))^2{{\mathrm{sn}} }^2[({\lambda}_1(z)+{\lambda}_2(z)) x,m].$$ This case, but in the presence of external trap potential is illustrated in Fig. \[figone\]c, where by means of increase of the magnetic field a periodic wave is transformed into a lattice of dark solitons. In physical units, the cases depicted in Fig. \[figone\] correspond to (a) ${\cal N}= 1.4\cdot 10^4$ $^7Li$ atoms in a trap with $a_\bot\approx 7\,\mu$m, and $a_0\approx 230\,\mu$m, (b) ${\cal N}= 2\cdot 10^4$ $^7Li$ atoms in a trap with $a_\bot\approx 6\,\mu$m, and $a_0\approx 416\,\mu$m, and (c) ${\cal N}= 10^4$ $^{23}Na$ atoms in a trap with $a_\bot=3.4\,\mu$m, and $a_0=264\,\mu$m. In the last case, however one observes shifts of the soliton positions as well as decrease of a density of particles located about the potential minimum because of weak oscillations of the condensate in the trap potential (the expanding phase is depicted in the figure). Thus, although effective $\tau$ corresponding to physical time [@Hulet] $t_0=40\,$ms (used in all simulations) is not large characteristic amplitudes of solitons placed in the center of the trap potential (in the cases (a) and (b) match well the amplitude values following from the adiabatic approximation developed for a homogeneous NLS and for large $\tau$, and in the case (c) one observes qualitative agreement. No instabilities of periodic waves are observed during periods of soliton train formation. The developed analytical approach can be generalized to the NLS equation with linear damping, when the right hand side of Eq. (\[NLS\]) is equal to $-i\gamma u$, $\gamma$ being the damping constant, the substitution (\[u-v\]) yields again Eq. (\[modNLS\]) but now with modified value of ${\varepsilon}$: ${\varepsilon}\to{\varepsilon}-\gamma$. Hence, the equations for ${\lambda}_1$ and ${\lambda}_2$ hold their form with $z(t)$ defined as $z(t)=\ln g(t)-2\gamma t$, so that the only change in Eqs. (\[cn-dens\]), (\[dn-dens\]), and (\[sn-dens\]) is multiplication of their right hand sides by $\exp(-2\gamma t)$. Another generalization corresponds to moving soliton trains what may be useful for treatment of BEC in ring traps. The above method remains actually the same, but the initial condition includes a phase factor linearly depending on the space coordinate. Then the results presented in Fig. \[figone\] can be viewed as plots of the respective current as functions of time. In practice, solitons can be put also in motion by imposing a proper external potential. One of applications of such moving soliton trains could be a “laser” of matter solitons. The consideration provided above implies that one starts the adiabatic deformation with an initially periodic solution. A question arises about possibility of creation of such a state experimentally. A natural approach to solving this problem would be the use of an optical trap [@opt_trap]. In such a trap it is possible to create a nonlinear periodic distribution of a BEC [@Bronski; @Konotop]. Then, switching off the laser beams, producing the trap, will result in a periodic distribution of the condensate. However, it is not stable without the trap since it is not a solution of the respective GP equation. This difficulty can be overcome if simultaneously with switching off the optical trap one abruptly changes the scattering length (or alternatively provides change of the number of particles) in such a way that the existing distribution will satisfy (\[NLS\]). To be more specific, let us consider an example of a BEC with a positive scattering length in an optical trap given by [@Bronski]: $V(x)=-2V_0{{\mathrm{sn}} }(2\kappa x, m)$, where $V_0$ is the potential amplitude, $\kappa=\lambda_1+\lambda_2$ and $m$ is the same as in (\[sn-dens\]). The equation describing BEC evolution now admits a solution $$\begin{aligned} u(x,t)=\sqrt{V_0+({\lambda}_1-{\lambda}_2)^2}\,e^{2i({\lambda}_1^2+{\lambda}_2^2)t}{{\mathrm{sn}} }[({\lambda}_1+{\lambda}_2)x,m].\end{aligned}$$ This last function solves also (\[NLS\]) with $\sigma=1$ and $ g=\frac{V_0+({\lambda}_1-{\lambda}_2)^2}{({\lambda}_1-{\lambda}_2)^2} $ and thus by switching off the potential $V(x)$ with simultaneous changing the scattering length by $\Delta a_s=\frac{a_0 V_0}{({\lambda}_1-{\lambda}_2)^2}$ one achieves the desired initial state. Notice that although experimentally a sn-potential is not easily reachable in a general case, for a large range of $m$ it is approximated very well by only a few first Fourier harmonics. For example, for a situation described in Fig. \[figone\] one has $V(x)\approx -2V_0[1.47\sin(0.78 x)+0.15 \sin(2.3 x)]$ with the accuracy about 1%. In order to estimate characteristic scales of adiabatic deformations we introduce an “aspect ratio" defined as $\delta=|\Delta A|/L$, where $\Delta A$ is a total variation of the amplitude of the periodic wave and $L$ is its wavelength. Then the cases $\delta\gg 1$ and $\delta \ll 1$ correspond to a solitonic lattice and to a modulated plane wave. Dependence of the scattering length on time can be simulated by $g(t)=e^{t/t_0}$ (physical units). Taking as an example the solution depicted in Fig. \[figone\]b, where $\delta\approx 1$ at $t=0$, we find that already at $t\approx 30$ms the aspect ratio becomes $\delta\approx 100$. The adiabaticity of the process means here that $\frac{\pi}{\omega_0}\gg t \gg \frac{\pi}{\omega_\bot}$ (physical units). It is satisfied well enough for traps with $\omega_\bot \gtrsim 2\pi\times 200$ Hz and $\omega_0 \lesssim 2\pi\times 5$ Hz. We notice that the imposed condition rules out a possibility of resonant phenomena which might be related to coincidence of a linear oscillator period, originated by one of the trap dimensions, on the one hand and the characteristic time of change of $a_s$ on the other hand. Stability of the above solutions has been studied numerically in [@CKR], where it has been found that soliton trains are stable in the case of positive scattering length and are also stable in the case of negative scattering length for special choice of the parameters. In this context, the adiabatic variation of the scattering length, which results in the change of the wave parameters, can be used for stabilizing (or destabilizing) respective periodic solutions. To conclude, we have outlined the main idea of management of periodic nonlinear waves in BEC’s. The theory, although being developed for a homogeneous NLS equation, give an accurate estimate of the central part of a BEC placed in a magnetic trap, the later being studied numerically. The existence of trap can also be taken into account in the framework of more sophisticated theory recently developed in [@kamch02]. Application of this method to the theory of a BEC will be presented elsewhere. Work of A.M.K. in Lisbon has been supported by the Senior NATO Fellowship. A.M.K. thanks also RFBR (Grant 01–01–00696) for partial support. V.V.K. acknowledges support from the European grant, COSYC n.o. HPRN-CT-2000-00158. Work of V.A.B. has been supported by the FCT fellowship SFRH/BPD/5632/2001. Cooperative work has been supported by the NATO-Linkage grant No. PST.CLG.978177. [99]{} S. Burger, et. al. Phys. Rev. Lett. [**83**]{}, 5198 (1999); J. Denschlag et al. Science [**287**]{}, 97 (2000). L. Khaykovich et al. Science, [**296**]{}, 1290 (2002). K.E. Strecker et al. Nature, [**417**]{}, 150 (2002). V.V. Konotop and M. Salerno, Phys. Rev. A [**65**]{}, 021602 (2002). F.Kh. Abdullaev, S.A. Darmanyan and J. Garnier, In [*Progress in Optics*]{}, Edit. E. Wolf, [**44**]{}, 303 (2002). A.J. Moerdijk, B.J. Verhaar, and A. Axelsson, Phys.Rev. A [**51**]{}, 4852 (1995). S. Inouye, et al., Nature (London), [**392**]{}, 151 (1998). P. Courteille et al., Phys. Rev. Lett., [**81**]{}, 69 (1998). B.P. Anderson, K. Dholakia and E.M. Wright, eprint: cond-mat/0209252. R. Kanamoto, H. Saito, and M. Ueda, eprint: cond-mat/0210229. Our choice of the parameters is dictated by simplicity of the obtained equations. Physical parameters can be expressed in terms of $\lambda_i$ (see (\[cn\]), (\[dn\]) or (\[sn\])). A.M. Kamchatnov, [*Nonlinear Periodic Waves and Their Modulations—An Introductory Course,*]{} (World Scientific, Singapore, 2000) B.P. Anderson and M. Kasevich, Science [**282**]{}, 1686 (1998); S. Burger et al., Phys. Rev. Lett. [**86**]{}, 4447 (2001); O. Morsch, et al., Phys. Rev. Lett. [**87**]{}, 140402 (2001). J.C. Bronski, et al., Phys. Rev. Lett., [**86**]{}, 1402 (2001). L. D. Carr J. N. Kutz, and W. P. Reinhardt, Phys. Rev. E [**63**]{}, 066604 (2001). A.M. Kamchatnov, eprint: nlin.SI/0302027.

ArXiv

--- abstract: 'In Statistical Learning, the Vapnik-Chervonenkis (VC) dimension is an important combinatorial property of classifiers. To our knowledge, no theoretical results yet exist for the VC dimension of edited nearest-neighbour (1NN) classifiers with reference set of fixed size. Related theoretical results are scattered in the literature and their implications have not been made explicit. We collect some relevant results and use them to provide explicit lower and upper bounds for the VC dimension of 1NN classifiers with a prototype set of fixed size. We discuss the implications of these bounds for the size of training set needed to learn such a classifier to a given accuracy. Further, we provide a new lower bound for the two-dimensional case, based on a new geometrical argument.' author: - | Iain A. D. Gunn [email protected]\ Ludmila I. Kuncheva [email protected]\ School of Computer Science,\ Bangor University,\ Dean Street,\ Bangor, Gwynedd,\ Wales LL57 2NJ,\ UK title: Bounds for the VC Dimension of 1NN Prototype Sets --- Machine learning, classification, VC dimension, prototype generation, statistical learning theory, nearest neighbour Introduction ============ The VC dimension is a measure of what is called the “capacity” of a classification algorithm: that is, roughly speaking, its flexibility or expressive power. The practical interest of this quantity arises from its role in the *Probably Approximately Correct* learning model and related models (see e.g. @Shais [ch. 3], @Holden95), where it appears in results relating the size of the training set to the classifier’s accuracy. The VC dimension is defined in terms of the concept of *shattering* [@Vapnik98]. For a two-class problem in $d$-dimensional space $\mathbb{R}^d$, a set $A$ of functions $\mathbb{R}^d \to \{+1,-1\}$ is said to shatter a set of points (in $\mathbb{R}^d$) if any labelling of that set can be given by an element of $A$. The VC dimension of $A$ is the size of the largest set which can be shattered by $A$. The nearest-neighbour (1NN) classification rule is a classic technique of supervised learning, first introduced by [@Fix52]. The 1NN rule determines a label for (“classifies”) a given point in a metric space by assigning it the label of the nearest point in a previously determined set of labelled reference points. In the original and simplest algorithm, the reference set is the set of all the training data. If the amount of training data is *a priori* unbounded, then the VC dimension of the set of associated 1NN-rule classifying functions is infinite: trivially, a set of points of any size is labelled correctly by a classifier whose reference set is the same set of labelled points. However, it is often not practical to store all the training data, especially in an era of “big data”. Therefore many algorithms have been proposed which learn a smaller reference set from the training data: [@Garcia12] and [@Triguero12] survey more than 75 such algorithms between them. What is the VC dimension associated with these “editing” NN algorithms? If the reference set may grow without bound, then the VC dimension is infinite, as for the na[ï]{}ve classifier described above. But if the reference set is constrained not to exceed a given size, then the VC dimension is finite. The main results of the present work are lower and upper bounds for the VC dimension of the set of all 1NN-rule classifiers which use a reference set of given fixed size, in Euclidean space. It should be noted that, in general, the size of the reference set to be formed by an editing algorithm is not fixed *a priori*. There are exceptions, an important one being the case where a classifier for a data stream is kept current by using a fixed window of the most recent $N$ points as its reference set [@GunnNeurocomputing]. But it is very common to impose a *maximum* size limit on the reference set, if only by hand in an *ad hoc* fashion. Our upper bounds will apply to any algorithm for which the size of the reference set is bounded, though our lower bounds will apply only where the size is prescribed, and the algorithm is capable of returning any reference set of that size. We believe that there is significant interest in the VC dimension of prototype classifiers with fixed-size reference sets, and that this is shown by the fact that an incorrect purported result [@Karacali Proposition 2] has been cited more than 50 times as giving the VC dimension of such classifier sets. Our study corrects the record. We feel there is significant value in bringing together the existing theoretical results, some of which seem otherwise likely to remain obscure to practitioners. Beyond our deductions from existing theory, our novel contributions are 1) A new lower bound for the two-dimensional case, higher than that implied by previous results for polytopes (Proposition \[OneBetter\]), and 2) an upper bound for all dimensions, slightly less tight than the best that can be deduced from existing theory, but avoiding the use of exotic functions and thereby facilitating a discussion of the asymptotic behaviour of the limits (Corollary \[LooseCor\]). We will discuss the theoretical framework in section \[NNinVC\], and briefly review the relevant literature in section \[lit\]. In sections \[lower2d\] and \[lowerdd\] we derive lower bounds for the VC dimension, and in section \[upper\] we determine upper bounds. Our results are summarised in section \[conclusions\]. 1NN classifiers in the VC theory {#NNinVC} ================================ We use “classifier” to mean a function from the feature space to the set of classes, which is used to classify unlabelled examples. An algorithm which learns such a function from training data is a “classification algorithm”. The set of all classifiers which the algorithm might produce in response to all sets of training data is called the “hypothesis class” of that algorithm. (For example, the original Rosenblatt Perceptron selects among all functions which map one half-space to one class, and the other half-space to the other class: this set of functions is the hypothesis class of the Perceptron.) When we speak of the VC dimension of a classification algorithm, we mean the VC dimension of the hypothesis class from which that algorithm learns a classifier. NN classification algorithms are not usually thought of as selecting a classifier from a fixed hypothesis class in this way [see e.g. @Shais ch. 19]. This is because the great practical advantage of 1NN classification is that a classifier function does not need to be explicitly evaluated when classifying an unlabelled example; the new point is simply assigned the class of the nearest prototype in the reference set, which can be identified efficiently using a $k$-d tree. However, this process is equivalent to classifying the new point according to a classifier function which maps the Voronoi cell of each prototype to the label of that prototype. Thus, the hypothesis class of the 1NN rule with a reference set of $m$ prototypes in $d$-dimensional space is the set of all labellings of all $m$-cell Voronoi diagrams in the space; it is parameterised by the $md$ co-ordinates of the $m$ prototypes, and the $m$ choices of label. We consider only features which take real values, thus classifiers whose domain is $\mathbb{R}^d$, and consider only the 1NN rule with the Euclidean metric. $\mathrm{1NN}(d,m)$ will denote the set of all classifiers $g:\mathbb{R}^d \to \{+1,-1\}$ which use the (Euclidean) nearest-neighbour rule with a reference set of size $m$. $\mathrm{VCdim}(A)$ will denote the VC dimension of a set $A$ of classifiers. The purpose of this paper is to give lower and upper bounds for $\mathrm{VCdim}(\mathrm{1NN}(d,m))$. Related Work {#lit} ============ A number of existing theoretical results have implications for the VC dimension of the NN classifier with arbitrary reference set of fixed size. However, these results are scattered in the literature and their implications have not been made explicit. In particular, the results we use to establish lower bounds were developed for a class of polytope classifiers, without mention of NN classifiers. In the present section we give a brief overview of relevant theoretical work, starting with some brief historical context and going on to include the work whose implications we will directly use in subsequent sections. Questions of “separating capacities” of families of decision surfaces were already considered before the advent of the Vapnik-Chervonenkis theory (see e.g. [@Cover65] and references therein). During and shortly after the years of the initial development of the VC theory, several authors used approaches from classical combinatorial geometry to derive results about the VC dimension (or related separability properties) of several simple sets. Readers interested in this literature will need to be aware of the distinction between the VC dimension of a set of classifiers, as we have defined it above, and the VC dimension of a set of subsets of the Euclidean space in question (called by some authors a “concept class”). See our discussion at the start of section \[upper\], and @Devroye96 [pp. 196, 199, and 215]. To give an example of the results obtained, @Dudley79 reports that the set of balls in $\mathbb{R}^d$ has VC dimension $d+1$. (N.B. Dudley’s quantity $V$ is one greater than the quantity defined as the VC dimension in more recent literature.) Similarly, it may be shown that the set of all half-spaces in $\mathbb{R}^d$ has VC dimension $d+1$ [see e.g. @Devroye96 ch. 13]. More recent authors have considered the intersection or union of half-spaces, which is to say, sets which are the interior or exterior of (possibly unbounded) polytopes. @Blumer89 show that the set of interiors of $N$-gons has VC dimension $2N + 1$. This result is quoted by @TakacsPataki as the starting point for their work in which they find upper and lower bounds for the VC dimension of convex polytope classifiers, from which we will derive a lower bound in section \[lowerdd\]. NN classifiers, like convex polytope classifiers, have decision boundaries which are the union of subsets of hyperplanes. However, the decision region for a NN classifier is not in general a simple intersection or union of half-spaces; it may be a complicated union of the interiors of polytopes formed by such intersections. This may explain why the question of the VC dimension of the NN classifier with an arbitrary reference set of fixed size $m$ has not previously, to our knowledge, been addressed, other than in [@Karacali]. (@Devroye96 consider the closely related property of the *shatter coefficient* of such classifiers; we will make explicit the implications of this work in section \[upper\].) @Karacali [Proposition 2] claim that the VC dimension of the NN classifier with reference set of size $m$, $\mathrm{VCdim}(\mathrm{1NN}(d,m))$ in our notation, is exactly $m$. Their argument consists of exhibiting a set of $m+1$ points which cannot be correctly classified by $m$ prototypes. They do not argue that a general set of $m+1$ points cannot be shattered. This apparently reflects a misunderstanding of the definition of VC dimension. The VC dimension is the largest number for which some set of that size can be shattered, not the largest number such that all sets of that size can be shattered. The latter quantity is called the Popper dimension, a quantity which thus far has not found a rôle in statistical learning theory [@Corfield09]. Lower bounds - two dimensions {#lower2d} ============================= Tak[á]{}cs and Pataki [@Takacs; @TakacsPataki] prove bounds for the VC dimension of sets of classifiers whose decision boundaries are convex polytopes. These sets can easily be related to sets of nearest-neighbour classifiers, giving lower bounds for the latter. This is because a decision boundary which is an $N$-faceted convex polytope can be obtained as the decision boundary of a 1NN classifier with $N+1$ prototypes, by placing a prototype of one label inside the polytope, and the remaining $N$ prototypes, with the opposite label, as the reflections of the first prototype in the $N$ facets of the desired decision boundary. See Figure \[fig:PolygonVoronoi\] for an example construction with $N=5$, $d=2$. We formalise this observation as Proposition \[subset\] below, and give a formal proof in the appendices. ![Illustration of the construction of a data set whose classification region is determined by a given convex polygon $P$. The required prototype set (filled and empty circles) is constructed by reflecting an arbitrary interior point in (the lines containing) the edges of $P$. Voronoi boundaries are shown; the Voronoi cell of the interior point is the decision region of the classifier, coincident with $P$.[]{data-label="fig:PolygonVoronoi"}](PolygonVoronoiRegion){width="0.6\linewidth"} We will use $G(d,N)$ to denote the set of classifiers $g:\mathbb{R}^d \to \{+1,-1\}$ whose decision boundary is a convex $N$-faceted polytope. \[subset\] The set of convex $N$-faceted polytope classifiers is a subset of the set of NN classifiers with reference set of size $N+1$, in the same Euclidean space. That is, $$G(d,N) \subseteq \mathrm{1NN}(d,N+1).$$ \[impliesLowerBound\] A lower bound for the VC dimension of $G(d,N)$ is also a lower bound for the VC dimension of $\mathrm{1NN}(d,N+1)$. The corollary follows immediately: any set of points which can be shattered by an element of $G(d,N)$ can be shattered by an element (the same element) of $\mathrm{1NN}(d,N+1)$. Tak[á]{}cs gives the following result for two-dimensional Euclidean space: (Tak[á]{}cs) \[Takacs1\] $$h(G(2,N)) \geq 2N + 2,$$ for $N \geq 2$. We give a sketch of the proof given in [@Takacs] in the appendices. By a more complicated argument, Tak[á]{}cs establishes that this lower bound is also an upper bound for the VC dimension of polytope classifiers. In general, upper bounds for polytope classifiers are of no relevance to the more general set of NN-rule classifiers. But the $N=2$ case is an exception, and we make the following brief side remark about this case: With three prototypes, the only decision surfaces an NN-rule classifier can form are an open 2-gon, or a pair of parallel lines. But any finite set of points which can be correctly dichotomised by a pair of parallel lines can also be correctly split by a digon formed by a suitably small adjustment of the lines such that they are non-parallel. So for this case, the set of NN-rule classifiers has no greater separating power than the related set of polygon classifiers. Thus a precise value for the VC dimension of this set of NN-rule classifiers is established: $$\mathrm{VCdim}(\mathrm{1NN}(2,3)) = 6.$$ Now, for general $m$, Lemma \[Takacs1\], with Corollary \[impliesLowerBound\], implies a lower bound for $\mathrm{VCdim}(\mathrm{1NN}(2,m))$: $$\mathrm{VCdim}(\mathrm{1NN}(2,m)) \geq 2m$$ However, we can do better than this. For $m \geq 4$, the NN-rule classifier can create a larger class of decision surfaces than a single convex polytope. In Appendix \[appendix1\] we present a new argument, inspired by the elementary geometrical approach of Tak[á]{}cs but considerably more involved, demonstrating a stronger lower bound for $\mathrm{VCdim}(\mathrm{1NN}(2,m))$, $m \geq 4$: \[OneBetter\] $$\mathrm{VCdim}(\mathrm{1NN}(2,m)) \geq 2m + 1.$$ Though this improvement is the smallest possible, it establishes the principle that the VC dimension of 1NN classifiers with $m$ prototypes in $\mathbb{R}^2$ is larger than that of the relevant comparable class of polygon decision boundaries (recall that Tak[á]{}cs’ lower bound, Lemma \[Takacs1\], is also an upper bound for that class). That is, Proposition \[OneBetter\] establishes that 1NN classifiers in $\mathbb{R}^2$ gain in expressivity from their ability (for $m \geq 4$) to form boundaries other than convex polygons. Lower bound – higher dimensions {#lowerdd} =============================== Tak[á]{}cs’ result was extended by Tak[á]{}cs and Pataki to Euclidean spaces of dimension higher than two: ([@TakacsPataki]) \[Takacs2\] $$\mathrm{VCdim}(G(d,N)) \geq dN + 2,$$ for $d \geq 2$, $N \geq 2$. The geometrical arguments are less simple than for the two-dimensional case; we refer readers to [@TakacsPataki] for the proof. Tak[á]{}cs and Pataki also offer slightly stronger lower bounds for the special cases $d=3$ and $d=4$: $h(G(3,N)) \geq 3N + 3$ and $h(G(4,N)) \geq 4N + 5$; but these require a higher minimum value for $N$ than Lemma \[Takacs2\] does. As in the two-dimensional case, we can deduce a lower bound for the VC dimension of the NN classifier: \[LowerBoundProp\] $$\mathrm{VCdim}(\mathrm{1NN}(d,m)) \geq dm + 2 - d,$$ for $d \geq 2$, $m \geq 3$. The relation does not hold for $d=2$, $m=2$: an NN-rule classifier with two prototypes in the plane has for its decision boundary a line, so cannot shatter four points: linear classifiers are, famously, unable to solve the XOR problem. Upper bounds {#upper} ============ The *shatter coefficient* of a family of sets $B$ (for our purposes, $B$ is a set of subsets of $\mathbb{R}^d$) is a number closely related to VC dimension. It is called the “growth function” by some authors, but other authors give that term a different definition. The $n$th shatter coefficient of $B$, denoted $S(B,n)$, is the maximum number of different subsets of $n$ points that can be formed by intersection of the $n$ points with elements of $B$: that is, the number of subsets that can be “picked out” using elements of $B$. (The maximum is taken over all sets of $n$ points.) The VC dimension of $B$ is then the largest $n$ such that $S(B,n) = 2^n$. This is an expression of the concept of shattering for families of sets $B$ rather than classifiers: if the set of all subsets of the $n$ points which can be formed by intersection of the $n$ points with elements of $B$ is all $2^n$ possible subsets of the $n$ points, then the $n$ points are shattered by $B$. The VC dimension of a family of classifiers, as we defined it in section \[introduction\], is equal to the VC dimension of the associated family of decision regions, as just defined, and the shatter coefficient of a family of classifiers is defined equal to the shatter coefficient of the associated family of decision regions see [see @Devroye96 ch.12]. Devroye et al. give upper bounds for the shatter coefficients of the class $C(d,m)$: (Devroye, Gy[ö]{}rfi, and Lugosi) For $m \geq 3$, \[DevroyeLemma\] $$\begin{aligned} S(C,n) &\leq 2^m n^{9(m-2)} &\text{for } d=2\\ S(C,n) &\leq 2^m n^{(d+1)m(m-1)/2} &\text{for } d \geq 3. \end{aligned}$$ See [@Devroye96 p. 312] As before, $d$ is the dimension of the space and $m$ is the number of prototypes used by the classifier. These upper bounds are based simply on the observation that for a reference set with $m$ prototypes there are at most $m(m-1)/2$ Voronoi cell boundaries, so the number of points which can be shattered by the set of Voronoi diagrams with $m$ centres is bounded above by the number of points which can be shattered by $m(m-1)/2$ hyperplanes. The stronger result for $d=2$ comes from a restriction on the number of edges of a planar graph, applied to the Delaunay triangulation which is the dual of the Voronoi diagram. These bounds imply the following result for the VC dimension of the NN classifier: \[Wbound\] For $m \geq 3$, $$\mathrm{VCdim}(\mathrm{1NN}(d,m)) \leq -\frac{q}{\log{2}} W_{-1} \left(-\frac{\log 2}{q} 2^{-\frac{m}{q}} \right),$$ where $$\begin{aligned} q &= 9(m-2) &\text{for } d = 2\\ q &= (d+1)m(m-1)/2 &\text{for } d \geq 3.\end{aligned}$$ $W$ is the Lambert W function; the $W_{-1}$ branch is the relevant branch. The logarithms are natural logarithms. See appendix \[LambertResult\]. We can obtain from this a looser but more easily interpretable upper bound by using a recent result which gives a lower bound for the $W_{-1}$ branch of the Lambert function: (Chatzigeorgiou) $$W_{-1}(-e^{-u-1}) > -1 - \sqrt{2u} - u,$$ for $u>0$. See [@Chatz13]. Using this result with proposition \[Wbound\] gives the following looser bound on the VC dimension: \[LooseCor\] Let $q' = q/\log{2}$, where $q$ is defined as in proposition \[Wbound\]. Then for $m \geq 3$, $$\label{looseupper} \mathrm{VCdim}(\mathrm{1NN}(d,m)) < q' \left( \sqrt{2\left(\frac{m}{q'} + \log{q'} -1\right)} + \frac{m}{q'} + \log{q'} \right).$$ This upper bound enables us to bound the the asymptotic rate of growth of $C$. Now, $q'$ grows monotonically with $m$ and with $d$, and grows at least as fast as $O(m)$ for increasing $m$. So whether considering growth with increasing $m$ or growth with increasing $d$, the fastest-growing term within the brackets of (\[looseupper\]) is $\log{q'}$. So we have $$\label{simq} \mathrm{VCdim}(\mathrm{1NN}(d,m)) \lesssim q' \log{q'},$$ for large $m$ or $d$. Table \[tab:results\] summarises the VC dimension bounds derived in this study. Discussion of asymptotic behaviour {#discussion} ================================== Considering first the rate of growth of the VC dimension with $m$ for fixed $d$, equation \[simq\] implies $$\label{Om2} \mathrm{VCdim}(\mathrm{1NN}(D,m)) \in O(m^2 \log(m^2)) = O(m^2 \log(m)),$$ for fixed $d=D>2$, with the better result $$\mathrm{VCdim}(\mathrm{1NN}(2,m)) \in O(m \log(m)),$$ for $d = 2$. Figure \[fig:VCBoundsFigure\] illustrates the upper bounds as a function of $m$, for the cases $d = 2$ and $d = 3$. ![Upper bounds for the VC dimension of NN-rule classifiers. “Accurate” upper bounds are the best we have obtained, given in Proposition \[Wbound\]. “Approximate” upper bounds are the looser bounds given in Corollary \[looseupper\].[]{data-label="fig:VCBoundsFigure"}](VCBoundsFigure){width="0.65\linewidth"} It is interesting to compare the log-linear growth in $m^2$ given by equation (\[Om2\]) with recent results for neural networks given by [@Harvey]. Consider a neural network with $d$ input neurons (the real coordinates of the feature space), and $N$ neurons in a single hidden layer with binary threshold activation functions. Each neuron in the hidden layer of such a network encodes a hyperplane decision boundary: the neuron will be in one or the other of its binary states depending on which side the input vector lies of a plane normal to the vector of weights of that neuron. So a network with $N=\frac{1}{2}m(m-1)$ hidden neurons is comparable to 1NN classifiers with $m$ prototypes, in the sense that both build their decision boundaries from parts of $\frac{1}{2}m(m-1)$ hyperplanes. [@Harvey] prove that a piecewise-linear neural network with $W$ parameters has VC dimension with asymptotic growth $O(W \log W)$ in the number of parameters. For the network just described, the number of parameters is $ \sim Nd = \frac{1}{2}m(m-1)d$, meaning the asymptotic growth of its VC dimension is $O(m^2d \log m^2d)$, just as relation (\[simq\]) gives as an upper bound for our 1NN classifiers. That is, the neural network achieves the (asymptotically) highest value it can for its VC dimension, given the number of hyperplane decision surfaces it has to work with. It is an interesting open question whether the same is true for 1NN classifiers. Turning now to consider the rate of growth with $d$ for fixed $m=M$, equation (\[simq\]) implies $$\mathrm{VCdim}(\mathrm{1NN}(d,M)) \in O(d \log(d)).$$ Thus, the VC dimension grows polynomially in $d$ (asymptotically slower than $d^2$). This has implications for learnability: for example, polynomial growth of the VC dimension of a class with the dimension of the space is a necessary condition for the class to be properly polynomially learnable [@Blumer89 Theorem 3.1.1.]. Conclusions =========== Type of bound Dimensionality Expression --------------- ---------------- --------------------------------------------------------------------------------------------------------------- Lower $d = 2$ $2m + 1$ Lower $d \geq 2$ $dm + 2 - d$ Upper $d = 2$ $-\frac{9(m-2)}{\log{2}} W_{-1} \left(-\frac{\log 2}{9(m-2)} 2^{-\frac{m}{9(m-2)}} \right)$ Upper $d \geq 2$ $-\frac{(d+1)m(m-1)}{2\log{2}} W_{-1} \left(-\frac{2\log 2}{(d+1)m(m-1)} 2^{-\frac{2m}{(d+1)m(m-1)}} \right)$ : Summary of the VC dimension bounds. $m$ denotes the number of prototypes in the reference set, and $d$ denotes dimensionality. $W_{-1}$ denotes the $-1$ branch of the Lambert W function.[]{data-label="tab:results"} The VC dimension for the set of all NN-rule classifiers in $d$-dimensional Euclidean space with a reference set of size $m$ grows at least as fast as $dm$ and not faster than $O(m^2 \log{m}$) as $m$ increases. For the case of two-dimensional Euclidean space, the VC dimension grows not faster than $O(m \log{m}$). Considering instead growth with $d$, the VC dimension for the set of all NN-rule classifiers in $d$-dimensional Euclidean space with a reference set of size $m$ grows at least as fast as $d(m-1)$ and not faster than $O(d \log{d}$) as $d$ increases. Precise lower and upper bounds for this VC dimension are given in our Corollary \[LowerBoundProp\] and Proposition \[Wbound\] respectively, and summarised in Table \[tab:results\]. The consequence of these bounds that is of interest to practitioners is the implication for the size of sample needed to learn an accurate classifier. In the Probably Approximately Correct learning model, the *sample complexity* is the number of training examples needed to learn a classifier of given accuracy with given probability. The sample complexity of a family of classifiers is known to depend linearly on the VC dimension [@Shais Theorem 6.8]. Therefore, the bounds we give above for the asymptotic growth of the VC dimension are also bounds on the asymptotic growth of the size of the training data set needed to learn (with given probability) an accurate NN classifier with reference set of given size. The lower bound applies only to classification algorithms which produce reference sets of given fixed size (and can produce *any* reference set of that size). The upper bound is significantly more broadly applicable. The upper bound applies to any NN classification algorithm which may not have more than $m$ points in its reference set. Thus we may conclude: the size of the training set needed to learn an accurate NN-rule classifier with reference set of size $m$ in $d$-dimensional Euclidean space grows not faster than $O(m^2 \log{m}$) as $m$ increases. For the case of two-dimensional Euclidean space, the size of the training set required grows not faster than $O(m \log{m}$). Considering instead growth with $d$ for fixed $m$, the size of the training set required grows not faster than $O(d \log{d}$). The fact that the growth rate of the upper bound is asymptotically faster (with $m$) for $d \geq 3$ than for $d=2$ raises the interesting possibility that there may be something fundamentally different about the behaviour of NN-rule classifiers in 3 dimensions and higher from their behaviour in two-dimensional space. If future work were to establish an $O(m^2)$ lower bound for $d \geq 3$, this would be confirmed. If, instead, an $O(m \log(m))$ upper bound were established, implying the same behaviour for the 1NN classifier in higher dimensions as in 2 dimensions, this would imply instead an interesting discrepancy between the behaviour of 1NN classifiers and the behaviour of neural networks with access to an equal number of hyperplanes from which to construct their decision boundaries, as discussed in section \[discussion\]. Acknowledgment {#acknowledgment .unnumbered} ============== This work was done under project RPG-2015-188 funded by The Leverhulme Trust, UK. While preparing the paper for publication, IG received support from the European Union’s Horizon 2020 research and Innovation programme under grant agreement No 731593. [17]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{} A Blumer, A Ehrenfeucht, D Haussler, and K Warmuth. Learnability and the [V]{}apnik-[C]{}hervonenkis dimension. *J. Assoc. Computing Machinery*, 360 (4):0 929–965, 1989. Ioannis Chatzigeorgiou. Bounds on the [L]{}ambert function and their application to the outage analysis of user cooperation. *IEEE Communications Letters*, 170 (8):0 1505–1508, 2013. D. Corfield, B. Sch[ö]{}lkopf, and V. Vapnik. Falsificationism and statistical learning theory: Comparing the [P]{}opper and [V]{}apnik-[C]{}hervonenkis dimensions. *J. General Philosophy of Science*, 40:0 51–58, 2009. T. M. Cover. Geometrical and statistical properties of systems of linear inequalities with applications in pattern recognition. *IEEE Trans. Electronic Computers*, 3:0 326–334, 1965. L Devroye, L Gy[ö]{}rfi, and G Lugosi. *A probabilistic theory of pattern recognition*. Springer-Verlag, New York, 1996. R Dudley. Balls in $\mathbb{R}^k$ do not cut all subsets of $k+2$ points. *Advances in Math.*, 31:0 306–308, 1979. E. Fix and J. L. Hodges. Discriminatory analysis : Non parametric discrimination : Small sample performance. Technical Report Project 21 - 49 - 004 (11), USAF School of Aviation Medicine, Randolph Field, Texas, 1952. Salvador Garcia, Joaqu[í]{}n Derrac, Jos[é]{} Ram[ó]{}n Cano, and Francisco Herrera. Prototype selection for nearest neighbor classification: Taxonomy and empirical study. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 340 (3):0 417–435, 2012. Iain A. D. Gunn, Álvar Ãrnaiz-González, and Ludmila I. Kuncheva. A taxonomic look at instance-based stream classifiers. *Neurocomputing*, under review. Nick Harvey, Christopher Liaw, and Abbas Mehrabian. Nearly-tight [VC]{}-dimension bounds for piecewise linear neural networks. *Proceedings of Machine Learning Research*, 65, 2017. S. B. Holden and M. Niranjan. On the practical applicability of [VC]{} dimension bounds. *Neural Computation*, 70 (6):0 1265–1288. Bilge Karaçali and Hamid Krim. Fast minimization of structural risk by nearest neighbor rule. *IEEE Trans. Neural Networks*, 140 (1):0 127–137, 2002. Shai Shalev-Shwartz and Shai Ben-David. *Understanding Machine Learning*. Cambridge University Press, New York, 2014. G[á]{}bor Tak[á]{}cs. The [V]{}apnik-[C]{}hervonenkis dimension of convex $n$-gon classifiers. *Hungarian Electronic J. Sciences*, 2007. G[á]{}bor Tak[á]{}cs and B[é]{}la Pataki. Lower bounds on the [V]{}apnik-[C]{}hervonenkis dimension of convex polytope classifiers. *Proc. IEEE Int’l Conf. Intelligent Eng. Systems*, pages 145–148, 2007. Isaac Triguero, Joaqu[í]{}n Derrac, Salvador Garcia, and Francisco Herrera. A taxonomy and experimental study on prototype generation for nearest neighbor classification. *IEEE Trans. Systems, Man, and Cybernetics, Part C: Applications and Reviews*, 420 (1):0 86–100, 2012. V. N. Vapnik. *Statistical Learning Theory*. Wiley-Interscience, New York, 1995. Proof of Lemma \[subset\] ========================= We will show by construction that for an arbitrary convex polytope $P$ with $N$ facets, there is a labelled set of $N+1$ points such that an NN classifier using this set as a reference set will have $P$ for its decision boundary. The reference set is constructed as follows: place one point, with the label of the interior region, anywhere in the interior of the polytope $P$. The remaining $N$ points, with the opposite label, are placed as the reflections of this point in the $N$ hyperplanes which contain the $N$ facets of $P$. The $i$th hyperplane is therefore the locus of points equidistant from the interior point and the $i$th exterior point. As the polytope is convex, no part of the interior of the polytope lies on the side of the $i$th hyperplane closer to the exterior point (in particular, the placement of the $i$th exterior point does not impact any boundary facet other than that formed by the $i$th hyperplane). Conversely, if a point is on the interior side of all $N$ hyperplanes, then it is in the interior of the polytope. Thus, by construction, the set of points closer to the interior point than to any exterior point is the intersection of the $N$ half-spaces which are closer to the interior point than to the respective exterior points. That is, the Voronoi cell of the interior point is the polytope $P$: the decision boundary using the NN classifier is $P$. Each convex $N$-faceted $d$-tope classifier is therefore a NN classifier with a reference set of size $N+1$. The set of all such polytope classifiers is therefore a subset of the set of all such NN classifiers. Proof of Lemma \[Takacs1\] {#TakacsAppendix} ========================== To prove that the VC dimension of the $N$-gon classifier is at least $2N+2$, we must show that there exists an arrangement of $2N+2$ points which can be correctly labelled using an $N$-gon decision boundary, for any possible partitioning of the points into two classes (i.e., the $2N+2$ points can be shattered). The $2N+2$ points to be labelled are arranged as follows: place $2N + 1$ points on a circle, with the final point in the centre. Then an $N$-gon can be constructed to include any subset of the points on the circle, together with the centre point, excluding the other points. The worst case is that, going round the circle, there are $N$ separate groups of points (of size one or two) of the opposite label to the centre point; the $N$ edges of the $N$-gon may be arranged to exclude one of these $N$ groups each. It is always possible to place a line separating a sequence of points around the circle from the rest of the points on the circle, as the points on the circle are in convex position. The case where a sequence of points of opposite label to the centre point extends more than half-way round the circle, as in Figure \[fig:Takacs\](a), must be considered separately but presents no difficulties. ![Illustration of an arrangement of 6 points which can be shattered by an unbounded “polygon” with $N=2$ sides. Plots (a), (b), and (c) illustrate examples of, respectively, a 1/5, a 2/4, and a 3/3 partition.[]{data-label="fig:Takacs"}](TakacsFigure "fig:"){width="0.9\linewidth"} (a) ------------------------------------------------------------------------ (b) ------------------------------------------------------------------------ (c) In the $N=2$ case the decision boundary is two half-lines radiating from one point; in Tak[á]{}cs’ terminology a “polygon” need not be bounded. Figure \[fig:Takacs\] illustrates examples of partitionings in the case of $N=2$ and six points. Any partition of the points into two groups can be constructed with the V-shaped border. @Devroye96 [p. 224] make a similar construction with polygons and points on a circle, but use it only to prove that the class of all convex polygons has infinite VC dimension. Proof of proposition \[OneBetter\] {#appendix1} ================================== We begin with some geometrical preliminaries, as illustrated in Figure \[fig:PolygonGeometry\]. A diagonal of a polygon is a line segment joining two non-adjacent vertices. The vertices of a regular polygon all lie on the same circle, called the circumcircle of the polygon, and are equally spaced around it. Each vertex of a regular $(2m-1)$-gon (for $m \geq 3$) is associated with two “longest diagonals”, which connect the vertex in question to the vertices which are furthest away from it (both around the perimeter and in Euclidean metric). The two longest diagonals are reflections of each other in the diameter of the circumcircle which passes through the given vertex. ![Geometrical illustration for Lemma \[geometry1\].[]{data-label="fig:PolygonGeometry"}](PolygonGeometry){width="0.6\linewidth"} \[geometry1\] No vertices of a regular $(2m-1)$-gon lie between the line of a longest diagonal of the $(2m-1)$-gon and a line through the centre of the polygon which is parallel to this diagonal. Recall that vertices of a regular polygon must all lie on the same circumcircle. If there were a vertex between the longest diagonal and a diameter of the circumcircle which does not intersect it, then a line from this vertex to the end of the longest diagonal further from it would be a diagonal longer than the longest diagonal. We will now introduce the arrangement of points which we will subsequently argue can be shattered by a NN-rule classifier with $m$ prototypes. \[arrangement1\] (Figure \[LemmaArrangement\].) Place $2m-1$ points as the vertices of a regular $(2m-1)$-gon, $m \geq 4$. Two further points are placed on a line which runs through the centre of the $(2m-1)$-gon, and which is perpendicular to a line from one vertex (marked ‘$A$’) through the centre. The two points are placed on this line at an equal distance either side of the centre. This distance is sufficiently small that neither of the two points is separated from the centre by any of the diagonals of the $(2m-1)$-gon. This is equivalent to requiring that the two points lie within the smaller $(2m-1)$-gon formed by the central sections of the longest diagonals of the original $(2m-1)$-gon, between their mutual intersections. ![Arrangement of 9 points which can be shattered by a NN-rule classifier with $m=4$ prototypes, but not by the set of classifiers whose decision boundary is a triangle. 7 of the points are vertices of a (convex) regular heptagon, which is shown together with its circumcircle and longest diagonals. Also shown are segments of two lines used in the construction of the remaining two points, and the centre point.[]{data-label="LemmaArrangement"}](LemmaArrangement){width="0.55\linewidth"} We now have two lemmata which establish some geometric properties of this arrangement of points. Lemma \[arrangement\] will show that there are certain sets of three points, which include one of the two centre points but not the other, which can be separated from the remaining points by two parallel lines. This will be important when we come to demonstrate that the set of points can be shattered, because the labellings which are difficult to achieve are those for which the central points do not both have the same label. \[arrangement\] Let $2m+1$ points be placed as Arrangement \[arrangement1\], $m \geq 4$. Call one of the inner points $B$, and the other $W$. Let $C$ be the vertex nearest to $W$. Then, given any choice $D$ of vertex other than $C$, there exist vertices $D_1, D_2$, not adjacent to $D$, such that $\{D,D_i,B\}$ can be separated from all other points by two parallel lines. (Figure \[fig:LemmaArrangementParallelLines\].) -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ![Given a choice B from the two centre points, and a choice D of any vertex apart from C, there are two sets of three points, containing the two chosen points, which can be separated from the remaining points by two parallel lines. The heptagon illustrated in (b) shows the worst case for the separation of these lines, at just under two-thirds of the circumradius. The lines may be placed closer together if $m$ is greater or if $D_1$ is chosen instead of $D_2$. Note that $l'$ (resp. $l''$) may be placed arbitrarily close to diagonal $DD_1$ (resp. $DD_2$).[]{data-label="fig:LemmaArrangementParallelLines"}](LemmaArrangementParallelLines1 "fig:"){width="0.5\linewidth"} ![Given a choice B from the two centre points, and a choice D of any vertex apart from C, there are two sets of three points, containing the two chosen points, which can be separated from the remaining points by two parallel lines. The heptagon illustrated in (b) shows the worst case for the separation of these lines, at just under two-thirds of the circumradius. The lines may be placed closer together if $m$ is greater or if $D_1$ is chosen instead of $D_2$. Note that $l'$ (resp. $l''$) may be placed arbitrarily close to diagonal $DD_1$ (resp. $DD_2$).[]{data-label="fig:LemmaArrangementParallelLines"}](LemmaArrangementParallelLines2 "fig:"){width="0.5\linewidth"} \(a) $\{D,B,D_1\}$ separated. \(b) $\{D,B,D_2\}$ separated. -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Denote the diameter of the circumcircle which passes through point $D$ by $d$. One of the two longest diagonals from $D$ passes on the same side of $d$ as $B$, and the other passes on the side of $W$. One of the two points which we will prove has the desired property is the point at the opposite end of that longest diagonal from D which passes on the side of $B$; call this point $D_1$. We will construct two parallel lines such that $D$, $B$, and $D_1$ are between the lines, and the $2m-2$ remaining points are not (Figure \[fig:LemmaArrangementParallelLines\](a)). The required lines are parallel to $DD_1$. One of the lines is a diameter, $d'$. By Lemma \[geometry1\] there is no vertex between $d'$ and $DD_1$. The second parallel line, $l'$, is placed on the far side of $DD_1$ from the centre. It can be placed arbitrarily close to $DD_1$, so it is always possible to place it such that that no vertices lie between $l'$ and $DD_1$. Therefore the only points between $d'$ and $l'$ are $B$, $D$, and $D_1$. $D_2$ is the vertex adjacent to $D_1$ on the side of $B$ (Figure \[fig:LemmaArrangementParallelLines\]). The set $\{D,D_2,B\}$ is separated from the rest of the points in the construction by two parallel lines as follows. Consider the line of symmetry $s$ of the polygon for which $D_2$ is the reflection of $D$. The two lines required are perpendicular to $s$ (unless $s$ passes through the vertex defined in Arrangement \[arrangement1\] as $A$, which case we discuss shortly). The first line, $d''$ passes through the centre of the polygon. The second line, $l''$, is placed further from the centre than $DD_2$, a sufficiently small distance from $DD_2$ that no vertices lie between it and $DD_2$. We now argue that no vertices lie between $d''$ and $DD_2$. Consider the longest diagonal from $D_1$ which does not go to $D$. Call this longest diagonal $c''$. Now suppose there were a vertex on the arc of the circumcircle between $D_1$ and $D_2$: then $D_2$ would not be an adjacent vertex to $D_1$, which it is by definition. Suppose instead there were a vertex on the arc of the circle between $D$ and the vertex at the other end of $c''$ from $D_1$. Then a line segment from this vertex to $D_1$ would be a diagonal longer than the longest diagonal. Therefore there are no vertices between $c''$ and $DD_2$. Now, $d''$ is parallel to $DD_2$, therefore, by the symmetry of the regular $(2m-1)$-gon in $s$, it is parallel to $c''$, and it is closer to $DD_2$ than $c''$ is, since it passes through the centre. So the region between $d''$ and $DD_2$ is entirely contained in the region between $c''$ and $DD_2$, so the fact that there is no vertex in the latter region implies there is no vertex in the former region. If $BW$ is perpendicular to $s$, then $d''$ and $l''$ are placed at a suitably small angle (it may be arbitrarily small) to the perpendicular to the line of symmetry, such that $B$ falls between the lines and $W$ does not. We shall need to know the width of the strip constructed in the previous lemma to contain $D$, $B$, and $D_2$. \[widthlemma\] Let $2m+1$ points be placed as Arrangement \[arrangement1\], $m \geq 4$. Let $D$ and $D_2$ be defined as in Lemma \[arrangement\] (Figure \[fig:LemmaWidth\]). Then the distance from the centre of the polygon to $DD_2$ is less than 0.63 times the circumradius. ![Greatest possible distance from $DD_2$ to centre[]{data-label="fig:LemmaWidth"}](LemmaWidth){width="0.7\linewidth"} By an elementary geometrical argument, the angle between the line passing through the centre and $D$, and $d''$, is $$\theta = \dfrac{3 \pi}{2(2m-1)}.$$ Therefore the distance between $DD_2$ and the centre is $$\mathrm{Width} = R\sin\dfrac{3 \pi}{2(2m-1)},$$ where $R$ is the circumradius. This width decreases monotonically with $m$. The largest value it may take in our construction is therefore associated with the smallest value of $m$ we consider, $m=4$. In this case, $$\mathrm{Width} = R\sin \dfrac{3 \pi}{14} < 0.63R.$$ We are now in a position to prove the proposition. $$\mathrm{VCdim}(\mathrm{1NN}(2,m)) \geq 2m + 1$$ for $m \geq 4$. We shall prove that the set of $2m+1$ points placed as Arrangement \[arrangement1\] can be shattered by NN-rule classifiers using $m$ prototypes. That is, we will prove that all labellings of these points can be correctly classified by a NN-rule classifier with $m$ prototypes in the reference set. If the two inner points have the same label, then the situation is the same as described for the polygon classifier in Appendix \[TakacsAppendix\]: an $(m-1)$-gon can be constructed containing the two central points and all the points on the circle of the same label, just as if there were only one central point. Consider now the case where the two interior points are of different labels. We will refer to the class less-represented among a given labelling of the $2m+1$ points as “black”, and the other class as “white”. The black class, by this definition, contains at most $m$ points. It remains to prove that all labellings in which the interior points have different colours can be labelled by a NN-rule classifier with $m$ prototypes. Identify the black and white interior point with the points B and W respectively of Lemma \[arrangement\]. Point C is then defined as in that lemma. Suppose there are $m$ points in the “black” class, of which $m-1$ are vertices of the polygon. Since $m \geq 4$, two of the $m$ black points must then be vertices of the polygon other than C. Let one of these these vertices be P, and define $P_1$ and $P_2$ to have the same relation to P as $D_1$ and $D_2$ have to $D$ in Lemma \[arrangement\] (Figure \[fig:PropositionSeparation\]). If either $P_1$ or $P_2$ is black, then that point, with P and the black centre point, can be separated from the remaining points by two parallel lines as described in Lemma \[arrangement\]. If both $P_1$ and $P_2$ are black, pick one of them arbitrarily to make the construction. By Lemma \[widthlemma\], the distance between these parallel lines is less than two thirds of the circumradius. These parallel lines can therefore be achieved as a decision surface by three prototypes all of which are inside the circle: a black prototype is placed on the line of the symmetry of the polygon which is perpendicular to these two lines, halfway between them, and therefore less than one third of the circumradius away from the centre of the circle. Two white prototypes are placed as the reflections of this black prototype in the two parallel lines; since the black prototype is less than one-third of the radius away from the centre, the white prototypes are within the circle. Each of the remaining $m-3$ black vertices can be separated from the rest of the points by a single line (due to the convexity of the polygon, as in the argument for the polygon classifier). This line can be achieved as a decision surface by placing a black prototype as the reflection in this line of the nearer white prototype. All vertices not so separated and not contained between the two parallel lines first constructed will be classified as white by the two white prototypes. Thus the $2m+1$ points are classified correctly by $m$ prototypes, 3 within the polygon and $m-3$ outside it. ![Example labelling of 9 points by 4 prototypes. $P$, $B$, and $P2$ are labelled by a black prototype inside the circle. They are separated from all other points by parallel decision surfaces created by two white prototypes which are also inside the circle. The remaining black vertices (one such, in this case) are labelled by one black prototype each, placed outside the circle so as to be the nearest prototype only for the relevant vertex.[]{data-label="fig:PropositionSeparation"}](PropositionSeparationFixed){width="0.65\linewidth"} Now consider the case in which neither $P_1$ nor $P_2$ is black. In this case, $m-1$ black labels must be distributed among the $2m-3$ vertices which are neither $P_1$ nor $P_2$. Consider the case where no two adjacent vertices are both black. If no two adjacent vertices are both black, then there is only one possible arrangement (since $P_1$ and $P_2$ are adjacent): The vertices either side of the $\{P_1,P_2\}$ pair must both be black, and every second vertex between them (going around the circle) must be black. Let $Q$ be a black vertex in this arrangement which is neither $P$ nor $C$ (recall that there must be at least one such vertex, because $m \geq 4$). Let the two points $Q_1$, $Q_2$ be the two points which have the same relation to $Q$ as $D_1$ and $D_2$ have to $D$ in the proof of Lemma \[arrangement\]. At most one of $Q_1$ and $Q_2$ can be an element of $\{P_1,P_2\}$, because $Q$ is not $P$. Therefore, since $Q_1$ and $Q_2$ are adjacent vertices, one of them must be black, and can therefore be separated along with $Q$ and the black centre point by two parallel lines, allowing correct classification of all the points as in the case where $P_1$ or $P_2$ was black. Thus, either the points can be classified this way, or there are two adjacent vertices which are both black. Two cases now remain to be considered. First, the case in which neither $P_1$ nor $P_2$ is black, but the remaining vertices do not have the only labelling which ensures no two adjacent vertices are black. Second, the case in which there are fewer than $m$ points in the black class. The construction is essentially the same for both of these cases. The construction proceeds as previously, but it is only necessary to contain two black points between the parallel lines. $B$ and one of the black vertices not part of the adjacent pair (if there is only one pair, or not the common vertex between two pairs if there are precisely three black vertices forming a run of three) are separated by two parallel lines (an arbitrarily small distance either side of the line containing these two points); these lines are achieved as a decision surface by three prototypes within the circle, a black prototype between two white prototypes as before. The remaining $m-3$ prototypes are black, and are placed as reflections of the nearer white prototype in a line separating a single black vertex or an adjacent pair of black vertices from the other points. Not all $m-3$ prototypes may be needed for this purpose, but unneeded prototypes may be placed at a large distance from the points so as not to affect the classification. Proof of Proposition \[Wbound\] {#LambertResult} =============================== From the definition of VC dimension in terms of shatter coefficient, $h(C)$ is given by the largest $n$ solving $$S(C,n) = 2^n. \label{VCdef}$$ Now, if a continuous function $f(n)$ satisfies $$f(n) \geq S(C,n) \quad \forall n,$$ and also $f(n)$ grows more slowly than $2^n$ asymptotically, then the largest $n$ solving $$f(n) = 2^n \label{feq}$$ must be at least as large as the largest $n$ solving (\[VCdef\]). The bounds of Lemma \[DevroyeLemma\] satisfy these conditions, so we may substitute them into (\[feq\]) to find that $h(C)$ is bounded above by the largest $n$ solving $$2^m n^q = 2^n. \label{VCeq}$$ Equation (\[VCeq\]) is solved by $$n = -\frac{q}{\log{2}} W \left(-\frac{\log 2}{q} 2^{-\frac{m}{q}} \right), \label{Wresult}$$ where the solution on the $W_{-1}$ branch is the largest real solution, which is therefore the upper bound.

ArXiv

--- bibliography: - 'MoDMaps.bib' --- [**Mapping the Space of Genomic Signatures** ]{}\ Lila Kari$^{1,\ast}$, Kathleen A. Hill$^{1, 2}$, Abu S. Sayem$^1$, Rallis Karamichalis$^1$, Nathaniel Bryans$^1$, Katelyn Davis$^2$, Nikesh S. Dattani$^3$\ **[1]{} Department of Computer Science, University of Western Ontario, London, ON, N6A 5B7, Canada\ **[2]{} Department of Biology, University of Western Ontario, London, ON, N6A 5B7, Canada\ **[3]{} Physical and Theoretical Chemistry Laboratory, Department of Chemistry, Oxford University, Oxford, OX1 3QZ, United Kingdom\ $\ast$ E-mail: [email protected]****** Abstract {#abstract .unnumbered} ======== We propose a computational method to measure and visualize interrelationships among any number of DNA sequences allowing, for example, the examination of hundreds or thousands of complete mitochondrial genomes. An “image distance” is computed for each pair of graphical representations of DNA sequences, and the distances are visualized as a Molecular Distance Map: Each point on the map represents a DNA sequence, and the spatial proximity between any two points reflects the degree of structural similarity between the corresponding sequences. The graphical representation of DNA sequences utilized, Chaos Game Representation (CGR), is genome- and species-specific and can thus act as a genomic signature. Consequently, Molecular Distance Maps could inform species identification, taxonomic classifications and, to a certain extent, evolutionary history. The image distance employed, Structural Dissimilarity Index (DSSIM), implicitly compares the occurrences of oligomers of length up to $k$ (herein $k=9$) in DNA sequences. We computed DSSIM distances for more than 5 million pairs of complete mitochondrial genomes, and used Multi-Dimensional Scaling (MDS) to obtain Molecular Distance Maps that visually display the sequence relatedness in various subsets, at different taxonomic levels. This general-purpose method does not require DNA sequence homology and can thus be used to compare similar or vastly different DNA sequences, genomic or computer-generated, of the same or different lengths. We illustrate potential uses of this approach by applying it to several taxonomic subsets: phylum Vertebrata, (super)kingdom Protista, classes Amphibia-Insecta-Mammalia, class Amphibia, and order Primates. This analysis of an extensive dataset confirms that the oligomer composition of full mtDNA sequences can be a source of taxonomic information. This method also correctly finds the mtDNA sequences most closely related to that of the anatomically modern human (the Neanderthal, the Denisovan, and the chimp), and that the sequence most different from it belongs to a cucumber. Introduction {#introduction .unnumbered} ============ Every year biologists discover and classify thousands of new species, with an average of 18,000 new species named each year since 1940 [@IISE2012], [@NationalGeographic2014]. Other findings, [@Mora2011], suggest that as many as 86% of existing species on Earth and 91% of species in the oceans have not yet been classified and catalogued. In the absence of a universal quantitative method to identify species’ relationships, information for species classification has to be gleaned and combined from several sources, morphological, sequence-alignment-based phylogenetic anaylsis, and non-alignment-based molecular information. We propose a computational process that outputs, for any given dataset of DNA sequences, a concurrent display of the structural similarities among all sequences in the dataset. This is obtained by first computing an “image distance” for each pair of graphical representations of DNA sequences, and then visualizing the resulting interrelationships in a two-dimensional plane. The result of applying this method to a collection of DNA sequences is an easily interpretable [*Molecular Distance Map*]{} wherein sequences are represented by points in a common Euclidean plane, and the spatial distance between any two points reflects the differences in their subsequence composition. The graphical representation we use is [*Chaos Game Representation*]{} (CGR) of DNA sequences, [@Jeffrey1990; @Jeffrey1992], that simultaneously displays all subsequence frequencies of a given DNA sequence as a visual pattern. CGR has a remarkable ability to differentiate between genetic sequences belonging to different species, and has thus been proposed as a [*genomic signature*]{}. Due to this characteristic, a Molecular Distance Map of a collection of genetic sequences may allow inferrences of relationships between the corresponding species. Concretely, to compute and visually display relationships in a given set $S = \{s_1, s_2, ..., s_n \}$ of $n$ DNA sequences, we propose the following computational process: \(i) [*Chaos Game Representation*]{} (CGR), to graphically represent all subsequences of a DNA sequence $s_i$, $1\leq i\leq n$, as pixels of one image, denoted by $c_i$; \(ii) [*Structural Dissimilarity Index*]{} (DSSIM), an “image-distance” measure, to compute the pairwise distances $\Delta (i, j)$, $1\leq i, j \leq n$, for each pair of CGR images $(c_i, c_j)$, and to produce a distance matrix; \(iii) [*Multi-Dimensional Scaling*]{} (MDS), an information visualization technique that takes as input the distance matrix and outputs a Molecular Distance Map in 2D, wherein each plotted point $p_i$ with coordinates $(x_i, y_i)$ represents the DNA sequence $s_i$ whose CGR image is $c_i$. The position of the point $p_i$ in the map, relative to all the other points $p_j$, reflects the distances between the DNA sequence $s_i$ and the other DNA sequences $s_j$ in the dataset. We apply this method to analyze and visualize several different taxonomic subsets of a dataset of 3,176 complete mtDNA sequences: phylum Vertebrata, (super)kingdom Protista, classes Amphibia-Insecta-Mammalia, class Amphibia only, and order Primates. We illustrate the usability of this approach by discussing, e.g., the placement of the genus [*Polypterus*]{} within phylum Vertebrata, of the unclassified organism [*Haemoproteus*]{} sp. jb1.JA27 (\#1466) within the (super)kingdom Protista, and the placement of the family Tarsiidae within the order Primates. We also provide an interactive web tool, [*MoD-Map*]{}, that allows an in-depth exploration of all Molecular Distance Maps in the paper, complete with zoom-in features, search options, and easily accessible additional information for each species-point sequence. Overall, this method groups mtDNA sequences in correct taxonomic groups, from the kingdom level down to the order and family level. These results are of interest both because of the size of the dataset analyzed and because this information was extracted from DNA sequences that normally would not be considered homologous. Our analysis confirms that sequence composition (presence or absence of oligomers) contains taxonomic information that could be relevant to species identification, taxonomic classification, and identification of large evolutionary lineages. Last but not least, the appeal of this method lies in its simplicity, robustness, and generality, whereby the exact same measuring tape is able to automatically yield meaningful measurements between non-specific DNA sequences of species as distant as those of the anatomically modern human and a cucumber, and as close as those of the anatomically modern human and the Neanderthal. Methods {#methods .unnumbered} ======= A CGR [@Jeffrey1990; @Jeffrey1992] associates an image to each DNA sequence as follows. Begin with a unit square with corners labelled [*A, C, G,*]{} and [*T*]{}, clockwise starting from the bottom-left corner. The first point of any CGR plot is the center of the square. To plot the CGR corresponding to a given DNA sequence, start reading the letters of the sequence from left to right, one by one. The point corresponding to the first letter is the point plotted in the middle of the segment determined by the center of the square and the corner labelled by the first letter. For example, if the center of the square is labelled “O” and the first letter of the sequence is “A”, then the point of the plot corresponding to the first “A” is the point situated halfway between O and the corner A. Subsequent letters are plotted iteratively as the middle point between the previously-drawn-point and the corner labelled by the letter currently being read. CGR images of genetic DNA sequences originating from various species show rich fractal patterns containing various motifs such as squares, parallel lines, rectangles, triangles and diagonal crosses, see, e.g., Figure \[fig:CGR\]. CGRs of genomic DNA sequences have been shown to be genome and species specific, [@Jeffrey1990; @Jeffrey1992; @Hill1992; @Hill1997; @Deschavanne1999; @Deschavanne2000; @Wang2005]. Thus, sequences chosen from each genome as a basis for computing “distances” between genomes do not need to have any relation with one another from the point of view of their position or information content. In addition, this graphical representation facilitates easy visual recognition of global string-usage characteristics: Prominent diagonals indicate purine or pyrimidine runs, sparseness in the upper half indicates low G+C content, etc., see, e.g., [@Deschavanne1999]. If the generated CGR image has a resolution of $2^k\times 2^k$ pixels, then every pixel represents a distinct DNA subsequence of length $k$: A pixel is black if the subsequence it represents occurs in the DNA sequence, otherwise it is white. In this paper, for the CGR images of all 3,176 complete mtDNA sequences in our dataset, we used the value $k = 9$, that is, occurrences of subsequences of lengths up to 9 were being taken into consideration. In general, a length of the DNA sequence of 4,000 bp is necessary to obtain a sharply defined CGR, but in many cases 2,000 bp give a reasonably good approximation, [@Jeffrey1990]. In our case, we used the full length of all analyzed mtDNA sequences, which ranged from 288 bp to 1,555,935 bp, with an average of 28,000 bp. Other visualizations of genetic data include the 2D rectangular walk [@Gates1986] and methods similar to it in [@Nandy1994],[@Leong1995], vector walk [@Liao2005], cell [@Yao2004], vertical vector [@Yu2010], Huffman coding [@Qi2011], and colorsquare [@Zhang2012] methods. Three-dimensional representations of DNA sequences include the tetrahedron [@Randic2000], 3D-vector [@Yuan2003], and trinucleotide curve [@Yu2009] methods. Among these visualization methods, CGR images arguably provide the most immediately comprehensible “signature” of a DNA sequence and a desirable genome-specificity, [@Jeffrey1990; @Deschavanne2000]. In addition, the images produced using CGR are easy to compare, visually and computationally. Coloured versions of CGR, wherein the colour of a point corresponds to the frequency of the corresponding oligomer in the given DNA sequence (from red for high frequency, to blue for no occurrences) have also been proposed [@Makula2009; @Hao2000]. Note that other alignment-free methods have been used for phylogenetic analysis of DNA strings, such as computing the Euclidean distance between frequencies of $k$-mers ($k \leq 5$) for the analysis of 125 GenBank DNA sequences from 20 bird species and the American alligator, [@Edwards2002]. Another study, [@Deschavanne2010], analyzed 459 dsDNA bacteriophage genomes and compared them with their host genomes to infer host-phage relationships, by computing Euclidean distances between frequencies of $k$-mers for $k=4$. In [@Pandit2010], 75 complete HIV genome sequences were compared using the Euclidean distance between frequencies of 6-mers ($k=6$), in order to group them in subtypes. In [@Pride2003], 27 microbial genomes were analyzed to find implications of 4-mer frequencies ($k=4$) on their evolutionary relationships. In [@Li2004], 20 mammalian complete mtDNA sequences were analyzed using the “similarity metric”. Our method uses a larger dataset (3,176 complete mtDNA sequences), an “image distance” measure that was designed to capture structural similarities between images, as well as a value of $k=9$. Structural Similarity (SSIM) index is an image similarity index used in the context of image processing and computer vision to compare two images from the point of view of their structural similarities [@Wang2004]. SSIM combines three parameters - luminance distortion, contrast distortion, and linear correlation - and was designed to perform similarly to the human visual system, which is highly adapted to extract structural information. Originally, SSIM was defined as a similarity measure $s(A,B)$ whose theoretical range between two images $A$ and $B$ is $[-1,1]$ where a high value amounts to close relatedness. We use a related [*DSSIM distance*]{} $\Delta(A,B) = 1 - s(A,B) \in [0,2]$, with the distance being 0 between two identical images, 1 between e.g. a black image and a white image, and 2 if the two images are negatively correlated; that is, $\Delta(A,B) = 2$ if and only if every pixel of image $A$ has the inverted value of the corresponding pixel in image $B$ while both images have the same luminance (brightness). For our particular dataset of genetic CGR images, almost all (over 5 million) distances are between 0 and 1, with only half a dozen exceptions of distances between 1 and 1.0033. MDS has been used for the visualization of data relatedness based on distance matrices in various fields such as cognitive science, information science, psychometrics, marketing, ecology, social science, and other areas of study [@BorgGroenen2010]. MDS takes as input a distance matrix containing the pairwise distances between $n$ given items and outputs a two-dimensional map wherein each item is represented by a point, and the spatial distances between points reflect the distances between the corresponding items in the distance matrix. Notable examples of molecular biology studies that used MDS are [@Lessa1990] (where it was used for the analysis of geographic genetic distributions of some natural populations), [@Hebert2003] (where it was used to provide a graphical summary of the distances among CO1 genes from various species), and [@Hillis2005] (where it was used to analyze and visualize relationships within collections of phylogenetic trees). Classical MDS, which we use in this paper, receives as input an $n\times n$ distance matrix $(\Delta(i, j))_{1\leq i, j \leq n}$ of the pairwise distances between any two items in the set. The output of classical MDS consists of $n$ points in a $q$-dimensional space whose pairwise spatial (Euclidean) distances are a linear function of the distances between the corresponding items in the input distance matrix. More precisely, MDS will return $n$ points $p_{1},p_{2},\ldots,p_{n}\in \mathbb{R}^{q}$ such that $d(i, j)= ||p_{i}-p_{j}||\thickapprox f(\Delta(i,j))$ for all $i,j\in \{1, \ldots, n \}$ where $d(i, j)$ is the spatial distance between the points $p_i $ and $p_j$, and $f$ is a function linear in $\Delta(i, j)$. Here, $q$ can be at most $n-1$ and the points are recovered from the eigenvalues and eigenvectors of the input $n\times n$ distance matrix. If we choose $q=2$ (respectively $q=3$), the result of classic MDS is an approximation of the original $(n-1)$-dimensional space as a two- (respectively three-) dimensional map. In this paper all Molecular Distance Maps consist of coloured points, wherein each point represents an mtDNA sequence from the dataset. Each mtDNA sequence is assigned a unique numerical identifier retained in all analyses, e.g., \#1321 is the identifier for the [*Homo sapiens sapiens*]{} mitochondrial genome. The colour assigned to a sequence-point may however vary from map to map, and it depends on the taxon assigned to the point in a particular Molecular Distance Map. For example, in Figure \[vertebrates\] all mammalian mtDNA sequence-points are coloured red, while in Figure \[primatesonly\] the red points represent mtDNA sequences from the primate suborder Haplorhini and the green points represent mtDNA sequences from the primate suborder Strepshirrini. For consistency, all maps are scaled so that the $x$- and the $y$-coordinates always span the interval $[-1, 1]$. The formula used for scaling is $x_{\rm{sca}} =2 \cdot (\frac{x - x_{\rm{min}}}{x_{\rm{max}} - x_{\rm{min}}}) - 1$, $y_{\rm{sca}} =2 \cdot (\frac{y - y_{\rm{min}}}{y_{\rm{max}} - y_{\rm{min}}}) - 1$, where $x_{\rm{min}}$ and $x_{\rm{max}}$ are the minimum and maximum of the $x$-coordinates of all the points in the original map, and similarly for $y_{\rm{min}}$ and $y_{\rm{max}}$. Each Molecular Distance Map has some error, that is, the spatial distances $d_{i, j}$ are not exactly the same as $f(\Delta(i,j))$. When using the same dataset, the error is in general lower for an MDS map in a higher-dimensional space. The [*Stress-1*]{} (Kruskal stress, [@Kruskal64]), is defined in our case as $$\mbox{{\it Stress-1}} = \sigma_1 = \sqrt{\frac{\Sigma_{i < j} [f(\Delta(i,j)) - d_{i,j}]^2}{\Sigma_{i < j} d_{i,j}^2}}$$ where the summations extend over all the sequences considered for a given map, and $f(\Delta(i,j)) = a \times \Delta(i, j) + b$ is a linear function whose parameters $a, b \in \mathbb{R}$ are determined by linear regression for each dataset and corresponding Molecular Distance Map. A benchmark that is often used to assess MDS results is that [*Stress-1*]{} should be in the range $[0, 0.20]$, see [@Kruskal64]. The dataset consists of the entire collection of complete mitochondrial DNA sequences from NCBI as of 12 July, 2012. This dataset consists of 3,176 complete mtDNA sequences, namely 79 protists, 111 fungi, 283 plants, and 2,703 animals. This collection of mitochondrial genomes has a great breadth of species across taxonomic categories and great depth of species coverage in certain taxonomic categories. For example, we compare sequences at every rank of taxonomy, with some pairs being different at as high as the (super)kingdom level, and some pairs of sequences being from the exact same species, as in the case of [*Silene conica*]{} for which our dataset contains the sequences of 140 different mitochondrial chromosomes [@Sloan2012]. The prokaryotic origins and evolutionary history of mitochondrial genomes have long been extensively studied, which will allow comparison of our results with known relatedness of species. Lastly, this genome dataset permits testing of both recent and deep rooted species relationships, providing fine resolution of species differences. The creation of the datasets, acquisition of data from NCBI’s GenBank, generation of the CGR images, calculation of the distance matrix, and calculation of the Molecular Distance Maps using MDS, were all done (and can be tested with) the free open-source MATLAB program OpenMPM [@Dattani2013]. This program makes use of an open source MATLAB program for SSIM, [@Wang2003], and MATLAB’s built-in MDS function. The interactive web tool [*MoD-Map*]{} ([*Mo*]{}lecular [*D*]{}istance [*Map*]{}), [@MoDMap], allows an in-depth exploration and navigation of the Molecular Distance Maps in this paper$^1$. Results and Discussion {#results-and-discussion .unnumbered} ====================== The Molecular Distance Maps we analyzed, of several different taxonomic subsets (phylum Vertebrata, (super)kingdom Protista, classes Amphibia-Insecta-Mammalia, class Amphibia only, and order Primates), confirm that the presence or absence of oligomers in mtDNA sequences may contain information that is relevant to taxonomic classifications. These results are of interest both because of the large dataset considered and because this information has been extracted from DNA sequences that, by normal criteria, would be considered nonhomologous. The main contributions of the paper are the following: - The use of an “image distance” (designed to detect structural similarities between images) to compare the graphic signatures of two DNA sequences. For any given $k$, this distance simultaneously compares the occurrences of all subsequences of length up to $k$ of the two sequences. In all computations of this paper we use $k=9$. This image distance (with parameter set to $k=9$) is highly sensitive and succeeds to successfully group hundreds of CGRs that are visually similar, such as the ones in Figure \[fig:CGR\](left) and Figure \[fig:CGR\](right), into correct taxonomic categories. - The use of an information visualization technique to display the results as easily interpretable Molecular Distance Maps, wherein the spatial position of each sequence-point in relation to all other sequence-points is quantitatively significant. This is augmented by an interactive web tool which allows an in-depth exploration of the Molecular Distance Maps in this paper, with features such as zoom-in, search by scientific name or NCBI accession number, and quick access to complete information for each of the full mtDNA sequences in the map. - A method that is general-purpose, simple, computationally efficient and scalable. Since the compared sequences need not be homologous or of the same length, this method can be used to provide comparisons among any number of completely different DNA sequences: within the genome of an individual, across genomes within a single species, between genomes within a taxonomic category, and across taxa. - The use of a large dataset of 3,176 complete mitochondrial DNA sequences. - An illustration of potential uses of this approach by the discussion of several case studies such as the placement of the genus [*Polypterus*]{} within phylum Vertebrata, of the unclassified organism [*Haemoproteus*]{} sp. jb1.JA27 (\#1466) within the (super)kingdom Protista, and the placement of the family Tarsiidae within the order Primates. This method could complement information obtained by using DNA barcodes [@Hebert2003] and Klee diagrams [@Sirovich2010], since it is applicable to cases where barcodes may have limited effectiveness: plants and fungi for which different barcoding regions have to be used [@Kress2005], [@Hollingsworth2009], [@Schoch2011]; protists where multiple loci are generally needed to distinguish between species [@Hoef2012]; prokaryotes [@Unwin2003]; and artificial, computer-generated, DNA sequences. This method may also complement other taxonomic analyses by bringing in additional information gleaned from comparisons of non-homologous and non-coding sequences. An example of the CGR/DSSIM/MDS approach is the Molecular Distance Map in Figure \[vertebrates\] which depicts the complete mitochondrial DNA sequences of all 1,791 jawed vertebrates in our dataset. (In the legends of Figures 2-6, the number of represented mtDNA sequences in each category is listed in paranthesis after the category name.) Note that the position of each point in a map is determined by [*all*]{} the distances between the sequence it represents and the other sequences in the dataset. In the case of Figure \[vertebrates\], the position of each sequence-point is determined by the 1,790 numerical distances between its sequence and all the other mtDNA vertebrate sequences in that dataset. Observe that all five different subphyla of jawed vertebrates are separated in non-overlapping clusters, with very few exceptions. Examples of fish species bordering or slightly mixed with the amphibian cluster include [*Polypterus ornatipinnis*]{} (\#3125, ornate bichir), [*Polypterus senegalus*]{} (\#2868, Senegal bichir), both with primitive pairs of lungs; [*Erpetoichthys calabaricus*]{} (\#2745, reedfish) who can breathe atmospheric air using a pair of lungs; and [*Porichtys myriaster*]{} (\#2483, specklefish midshipman) a toadfish of the order Batrachoidiformes. It is noteworthy that the question of whether species of the [*Polypterus*]{} genus are fish or amphibians has been discussed extensively for hundreds of years [@Hall2001]. Interestingly, all four represented lungfish (a.k.a. salamanderfish), are also bordering the amphibian cluster: [*Protopterus aethiopicus*]{} (\#873, marbled lungfish), [*Lepidosiren paradoxa*]{} (\#2910, South American lungfish), [*Neoceratodus forsteri*]{} (\#2957, Australian lungfish), [*Protopterus doloi*]{} (\#3119, spotted African lungfish). Note that, in answer to the hypothesis in [@Edwards2002] regarding the diversity of signatures across vertebrates, in Figure \[vertebrates\], the avian mtDNA signatures cluster neither with the mammals nor with the reptiles, and form a completely separate cluster of their own (albeit closer to reptiles than to mammals). We applied our method to visualize the relationships among all represented species from the (super) kingdom Protista whose taxon, as defined in the legend of Figure \[AllProtists\], had more than one representative. As expected, the maximum distance between pairs of sequences in this map was higher than the maximum distances for the other maps in this paper, all at lower taxonomic levels. The most obvious outlier in Figure \[AllProtists\] is [*Haemoproteus*]{} sp. jb1.JA27 (\#1466), sequenced in [@Beadell2005] (see also [@Valkiunas2010]), and listed as an [*unclassified*]{} organism in the NCBI taxonomy. Note first that this species-point belongs to the same kingdom (Chromalveolata), superphylum (Alveolata), phylum (Apicomplexa), and class (Aconoidasida), as the other two species-points that appear grouped with it, [*Babesia bovis*]{} T2Bo (\#1935), and [*Theileria parva*]{} (\#3173). This indicates that its position is not fully anomalous. Moreover, as indicated by the high value of [*Stress-1*]{} for this figure, an inspection of DSSIM distances shows that this species-point may not be a true outlier, and its position may not be as striking in a higher dimensional version of the Molecular Distance Map. Overall, this map shows that our method allows an exploration of diversity at the level of super kingdom, obtains good clustering of known subtaxonomic groups, while at the same time indicating a lack of genome sequence information and paucity of representation that complicates analyses for this fascinating taxonomic group. We then applied our method to visualize the relationships between all available complete mtDNA sequences from three classes, Amphibia, Insecta and Mammalia (Figure \[insects\_mammals\_amphibians\]), as well as observe relationships within class Amphibia and three of its orders (Figure \[amphibians\]). Note that a feature of MDS is that the points $p_{i}$ are not unique. Indeed, one can translate or rotate a map without affecting the pairwise spatial distances $d(i, j) = ||p_{i}-p_{j}||$. In addition, the obtained points in an MDS map may change coordinates when more data items are added to or removed from the dataset. This is because the output of the MDS aims to preserve only the pairwise spatial distances between points, and this can be achieved even when some of the points change their coordinates. In particular, the $(x, y)$-coordinates of a point representing an amphibian species in the amphibians-insects-mammals map (Figure \[insects\_mammals\_amphibians\]) will not necessarily be the same as the $(x, y)$-coordinates of the same point when only amphibians are mapped (Figure \[amphibians\]). In general, Molecular Distance Maps are in good agreement with classical phylogenetic trees at all scales of taxonomic comparisons, see Figure \[amphibians\] with [@Pyron2011], and Figure \[primatesonly\] with[@Shoshani1996]. In addition, our approach may be able to weigh in on conflicts between taxonomic classifications based on morphological traits and those based on more recent molecular data, as in the case of tarsiers, as seen below. Zooming in, we observed the relationships within an order, Primates, with its suborders (Figure \[primatesonly\]). Notably, two extinct species of the genus [*Homo*]{} are represented: [*Homo sapiens neanderthalensis*]{} and [*Homo sapiens ssp. Denisova*]{}. Primates can be classified into two groups, Haplorhini (dry-nosed primates comprising anthropoids and tarsiers) and Strepsirrhini (wet-nosed primates including lemurs and lorises). The map shows a clear separation of these suborders, with the top-left arm of the map in Figure \[primatesonly\], comprising the Strepsirrhini. However, there are two Haplorhini placed in the Strepsirrhini cluster, namely [*Tarsius bancanus*]{} (\#2978, Horsfield’s tarsier) and [*Tarsius syrichta*]{} (\#1381, Philippine tarsier). The phylogenetic placement of tarsiers within the order Primates has been controversial for over a century, [@Jameson2011]. According to [@Chatterjee2009], mitochondrial DNA evidence places tarsiiformes as a sister group to Strepsirrhini, while in contrast, [@Perelman2011] places tarsiers within Happlorhini. In Figure \[primatesonly\] the tarsiers are located within the Strepsrrhini cluster, thus agreeing with [@Chatterjee2009]. This may be partly because both this study and [@Chatterjee2009] used mitochondrial DNA, whose signature may be different from that of chromosomal DNA as seen in Figure \[fig:CGR\](left) and Figure \[fig:CGR\](center). The DSSIM distances computed between all pairs of complete mtDNA sequences varied in range. The minimum distance was 0, between two pairs of identical mtDNA sequences. The first pair comprised the mtDNA of [*Rhinomugil nasutus*]{} (\#98, shark mullet, length 16,974 bp) and [*Moolgarda cunnesius*]{} (\#103, longarm mullet, length 16,974 bp). A base-to-base sequence comparison between these sequences (\#98, NC\_017897.1; \#103, NC\_017902.1) showed that the sequences were indeed identical. However, after completion of this work, the sequence for species \#103 was updated to a new version (NC\_017902.2), on 7 March, 2013, and is now different from the sequence for species \#98 (NC\_017897.1). The second pair comprises the mtDNA sequences \#1033 and \#1034 (length 16,623 bp), generated by crossing female [*Megalobrama amblycephala*]{} with male [*Xenocypris davidi*]{} leading to the creation of both diploid (\#1033) and triploid (\#1034) nuclear genomes, [@Hu2012], but identical mitochondrial genomes. The maximum distance was found to be between [*Pseudendoclonium akinetum*]{} (\# 2656, a green alga, length 95,880) and [*Candida subhashii*]{} (\#954, a yeast, length 29,795). Interestingly, the pair with the maximum distance $\Delta (\# 2656, \# 954) = 1.0033$ featured neither the longest mitochondrial sequence, with the darkest CGR ([*Cucumis sativus*]{}, \#533, cucumber, length 1,555,935 bp), nor the shortest mitochondrial sequence, with the lightest CGR ([*Silene conica*]{}, \#440, sand catchfly, a plant, length 288 bp). An inspection of the distances between [*Homo sapiens sapiens*]{} and all the other primate mitochondrial genomes in the dataset showed that the minimum distance to [*Homo sapiens sapiens*]{} was $\Delta (\# 1321, \# 1720)$ $= 0.1340$, the distance to [*Homo sapiens neanderthalensis*]{} (\#1720, Neanderthal), with the second smallest distance to it being $\Delta(\# 1321, \# 1052) = 0.2280$, the distance to [*Homo sapiens ssp. Denisova*]{} (\#1052, Denisovan). The third smallest distance was $\Delta(\# 1321, \# 3084) = 0.5591$ to [*Pan troglodytes*]{} (\#3084, chimp). Figure \[human\_graph\] shows the graph of the distances between the [*Homo sapiens sapiens*]{} mtDNA and each of the primate mitochondrial genomes. With no exceptions, this graph is in full agreement with established phylogenetic trees, [@Shoshani1996]. The largest distance between the [*Homo sapiens sapiens*]{} mtDNA and another mtDNA sequence in the dataset was 0.9957, the distance between [*Homo sapiens sapiens*]{} and [*Cucumis sativus*]{} (\#533, cucumber, length 1,555,935 bp). In addition to comparing real DNA sequences, this method can compare real DNA sequences to computer-generated sequences. As an example, we compared the mtDNA genome of [*Homo sapiens sapiens*]{} with one hundred artificial, computer-generated, DNA sequences of the same length and the same trinucleotide frequencies as the original. The average distance between these artificial sequences and the original human mitochondrial DNA is 0.8991. This indicates that all “human” artificial DNA sequences are more distant from the [*Homo sapiens sapiens*]{} mitochondrial genome than [*Drosophila melanogaster* ]{} (\#3120, fruit fly) mtDNA, with $\Delta (\# 3120, \# 1321) = 0.8572$. This further implies that trinucleotide frequencies may not contain sufficient information to classify a genetic sequence, suggesting that Goldman’s claim [@Goldman1993] that “CGR gives no futher insight into the structure of the DNA sequence than is given by the dinucleotide and trinucleotide frequencies” may not hold in general. The [*Stress-1*]{} values for all but one of the Molecular Distance Maps in this paper were in the “acceptable” range $[0, 0.2]$. The exception was Figure \[AllProtists\] with [*Stress-1*]{} equal to 0.26. Note that [*Stress-1*]{} generally decreases with an increase in dimensionality, from $q = 2$ to $q = 3, 4, 5...$. Note also that, as suggested in [@BorgGroenen2010], the [*Stress-1*]{} guidelines are not absolute: It is not always the case that only MDS representations with [*Stress-1*]{} under $0.2$ are acceptable, nor that all MDS representations with [*Stress-1*]{} under $0.05$ are good. In all the calculations in this paper, we used the full mitochondrial sequences. However, since the length of a sequence can influence the brightness of its CGR and thus its Molecular Distance Map coordinates, further analysis is needed to elucidate the effect of sequence length on the positions of sequence-points in a Molecular Distance Map. The choice of length of DNA sequences used may ultimately depend on the particular dataset and particular application. We now discuss some limitations of the proposed methods. Firstly, DSSIM is very effective at picking up subtle differences between images. For example, all vertebrate CGRs present the triangular fractal structure seen in the human mtDNA, and are visually very similar, as seen in Figure \[fig:CGR\](left) and Figure \[fig:CGR\](right). In spite of this, DSSIM is able to detect a range of differences that is sufficient for a good positioning of all 1,791 mtDNA sequences relative to each other. This being said, DSSIM may give too much weight to subtle differences, so that small and big differences in images produce distances that are numerically very close. This may be a useful feature for the analysis of datasets of closely related sequences. For large-scale taxonomic comparisons however, refinements of DSSIM or the use of other distances needs to be explored, that would space further apart the values of distances arising from small differences versus those arising from big-pattern differences between images. Secondly, MDS always has some errors, in the sense that the spatial distance between two points does not always reflect the original distance in the distance matrix. For fine analyses, the placement of a sequence-point in a map has to be confirmed by checking the original distance matrix. Possible solutions include increasing the dimensionality of the maps to three-dimensional maps, which are still easily interpretable visually and have been shown in some cases to separate clusters which seemed incorrectly intermeshed in the two-dimensional version of the map. Other possibilities include a colour-scheme that would colour points with low stress-per-point differently from the ones with high stress-per-point, and thus alert the user to the regions where discrepancies between the spatial distance and the original distance exist. Thirdly, we note that the use of the particular distance measure (DSSIM) or particular scaling technique (classical MDS) does not mean that these are the optimal choices in all cases. Lastly, as seen in Figure \[fig:CGR\](left) and Figure \[fig:CGR\](center), the genomic signature of mtDNA can be very different from that of nuclear DNA of the same species and care must be employed in choosing the dataset and interpreting the results. Conclusions {#conclusions .unnumbered} =========== Our analysis suggests that the oligomer composition of mitochondrial DNA sequences can be a source of taxonomic information. These results are of interest both because of the large dataset considered (see, e.g., the correct grouping in taxonomic categories of 1,791 mitochondrial genomes in Figure \[vertebrates\]), and because this information is extracted from DNA sequences that, by normal criteria, would not be considered homologous. Potential applications of Molecular Distance Maps - when used on a dataset of genomic sequences, whether coding or non-coding, homologous or not homologous, of the same length or vastly different lengths – include identification of large evolutionary lineages, taxonomic classifications, species identification, as well as possible quantitative definitions of the notion of species and other taxa. Possible extensions include generalizations of MDS, such as 3-dimensional MDS, for improved visualization, and the use of increased oligomer length (higher values of $k$) for comparisons of longer subsequences in case of whole chromosome or whole genome analyses. We note also that this method can be applied to analyzing sequences over other alphabets. For example binary sequences could be imaged using a square with vertices labelled 00, 01, 10, 11, and then DSSIM and MDS could be employed to compare and map them. Acknowledgments {#acknowledgments .unnumbered} =============== We thank Ronghai Tu for an early version of our MATLAB code to generate CGR images, Tao Tao for assistance with NCBI’s GenBank, Steffen Kopecki for generating artificial sequences and discussions. We also thank Andre Lachance, Jeremy McNeill, and Greg Thorn for resources and discussions on taxonomy. We thank the Oxford University Mathematical Institute for the use of their Windows compute server Pootle/WTS. Figures {#figures .unnumbered} ======= \ For a complete version, including figures, please see\ <http://www.csd.uwo.ca/~lila/pdfs/Space_of_Genomic_Signatures.pdf>

ArXiv

--- abstract: | This paper is concerned with the impermeable wall problem for an ideal polytropic model of non-viscous and heat-conductive gas in one-dimensional half space. It is shown that the 3-rarefaction wave is stable under some smallness conditions. The proof is given by an elementary energy method and the key point is to control the boundary terms due to the less dissipativity of the system. [**Keywords**]{}: non-viscous; impermeable problem; rarefaction wave; address: - '1.School of mathematical sciences, University of Chinese Academy of Sciences, Beijing, 100049, China' - '2. Institute of Applied Mathematics, AMSS, Beijing 100190, China Academy of Mathematics and Systems Science, Academia Sinica, Beijing 100190, China' author: - 'Meichen Hou$^{1,2}$' title: 'Asymptotic stability of the rarefaction wave for the non-viscous and heat-conductive ideal gas in half space' --- Introduction ============ In this paper we consider the one-dimensional initial boundary value problem (IBVP) for the equation of heat-conductive ideal gas without viscosity, which is modelled by following $$\label{1.1} \left\{ \begin{aligned} &\rho_{\tilde{t}}+(\rho u)_{\tilde{x}}=0,\\ &(\rho u)_{\tilde{t}}+(\rho u^2+p)_{\tilde{x}}=0,\\ &(\rho(e+\frac{u^2}{2}))_{\tilde{t}}+(\rho u(e+\frac{u^2}{2})+pu)_{\tilde{x}}=k\theta_{\tilde{x}\tilde{x}}.\\ \end{aligned} \right.$$ Here $(\tilde{t},\tilde{x})\in \mathbb{R}^+\times \mathbb{R}^+$ are Eulerian coordinates. And $\rho(\tilde{t},\tilde{x})>0,u(\tilde{t},\tilde{x}),e(\tilde{t},\tilde{x})>0,\theta(\tilde{t},\tilde{x})>0$ and $p(\tilde{t},\tilde{x})$ are density, fluid velocity, internal energy, absolute temperature and pressure respectively, while $k>0$ is the coefficient of the heat conduction, and it is a constant in this paper. We consider the ideal polytropic fluids so that $$\label{1.2} p=R\rho\theta=A\rho^{\gamma}\exp(\frac{{\gamma}-1}{R}s),\quad e=\frac{R}{{\gamma}-1}\theta,$$ where ${\gamma}>1, s$ are the adiabatic exponent of the gas and the specific entropy of the fluid respectively. For the IBVP problem of (\[1.1\]) in the half space $\tilde{x}>0$, the initial data is given by $$\label{1.3} \left\{ \begin{array}{l} (\rho_0,u_0,\theta_0)(\tilde{x})\triangleq (\rho,u,\theta)(0,\tilde{x})\rightarrow (\rho_+,u_+,\theta_+) =z_+,\quad as \quad \tilde{x}\rightarrow +\infty,\\[2mm] \inf_{\tilde{x}\in {\bf R_+}}(\rho_0,\theta_0)(\tilde{x})>0, \end{array} \right.$$ where $\rho_+>0$, $\theta_+>0$ and $u_+\in \mathbb{R}$ are given constants. Rewriting (\[1.1\]) as following form $$\begin{aligned} \label{1.4} \left\{ \begin{array}{ll} \rho_{\tilde{t}}+u\rho_{\tilde{x}}+\rho u_{\tilde{x}}=0,\\[2mm] u_t+uu_{\tilde{x}}+\frac{p}{\rho^2}\rho_{\tilde{x}}=-R\theta_{\tilde{x}},\\[2mm] \frac{R}{\gamma-1}\theta_{\tilde{t}} -\frac{\kappa }{\rho}\theta_{\tilde{x}\tilde{x}}=-\frac{R}{\gamma-1}u \theta_{\tilde{x}}-R\theta u_{\tilde{x}}, \end{array} \right.\end{aligned}$$ one sees that the eigenvalues of the hyperbolic part are $$\begin{aligned} \label{1.5} \widetilde{\lambda}_1=u-\widetilde{c}_s(\theta),\quad \widetilde{\lambda}_2=u+\widetilde{c}_s(\theta),\end{aligned}$$ where $$\begin{aligned} \label{1.6} \widetilde{c}_s(\theta):=\sqrt{\frac{ p}{\rho}}=\sqrt{ R \theta}.\end{aligned}$$ As pointed out by ([@MPPT]), the boundary conditions of (\[1.1\]) depend on the sign of $\widetilde{\lambda}_1,\widetilde{\lambda}_2$. Because $\widetilde{\lambda}_1,\widetilde{\lambda}_2$ depend the solution itself, the boundary condition of (\[1.1\]) should be proposed carefully. To simplify this, the global solution $z=(\rho,u,\theta)$ of (\[1.1\]) is considered in a small neighborhood $U(z_+)$ of the far field states $z_+=(\rho_+,u_+,\theta_+)$ such that the sign of $\widetilde{\lambda}_i(z)(i=1,2)$ is same as the sign of $\widetilde{\lambda}_i(z_+).$ Hence, the phase space is divided into following regions: $$\begin{aligned} & \widetilde{\Omega}^+_{sub}:=\left\{(\rho,u,\theta);\ 0<u<\widetilde{c}_s(\theta)\right\},\quad \widetilde{\Omega}^-_{sub}:=\left\{(\rho,u,\theta);\ -\widetilde{c}_s(\theta)<u<0\right\};\\[2mm] & \widetilde{\Omega}^+_{supper}:=\left\{(\rho,u,\theta);\ u>\widetilde{c}_s(\theta)\right\},\quad \widetilde{\Omega}^-_{supper}:=\left\{(\rho,u,\theta);\ u<-\widetilde{c}_s(\theta)\right\};\\[2mm] & \widetilde{\Gamma}^+_{trans}:=\left\{(\rho,u,\theta);\ u=\widetilde{c}_s(\theta)\right\},\quad \widetilde{\Gamma}^-_{trans}:=\left\{(\rho,u,\theta);\ u=-\widetilde{c}_s(\theta)\right\};\\[2mm] &\widetilde{\Gamma}^0_{sub}:=\left\{(\rho,u,\theta);\ u=0\right\} \end{aligned}$$ and there are three cases for the boundary conditions of (\[1.1\]):\ Case (1): If $z_+=(\rho_+,u_+,\theta_+) \in \widetilde{\Omega}^-_{supper}$, in the neighborhood of $U(z_+)$, $\widetilde{\lambda}_1(z)<0$, $\widetilde{\lambda}_2(z)<0$, the boundary condition is $$\begin{aligned} \label{1.7} \theta(\tilde{t},0)=\theta_-.\end{aligned}$$ Case (2): If $z_+=(\rho_+,u_+,\theta_+) \in \widetilde{\Omega}_{sub}^+\bigcup\widetilde{\Omega}_{sub}^- \bigcup\widetilde{\Gamma}^0_{sub}$, in the neighborhood of $U(z_+)$, $\widetilde{\lambda}_1(z)<0$, $\widetilde{\lambda}_2(z)>0$, the boundary condition is $$\begin{aligned} \label{1.8} u(\tilde{t},0)=u_-,\quad \theta(\tilde{t},0)=\theta_-.\end{aligned}$$ Case (3): If $z_+=(\rho_+,u_+,\theta_+) \in \widetilde{\Omega}^+_{supper}$, in the neighborhood of $U(z_+)$, $\widetilde{\lambda}_1(z)>0$, $\widetilde{\lambda}_2(z)>0$, the boundary condition is $$\begin{aligned} \label{1.9} \rho(\tilde{t},0)=\rho_-,\quad u(\tilde{t},0)=u_-,\quad \theta(\tilde{t},0)=\theta_-.\end{aligned}$$ Here $\rho_-,u_-,\theta_-$ are given constants. Above cases tell us that the IBVP problem of (\[1.1\]) is very different from the IBVP problem of the viscous and heat-conductive ideal gas. For the latter, the system is $$\label{1.10} \left\{ \begin{aligned} &\rho_{\tilde{t}}+(\rho u)_{\tilde{x}} = 0, \\ &(\rho u)_{\tilde{t}}+(\rho u^2+p)_{\tilde{x}} = \mu u_{\tilde{x}\tilde{x}}, \\ &(\rho (e+\frac{u^2}{2}))_{\tilde{t}}+(\rho u(e+\frac{u^2}{2})+pu)_{\tilde{x}} = k\theta_{\tilde{x}\tilde{x}}+(\mu uu_{\tilde{x}})_{\tilde{x}}, \end{aligned} \right.$$ where $\mu>0$ stands for the coefficient of viscosity. By [@Matsumura], because the first equation is of hyperbolic, the second and third equations are of parabolic, then the boundary conditions of (\[1.10\]) can be divided into three cases $$\label{1.11} \left\{ \begin{aligned} &\text{Case}\quad {\uppercase\expandafter{\romannumeral 1\relax}}: u(\tilde{t},0)=u_-<0,\quad\quad \theta(\tilde{t},0)=\theta_-,\\ &\text{Case}\quad {\uppercase\expandafter{\romannumeral 2\relax}}: u(\tilde{t},0)=u_-=0,\quad\quad \theta(\tilde{t},0)=\theta_-,\\ &\text{Case}\quad {\uppercase\expandafter{\romannumeral 3\relax}}: u(\tilde{t},0)=u_->0,\quad\quad \rho(\tilde{t},0)=\rho_-,\quad\quad \theta(\tilde{t},0)=\theta_-. \end{aligned} \right.$$ Note that in Case ${\uppercase\expandafter{\romannumeral 1\relax}}-{\uppercase\expandafter{\romannumeral 2\relax}}$, the density $\rho$ on the boundary $\tilde{x}=0$ is unknown. Usually, Case ${\uppercase\expandafter{\romannumeral 1\relax}}-{\uppercase\expandafter{\romannumeral 3\relax}}$ are called outflow problem ($u(\tilde{t},0)<0$), impermeable wall problem ($u(\tilde{t},0)=0$), inflow problem ($u(\tilde{t},0)>0$) respectively. Compare the boundary conditions ((\[1.7\])-(\[1.9\])) of (\[1.1\]) with (\[1.11\]), we immediately know that the IBVP problems of (\[1.1\]) is more complicated than the related problems of (\[1.10\]) due to the less dissipativity. There are many works to the large time behavior of solutions to the Cauchy problem of systems (\[1.1\]) and (\[1.10\]), see ([@F-M],[@Huang-L-M], [@Huang-Matsumura],[@Huang-Matsumura-Xin], [@H-T-X],[@T.P.Liu-1],[@N-T-Z],[@Wang-Zhao]), all this works tell us that the large time behavior of solutions to the Cauchy problem is governed by the Riemann solutions to the corresponding compressible Euler equations. For the IBVP problems of (\[1.1\]) and (\[1.10\]), the boundary layer maybe appear due to the boundary effect, see ([@Matsumura],[@M-Nishihara]). Therefore, it attracts many authors to study the related problems. For the inflow problems of (\[1.10\]), Qin-Wang ([@Qin-Wang],[@Qin-Wang-2011]) studied the superposition of multiple waves, i.e, rarefaction waves, boundary layer and contact wave. Recently, for the outflow problems of (\[1.1\]), Nakamura-Nishibata [@Naka-Nishi1] studied the existence and stability of the boundary layer for the general symmetric hyperbolic-parabolic systems and their results contained the supersonic case of (\[1.1\]), i.e., $z\in\widetilde{\Omega}_{supper}^-$ (Case (1)). And for the inflow problems of (\[1.1\]), we have showned the existence of the boundary layer for $u_+>0$, moreover, the stability of the composition of rarefaction wave and boundary layer in supersonic case has been proved, i.e., $z\in\widetilde{\Omega}_{supper}^+$ (Case (3)), which will be appear, see ([@H-F]). For other related results, we refer to ([@F-H-W-Z],[@Huang-L-S], [@H-M-S-1],[@H-M-S],[@Huang-Qin],[@F-Z]-[@Kawashima-Zhu], [@M-Nishihara]-[@Nakamura2],[@Qin],[@Wan-Wang-Zou]) and some references therein. For the impermeable wall problem, Matsumura-Mei [@Matsumura-Mei] studied the asymptotic stability of the viscous shock wave for the isentropic gas ($p=A\rho^{\gamma}$). Matsumura-Nishihara [[@M-Nishihara1]]{} proved the stability of the rarefaction wave for the same model. Lately, Min-Qin [@Min-Qin] proved the asymptotic stability of the rarefaction wave of (\[1.10\]) for some large perturbation. Motivated by this, in this paper, we turn to study the stability of the rarefaction wave for the impermeable wall problem of (\[1.1\]). For this problem, $z_+\in\Omega_{sub}^+,z\in U(z_+),$ then $\widetilde{\lambda}_1(z)<0,\widetilde{\lambda}_2(z)>0$ immediately tell us that the boundary condition of this problem should be $(u,\theta)(t,0)=(u_-,\theta_-)$, here $u_-=0$(see Case (2)). Hence, the IBVP problem we consider here is (\[1.1\]), (\[1.3\]) and (\[1.8\]), where $u_-=0$. There are two main difficulties of this problem to overcome. The one is that both the density and the velocity equations are of hyperbolic, we need more higher order derivative estimates to recover the dissipativity of the hyperbolic part. The other is that how to control the higher order derivatives of boundary terms (see ($H_2(\tau,0)$ in (\[3.51\])). In order to solve the second difficulty, we use the relationship (\[3.56\]) between the boundary terms in our estimates. This paper is organized as follows. In Section 2, we do the transformation of coordinates for (\[1.1\]) and list some properties of the smooth approximation of rarefaction wave, then we state our main Theorem 2.1. In Section 3, we prove Theorem 2.1. 0.5in Notations. Throughout this paper, $c$ and $C$ denote some positive constants (generally large). For function spaces, $L^p(\mathbb{R}_+)(1\leq p\leq \infty)$ denotes the usual Lebesgue space on $\mathbb{R}_+$ with norm $\|{\cdot}\|_{L^p}$ and $H^k(\mathbb{R}_+)$ the usual Sobolev space in the $L^2$ sense with norm $\|\cdot\|_p$. We note $\|\cdot\|=\|\cdot\|_{L^2}$ for simplicity. And $C^k(I; H^p)$ is the space of $k$-times continuously differentiable functions on the interval $I$ with values in $H^p(\mathbb{R}_+)$ and $L^2(I; H^p)$ the space of $L^2$-functions on $I$ with values in $H^p(\mathbb{R}_+)$. Preliminaries and Main results ============================== To simplify the system (\[1.1\]), (\[1.3\]) and (\[1.8\]), we change the Eulerian coordinates $(\tilde{t},\tilde{x})$ into Lagrangian coordinates $(t,x)$ $$\label{2.1} t=\tilde{t},\quad\quad x=\int_{(0,0)}^{(\tilde{t},\tilde{x})}\rho(\tau,y)dy-(\rho u)(\tau,y)d\tau.$$ Since $u_-=0,$ we let $v=\frac{1}{\rho}>0$ which is the specific volume of the gas, then the impermeable wall problem for (\[1.1\]) becomes following $$\label{2.2} \left\{ \begin{aligned} &v_t-u_x=0, \quad\quad t>0,\quad x\in \mathbb{R}_+,\\ &u_t+p_x=0,\\ &(\frac{R}{{\gamma}-1}\theta+\frac{u^2}{2})_t+(pu)_x=k(\frac{\theta_x}{v})_x,\\ &(v_0,u_0,\theta_0)(x)\triangleq (v,u,\theta)(0,x)\rightarrow (v_+,u_+,\theta_+)\quad\quad x\rightarrow +\infty, \\ &\inf_{x\in \mathbb{R}_+}(v_0,\theta_0)(x)>0,\quad u(t,0)=0,\quad\theta(t,0)=\theta_-, \end{aligned} \right.$$ where $$\label{2.2-1} p=p(v,\theta)=\frac{R\theta}{v} =A{v}^{-{\gamma}}\exp(\frac{{\gamma}-1}{R}s) =p(v,s),$$ and the initial data satisfy $u_0(0)=0,\theta_0(0)=\theta_-$ as compatibility conditions. We have known that the corresponding hyperbolic system of (\[2.2\]) has three characteristic speeds $$\label{2.3} \lambda_1(v,\theta)=-\frac{\sqrt{R\gamma\theta}}{v}<0,\quad\quad \lambda_2=0,\quad\quad \lambda_3(v,\theta)=\frac{\sqrt{R\gamma\theta}}{v}>0.$$ Now we turn to list some properties of the 3-rarefaction wave. The 3-rarefaction wave curve through the right-hand side state $(v_+,u_+,\theta_+)$ is $$\begin{aligned} \label{2.4-1} R_3(v_+,u_+,\theta_+):&=&\big\{(v,u,\theta): v>v_+, v^{\gamma-1}\theta=v^{\gamma-1}_+\theta_+,\\[2mm] &&u=u_+-\int^v_{v_+}\sqrt{R\gamma v_+^{\gamma-1}\theta_+}\xi^{-\frac{\gamma+1}{2}}d\xi\big\}.\end{aligned}$$ Precisely to say, if $v_+,u_+,\theta_+,\theta_-$ satisfy following condition $$\begin{aligned} \label{2.4} 0<u_+=\int^{(\frac{\theta_+}{\theta_-})^{\frac{1}{\gamma-1}}v_+}_{v_+}\sqrt{R\gamma v_+^{\gamma-1}\theta_+}\xi^{-\frac{\gamma+1}{2}}d\xi,\end{aligned}$$ then the 3-rarefaction wave $(v^r,u^r,\theta^r)(\frac{x}{t})$ is the unique weak solution which is global in time to the following Riemann problem $$\label{2.5} \left\{ \begin{aligned} &v_t^r-u_x^r=0, \\ &u_t^r+p_x^r=0, \\ &(\frac{R}{{\gamma}-1}\theta^r+\frac{u^{r^2}}{2})_t+(p^ru^r)_x=0, \\ &(v^r,u^r,\theta^r)(t,0)= \begin{cases} &(v_-,0,\theta_-),\quad x<0, \\ &(v_+,u_+,\theta_+),\quad x>0. \end{cases} \end{aligned} \right.$$ Here $p^r=\frac{R\theta^r}{v^r}$, $v_->v_+,\theta_-<\theta_+,$ and $v_-=(\frac{\theta_+}{\theta_-})^{\frac{1}{{\gamma}-1}}v_+.$ To study the large time behavior of the solutions to the impermeable wall problem (\[2.2\]), we construct a smooth approximation $(\tilde{v},\tilde{u},\tilde{\theta})(t,x)$ of 3-rarefaction wave $(v^r,u^r,\theta^r)(\frac{x}{t})$. Same as [@H-M-S-2], we consider the following Cauchy problem $$\begin{aligned} \label{2.6} \left\{ \begin{array}{ll} w_t+ww_x=0,\quad\quad\\[2mm] w(x,0)=\left\{ \begin{array}{ll} w_-, \quad x<0,\\[2mm] w_-+C_q\delta^r \int^{\epsilon x}_0y^qe^{-y}dy, \quad x>0. \end{array} \right. \end{array} \right.\end{aligned}$$ Here $\delta^r=w_+-w_->0$, $0<\epsilon\leq 1$ is a constant which will be determined later and $q\geq 10, C_q$ are two constants such that $C_q \int^{\infty}_0y^qe^{-y}dy=1$. Then the smooth approximation $(\tilde{v},\tilde{u},\tilde{\theta})(t,x)$ is constructed in the following way $$\begin{aligned} \label{2.7} \left\{ \begin{array}{ll} (\frac{\sqrt{R\gamma \tilde{\theta}}}{\tilde{v}})(t,x)=w(1+t,x),\\[2mm] (\tilde{v}^{\gamma-1}\tilde{\theta})(t,x)=v^{\gamma-1}_+\theta_+,\quad x \in \mathbb{R},\quad t>0.\\[2mm] \tilde{u}=u_+-\int^{\tilde{v}}_{v_+}\sqrt{R\gamma v^{\gamma-1}_+\theta_+}\xi^{-\frac{\gamma+1}{2}}d\xi. \end{array} \right.\end{aligned}$$ Because $\lambda_3(v_-,\theta_-)>0$, we immediately know that both the 3-rarefaction wave $(v^r,u^r,\theta^r)$ and its smooth approximation $(\tilde{v},\tilde{u},\tilde{\theta})$ are constants on $(t,x)\in(0,+\infty)\times R_-.$ Then we use $(\tilde{v},\tilde{u},\tilde{\theta})(t,x)$ to represent $(\tilde{v},\tilde{u},\tilde{\theta})(t,x) |_{x\geq0}$, we have $$\label{2.8} \left\{ \begin{aligned} &\tilde{v}_t-\tilde{u}_x=0, \\ &\tilde{u}_t+\tilde{p}_x=0,\quad\quad t>0,\quad x\in \mathbb{R}_+ \\ &(\frac{R}{{\gamma}-1}\tilde{\theta}+\frac{\tilde{u}^{2}}{2})_t+(\tilde{p}\tilde{u})_x=0, \\ &(\tilde{v},\tilde{u},\tilde{\theta})(t,0)=(v_-,0,\theta_-),\\ &(\tilde{v}_0,\tilde{u}_0,\tilde{\theta}_0)(x)\rightarrow \begin{cases} &(v_-,0,\theta_-),\quad\quad x\rightarrow 0_+, \\ &(v_+,u_+,\theta_+),\quad\quad x\rightarrow +\infty, \end{cases} \end{aligned} \right.$$ where $\tilde{p}=p(\tilde{v},\tilde{\theta})=\frac{R\tilde{\theta}}{\tilde{v}}$. We get following Lemma. \[L1\](Smooth rarefaction wave)([@H-M-S-2])$(\tilde{v},\tilde{u},\tilde{\theta})(t,x)$ satisfies (1)$0\leq -\tilde{v}_x(t,x),\frac{R}{{\gamma}-1}\tilde{\theta}_x(t,x)\leq C\tilde{u}_x(t,x)$; (2)For any p($1\leq p\leq +\infty$), there exists a constant $C_{pq}$ such that $$\label{2.17} \begin{aligned} &\|(\tilde{v}_x,\tilde{u}_x,\tilde{\theta}_x)(t)\|_{L^p}\leq C_{pq}\min\{\delta^r{\epsilon}^{1-\frac{1}{p}}, (\delta^r)^\frac{1}{p}(1+t)^{-1+\frac{1}{p}}\}, \\ &\|(\tilde{v}_{xx},\tilde{u}_{xx},\tilde{\theta}_{xx})(t)\|_{L^p} \leq C_{pq}\min \{\delta^r{\epsilon}^{2-\frac{1}{p}},((\delta^r)^{\frac{1}{p}}+(\delta^r)^{\frac{1}{q}})(1+t)^{-1+\frac{1}{q}}\}, \end{aligned}$$ (3)If $x<\lambda_3(v_-,\theta_-)(1+t),$ then $(\tilde{v},\tilde{u},\tilde{\theta})(t,x)\equiv(v_-,u_-,\theta_-),$ (4)$\lim_{t\rightarrow+\infty}\sup_{\xi\in R_+}|(\tilde{v},\tilde{u},\tilde{\theta})(t,x)-(v^r,u^r,\theta^r)(\frac{x}{1+t})|=0.$ We assume that the initial data satisfy $$\label{2.18} (v_0-\tilde{v}_0,u_0-\tilde{u}_0,\theta_0-\tilde{\theta}_0)\in H^2(\mathbb{R}_+),\quad (v_t-\tilde{v}_t,u_t-\tilde{u}_t,\theta_t-\tilde{\theta}_t)(0,x)\in H^1(\mathbb{R}_+).$$ Then the stability of 3-rarefaction wave is listed as follows \[t1\] Assume that ${\gamma}>1$ and the relationship between $v_+,u_+,\theta_+,$ and $\theta_-$ satisfy $$\label{2.19} u_+=\int^{(\frac{\theta_+}{\theta_-})^{\frac{1}{\gamma-1}}v_+}_{v_+}\sqrt{R\gamma v_+^{\gamma-1}\theta_+}\xi^{-\frac{\gamma+1}{2}}d\xi,\quad\quad \theta_+-\theta_->0.$$ If the initial data satisfy (\[2.18\]), then there exist positive constants ${\epsilon}_1,{\varepsilon}_0$ such that if ${\epsilon}\leq{\epsilon}_1$ and $$\label{2.20} \|v_0-\tilde{v}_0,u_0-\tilde{u}_0,\theta_0-\tilde{\theta}_0\|_2+\|(v_t-\tilde{v}_t,u_t-\tilde{u}_t,\theta_t-\tilde{\theta}_t)(0)\|_1\leq {\varepsilon}_0,$$ the impermeable wall problem (\[2.2\]) has a unique global solution $(v,u,\theta)(t,x)$ which satisfy $$\label{2.21} \left\{ \begin{aligned} &(v-\tilde{v},u-\tilde{u},\theta-\tilde{\theta})\in C([0,+\infty);H^2(R_+)), (v-\tilde{v},u-\tilde{u},\theta-\tilde{\theta})_t\in C([0,+\infty);H^1(R_+)), \\ &(v-\tilde{v},u-\tilde{u})_x\in L^2(0,+\infty;H^1(R_+)),(\theta-\tilde{\theta})_x\in L^2(0,+\infty;H^2(R_+)), \\ &(\theta-\tilde{\theta})_t\in L^2(0,+\infty;H^2(R_+)). \end{aligned} \right.$$ Moreover, $$\label{2.27} \sup_{x\geq 0}|(v,u,\theta)(t,x)-(v^r,u^r,\theta^r)(\frac{x}{t})|\rightarrow 0, \quad as \quad t\rightarrow +\infty.$$ Proof of Theorem 2.1 ==================== This section we mainly proof Theorem \[t1\]. In Section 3.1, we shall reformulate system (\[2.2\]) to a new perturbed system and show the local existence of the solution, see Proposition 3.1. In Section 3.2, a priori estimates will be investigated, see Proposition 3.2. Finally ,we will combine this two Propositions to get the stability of the 3-rarefaction wave. Local existence of the solution ------------------------------- Now we turn to reformulate the system (\[2.2\]). Put the perturbation $(\phi,\psi,\xi)(t,x)$ by $$\label{3.1} (\phi,\psi,\xi)(t,x)=(v,u,\theta)(t,x)-(\tilde{v},\tilde{u},\tilde{\theta})(t,x),$$ then the reformulated problem is $$\label{3.2} \left\{ \begin{aligned} &\phi_t-\psi_x=0,\quad\quad t>0,\quad x\in \mathbb{R}_+\\ &\psi_t+(\frac{R\xi-\tilde{p}\phi}{v})_x=0,\\ &\frac{R}{\gamma-1}\xi_t+p\psi_x+\tilde{u}_x(p-\tilde{p}) =k(\frac{\theta_x}{v})_x,\\ &(\psi,\xi)(t,0)=(0,0), \\ &(\phi,\psi,\xi)(0,x)\triangleq(\phi_0,\psi_0,\xi_0)(x)\rightarrow (0,0,0),\quad as \quad x\rightarrow +\infty. \end{aligned} \right.$$ Define the solution space as $$\label{3.3} \begin{aligned} &\mathbb{X}_{m_1,m_2,M}(0,T):=\bigg\{(\phi,\psi,\xi)\in C([0,T];H^2(\mathbb{R}_+)), (\phi,\psi,\xi)_t\in C([0,T];H^1(\mathbb{R}_+)),\\ &(\phi,\psi)_x\in L^2(0,T;H^1(\mathbb{R}_+)),\xi_x\in L^2(0,T;H^2(\mathbb{R}_+)), \xi_t\in L^2(0,T;H^2(\mathbb{R}_+)),\\ &\text{with}\quad\sup_{[0,T]}\|(\phi,\psi,\xi)(t)\|_2+\|(\phi_t,\psi_t,\xi_t)(t)\|_1\leq M, \inf_{\mathbb{R}_+\times[0,T]}(\tilde{v}+\phi)(t,x)\geq m_1,\\ &\inf_{\mathbb{R}_+\times[0,T]}(\tilde{\theta}+\xi)(t,x)\geq m_2\bigg\} \end{aligned}$$ for some positive constants $m_1,m_2,M$. \[p1\] (Local existence) For any given initial data $(\phi_0,\psi_0,\xi_0)\in H^2(\mathbb{R}_+)$ and $(\phi_t,\psi_t,\xi_t)(0,x)\in H^1(\mathbb{R}_+)$, there exists positive constant $M_0$ such that if $\|(\phi_0,\psi_0,\xi_0)\|_2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1\leq M(\tilde{C}M\leq M_0)$and $\inf_{\mathbb{R}_+\times[0,T]}(\tilde{v}+\phi_0)\geq \frac{3v_+}{8},\inf_{\mathbb{R}_+\times[0,T]}(\tilde{\theta}+\xi_0)\geq \frac{3\theta_-}{8}$, then there exists $t_0=t_0(M_0)>0$ such that (\[3.2\]) has a unique solution $(\phi,\psi,\xi)\in\mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\tilde{C}M}(0,t_0).$ Proof. Rewrite $(\ref{3.2})$ with following forms, $$\label{3.4} \left\{ \begin{aligned} &\phi_t-\psi_x=0,\\ &\psi_t-\frac{p}{v}\phi_x =g_1:=g_1(\phi,\xi,\xi_x),\\ &\frac{R}{\gamma-1}\xi_t-k\frac{\xi_{xx}}{\tilde{v}+\phi}=g_2:=g_2(\phi,\xi,\phi_x,\psi_x,\xi_x),\\ &\psi(t,0)=0,\xi(t,0)=0,\\ &(\phi(0,x),\psi(0,x),\xi(0,x))=(\phi_0(x),\psi_0(x),\xi_0(x))\rightarrow (0,0,0),\quad as \quad x\rightarrow +\infty, \end{aligned} \right.$$ where $$\label{3.5} \begin{aligned} &g_1(\phi,\xi,\xi_x)=\frac{\tilde{p}_x\phi}{\tilde{v}+\phi}-\frac{\tilde{p}\tilde{v}_x\phi} {(\tilde{v}+\phi)^2}-\frac{R\xi_x}{\tilde{v}+\phi}+\frac{R\xi\tilde{v}_x}{(\tilde{v}+\phi)^2},\\ &g_2(\phi,\xi,\phi_x,\psi_x,\xi_x)=-k\frac{\xi_x(\tilde{v}+\phi)_x}{(\tilde{v}+\phi)^2} -p\psi_x-\tilde{u}_x(p-\tilde{p})+k(\frac{\tilde{\theta}_x}{\tilde{v}+\phi})_x. \end{aligned}$$ Now we approximate the initial data $(\phi_0,\psi_0,\xi_0)$ by $ (\phi_{0j},\psi_{0j},\xi_{0j})\in H^m\cap H^2$, $m\geq 6$ such that $$\label{3.6} (\phi_{0j},\psi_{0j},\xi_{0j})\rightarrow (\phi_0,\psi_0,\xi_0)\quad \text{strongly}\quad\quad \text{in}\quad\quad H^m$$ as $j\rightarrow\infty$ and $$\label{3.6-1} \|(\phi_{0j},\psi_{0j})\|_2+\|\xi_{0j}\|_3\leq\frac{3}{2}M,$$ furthermore, $\inf_{\mathbb{R}_+}(\tilde{v}+\phi_{0j})(t,x)\geq \frac{3v_+}{8}, \inf_{\mathbb{R}_+}(\tilde{\theta}+\xi_{0j})(t,x)\geq \frac{3\theta_-}{8}$ hold for any $j\geq 1.$ (Here because our initial data satisfy $\|(\phi_0,\psi_0,\xi_0)\|_2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1\leq M$, by system (\[3.2\]), we know that the norm is equivalent to $\|(\phi_0,\psi_0)\|_2+\|\xi_0\|_3$, that’s why we construct approximated sequence satisfy (\[3.6-1\])). We will use the iteration method to prove our Proposition 3.1. Define the sequence $\{(\phi^{(n)}_j(t,x),\psi^{(n)}_j(t,x),\xi^{(n)}_j(t,x))\}$ for each $j$ so that $$\label{3.7} (\phi^{(0)}_j,\psi^{(0)}_j,\xi^{(0)}_j)(t,x)=(\phi_{0j},\psi_{0j},\xi_{0j})(x),$$ and for a given $(\phi^{(n-1)}_j,\psi^{(n-1)}_j,\xi^{(n-1)}_j)(t,x),$ $(\phi^{(n)}_j,\psi^{(n)}_j,\xi^{(n)}_j)(t,x)$ is a solution to following equation $$\label{3.8} \left\{ \begin{aligned} &\phi_{jt}^{(n)}-\psi_{jx}^{(n)}=0\\ &\psi_{jt}^{(n)}-(\frac{R(\tilde{\theta}+\xi^{(n-1)}_j)}{(\tilde{v}+\phi_j^{(n-1)})^2})\phi^{(n)}_{jx} =g_1^{(n-1)}=g_1^{(n-1)}(\phi^{(n-1)}_j,\xi^{(n-1)}_j,\xi^{(n-1)}_{jx})\\ &\frac{R}{\gamma-1}\xi_{jt}^{(n)}-k\frac{\xi_{jxx}^{(n)}}{\tilde{v}+\phi_{j}^{(n-1)}}=g_2^{(n-1)} =g_2^{(n-1)}(\phi_{j}^{(n-1)},\xi_j^{(n-1)},\phi_{jx}^{(n-1)},\psi_{jx}^{(n-1)},\xi_{jx}^{(n-1)})\\ &\psi^{(n)}_j(t,0)=0,\quad\xi^{(n)}_j(t,0)=0\\ &(\phi^{(n)}_j,\psi^{(n)}_j,\xi^{(n)}_j)(0,x)=(\phi_{0j},\psi_{0j},\xi_{0j})(x), \end{aligned} \right.$$ where $$\label{3.9} \begin{aligned} &g_1^{(n-1)}(\phi^{(n-1)}_j,\xi^{(n-1)}_j,\xi^{(n-1)}_{jx})\\ &=\frac{\tilde{p}_x\phi^{(n-1)}}{\tilde{v}+\phi^{(n-1)}_j}-\frac{\tilde{p}\tilde{v}_x\phi^{(n-1)}_j} {(\tilde{v}+\phi^{(n-1)}_j)^2}-\frac{R\xi^{(n-1)}_{jx}}{\tilde{v}+\phi^{(n-1)}_j} +\frac{R\xi^{(n-1)}_j\tilde{v}_x}{(\tilde{v}+\phi^{(n-1)}_j)^2}\\ &g_2^{(n-1)}(\phi_j^{(n-1)},\xi_j^{(n-1)},\phi_{jx}^{(n-1)},\psi_{jx}^{(n-1)},\xi_{jx}^{(n-1)})\\ &=-k\frac{\xi^{(n-1)}_{jx}(\tilde{v}_x+\phi^{(n-1)}_{jx})}{(\tilde{v}+\phi^{(n-1)}_j)^2} -\tilde{p}\psi_{jx}^{(n-1)}-(\tilde{u}_x+\psi_{jx}^{(n-1)})(\frac{R\xi^{(n-1)}_j-\tilde{p}\phi^{(n-1)}_j} {\tilde{v}+\phi^{(n-1)}_j})+k(\frac{\tilde{\theta}_x}{\tilde{v}+\phi^{(n-1)}_j})_x. \end{aligned}$$ We now assume that $M_0$ suitably small, if $g_2^{(n-1)}\in C(0,t_0;H^{m-1})$, and $\xi_{0j}\in H^{m},$ then there exists a unique local solution $\xi_{j}^{(n)}$ to (\[3.8\]) satisfying $$\label{3.11} \xi^{(n)}_j\in C(0,T_0;H^m)\cap C^1(0,T_0;H^{m-2})\cap L^2(0,T_0;H^{m+1}).$$ Making use of this, if $(\phi^{(n-1)}_j(t,x),\psi^{(n-1)}_j(t,x),\xi^{(n-1)}_j(t,x))\in\mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\tilde{C}M},$ from system (\[3.8\]), we immediately get that $$\label{3.12} \begin{aligned} &\|\xi^{(n)}_j(t)\|_2^2+\|\xi^{(n)}_{jt}(t)\|_1^2+ \int_0^{t_0}\|\xi^{(n)}_{jx}(\tau)\|_2^2+\|\xi^{(n)}_{jt}(\tau)\|_2^2d\tau\\ &\leq Ce^{C(v_+,\theta_-,M_0)t_0} (\|\xi_{0j}\|_2^2+\|\xi_{jt}(0)\|_1^2+C(v_+,\theta_-,M_0)t_0). \end{aligned}$$ Then a direct computation on $(\ref{3.8})_{1,2}$ with (\[3.12\]) also tell us $$\label{3.13} \begin{aligned} &\|(\phi^{(n)}_j,\psi^{(n)}_j)(t)\|_2^2+\|(\phi^{(n)}_{jt},\psi^{(n)}_{jt})(t)\|_1^2\leq Ce^{C(v_+,\theta_-,M_0)t_0}(\|\phi_{0j},\psi_{0j}\|_2^2 +\|(\phi_{jt},\psi_{jt})(0)\|_1^2\\ &+C(v_+,\theta_-,M_0)t_0 +\int_0^{t_0}\|\xi^{(n-1)}_{jxxx}(\tau)\|^2+\|\xi^{(n-1)}_{jtxx}(\tau)\|^2d\tau). \end{aligned}$$ Combing (\[3.12\]) and (\[3.13\]), as long as $t_0$ suitably small, we finally get $$\label{3.14} \|(\phi^{(n)}_j,\psi^{(n)}_j,\xi^{(n)}_j)(t)\|_2+\|(\phi_{jt}^{(n)},\psi_{jt}^{(n)},\xi_{jt}^{(n)})(t)\|_1\leq \tilde{C}M.$$ And $\inf_{\mathbb{R}_+\times[0,T]}(\tilde{v}+\phi_j^{(n)})(t,x)\geq \frac{3v_+}{8}, \inf_{\mathbb{R}_+\times[0,T]}(\tilde{\theta}+\xi_j^{(n)})(t,x)\geq \frac{3\theta_-}{8}$. That is to say the sequence $(\phi^{(n)}_j,\psi^{(n)}_j,\xi^{(n)}_j)$ is uniformly bounded in the function space $\mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\tilde{C}M}(0,t_0).$ By using the same method in ([@Kawa]), we can finally prove that $(\phi^{(n)}_j,\psi^{(n)}_j,\xi^{(n)}_j)$ has a subsequence $(\phi^{(n')}_j,\psi^{(n')}_j,\xi^{(n')}_j)\rightarrow (\phi_j,\psi_j,\xi_j)\in \mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\tilde{C}M}(0,t_0)$ as $n'\rightarrow+\infty$. Again, we let $j\rightarrow\infty,$ we obtain the desired unique local solution $(\phi,\psi,\xi)(t,x)\in \mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\tilde{C}M}$ as long as $t_0$ so small. Thus Proposition \[p1\] has been proved. Note that we could let $M_0(\leq\min\{\frac{5v_+}{8},\frac{5\theta_-}{8}\})$ so small, then from the proof of Proposition \[p1\] and the Sobolev’s inequality, the lower bounds of $v, \theta$ satisfy following condition obviously. $$\label{3.15-1} \inf_{\mathbb{R}_+\times[0,t_0]}(\tilde{v}+\phi(t,x))\geq \frac{3v_+}{8}, \quad\quad\inf_{\mathbb{R}_+\times[0,t_0]}(\tilde{\theta}+\xi(t,x))\geq \frac{3\theta_-}{8}.$$ Stability of the 3-rarefaction wave ----------------------------------- In this section, a priori estimates will be listed, see Proposition \[p2\]. As long as this Proposition be proved, then Proposition \[p1\], Proposition \[p2\] and Lemma \[L1\] imply Theorem \[t1\] immediately. \[p2\] (A priori estimates) For $0<{\epsilon}\leq 1$ listed in (\[2.6\]), if ${\epsilon}\leq{\epsilon}_1$ for small positive constant ${\epsilon}_1$, there exist positive constants ${\varepsilon}_0(\tilde{C}{\varepsilon}_0\leq M_0),$ and $C_0$ which depend on only the initial data such that the following statements hold. If $(\phi,\psi,\xi)(t,x)\in \mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},{\varepsilon}_0}(0,T)$ is a solution to (\[3.2\]) for some $T>0$, then it holds that $$\label{3.15} \begin{aligned} &\|(\phi,\psi,\xi)(t)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(t)\|_1^2+\int_0^t \|(\phi_x,\psi_x)(\tau)\|_1^2+\|\xi_x(\tau)\|_2^2+\|\xi_t(\tau)\|_2^2\,d\tau\\ &\leq C_0(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2 +{\epsilon}^{\frac{1}{4}}). \end{aligned}$$ Once Proposition \[p2\] is proved, choosing ${\epsilon},\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2$ suitably small such that $$\label{3.16} \begin{aligned} &\|(\phi_0,\psi_0,\xi_0)\|_2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1\leq\frac{{\varepsilon}_0}{\tilde{C}},\\ &C_0(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2+{\epsilon}^{\frac{1}{4}})\leq(\frac{{\varepsilon}_0}{\tilde{C}})^2. \end{aligned}$$ Then combing Proposition \[p1\] and Proposition \[p2\], we can construct the global solution $(\phi,\psi,\xi)(t,x)\in\mathbb{X}_{\frac{3v_+}{8},\frac{3\theta_-}{8},\frac{{\varepsilon}_0}{\tilde{C}}}(0,+\infty).$ Moreover, it satisfy the estimate (\[3.15\]) for all $t>0.$ Then from (\[3.15\]) and the system (\[3.2\]), we can get $$\label{3.17} \begin{aligned} \int_0^{+\infty}\|(\phi_x,\psi_x,\xi_x)(\tau)\|^2d\tau +\int_0^{+\infty}\frac{d}{dt}\|(\phi_x,\psi_x,\xi_x)(\tau)\|^2<+\infty, \end{aligned}$$ which together with the Sobolev’s inequality leads to the asymptotic behavior of the solution $$\label{3.18} \lim_{t\rightarrow+\infty}\sup_{x\in\mathbb{R}_+}|(\phi,\psi,\xi)(t,x)|= 0.$$ Combing Lemma \[L1\], we finally get Theorem \[t1\]. Now we mainly proof Proposition \[p2\].\ Proof of Proposition \[p2\]:\ First, we show the basic estimates. Under the same assumptions listed in Proposition \[p2\], if ${\epsilon}, {\varepsilon}_0$ are suitably small, it holds that $$\label{3.19} \begin{aligned} &\|(\phi,\psi,\xi)(t)\|^2+{\int_0^t}\|\xi_x(\tau)\|^2 \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\zeta_0)\|^2+{\epsilon}^{\frac{1}{4}}[{\int_0^t}\|\phi_x(\tau)\|^2d\tau+1]), \end{aligned}$$ where $C(v_+,\theta_-)$ is a positive constant which depends on $v_+,\theta_-$, and it is always true without illustration in below. Proof. Define the energy form $$\label{3.20} E=R\hat{\theta}\Phi(\frac{v}{\tilde{v}})+\frac{\psi^2}{2}+\frac{R}{\gamma-1}\tilde{\theta}\Phi(\frac{\theta}{\tilde{\theta}}),$$ where $\Phi(s)=s-1-\ln s,$ obviously, there exists a positive constant $C(s)$ such that $$\label{3.21} C(s)^{-1}s^2\leq\Phi(s)\leq C(s) s^2.$$ Then after computation, we get $$\label{3.22} \begin{aligned} &E_t+((p-\tilde{p})\psi)_x=k(\frac{\xi\xi_x}{v\theta})_x-k\frac{\tilde{\theta}}{v\theta^2}\xi_{x}^2+H \end{aligned}$$ where $$\label{3.23} \begin{aligned} &H=\{-\tilde{p}\tilde{u}_x(\frac{\phi^2}{v\tilde{v}}+\frac{(p-\tilde{p})\xi}{\tilde{p}\theta}-\frac{\xi^2}{\tilde{\theta}\theta}+({\gamma}-1)\Phi(\frac{v}{\tilde{v}})+\Phi(\frac{\theta}{\tilde{\theta}}))\} \\ &+\{-k\frac{\tilde{\theta}_x\phi_x\xi}{v^2\theta}+k\frac{\tilde{\theta}_x\xi\xi_x}{v\theta^2}+k\frac{\tilde{\theta}_{xx}\xi}{v\theta} -k\frac{\tilde{\theta}_x\tilde{v}_x\xi}{v^2\theta}\}\\ &=H_1+H_2 \end{aligned}$$ The form of $p(v,s)$ in (\[2.2-1\]) tell us $$\label{3.24} \begin{aligned} &0\leq p(v,s)-p(\tilde{v},\tilde{s})-p_v(\tilde{v},\tilde{s})\phi-p_s(\tilde{v},\tilde{s})(s-\tilde{s})\\ &=\tilde{p}(\frac{\phi^2}{v\tilde{v}}+\frac{(p-\tilde{p})\xi}{\tilde{p}\theta}-\frac{\xi^2}{\tilde{\theta}\theta}+({\gamma}-1)\Phi(\frac{v}{\tilde{v}}) +\Phi(\frac{\theta}{\tilde{\theta}})), \end{aligned}$$ since $\tilde{u}_x\geq0$, $$\label{3.25} \int_{0}^{t}\int_{\mathbb{R}_+}H_1dxd\tau\leq 0.$$ Besides that, $$\label{3.26} \begin{aligned} &\int_{0}^{t}\int_{\mathbb{R}_+}H_2dxd\tau \\ &\leq c(v_+,\theta_-){\epsilon}^{-\frac{1}{4}}\int_{0}^{t}\|\tilde{\theta}_x\|_{L^{\infty}}^2\|\xi\|^2d\tau +c(v_+,\theta_-)\int_{0}^{t}\|\xi\|_{L^{\infty}}(\|\tilde{u}_x\|^2+\|\tilde{\theta}_{xx}\|_{L^1})d\tau\\ &+\frac{1}{8}k\int_{0}^{t}\int_{\mathbb{R}_+}\frac{\tilde{\theta}}{v\theta^2}\xi_{x}^2dxd\tau+{\epsilon}^{\frac{1}{4}}\int_{0}^{t}\|\phi_x(\tau)\|^2d\tau\\ &\leq c(v_+,\theta_-)({\epsilon}^{\frac{1}{4}}\int_0^t(1+\tau)^{-\frac{3}{2}}\|\xi\|^2d\tau+ \int_{0}^{t}\|\xi\|^{\frac{2}{3}}(\|\tilde{u}_x\|^{\frac{8}{3}}+\|\tilde{\theta}_{xx}\|_{L^1}^{\frac{4}{3}})d\tau\\ &+\frac{1}{4}k\int_{0}^{t}\int_{\mathbb{R}_+}\frac{\tilde{\theta}}{v\theta^2}\xi_{x}^2dxd\tau+{\epsilon}^{\frac{1}{4}}\int_{0}^{t}\|\phi_x(\tau)\|^2d\tau,\\ \end{aligned}$$ where $$\label{3.27} \begin{aligned} &\int_{0}^{t}\|\xi\|^{\frac{2}{3}}(\|\tilde{u}_x\|^{\frac{8}{3}}+\|\tilde{\theta}_{xx}\|_{L^1}^{\frac{4}{3}})d\tau \leq C{\epsilon}^{\frac{1}{4}}\int_{0}^{t}(1+\tau)^{(\frac{13}{12})(-1+\frac{1}{q})}(1+\|\xi\|^2)d\tau.\\ \end{aligned}$$ Here we use the fact that $|\tilde{\theta}_x|\leq C({\gamma}-1)|\tilde{u}_x|,|\tilde{v}_x|\leq C|\tilde{u}_x|.$ Combing the results (\[3.25\])-(\[3.27\]), choosing $q$ suitably large such that $\frac{13}{12}(-1+\frac{1}{q})>1,$ then integrating (\[3.22\]) over $[0,t]\times\mathbb{R}_+,$ we immediately get that $$\label{3.28} \begin{aligned} &\|(\phi,\psi,\xi)(t)\|^2+\int_{0}^{t}\|\xi_x(\tau)\|^2\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|^2+{\epsilon}^{\frac{1}{4}}\int_{0}^{t}\|\phi_x(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}\|\xi(t)\|^2+{\epsilon}^{\frac{1}{4}}). \end{aligned}$$ Choosing ${\epsilon}^{\frac{1}{4}}$ suiably small, we finally get $$\label{3.29} \|(\phi,\psi,\xi)(t)\|^2+\int_{0}^{t}\|\xi_x(\tau)\|^2 \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|^2+{\epsilon}^{\frac{1}{4}}\int_{0}^{t}\|\phi_x(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}).$$ That is (\[3.19\]) in our Lemma 3.1. Secondly, we show the one derivative estimates in $L^2.$ Under the same assumptions listed in Proposition \[p2\], if ${\epsilon}, {\varepsilon}_0$ are suitably small, it holds that $$\label{3.30} \begin{aligned} &\|(\phi_x,\psi_x,\xi_x)(t)\|^2+\int_0^t\|\xi_{xx}(\tau)\|^2d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}+\eta_1)\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|^2d\tau +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}). \\ \end{aligned}$$ where $\eta_1$ is some small positive constant which will be determined later. Proof. Let $(\ref{3.2})_{1x}\times\frac{\tilde{p}}{v}\phi_x, (\ref{3.2})_2\times-\psi_{xx}, (\ref{3.2})_3\times-\frac{\xi_{xx}}{\theta}$, then $$\begin{aligned}\label{3.31} &(\frac{1}{2}\frac{\tilde{p}}{v}\phi_x^2+\frac{1}{2}\psi_x^2+\frac{1}{2}\frac{R}{\gamma-1} \frac{\xi_x^2}{\theta})_t -(\psi_t\psi_x+\frac{R}{\gamma-1}\xi_t\frac{\xi_x}{\theta}+\frac{R\xi_x\psi_x}{v})_x\\ &-\frac{1}{2}(\frac{\tilde{p}}{v})_t\phi_x^2 +\frac{R}{\gamma-1}(\frac{1}{\theta})_x\xi_t\xi_x -\frac{1}{2}\frac{R}{\gamma-1}(\frac{1}{\theta})_t\xi_x^2-\tilde{u}_x(p-\tilde{p})\frac{\xi_{xx}}{\theta}\\ &=-k(\frac{\xi_x}{v})_x \frac{\xi_{xx}}{\theta}-k(\frac{\tilde{\theta}_x}{v})_x\frac{\xi_{xx}}{\theta}+F_1, \end{aligned}$$ where $F_1$ is following: $$\label{3.32} F_1=-(\frac{R}{v})_x\psi_x\xi_x+(\frac{R}{v})_x\xi\psi_{xx}-(\frac{\tilde{p}}{v})_x\phi\psi_{xx}.$$ Integrating (\[3.31\]) over $[0,t]\times\mathbb{R}_+,$ and by the Sobelov’s inequality, $$\label{3.33} \begin{aligned} &\|(\phi_x,\psi_x,\xi_x)(t)\|^2+\int_{0}^{t}\|\xi_{xx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+\int_{0}^{t}|\xi_x\psi_x|(\tau,0)d\tau+ ({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_x,\psi_x,\xi_x)(\tau)\|_1^2\\ &+{\epsilon}^{-\frac{1}{4}}(\int_{0}^{t}\int_{\mathbb{R}_+}|\tilde{u}_x|^2(\phi^2+\xi^2)dxd\tau +\int_{0}^{t}\|\tilde{\theta}_{xx},\tilde{\theta}_x\tilde{v}_x\|^2d\tau)) \end{aligned}$$ Similar as (\[3.26\]), we immediately get that $$\label{3.34} \begin{aligned} &{\epsilon}^{-\frac{1}{4}}(\int_{0}^{t}\int_{\mathbb{R}_+}|\tilde{u}_x|^2(\phi^2+\xi^2)dxd\tau+\int_{0}^{t}\|\tilde{\theta}_{xx},\tilde{\theta}_x\tilde{v}_x\|^2d\tau)\\ &\leq C{\epsilon}^{\frac{1}{4}}(\int_{0}^{t}(1+\tau)^{-\frac{3}{2}}\|\phi,\xi\|^2d\tau +\int_{0}^{t}(1+\tau)^{-\frac{5}{2}}+(1+\tau)^{(\frac{5}{3})(-1+\frac{1}{q})}d\tau). \end{aligned}$$ For the boundary term, we deal with it in following way. On the one hand, $$\label{3.34-1} \begin{aligned} &\int_0^t|\xi_x\psi_x|(\tau,0)d\tau \\ &\leq C\int_0^t\|\xi_x(\tau)\|^2d\tau +\frac{1}{4}\int_{0}^{t}\|\xi_{xx}(\tau)\|^2d\tau +\eta_1\int_{0}^{t}|\psi_x|^2(\tau,0)d\tau, \end{aligned}$$ and $\eta_1$ is some small constant which will be determined later, on the other hand, $(\ref{3.2})_{1x}\times\psi_{x}+(\ref{3.2})_{2x}\times\phi_x$ tell us $$\label{3.34-2} (\phi_x\psi_x)_t-(\frac{1}{2}\psi_x^2+\frac{\tilde{p}}{2v}\phi_x^2)_x=F_2,$$ where $F_2$ is $$\label{3.34-3} F_2=2\frac{R\xi_xv_x}{v^2}\phi_x+\frac{3}{2}(\frac{\tilde{p}}{v})_x\phi_x^2+(\frac{\tilde{p}}{v})_{xx}\phi\phi_x-\frac{R\xi_{xx}}{v}\phi_x -(\frac{R}{v})_{xx}\xi\phi_x.$$ Integrating (\[3.34-2\]) over $[0,t]\times\mathbb{R}_+$, then $$\label{3.34-4} \begin{aligned} &\int_{0}^{t}\phi_x^2(\tau,0)+\psi_x^2(\tau,0)d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0)\|_1^2+\|(\phi_x,\psi_x)(t)\|^2+({\varepsilon}_0 +{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\xi_x,\phi_{xx})(\tau)\|^2\\ &+\int_0^t\|(\phi_x,\xi_{xx})(\tau)\|^2d\tau)+{\epsilon}^{-\frac{1}{4}}\int_{0}^{t} (\|\tilde{u}_{xx}\|_{L^\infty}^2+\|\tilde{u}_x\|^4_{L^\infty})\|\phi,\xi\|^2d\tau \end{aligned}$$ Simiar as before, we have $$\label{3.34-5} \begin{aligned} &{\epsilon}^{-\frac{1}{4}}\int_{0}^{t} (\|\tilde{u}_{xx}\|_{L^\infty}^2+\|\tilde{u}_x\|^4_{L^\infty})\|\phi,\xi\|^2d\tau \leq c{\epsilon}^{\frac{1}{4}}\int_{0}^{t}(1+\tau)^{-\frac{3}{2}}\|\phi,\xi\|^2d\tau \end{aligned}$$ Substituting (\[3.19\]), (\[3.34\])-(\[3.34-1\]) and (\[3.34-4\])-(\[3.34-5\]) into (\[3.33\]), choosing ${\epsilon}, {\varepsilon}_0$ suitably small, $q$ suitably large, we have $$\label{3.35} \begin{aligned} &\|(\phi_x,\psi_x,\xi_x)(t)\|^2+\int_0^t\|\xi_{xx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+\eta_1\|(\phi_x,\psi_x)(t)\|^2 +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|_1^2d\tau+{\epsilon}^{\frac{1}{4}}\\ &+\eta_1\int_0^t\|(\phi_x,\xi_{xx})(\tau)\|^2d\tau). \end{aligned}$$ Again let $\eta_1$ suitably small, we get $$\label{3.36} \begin{aligned} &\|(\phi_x,\psi_x,\xi_x)(t)\|^2+\int_0^t\|\xi_{xx}(\tau)\|^2d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}+\eta_1)\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|^2d\tau +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}). \\ \end{aligned}$$ This is our estimate $(\ref{3.30}).$ Under the same assumptions listed in Lemma 3.2, if ${\epsilon}, {\varepsilon}_0,\eta_1$ are suitably small, it holds that $$\label{3.37} \begin{aligned} &\int_0^t\|(\phi_x,\psi_x)(\tau)\|^2d\tau \\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}), \end{aligned}$$ Proof. $(\ref{3.2})_2\times-\frac{1}{2}\phi_x, (\ref{3.2})_3\times\frac{\psi_x}{p}$ imply that $$\label{3.38} \begin{aligned} &\frac{\tilde{p}\phi_x^2}{2v}+\frac{1}{2}[(\frac{\tilde{p}}{v})_x\phi\phi_x-(\frac{R\xi}{v})_x\phi_x]=\frac{1}{2}[(\psi\phi_x)_t-(\psi\phi_t)_x+\psi_x^2] \\ &(\frac{R}{{\gamma}-1}\frac{\psi_x}{p}\xi)_t-\frac{R}{{\gamma}-1}(\frac{1}{p})_t\xi\psi_x -(\frac{R}{{\gamma}-1}\frac{\psi_t\xi}{p})_x+\frac{R}{{\gamma}-1}(\frac{\xi}{p})_x\psi_t+\psi_x^2\\ &+\frac{\tilde{u}_x(p-\tilde{p})}{p}\psi_x=k(\frac{\theta_x}{v})_x\frac{\psi_x}{p} \end{aligned}$$ Integrating $(\ref{3.38})_1+(\ref{3.38})_2$ over $[0,t]\times\mathbb{R}_+,$ and by Sobelov’s inequality and previous results $(\ref{3.19}), (\ref{3.30})$, $$\label{3.39} \begin{aligned} &\int_0^t\|(\phi_x,\psi_x)(\tau)\|^2d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}+\eta_1)\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|^2d\tau+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx}, \psi_{xx})(\tau)\|^2+{\epsilon}^{\frac{1}{4}}). \end{aligned}$$ Again choosing ${\varepsilon}_0,{\epsilon},\eta_1$ suitably small, we finally get $$\label{3.41} \begin{aligned} &\int_0^t\|(\phi_x,\psi_x)(\tau)\|^2d\tau \\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}), \end{aligned}$$ which is our estimate (\[3.37\]). Combining Lemma 3.1-Lemma 3.3, our estimates are listed as follows $$\label{3.42} \begin{aligned} &\|(\phi,\psi,\xi)(t)\|_1^2+\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|^2+\|\xi_x(\tau)\|_1^2d\tau \\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+({\varepsilon}_0 +{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}} ) \end{aligned}$$ By our system (\[3.2\]), from (\[3.42\]) and previous computations, we can deduce that $$\label{3.43} \begin{aligned} &\int_0^t\|(\phi_t,\psi_t,\xi_t)(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+({\varepsilon}_0 +{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}} ) \end{aligned}$$ Under the same assumptions listed in Proposition 3.2, if ${\epsilon}, {\varepsilon}_0$ are suitably small, it holds that $$\label{3.44} \begin{aligned} &\|(\phi_t,\psi_t,\xi_t)(t)\|^2+\int_{0}^{t}\|\xi_{tx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_1^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t} \|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}) \end{aligned}$$ Proof. $(\ref{3.2})_{1t}\times\frac{\tilde{p}}{v}\phi_t,(\ref{3.2})_{2t}\times\psi_t, (\ref{3.2})_{3t}\times\frac{\xi_t}{\theta}$ $$\label{3.45} \begin{aligned} &(\frac{1}{2}\frac{\tilde{p}}{v}\phi_t^2+\frac{1}{2}\psi_t^2+\frac{1}{2}\frac{R}{\gamma-1}\frac{\xi_t^2}{\theta})_t+(\frac{R\xi_t-\tilde{p}\phi_t}{v}\psi_t)_x +(\tilde{u}_x(p-\tilde{p}))_t\frac{\xi_t}{\theta}\\ &=k((\frac{\theta_x}{v})_t\frac{\xi_t}{\theta})_x-k(\frac{\theta_x}{v})_t\frac{\xi_{tx}}{\theta}+F_3, \end{aligned}$$ where $F_3$ is $$\label{3.46} \begin{aligned} &F_3=\frac{1}{2}(\frac{\tilde{p}}{v})_t\phi_t^2+\frac{1}{2}\frac{R}{{\gamma}-1}(\frac{1}{\theta})_t\xi_t^2+(\frac{\tilde{p}}{v})_t\phi_x\psi_t-(\frac{R}{v})_t\xi_x\psi_t\\ &+[(\frac{\tilde{p}}{v})_{tx}\phi-(\frac{R}{v})_{tx}\xi]\psi_t -p_t\psi_x\frac{\xi_t}{\theta}-k(\frac{\theta_x}{v})_t(\frac{1}{\theta})_x\xi_t. \end{aligned}$$ Integrating (\[3.45\]) over $[0,t]\times\mathbb{R}_+$ and note that $\psi_t(\tau,0)=0, \xi_t(\tau,0)=0,$ then making use of (\[3.42\]), (\[3.43\]) and previous results, when ${\varepsilon}_0, {\epsilon}$ are suitably small, $$\label{3.47} \begin{aligned} &\|(\phi_t,\psi_t,\xi_t)(t)\|^2+\int_{0}^{t}\|\xi_{tx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_t,\psi_t,\xi_t)(0)\|^2+\|(\phi_0,\psi_0,\xi_0)\|_1^2 +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}) \end{aligned}$$ That is $(\ref{3.44}).$ Lastly, we show the higher order derivative estimates. Under the same assumptions listed in Proposition 3.2, if ${\epsilon}, {\varepsilon}_0$ are suitably small, it holds that $$\label{3.48} \begin{aligned} &\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+\int_{0}^{t}\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+{\epsilon}^{\frac{1}{4}}+ (\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &+\eta_3^{-1}\int_{0}^{t}\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ where $\eta_2, \eta_3$ are small positive constants which will be determined later. Proof. $(\ref{3.2})_{1x}\times-\frac{\tilde{p}}{v}\phi_{xxx}, (\ref{3.2})_{2xx}\times\psi_{xx}, (\ref{3.2})_{3x}\times-\frac{\xi_{xxx}}{\theta}$ tell us $$\label{3.49} \begin{aligned} &\frac{1}{2}(\frac{\tilde{p}}{v}\phi_{xx}^2+\psi_{xx}^2+\frac{R}{{\gamma}-1}\frac{1}{\theta}\xi_{xx}^2)_t-(\frac{\tilde{p}}{v}\phi_{tx}\phi_{xx}+\frac{R}{{\gamma}-1}\xi_{tx}\frac{\xi_{xx}}{\theta})_x -(\tilde{u}_x(p-\tilde{p}))_x\frac{\xi_{xxx}}{\theta}\\ &+k\frac{\xi_{xxx}^2}{v\theta}=F_4-2k(\frac{1}{v})_{x}\xi_{xx}\frac{\xi_{xxx}}{\theta}-k(\frac{1}{v})_{xx}\xi_x\frac{\xi_{xxx}}{\theta}-k(\frac{\tilde{\theta}_x}{v})_{xx}\frac{\xi_{xxx}}{\theta}, \end{aligned}$$ where $F_4$ is $$\label{3.50} \begin{aligned} &F_4=\frac{1}{2}(\frac{\tilde{p}}{v})_t\phi_{xx}^2+\frac{1}{2}(\frac{R}{{\gamma}-1}\frac{1}{\theta})_t\xi_{xx}^2 -(\frac{\tilde{p}}{v})_x\phi_{tx}\phi_{xx}-\frac{R}{{\gamma}-1}(\frac{1}{\theta})_x\xi_{tx}\xi_{xx}+p_x\psi_x\frac{\xi_{xxx}}{\theta}\\ &+3((\frac{\tilde{p}}{v})_{xx}\phi_x+(\frac{\tilde{p}}{v})_x\phi_{xx})\psi_{xx}+(\frac{\tilde{p}}{v})_{xxx}\phi\psi_{xx}-3((\frac{R}{v})_{xx}\xi_x+(\frac{R}{v})_x\xi_{xx})\psi_{xx}-(\frac{R}{v})_{xxx}\xi\psi_{xx}. \end{aligned}$$ Integrating (\[3.49\]) over $[0,t]\times\mathbb{R}_+,$ by $(\ref{3.42})$ and $(\ref{3.44})$ $$\label{3.51} \begin{aligned} &\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+\int_{0}^{t}\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+|\int_{0}^{t}H(\tau,0)d\tau|\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2+\|\xi_{xxx}(\tau)\|^2d\tau+{\epsilon}^{\frac{1}{4}}\\ &+\int_{0}^{t}\int_{\mathbb{R}_+}((\frac{\tilde{p}}{v})_{xxx}\phi-(\frac{R}{v})_{xxx}\xi)\psi_{xx}dxd\tau) . \end{aligned}$$ where $$H(\tau,0)=\frac{R}{{\gamma}-1}\xi_{tx}\xi_{xx}(\tau,0)+\frac{\tilde{p}}{v}\phi_{tx}\phi_{xx}(\tau,0)=H_1(\tau,0)+H_2(\tau,0)$$ On the one hand, using the relationship $\phi_{txx}=\psi_{xxx}$ and previous results, we can get $$\label{3.52} \begin{aligned} &\int_{0}^{t}\int_{\mathbb{R}_+}((\frac{\tilde{p}}{v})_{xxx}\phi-(\frac{R}{v})_{xxx}\xi)\psi_{xx}dxd\tau\\ &= \int_{0}^{t}\int_{\mathbb{R}_+}[(\frac{\tilde{p}_{xxx}}{v}+3(\frac{1}{v})_x\tilde{p}_{xx}+3(\frac{1}{v})_{xx}\tilde{p}_x)\phi +(\tilde{p}\phi-R\xi)(\frac{1}{v})_{xxx}]\psi_{xx}\\ &=\int_{0}^{t}\int_{\mathbb{R}_+}[(\frac{\tilde{p}_{xxx}}{v}+3(\frac{1}{v})_x\tilde{p}_{xx}+3(\frac{1}{v})_{xx}\tilde{p}_x)\phi +(\tilde{p}\phi-R\xi)(-\frac{\tilde{v}_{xxx}}{v^2}+6\frac{v_xv_{xx}}{v^3}-6\frac{v_x^3}{v^4})]\psi_{xx}\\ &-\int_{0}^{t}\int_{\mathbb{R}_+}(\tilde{p}\phi-R\xi)(\frac{\phi_{xxx}}{v^2})\psi_{xx}\\ &\leq C(v_+,\theta_-)(({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_x,\psi_x)(\tau) \|_1^2+{\epsilon}^{-\frac{1}{4}}\int_{0}^{t}\|\tilde{u}_x\|_{L^\infty}^2\|\phi,\xi\|^2d\tau)\\ &-\int_{0}^{t}\int_{\mathbb{R}_+}((\tilde{p}\phi-R\xi)\frac{\phi_{xx}}{v^2}\psi_{xx})_x +\int_{0}^{t}\int_{\mathbb{R}_+}(\frac{\tilde{p}\phi-R\xi}{v^2})_x\phi_{xx}\psi_{xx}dxd\tau\\ &+\int_{0}^{t}\int_{\mathbb{R}_+}((\frac{\tilde{p}\phi-R\xi}{2v^2})\phi_{xx}^2)_t-(\frac{\tilde{p}\phi-R\xi}{2v^2})_t\phi_{xx}^2dxd\tau\\ &\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+{\varepsilon}_0\|\phi_{xx}\|^2+{\epsilon}^{\frac{1}{4}} +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_x,\psi_x)(\tau)\|_1^2+\|\xi_{xxx}(\tau)\|^2d\tau\\ &+{\varepsilon}_0\int_{0}^{t}|\phi_{xx}|^2(\tau,0)d\tau), \end{aligned}$$ on the other hand, we turn to deal with the higher order boundary terms $H(\tau,0)$ in (\[3.51\]). By the conductivity of the equation $(\ref{3.2})_3$, we know that $$\label{3.53} \begin{aligned} \xi_{xx}(\tau,0)=(\frac{1}{k}(R\theta_-\psi_x-\tilde{u}_x\tilde{p}\phi)+\frac{\xi_xv_x}{v}-\tilde{\theta}_{xx}+\frac{\tilde{\theta}_xv_x}{v})(\tau,0). \end{aligned}$$ Hence $H_1(\tau,0)$ can be estimated in following way $$\label{3.54} \begin{aligned} &|\int_0^t\frac{R}{{\gamma}-1}\xi_{tx}\xi_{xx}(\tau,0)d\tau| \\ &=|\int_0^t\frac{R}{({\gamma}-1)}[(\frac{1}{k}(R\theta_-\psi_x-\tilde{u}_x\tilde{p}\phi)+\frac{\xi_xv_x}{v}-\tilde{\theta}_{xx}+\frac{\tilde{\theta}_xv_x}{v})(\tau,0) \xi_x(\tau,0)]_td\tau\\ &-\int_0^t\frac{R}{{\gamma}-1}(\frac{1}{k}(R\theta_-\psi_x-\tilde{u}_x\tilde{p}\phi)+\frac{\xi_xv_x}{v}-\tilde{\theta}_{xx}+\frac{\tilde{\theta}_xv_x}{v})_t(\tau,0)\xi_x(\tau,0)d\tau|\\ &\leq C(v_+,\theta_-) (\|\phi_0,\psi_0,\xi_0\|_2^2+\|\phi,\psi,\xi\|_1^2+(\eta_2+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\|\phi_{xx},\psi_{xx},\xi_{xx}\|^2+{\epsilon}^{\frac{1}{4}}\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|\phi_{x}(\tau),\psi_x(\tau)\|_1^2d\tau+\frac{1}{4}\int_{0}^{t}\|\xi_{xxx}\|^2d\tau +\int_0^t\|\xi_{x}(\tau)\|_1^2d\tau\\ &+(\eta_2+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t(\phi_{xx}^2+\psi_{xx}^2)(\tau,0)d\tau). \end{aligned}$$ Here $\eta_2$ is a small positive constant which will be determined below. For the other higher order boundary term $H_2(\tau,0)$ in $(\ref{3.51})$ which is one of the main difficulties in our estimates, we deal with it in following procedure. By $(\ref{3.2})_2$, $$\label{3.55} \begin{aligned} \psi_{tt}(\tau,0)+(\frac{R\xi-\tilde{p}\phi}{v})_{tx}(\tau,0)=0, \end{aligned}$$ and $\psi_{tt}(\tau,0)=0$ immediately gives us that $$\label{3.56} \begin{aligned} \frac{\tilde{p}}{v}\phi_{tx}(\tau,0)=\frac{R}{v}\xi_{tx}(\tau,0)-((\frac{\tilde{p}}{v})_{tx}\phi+(\frac{\tilde{p}}{v})_t\phi_x+(\frac{\tilde{p}}{v})_x\phi_t -(\frac{R}{v})_{t}\xi_x)(\tau,0). \end{aligned}$$ Using this, and by Young-inequality, it holds that $$\label{3.57} \begin{aligned} &|\int_0^t\frac{\tilde{p}}{v}\phi_{tx}\phi_{xx}(\tau,0)d\tau|\\ &\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+{\epsilon}^{\frac{1}{4}}+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}) \int_0^t(\phi_x^2+\psi_x^2+\xi_x^2+\phi_{xx}^2+\psi_{xx}^2)(\tau,0)d\tau\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+\eta_3\int_{0}^{t}\phi_{xx}^2(\tau,0)d\tau +\eta_3^{-1}\int_0^t\xi_{tx}^2(\tau,0)d\tau, \end{aligned}$$ where $\eta_3$ is some small positive constant to be determined. Substituting $(\ref{3.52}),(\ref{3.54})$, $(\ref{3.57})$ into $(\ref{3.51})$ and using (\[3.42\]), we have $$\label{3.58} \begin{aligned} &\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+\int_{0}^{t}\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+(\eta_2+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2\\ &+{\epsilon}^{\frac{1}{4}} +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &+(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}(\phi_{xx}^2+\psi_{xx}^2)(\tau,0)d\tau+\eta_3^{-1}\int_{0}^{t}\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ Similar as procedures (\[3.34-2\])-(\[3.34-4\]), let $(\ref{3.2})_{1xx}\times\psi_{xx}+(\ref{3.2})_{2xx}\times\phi_{xx}$, we have $$\label{3.59} \begin{aligned} (\phi_{xx}\psi_{xx})_t-\frac{1}{2}(\frac{\tilde{p}}{v}\phi_{xx}^2+\psi_{xx}^2)_x=F_5, \end{aligned}$$ where $F_5$ is $$\label{3.60} \begin{aligned} &F_5= \frac{5}{2}(\frac{\tilde{p}}{v})_x\phi_{xx}^2+3(\frac{\tilde{p}}{v})_{xx}\phi_x\phi_{xx}-3((\frac{R}{v})_{xx}\xi_x-\frac{R}{v}\xi_{xxx}\phi_{xx}.\\ &+(\frac{R}{v})_x\xi_{xx})\phi_{xx}+\{((\frac{\tilde{p}}{v})_{xxx}\phi-(\frac{R}{v})_{xxx}\xi)\phi_{xx}\} \end{aligned}$$ Integating $(\ref{3.59})$ over $[0,t]\times\mathbb{R}_+,$ we can get $$\label{3.61} \begin{aligned} &\int_{0}^{t}\phi_{xx}^2(\tau,0)+\psi_{xx}^2(\tau,0) d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0)\|_2^2+\|(\phi_{xx},\psi_{xx})(t)\|^2+{\epsilon}^{\frac{1}{4}}\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\psi_{xx},\xi_{xx})(\tau)\|^2d\tau +\int_{0}^{t}\|(\phi_{xx},\xi_{xxx})(\tau)\|^2d\tau). \end{aligned}$$ Here the estimate of the last term in (\[3.60\]) is similar as $(\ref{3.52}).$ Using (\[3.61\]), (\[3.58\]) turns to $$\label{3.62} \begin{aligned} &\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+\int_{0}^{t}\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+{\epsilon}^{\frac{1}{4}}\\ & +(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx},\xi_{xxx})(\tau)\|^2d\tau+\eta_3^{-1}\int_{0}^{t}\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ Choosing $\eta_2,\eta_3,{\varepsilon}_0,{\epsilon}$ suitably small such that $$\label{3.63} C(v_+,\theta_-)(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\leq \frac{1}{2},$$ we get $$\label{3.66} \begin{aligned} &\|(\phi_{xx},\psi_{xx},\xi_{xx})(t)\|^2+\int_{0}^{t}\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+{\epsilon}^{\frac{1}{4}}+ (\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_{0}^{t}\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &+\eta_3^{-1}\int_{0}^{t}\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ That is our estimate $(\ref{3.48}).$ Under the same assumptions listed in Proposition 3.2, if ${\epsilon}, {\varepsilon}_0$ are suitably small, it holds that $$\label{3.67} \begin{aligned} &\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2+{\epsilon}^{\frac{1}{4}}+\eta_4\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau). \end{aligned}$$ where $\eta_4$ is a small positive constant which will be determined later. Proof. Similar as Lemma 3.3, $(\ref{3.2})_{2x}\times-\frac{\phi_{xx}}{2},(\ref{3.2})_{3x}\times\frac{\psi_{xx}}{p}$ give us $$\label{3.68} \begin{aligned} &\frac{\tilde{p}}{2v}\phi_{xx}^2+(\frac{\tilde{p}}{v})_x\phi_x\phi_{xx}+(\frac{\tilde{p}}{2v})_{xx}\phi\phi_{xx} =\frac{1}{2}((\psi_x\phi_{xx})_t-(\psi_x\phi_{tx})_x+\psi_{xx}^2+(\frac{R\xi}{v})_{xx}\phi_{xx})\\ &(\frac{R}{{\gamma}-1}\frac{\psi_{xx}}{p}\xi_x)_t-\frac{R}{{\gamma}-1}(\frac{1}{p})_t\xi_x\psi_{xx} -(\frac{R}{{\gamma}-1}\frac{\psi_{tx}\xi_x}{p})_x+\frac{R}{{\gamma}-1}(\frac{\xi_x}{p})_x\psi_{tx}+\psi_{xx}^2+p_x\psi_x\frac{\psi_{xx}}{p}\\ &+(\tilde{u}_x(p-\tilde{p}))_x\frac{\psi_{xx}}{p}=k(\frac{\theta_x}{v})_{xx}\frac{\psi_{xx}}{p}. \end{aligned}$$ Then integrating $(\ref{3.68})_1+(\ref{3.68})_2$ over $[0,t]\times\mathbb{R}_+,$ we get $$\label{3.69} \begin{aligned} &\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_{x},\psi_{x},\xi_{x})(t)\|_1^2 +\int_{0}^{t}\|\xi_x(\tau)\|_2^2d\tau\\ &+{\epsilon}^{\frac{1}{4}} +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau)+|\int_0^t(\psi_x\phi_{tx})(\tau,0)d\tau|+ |\int_0^t(\frac{\psi_{tx}\xi_x}{p})(\tau,0)d\tau| \end{aligned}$$ For the boundary terms in $(\ref{3.69}),$ making use of $(\ref{3.56}), (\ref{3.34-4}), (\ref{3.42})$ and (\[3.48\]), (\[3.61\]) we have $$\label{3.70} \begin{aligned} &|\int_0^t(\psi_x\phi_{tx})(\tau,0)d\tau|\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+ {\epsilon}^{\frac{1}{4}}+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2\\ &+[({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\eta_3^{-1}+1]\int_0^t\xi_{tx}^2(\tau,0)d\tau)\\ \end{aligned}$$ and $$\label{3.71} \begin{aligned} &|\int_0^t(\frac{\psi_{tx}\xi_x}{p})(\tau,0)d\tau|= |\int_{0}^{t}(\frac{\psi_x\xi_x}{p})_t(\tau,0)-\int_0^t((\frac{\xi_x}{p})_t\psi_x)(\tau,0)|d\tau\\ &\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+\|\psi_x,\xi_x\|_1^2+{\epsilon}^{\frac{1}{4}}+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &+\int_{0}^{t}\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ Substituting $(\ref{3.70})$ and $(\ref{3.71})$ into $(\ref{3.69}),$ and using $(\ref{3.42})$, $(\ref{3.48})$, we have $$\label{3.72} \begin{aligned} &\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau \leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2 +{\epsilon}^{\frac{1}{4}}\\ &+(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau+\eta_3^{-1}\int_0^t\xi_{tx}^2(\tau,0)d\tau). \end{aligned}$$ Again choosing $\eta_2,\eta_3,{\varepsilon}_0,{\epsilon}$ suitably small such that $$\label{3.73} C(v_+,\theta_-)(\eta_2+\eta_3+{\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\leq \frac{3}{4},\quad C(v_+,\theta_-)\eta_3^{-1}\leq \bar{C}(v_+,\theta_-),$$ for some positive constant $\bar{C}(v_+,\theta_-)$ which is larger than $C(v_+,\theta_-),$ the condition (\[3.73\]) we proposed could be satisfied obviously. Then using (\[3.44\]), by Sobolev’s inequality and Young inequality, (\[3.72\]) becomes $$\label{3.74} \begin{aligned} &\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &\leq C_1(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2 +{\epsilon}^{\frac{1}{4}}\\ &+\int_0^t\|\xi_{tx}(\tau)\|^2d\tau+\eta_4\int_0^t\|\xi_{txx}(\tau)\|^2d\tau)\\ &\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2 +{\epsilon}^{\frac{1}{4}}\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}})\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2 +\eta_4\int_0^t\|\xi_{txx}(\tau)\|^2d\tau), \end{aligned}$$ where $\eta_4$ is a small positive constant to be determined later. Finally, because of the sufficient small constants ${\varepsilon}_0,{\epsilon}$ again, we see that $$\label{3.74-3} \begin{aligned} &\int_0^t\|(\phi_{xx},\psi_{xx})(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|\phi_0,\psi_0,\xi_0\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2 +{\epsilon}^{\frac{1}{4}}+\eta_4\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau). \end{aligned}$$ That is (\[3.67\]). Now combining Lemma 3.1-Lemma 3.6, we have following results: $$\label{3.75} \begin{aligned} &\|(\phi,\psi,\xi)(t)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(t)\|^2+\int_{0}^{t}\|(\phi_{x},\psi_{x},\xi_x)(\tau)\|_1^2d\tau\\ &+\int_0^t\|\xi_{tx}(\tau)\|^2+\|\xi_{xxx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2 +{\epsilon}^{\frac{1}{4}}+\eta_4\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau). \end{aligned}$$ In order to close our estimate (\[3.75\]), we have following Lemma. Under the same assumptions listed in Proposition 3.2, if ${\epsilon}, {\varepsilon}_0,\eta_4$ are suitably small, it holds that $$\label{3.76} \begin{aligned} &\|(\phi_{tx},\psi_{tx},\xi_{tx})(t)\|^2+\int_0^t\|\xi_{txx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2 +{\epsilon}^{\frac{1}{4}}) \end{aligned}$$ Proof. $(\ref{3.2})_{1tx}\times\frac{\tilde{p}}{v}\phi_{tx}, (\ref{3.2})_{2t}\times-\psi_{txx}, (\ref{3.2})_{3t}\times-\frac{\xi_{txx}}{\theta}$ imply that $$\label{3.77} \begin{aligned} &\frac{1}{2}(\frac{\tilde{p}}{v}\phi_{tx}^2+\psi_{tx}^2+\frac{R}{{\gamma}-1}\frac{\xi_{tx}^2}{\theta})_t-(\psi_{tt}\psi_{tx}+\frac{R}{{\gamma}-1}\xi_{tt}\frac{\xi_{tx}}{\theta}+\frac{R}{v}\psi_{tx}\xi_{tx})_x +(\frac{R}{v})_x\psi_{tx}\xi_{tx}\\ &-(\tilde{u}_x(p-\tilde{p}))_t\frac{\xi_{txx}}{\theta}=F_6-k(\frac{\xi_x}{v})_{tx}\frac{\xi_{txx}}{\theta}-k(\frac{\tilde{\theta}_x}{v})_{tx}\frac{\xi_{txx}}{\theta}, \end{aligned}$$ where $F_6$ is $$\label{3.78} \begin{aligned} &F_6=\{\frac{1}{2}(\frac{\tilde{p}}{v})_t\phi_{tx}^2+\frac{1}{2}(\frac{R}{{\gamma}-1})(\frac{1}{\theta})_t\xi_{tx}^2-\frac{R}{{\gamma}-1}(\frac{1}{\theta})_x\xi_{tt}\xi_{tx}+p_t\psi_x\frac{\xi_{txx}}{\theta}\}\\ &+\{[((\frac{R}{v})_{x}\xi_t+(\frac{R}{v})_t\xi_x+(\frac{R}{v})_{tx}\xi) -((\frac{\tilde{p}}{v})_t\phi_x+(\frac{\tilde{p}}{v})_x\phi_t+(\frac{\tilde{p}}{v})_{tx}\phi)]\psi_{txx}\}\\ &=F_{61}+F_{62}. \end{aligned}$$ Obviously, $$\begin{aligned}\label{3.78-1} &|F_{61}|\leq C(v_+,\theta_-)[({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}) (\phi_x^2+\psi_x^2+\xi_x^2+\phi_{xx}^2+\psi_{xx}^2+\xi_{xx}^2+\xi_{tx}^2+\xi_{txx}^2)\\ &+|\tilde{v}_x|^2(\phi^2+\xi^2)+|\tilde{\theta}_{txx}|^2+|\tilde{v}_{xx}\tilde{\theta}_x|^2 +|\tilde{v}_x|^6] \end{aligned}$$ Integrating $(\ref{3.77})$ over $[0,t]\times\mathbb{R}_+$ and using (\[3.75\]), (\[3.78\]),(\[3.78-1\]), note that $\psi_{tt}(\tau,0)=\xi_{tt}(\tau,0)=0$, we have $$\label{3.79} \begin{aligned} &\|(\phi_{tx},\psi_{tx},\xi_{tx})(t)\|^2+\int_0^t\|\xi_{txx}(\tau)\|^2d\tau\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2 +{\epsilon}^{\frac{1}{4}}\\ &+({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}+\eta_4)\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau+|\int_{0}^{t}\frac{R}{v}\psi_{tx}\xi_{tx}(\tau,0)d\tau|+\int_{0}^{t}\int_{\mathbb{R}_+}F_{62}dxd\tau.\\ \end{aligned}$$ By using the equations $\phi_{tx}=\psi_{xx}$, (\[3.34-4\]), (\[3.61\]) and (\[3.75\]), $F_{62}$ could be estimated by $$\label{3.80} \begin{aligned} &\int_0^t\int_{\mathbb{R}_+}[((\frac{R}{v})_{x}\xi_t+(\frac{R}{v})_t\xi_x+(\frac{R}{v})_{tx}\xi) -((\frac{\tilde{p}}{v})_t\phi_x+(\frac{\tilde{p}}{v})_x\phi_t+(\frac{\tilde{p}}{v})_{tx}\phi)]\psi_{txx}dxd\tau\\ &=\int_{0}^{t}\int_{\mathbb{R}_+}[((\frac{R}{v})_{x}\xi_t+(\frac{R}{v})_t\xi_x)\psi_{tx}]_x -((\frac{R}{v})_{x}\xi_t+(\frac{R}{v})_t\xi_x)_x\psi_{tx}\\ &-[((\frac{\tilde{p}}{v})_t\phi_x+(\frac{\tilde{p}}{v})_x\phi_t+\frac{\tilde{p}_{tx}}{v}+\tilde{p}_t(\frac{1}{v})_x+\tilde{p}_x(\frac{1}{v})_t\phi)\psi_{tx}]_x +(R\xi-\tilde{p}\phi)(\frac{1}{v})_{tx}\psi_{txx}\\ &+((\frac{\tilde{p}}{v})_t\phi_x+(\frac{\tilde{p}}{v})_x\phi_t+\frac{\tilde{p}_{tx}}{v}+\tilde{p}_t(\frac{1}{v})_x+\tilde{p}_x(\frac{1}{v})_t\phi)_x\psi_{tx}dxd\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2+{\epsilon}^{\frac{1}{4}}\\ &+({\epsilon}^{\frac{1}{4}}+{\varepsilon}_0)\int_0^t\phi_x^2(\tau,0)+\psi_x^2(\tau,0) +\phi_{xx}^2(\tau,0)d\tau+\eta_4\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2+{\epsilon}^{\frac{1}{4}}+\eta_4\int_{0}^{t}\|\xi_{txx}(\tau)\|^2d\tau, \end{aligned}$$ where $$\label{3.81} \begin{aligned} &\int_0^t\int_{\mathbb{R}_+}(R\xi-\tilde{p}\phi)(\frac{1}{v})_{tx}\psi_{txx}dxd\tau\\ &=\int_0^t\int_{\mathbb{R}_+}[(R\xi-\tilde{p}\phi)(2\frac{v_xv_t}{v^3} -\frac{\tilde{v}_{tx}}{v^2})\psi_{tx}]_x -[(R\xi-\tilde{p}\phi)(2\frac{v_xv_t}{v^3}-\frac{\tilde{v}_{tx}}{v^2})]_x\psi_{tx}\\ &+((\frac{\tilde{p}\phi-R\xi}{v^2})\frac{\psi_{xx}^2}{2})_t-(\frac{\tilde{p}\phi-R\xi}{v^2})_t\frac{\psi_{xx}^2}{2}dxd\tau. \end{aligned}$$ For the boundary term in the right hand side of $(\ref{3.79}),$ using above results $(\ref{3.34-4}), (\ref{3.61}), (\ref{3.75})$ again, we immediately have $$\label{3.82} \begin{aligned} &|\int_0^t\frac{R}{v}\psi_{tx}\xi_{tx}(\tau,0)d\tau|\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2+{\epsilon}^{\frac{1}{4}}\\ &+\int_0^t\phi_x^2(\tau,0)+\phi_{xx}^2(\tau,0) +\|\xi_{tx}(\tau)\|^2d\tau+\eta_4\int_0^t\|\xi_{txx}(\tau)\|^2d\tau)\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|^2 +{\epsilon}^{\frac{1}{4}}+\eta_4\int_0^t\|\xi_{txx}(\tau)\|^2d\tau)). \end{aligned}$$ Substituting $(\ref{3.80})-(\ref{3.82})$ into $(\ref{3.79})$, $(\ref{3.79})$ turns to $$\label{3.83} \begin{aligned} &\|(\phi_{tx},\psi_{tx},\xi_{tx})(t)\|^2+\int_0^t\|\xi_{txx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2 +{\epsilon}^{\frac{1}{4}} +({\varepsilon}_0+{\epsilon}^{\frac{1}{4}}+\eta_4)\int_0^t\|\xi_{txx}(\tau)\|^2d\tau). \end{aligned}$$ Choosing ${\varepsilon}_0,{\epsilon},\eta_4$ suitably small, we finally get $$\label{3.84} \begin{aligned} &\|(\phi_{tx},\psi_{tx},\xi_{tx})(t)\|^2+\int_0^t\|\xi_{txx}(\tau)\|^2d\tau\\ &\leq C(v_+,\theta_-)(\|(\phi_0,\psi_0,\xi_0)\|_2^2+\|(\phi_t,\psi_t,\xi_t)(0)\|_1^2 +{\epsilon}^{\frac{1}{4}}) \end{aligned}$$ That is (\[3.76\]). Combining Lemma 3.1-Lemma 3.7, we can get Proposition 3.2. [99]{} L. Fan, H. Liu, T. Wang and H. Zhao, Inflow problem for the one-dimensional compressible Navier-Stokes equations under large initial perturbation J. Differ. Equ. [**257**]{}(2014), 3521-3553. L. Fan and A. Matsumura, Asymptotic stability of a composite wave of two viscous shock waves for the equation of non-viscous and heat-conductive ideal gas. J. Differential Equations, [**258**]{}(2015), 1129-1157. C. Hou and L. Fan, Asymptotic stability for the inflow problem of the heat-conductive ideal gas without viscosity, Preprint. F. Huang, J. Li and X. Shi, Asymptotic behavior of solutions to the full compressible Navier-Stokes equations in the half space Commun. Math. Sci.[**8**]{}(2010),639-654. F. Huang, J. Li, A. Matsumura, Stability of the combination of the viscous contact wave and the rarefaction wave to the compressible Navier-Stokes equations. Arch. Rat. Mech. Anal. [**197**]{}(2010), 89-116. F. Huang and A. Matsumura, Stability of a Composite Wave of Two Viscous Shock Waves for the Full Compresible Navier-Stokes Equation. Comm. Math. Phys. [**289**]{}(2009), 841-861. F. Huang, A. Matsumura and X. Shi, Viscous shock wave and boundary layer solution to an inflow problem for compressible viscous gas. Comm. Math. Phys. [**239**]{}(2003), 261-285. F.Huang, A. Matsumura and X. Shi, A gas-solid free boundary problem for compressible viscous gas. SIAM J. Math. Anal. [**34**]{}(2003) no.6, 1331-1355 F. Huang, A. Matsumura and X. Shi, On the stability of contact discontinuity for compressible Navier-Stokes equations with free boundary Osaka J. Math. [**41**]{}(2004) 193-210. F. Huang, A. Matsumura and Z. Xin, Stability of Contact discontinuties for the 1-D Compressible Navier-Stokes equations. Arch. Ration. Mech.Anal. [**179**]{} (2005), 55-77. F. Huang and X. Qin, Stability of boundary layer and rarefactionwave to an outflow problem for compressible Navier-Stokes equations under large perturbation J. Diff. Eqns [**246**]{} (2009),4077-4096. F. Huang, T. Yang and Z. Xin, Contact discontinuity with general perturbations for gas motions. Adv. in Math. [**219**]{}(2008), 1246-1297 . F. Huang, H.J. Zhao, On the global stability of contact discontinuity for compressible Navier-Stokes equations, Rend. Sem. Mat. Univ. Padova [**109**]{}(2003) 283-305. S. Kawashima, Systems of a Hyperbolic-Parabolic Composite Type, with Applications to the Equations of Magnetohydrodynamics, doctoral thesis in Kyoto University, 1984. S. Kawashima, T. Nakamura, S. Nishibata and P. Zhu, Stationary waves to viscous heat-conductive gases in half space:existence, stability and convergence rate, Mathematical Models and Methods in Applied Sciences. [**20**]{}(2011), 2201-2235. S. Kawashima, S. Nishibata and P.Zhu, Asymptotic stability of the stationary solution to the compressible Navier-Stokes equations in the half space Commun. Math. Phys, [**240**]{}(2003) 483-500. S. Kawashima, P. Zhu, Asymptotic stability of rarefaction wave for the Navier-Stokes equations for a compressible fluid in the half space Arch. Ration. Mech. Anal. [**194**]{}(2009) 105-132. T. Liu, Shock wave for Compresible Navier-Stokes Equations are stable. Comm. Math. Phys. [**50**]{}(1986), 565-594. A. Matsumura, Inflow and outflow problems in the half space for a one-dimensional isentropic model system of compressible viscous gas, Methods Appl. Anal., [**8**]{}(2001),645-666. IMS Conference on Differential Equations from Mechanics (Hong Kong,1999). A. Matsumura, Large-time behavior of solutions for a one-dimensional system of non-viscous and heat-conductive ideal gas, private communication, 2016. A. Matsumura and M. Mei, Convergence to travelling fronts of solutions of the p-system with viscosity in the presence of a boundary. Arch. Ration. Mech. Anal. [**146**]{}(1999), 1-22. A. Matsumura, K. Nishihara, Asymptotic toward the rarefaction waves for solutions of viscous p-system with boundary effect. Q. Appl. Math., [**58**]{}(2000), 69¨C83. A. Matsumura, K. Nishihara, Large time behaviors of solutions to an inflow problem in the half space for a one-dimensional system of compressible viscous gas. Comm. Math. Phys., [**222**]{}(2001), 449-474. L. Min, X. Qin, Stability of rarefaction wave for compressible Navier-Stokes equations on the half line. Acta Mathematicae Applicatae Sinica, English Series. [**32**]{}(2016), 175-186. T.Nakamura, Degenerate boundary layers for a system of viscous conservation laws. Anal. Appl. (Singap.) [**14**]{}(2016), 75-99. T.Nakamura, S.Nishibata, Stationary wave associated with an inflow problem in the half line for viscous heat-conductive gas. J. Hyperbolic Differ. Equ. [**8**]{}(2011), 651-670. T.Nakamura, S.Nishibata, Boundary layer solution to system of viscous conservation laws in half line. Bull. Braz. Math. Soc. (N.S.) [**47**]{}(2016), 619-630. T.Nakamura, S.Nishibata, Existence and asymptotic stability of stationary waves for symmetric hyperbolic-parabolic systems in half-line. Mathematical Models and Methods in Applied Sciences. [**27**]{}(2017), 2071-2110. T.Nakamura, S.Nishibata, and N.Usami, Convergence rate of solutions towards the stationary solutions to symmetric hyperbolic-parabolic systems in half space, Preprint. T.Nakamura, S.Nishibata and T.Yuge, Convergence rate of solutions toward stationary solutions to the compressible Navier-Stokes equation in half line, J.Differential Equations. [**241**]{}(2007), 94-111. K. Nishihara, T. Yang, and H. Zhao, Nonlinear stability of strong rarefaction waves for compressible Navier-Stokes equations. SIAM J. Math. Anal., [**35**]{}(2004), 1561- 1597. X. Qin, Large-time behaviour of solution to the outflow problem of full compressible Navier-Stokes equations. Nonlinearity, [**24**]{}(2011), 1369-1394. X. Qin, Y. Wang, Stability of wave patterns to the inflow problem of full compressible Navier-Stokes equations SIAM J. Math. Anal. [**41**]{} 2009 2057-2087. X. Qin, Y. Wang, Large-time behavior of solutions to the inflow problem of full compressible Navier-Stokes equations SIAM J. Math. Anal, [**43**]{} (2011) 341-366. L. Wan, T. Wang and Q. Zou, Stability of stationary solutions to the outflow problem for full compressible Navier-Stokes equations with large initial perturbation, Nonlinearity, [**29**]{} (2016), 1329-1354. T. Wang and H. Zhao. One-dimensional compressible heat-conducting gas with temperature-dependent viscosity. Math. Models Methods Appl. Sci., [**26**]{}(2016), 2237-2275.

ArXiv